Research ArticleDynamic Early Warning Method for Major Hazard InstallationSystems in Chemical Industrial Park

Yaguang Kong Chenfeng Xie Song Zheng Peng Jiang Meng Guan and Fang Wang

Hangzhou Dianzi University China

Correspondence should be addressed to Chenfeng Xie hdu muskhotmailcom

Received 9 January 2019 Revised 21 March 2019 Accepted 15 April 2019 Published 20 May 2019

Academic Editor Roberto Natella

Copyright copy 2019 Yaguang Kong et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The production and storage of major hazard installations (MHIs) bring potential risks to chemical industrial park (CIP) In theproduction system of MHIs its dangerous degree is mainly determined by key parameters and abnormal key parameters often leadto accidents To predict the real-time risk values of MHIs and improve accident prevention ability of CIP we need a method thatcan combine dynamic prediction and assessment Quantitative risk assessment (QRA) is not capable of modelling risk variationsduring the operation of a processTherefore this paper adopts the data-driven approach Inspired by visual qualitative analysis andquantitative analysis a dynamic early warning method is proposed for MHIs We can get the future trend of these key parametersby using strongly correlation variables to predict key parameters Fuzzy evaluation analysis is performed on the risk levels of keyparameters and the dynamic evaluation index of these MHIs is obtained This method can be applied to the dynamic evaluationof MHIs system in CIP It can contribute to the safety of CIP in some aspects

1 Introduction

The demand for everyday life materials has stimulated thevigorous development of the chemical industry Large CIPhave been built in various countries and regions CIP havethe characteristics of high complexity strong associationand long production process It is a process with multiplevariables strong coupling Industrial process variables devi-ate from target values and cause a series of failure due toequipment failure production equipment aging personneloperation error raw material characteristics and changesin external environment [1] It may also result in industrialaccidents such as fire explosions and leaks In severe casesthis deviation may lead to casualties major economic lossesand serious environmental pollution [2] On December 31984 the release of methyl isocyanate gas killed some 3000people in Bhopal India This incident is considered to be theworldrsquos worst industrial chemical accident on July 10 1976as a result of a reactor runaway accident in a chemical plantin Seveso Italy Dispersed dioxin into the local atmosphereled to skin disease chloracne in about 250 people and causticsoda burns in about 450 people [3] On August 12 2015 a

hazardous chemical warehouse in Tianjin Binhai New Areaexploded leading to the death of 165 people 8 missing798 injured and 687-billion-yuan direct economic loss [4]Therefore accident prevention ability of CIP should beimproved

QRA is a conventional risk modeling method AlthoughQRAmethods are useful in modeling accident scenarios andquantifying risk but their static structure is not capable ofmodeling risk variations during the process operation [5]Dynamic early warning is a way to improve the preventiveability of the production system of MHIs in the chemicalindustry In the area of early warning and assessment withthe use of the Process Resilience Analysis Framework Jainet al [6] proposed a resilience-based approach to managinguncertainties to better predict process upsetsMeel and Seider[7] proposed that the failure probabilities of safety systemsand end-states are estimated using copulas andBayesian anal-ysis to ensure better predictions in plant-specific dynamicfailure assessment Cheng et al [8] presented amethod basedon Bayesian and Vine Copula to predict the frequency ofchemical accidents based on variations in different groups

HindawiComplexityVolume 2019 Article ID 6250483 18 pageshttpsdoiorg10115520196250483

2 Complexity

Although Bayesrsquo theorem helps to obtain posterior probabili-ties it necessitates identifying likelihood function which isa difficult task if it is not a conjugate distribution to priorprobability Abdolhamidzadeh et al [9] proposed the conceptof lsquolocal domino effectrsquo to deal with a chain of accidentsoccurring within a process unit achieving a good result tohandle the uncertainty and the complexity Pariyani et al[10] introduced a distributed control system through biasanalysis to calculate the probability of system failure therisks associated with low probability severe consequenceincidents being predicted more accurately Zhou et al [11]established a security risk assessment method based onweighted fuzzy Petri net This approach can easily modeldifferent relationships between the risk factors as well as theirimportance In terms of social factor Jain et al [12] proposeda novel framework Process Resilience Analysis Frameworkfor incorporating both technical and social factors in anintegrated approach resulting in a reduced frequency ofloss of containment events Aven and Ylonen [13] proposeda closer integration of the risk analysis and managementapproach and the sociotechnical perspective on safety can beused to improve risk and safety regulations These methodstake human factors social factors and domino effect intoaccount but they cannot dynamically assess the risk changesof hazards caused by abnormal parameter changes

The prediction of parameters is an important part ofdynamic early warning Other fields have been applied toinvestigate parameter prediction models [14ndash16] Howeverthese traditional methods face the background of big datawith large computational complexity multiple processes andmultivariable chemical processes [17] For instance due tothe ease in configuring the alarms in control systems thenumber of alarms in a plant has also gone up This has led tofrequent system problems increase in the operator workloaddue to alarm overload and industrial accidents [18] Hencesuchmethods are generally inapplicable Today big data havebecome an important strategic resource for a country [19]Goel et al [20] discuss the potential of big data analytics inthe area of process safety and risk management in the energyindustry A large amount of data is also available in the CIPIn the early warning of CIP conducting a reasonable analysisof big data and fully considering the correlation betweenkey parameters of various process industries is essential forprediction analysis [21] If all variables in the productionsystem are not filtered then the calculation is highly complexVisual dynamic risk assessment method is an instrumentthat can qualitatively analyze the correlation between high-dimensional variables Saatci [22] used Parallel Coordinatesto explore the correlation between respiratory signals andairflow relative temperature relative humidity and otherfactors and achieved satisfactory results Azhar and Rissanen[23] evaluated an application of Parallel Coordinates forinteractive filtering of alarm data by comparing its userperformance against typical alarm lists The results showedthat Parallel Coordinates reduce alarm filtering time andreduce human mistakes Visualization method can onlysimultaneously perform qualitative analysis and the corre-lation between variables should be quantitatively analyzedBerthold and Hoppner [24] investigated clustering time

series based on Euclidean distance and Pearson correlationand contributed to future research

The prediction of MHIs should be performed rapidlyExcessive delay will result in unexpected consequences Zhaoet al [25] used GPU to conduct an accelerated operationand effectively improve the prediction speed Ruiz et al [26]employed a genetic algorithm to optimize neural networkweights enhance prediction accuracy and reduce predictiontime Chemical process industry is associated with manyvariables To handle the high-dimensional variables Erkmenand Yıldırım [27] used principal components analysis asa feature extraction method to promote classification per-formance Support vector machine [28] back propagationneural network (BP neural network) [29] and probabilisticmodels are commonly used in traditional prediction models[30] However these methods can be implemented to staticprediction only When these methods are applied to theprediction of industrial process dynamics a dynamic timemodeling problem is transformed into a stationary spacemodeling problem which inevitably leads to a decline inprediction accuracy We need to construct the model ithas a feedback link which can later memorize the stateinformation of the key parameters of the chemical industryallowing the system to adapt to time-varying characteristicsFor example Qin et al [31] used an Elman neural networkto improve the smooth transition period of an autoregressivemodel and accurately predicted wind energy and speed

The present paper is divided as follows The overallprocess of Section 2 briefly introduced a structure of dynamicearly warning method Section 3 presents the dynamic pre-diction and evaluation method which combines variablecorrelation analysis and feature analysis and feature vec-tor acquisition to conduct strong correlation analysis anddimensionality reduction Construct a recurrent neural net-work to predict MHIs Finally we use the MHIs risk level todefining amethod for the dynamic evaluation Section 4 takesQuzhou City CIP as an example to perform a case study of themethod Section 5 provides some concluding remarks

2 Dynamic Prediction and Evaluation Methodfor Major Hazard Installations

Given that the chemical accidents are usually caused bysome parameter fluctuations therefore it is necessary toconstruct a dynamic prediction model The applications ofneural network in various fields have continuously emergedrecently Compared with support vector machine [28] andnonlinear regressions neural network possesses the robustmultivariable prediction performance and can constantlyadjust the weight among input layer output layer and hiddenlayer to achieve high-precision prediction [29] In traditionalneural network we add a state layer to construct the dynamicrecurrent network which can better adapt to the dynamicchange characteristics of chemical production parametersAnalyzing the strong correlation variables of key parametersis crucial due to numerous parameters and high complexityof the chemical process Correlation analysis work such asthat of Cheng et al [8] uses the Copula function to establish

Complexity 3

All correlation variables of key parameters

of chemical industry MHIsComposition matrix S

Dynamic risk assessment

output dynamic risk value of MHIs

Variable correlation analysisCarry out (a)Visual qualitative analysis(b) Quantitative

analysis of correlation

Inputmatrix S

OutputMatrix X consisting of strongly correlated variables

Dimension reduction optimizationAfter (a) feature analysis (b) feature vector acquisition

Inputmatrix X

Outputmatrix T after dimension reduction

Dynamic recurrent predictionA recurrent neural network structure with State layer is constructed

Inputmatrix T

OutputPredicted values of key parameters

Constructing two-level fuzzy comprehensive evaluation to

Figure 1 Structure block diagram of dynamic early warning method

the asymmetric correlation model of anomalous number ofevents considering the correlation among the four teamsUnder the multivariable condition of chemical productionengineering the process of establishing Copula function isconsiderably complicated and a fast and accurate methodof correlation analysis is necessary The visualization methodcan be used for fast qualitative analysis of relevant variablesand combined with quantitative analysis method effectivelyimplement the selection of strong correlation variables of keyparameter

The structural block diagram of the dynamic early warn-ing method is presented in Figure 1 The first part of thealgorithm is variable correlation analysis the second part isfeature analysis and feature vector acquisition and the thirdpart is dynamic recurrent prediction And the last part is thedynamic risk assessment The implementation process of thealgorithm is as follows

(1) All correlation variables of the key parameters ofMHIsin the CIP use the qualitative analysis of variable correlationby data visualization [32]

4 Complexity

(2) The average value and standard deviation of keyparameters and related variables are calculated

(3) The correlation coefficient r value between two vari-ables is obtained and the strong correlation variable (4)The covariance matrix of the strong correlation variable isalso calculated and its eigenvalue and the eigenvector areobtained

(5) The order of the eigenvalue from large to smallis selected and the maximum is chosen according to theestablished rule K eigenvalue constitutes the correspondingeigenvector matrix

(6)The samples of all relevant variables are projected intothe selected K feature vectors

(7) The strong correlation matrix is used for dynamicallyrecursive prediction

(8) The strong correlation variables used dynamic recur-rent prediction and evaluated the risk level using the dynam-ics prediction results of each parameter The assessmentstandard is divided into two levels ultimately determining therisk level of CIP

3 The Proposed Methodology

31 Correlation Analysis of Variables Firstly there is a qual-itative analysis of the correlative variables of key parametersof MHIs and these correlation variables are composed intoa high-dimensional data set Each variable is divided intoa coordinate axis that is equal in distance and parallel toeach other Each axis represents an attribute dimension andthe value of the variable corresponds to the correspondingposition on the axis In this way every chemical processvariable can be divided into a broken line on the parallel axisof n according to its attribute valueThe implementation stepsare described in the following

The associated variables of production key parameters inthe chemical industry MHIs system are combined into thefollowing scoring matrix S

S =[[[[[[[[

11987811 11987812 sdot sdot sdot 1198781p11987821 11987822 sdot sdot sdot 1198782119901 d1198781198991 1198781198992 sdot sdot sdot 119878119899119901

]]]]]]]] (1)

In the plane with the Descartes coordinate system eachrow in the score matrix S 119878119899 = [1198781198991 1198781198992 sdot sdot sdot 119878119899119901] cor-responds to a fold line connected by a k-1-line segment ina parallel coordinate system These data must be convertedfrom the Descartes coordinate system to the parallel coordi-nate system because each coordinate axis is equal in lengththus the scoring matrix is homogenized 0ndash1The relationshipbetween the processed data s1015840119899p and the original data s119899p iss1015840np = (S119899p minus S119899min)(s119899maxminusS119899min) where S119899min and S119899maxare the maximum and minimum values of the correlationvariable respectively

Parallel axes represent the correlation variables of eachkey parameter in MHIs which are connected by the param-eter values of these correlation variables If the line between

the two variables is crossed to show the X type then the twovariables are negatively correlated If the line between the twovariables is parallel then the two variables indicate positivecorrelation If the lines between variables are randomlycrossed then no unique relationship exists between the twovariables

Then quantitative analyzed the correlation variables ofkey parameters of MHIs [33] Strong and medium correla-tion variables are further divided to verify the visualizationmethod The correlation coefficient is reflected in the symbolldquorrdquo and the Pearson correlation coefficient is as follows

rAB = 1119899 minus 1119899sum119894=1

(119860 119894 minus Avg (A)120575A )(B119894 minus Avg (B)120575B ) (2)

where n is equal to the number of selected data itemsAvg(A) and Avg(B) are the average values of selected dataon attributes A and B respectively and 120575A 120575119861 are theircorresponding standard deviations

32 Feature Analysis and Feature Vector Acquisition Con-cerning the majority of chemical processes the key param-eters have numerous strong correlation variables Thesestrongly correlation variables constitute a high-dimensionaldata which must be reduced in order to improve theefficiency and accuracy of prediction of key parameters [34]n-dimensional features to K dimensions (kltn) The steps areas follows

(i) The strong correlation variables are selected fromvariable correlation analysis constituting the data matrix X

119883 =[[[[[[[[

x11 x12 sdot sdot sdot x1mx21 x22 sdot sdot sdot x2m d

xn1 xn2 sdot sdot sdot xnm

]]]]]]]] (3)

(ii)The dimensionality of each data matrix dimension hasa serious influence on the principal component which maycause a considerable difference between the number and thevalue of the principal component and the actual situationThe strong correlation variables constitute the data matrix Xand each variable is standardized reprocessed to obtained thestandardized data matrix Z znm = (xnmminusxn)Y2mThe overallmean of 119909119898 is xm = (1n) sumn

n=1 xnm and the total varianceof 119909119898 is Y2m = (1(n minus 1))sumn

n=1(xnm minus xn)2 The correlationmatrix C = (1n)ZTZ is calculated in which ZT is thetranspose of the standardized data matrix Z n represents therow of matrix Z m represents the column of matrix Z andthe calculated correlation matrix C comprises the D principalcomponent

(iii) Jacobirsquos method is employed to find the eigenvectorwi i = 1 2 m of the correlation matrix C and eigenvaluematrix Λ = diag(1205821 1205822 120582m) where 120582 is the eigenvalueand diag represents a diagonal matrix

(iv) The eigenvalues are arranged in order from large tosmall 1205821 gt 1205822 gt sdot sdot sdot gt 120582m The order of the eigenvectorcolumns is adjusted correspondingly allowing the maximum

Complexity 5

variance to be the first principal components the sublargevariance as the second principal components and the leastvariance as the D principal component

(v)The K principal component of the maximum varianceis selected to allow the K principal component to containmost of the original data information The cumulative vari-ance contribution of the commonly selected K principalcomponents is 85 larger than that of the total variance Thatis sumk

j=1 120582jsumdj=1 120582j ge 85

(vi) By selecting the eigenvectors wi i = 1 2 k of Kprincipal components k independent linear combination ofnew variables is obtained

1205851 = 119909111990811 + 119909211990821 + sdot sdot sdot + 11990911988911990811988911205852 = 119909111990812 + 119909211990822 + sdot sdot sdot + 1199091198891199081198892120585119896 = 11990911199081119896 + 11990921199082119896 + sdot sdot sdot + 119909119889119908119889119896

(4)

(vii) 1205851 1205852 120585119896 is the rebuilt K feature after the reduc-tion of dimension These features represent the principalcomponents of all strong correlation variables and thematrixT of the K principal components of the strong correlationvariable is as follows

119879 = [[[[[[[

x111990811 119909211990821 sdot sdot sdot 1199091198891199081198891119909111990812 119909211990822 sdot sdot sdot 1199091198891199081198892 d11990911199081119896 11990921199082119896 sdot sdot sdot 119909119889119908119889119896

]]]]]]] (5)

33 Dynamic Recurrent Prediction In the prediction modelin traditional neural network we add a state layer to constructthe dynamic recurrent network As in Figure 2 which showsfour layers namely input layer hidden layer state layer andoutput layer [35]

The mathematical model is as follows

119909 (119905) = 119891 (119908(1)119909119888 (119905) + 119908(2)119906 (119905 minus 1) + 120579(1)) 119909119888 (119905) = 119909 (119905 minus 1) y (t) = w(3)x (t) + 120579(2)

(6)

where y(t) is the output of the output layer u(t) is theexternal input of the input layer the matrix T of the Kprincipal component of the strong correlation variable isinput xc(t) is the output of the state layer x(t) is the outputof the hidden layer w(1) is the connection weight of the statelayer and the hidden layer and w(2) is the connection weightbetween the input layer and the hidden layer w(3) is theconnection weight of the hidden layer and the output layerthe 120579(1) is the hidden layer threshold and the 120579(2) is the outputthreshold value where 119891() represents the activation function

y(t)x(t)

hidden output layer

Input layer

state layer

u(t-1) layer

w(2)

w(3)

w(1)

Ｒ＝(t)

Figure 2 Recurrent neural network structures

of the neural network using the Sigmoid function presentedin formula (7)

f (x) = 11 + exp [minusx] 0 lt 119891 (x) lt 1 (7)

Formula (6) can be introduced as follows

xc (t) = f (w(1)kminus1xc (t minus 1) + w(2)kminus2xc (t minus 2)) (8)

where w(1)kminus1 and w(2)kminus2 represent the connection weight atprevious different times Formula (8) indicates that xc(t) isrelated to the weight of the connection at the previous timeand realizes the characteristic of dynamic recursion [36]

The input and output of input layer are respectively

I(1)120572 (t) = u120572 (t minus 1) (9)

O(1)120572 (t) = I(1)120572 (t) (10)

120572 = [1 2 sdot sdot sdot 1198641] where the input and output of thehidden layer are respectively

I(2)120573 (t) = E1sum120572=1

w(2)120573120572O(1)120572 (t) + E2sum

120573=1

w(1)120573120574O(c)120574 (t) + 120579(1) (11)

O(2)120573 (t) = f (I(2)120573 (t)) = 11 + 119890I(2)120573 (t) (12)

120573 = [1 2 sdot sdot sdot 1198642] The input and output of the statelayer are respectively as follows

I(C)120574 (t) = O(2)120573 (t minus 1) (13)

O(C)120574 (t) = I(C)120574 (t) = x(C) (t) (14)

6 Complexity

Valu

e y(t)

Time

G high alarm value

=

expert evaluating method

Second-level fuzzy comprehensive evaluation

Major hazard

Key parameter

First-level fuzzy comprehensive

evaluation

11 11

Major hazard installations 2 2

installations 1 1

Major hazard installations 3 3

Major hazard installations p p

Key parameter

21 22

Key parameter

31 33

Key parameter

11 pressure

p3 pressure 3

p2 pressure 2 p1 pressure 1

12 temperature

21 temperature31 temperature 1

32 temperature 2

33 temperature 3

13 level

23 level 2

22 level 1

11 pressure

22 pressure 33

pressure

pp pressure

Ｓ(Ｎ)minus

times 100

ppp1

１11

１12

１13

１21

１22

１23１32

１33

１31

11

12

13

21

22

23 31

33

32

＄1

＄1

＄1

＄1

B1

B2

B3

Bp

＄2

＄p

＄2

＄2 ＄2 ＄3

＄3

＄3

＄3

p = Apmiddot Dp

50⩽R lt 100

10⩽ R lt 50

R ⩽ 10

A11

A22

１11

１22

A33

１33

１p3

１p2

１p1

R ⩾ 100

p= p

middot １p

R = 1+2+3+ p

＄p Ap3

＄p Ap2

＄p Ap1

１ppApp

p

I Disastrous

II High risk

III Moderate risk

IV Low risk

Figure 3 Hierarchical chart of dynamic assessment of MHIs

120574 = [1 2 sdot sdot sdot 1198642] The input and output of the outputlayer are respectively

I(3)120583 (t) = s2sumj=1w(3)120583120573O

(2)120573 (t) + 120579(2) (15)

y(120583) (t) = O(3)120583 (t) = g ( I(3)120583 (t)) (16)

where 120583 = [1 2 sdot sdot sdot 1198643] Among them 1198641 1198642 and 1198643are the input hidden and output layers respectively and thenumber of layers of the state layer is the same as that of thehidden layer The final y(t) is the predictive value of the keyparameters

34 Dynamic Risk Assessment and Early Warning Fuzzycomprehensive evaluation is a comprehensive decision-making methodology of a multivariable problem-solvingcomplex decision process [37] It is developed from fuzzy

sets We through the first-level and second-level fuzzy com-prehensive evaluation to determine the risk R value of CIPHierarchical chart is presented in Figure 3 The specific stepsare as follows

Step 1 (construction of second-level fuzzy comprehensiveevaluation) The second-level fuzzy comprehensive eval-uation defines an evaluation matrix of key parametersQ119901120590119901 = [Q11205901 Q21205902 Q119901120590119901 ] 119901 is the number of MHIsand 1205901 1205902 120590119901 is the number of key parameters for eachMHIs they may be different values The percentage of thepredicted value of the key parameter y(t) exceeding the highalarm value G C119901120590119901 = (y(t) minus G)G times 100 multiplies thecorresponding elements between Q119901120590119901 and C119901120590119901 Then thesecond-level fuzzy comprehensive evaluation index matrix isobtained

A119901120590119901 = C119901120590119901 sdot Q119901120590119901 (17)

Complexity 7

Air

Compressor

oxygennitrogen

Choke valve

Air precooling system

Air purification system

Filter

Turboexpander

Air separation tower system

Air

Figure 4 Air separation distillation section

Table 1 Dynamic risk level

comment grade rangeDisastrous I R ge 100High risk II 50 le R lt 100Moderate risk III 10 le R lt 50Low risk IV R lt 10Step 2 (construction of the first-level fuzzy comprehensiveevaluation) According to the risk degree of each key param-eter in MHIs the weight of the first-level fuzzy comprehen-sive evaluation is determined D119901 = [D1D2 D119901] Thesecond-level fuzzy comprehensive evaluation matrix A119901120590119901 =[A11205901 A21205902 A119901120590119901 ] obtained by (6) and the weight of thefirst-level fuzzy comprehensive evaluation are conducted bypoint multiplication and the risk grade index of the MHIs inthe CIP is obtained as follows

B119901 = A119901120590119901 sdot D119901 (18)

where 119901 is the total number of MHIs in CIP

Finally the dynamic level of MHIs in CIP is determinedby the evaluation result The score of dynamic evaluation isdecided by the scoring method and the range of the grade isprovided in Table 1 R which is specified in Identification andControl of MHIs (GB18218) is used as a grading index [38]Here R obtains the value of hazard installations dynamicindex B119901The historical factors ofMHIsmay affect the overallrisk The data of all dynamic risk grades are derived fromthe last time the true value of the historical key parameterexceeds the high alarm value to return to the safety value orthe current value

Step 3 The final dynamic assessment of MHIs and earlywarning results are used to urge security personnel to excludethe potential risks of MHIs for the first time

4 Practical Application

CIP is composed of many MHIs the air separation distil-lation section is one of the typical representatives of MHIsThis practical application relies on the industrial data of alarge CIP in Quzhou City from July 2017 and a series ofmodel establishment correlation analysis prediction andassessment is completed to realize the dynamic early warningfunction of MHIs in the CIP

This practical application selects an industry of air sep-aration distillation section in the CIP as a dynamic earlywarning object of MHIs [39] a simplified graphic sketchof the air separation distillation section in Figure 4 Thissection simplifies and removes some industrial process fromthe air separation distillation section To make it easier tounderstand the industrial process we are based on data-driven method and can neglect to describe the mechanismof the chemical process industry The section can separateair into oxygen and nitrogen Risk degree of the section isaffected by several key links such as air separation towersystem air purification system and air precooling systemThese production lines are strictly dependent on pressureand temperature if pressure and temperature are abnormalthey may result in an explosion and fire Air separationtower system consists of Turboexpander and distillationcolumn etcWe removed several weakly parameters and linksfor ensuring factories data privacy but did not affect ourapplication result We take liquid air level in lower columnLI101 as the key parameter in air separation distillation

8 Complexity

Table 2

item DescriptionPI101 Air intake chamber pressure (MPa)PI102 Nitrogen Outlet Tower Pressure (MPa)PI103 Oxygen-enriched air outlet pressure (MPa)PI1201 Air outlet left adsorption bucket pressure (MPa)PI1202 Air outlet right adsorption barrel pressure(MPa)PI1203 Instrument air pressure (MPa)PI01 Pressure of air outlet main heat exchanger(MPa)PI02 Lower column pressure (MPa)PI03 Evaporator condenser pressure (MPa)PDI01 The resistance of lower column (MPa)PI401 Turboexpander1 Intake pressure (MPa)PI402 Turboexpander1Exhaust pressure (MPa)PI403 Turboexpander2Intake pressure (MPa)PI404 Turboexpander2Exhaust pressure (MPa)PI2001 Turboexpander1Bearing air pressure (MPa)PI2002 Turboexpander2 Bearing air pressure (MPa)PI2003 Turboexpander1Sealing pressure (MPa)PI2004 Turboexpander2Sealing pressure (MPa)TI101 Air temperature before entering tower (∘C)TI102 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI103 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI1201 Air outlet left adsorption bucket temperature (∘C)TI1202 Air outlet right adsorption bucket temperature (∘C)TE1204 Temperature of regenerated gas outlet adsorption bucket (∘C)TI01 Temperature of main air outlet heat exchanger (∘C)TI02 Turboexpander1 Intake temperature (∘C)TI03 Turboexpander2 Intake temperature (∘C)TI04 Turboexpander1 Exhaust temperature (∘C)TI05 Turboexpander2 Exhaust temperature (∘C)TI07 Exit temperature of precooler (∘C)FI1201 Regenerative gas flow rate (Nm3h)FI101 Air flow rate (Nm3h)FI102 Nitrogen flow rate (Nm3h)LI101 liquid air level in lower column (mm)LI01 Water cooled liquid level (mm)LI02 Liquid oxygen level in the main condenser (mm)SI402 Expander speed (RPM)LV02 Liquid-air throttle valve ()

section Since most of the top feed comes from the bottomof the tower the working conditions at the bottom of thetower play a decisive role in the air separation distillationsection The abnormal liquid air level in lower column willresult in the decrease of oxygen purity which poses a seriousthreat to the normal production of the system The sectionhas 38 liquid levels pressure temperature and other sensorvariables as shown in Table 2 in which the liquid air level inlower columnof the key parameter is LI101 In order to predictLI101 we need to find the strong correlation variables in 38variables The change trend of LI101 is predicted by the subtlechanges in these strong correlation variables It should benoted that LI101 is only one of many key parameters analysis

process of other key parameters is similar enough to analysisprocess of LI101 so that they are not repeated here

The 38 variables of air separation distillation section areanalyzed by visual correlation analysis The experimentaldatasets comprise 1000 sets as shown in Figure 5 The valueof LI101 exceeds the high alarm value when the broken linebetween two adjacent variables is red Conversely the blueline indicates that the LI101 is not exceeding the high alarmvalue

The 38 variables are not convenient to observe Soanalyze the correlation between 22 variables on the localgraph As shown in Figure 6 the lines between LI101 andadjacent variables PI404 and PI403 are randomly crossed

Complexity 9

Original dataset Parallel Coordinates

0700

PI1201

PI1202

PI1203PI02PI03

PI01PI2001PI2002

LI01LI02

LV02

TI01TI02TI03TI04TI05

TI101

TI103

PI101

LI101TI102FI102PI102

FI101

TI07SI402PI103

PI2003PI2004

PI401PI402PI403PI404

PDI01

TI1201

FI1201

TI1202

TE1204

1920minus13890 830

0 0 0 0 0 0 0 01200

1490000

001

1490

003

010002

002002

003009008006010290866289118260

15560153601685018550106

000078

07420040

114391068033010

1073055033041038073062063027030

40000109679162507840

126309280

1597010000

24189802501450

0093990073

30527386010003430

136528033

40802109

Figure 5 Visual correlation analysis of 38 variables

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0078010 0015914 002 002 0 0023 0089 0064078830

074 20040 114391 069 033 01 055 033 041 03810 073 062 063 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

LI101

Figure 6 Visual correlation analysis local graph A

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Complexity 3

All correlation variables of key parameters

of chemical industry MHIsComposition matrix S

Dynamic risk assessment

output dynamic risk value of MHIs

Variable correlation analysisCarry out (a)Visual qualitative analysis(b) Quantitative

analysis of correlation

Inputmatrix S

OutputMatrix X consisting of strongly correlated variables

Dimension reduction optimizationAfter (a) feature analysis (b) feature vector acquisition

Inputmatrix X

Outputmatrix T after dimension reduction

Dynamic recurrent predictionA recurrent neural network structure with State layer is constructed

Inputmatrix T

OutputPredicted values of key parameters

Constructing two-level fuzzy comprehensive evaluation to

Figure 1 Structure block diagram of dynamic early warning method

the asymmetric correlation model of anomalous number ofevents considering the correlation among the four teamsUnder the multivariable condition of chemical productionengineering the process of establishing Copula function isconsiderably complicated and a fast and accurate methodof correlation analysis is necessary The visualization methodcan be used for fast qualitative analysis of relevant variablesand combined with quantitative analysis method effectivelyimplement the selection of strong correlation variables of keyparameter

The structural block diagram of the dynamic early warn-ing method is presented in Figure 1 The first part of thealgorithm is variable correlation analysis the second part isfeature analysis and feature vector acquisition and the thirdpart is dynamic recurrent prediction And the last part is thedynamic risk assessment The implementation process of thealgorithm is as follows

(1) All correlation variables of the key parameters ofMHIsin the CIP use the qualitative analysis of variable correlationby data visualization [32]

4 Complexity

(2) The average value and standard deviation of keyparameters and related variables are calculated

(3) The correlation coefficient r value between two vari-ables is obtained and the strong correlation variable (4)The covariance matrix of the strong correlation variable isalso calculated and its eigenvalue and the eigenvector areobtained

(5) The order of the eigenvalue from large to smallis selected and the maximum is chosen according to theestablished rule K eigenvalue constitutes the correspondingeigenvector matrix

(6)The samples of all relevant variables are projected intothe selected K feature vectors

(7) The strong correlation matrix is used for dynamicallyrecursive prediction

(8) The strong correlation variables used dynamic recur-rent prediction and evaluated the risk level using the dynam-ics prediction results of each parameter The assessmentstandard is divided into two levels ultimately determining therisk level of CIP

3 The Proposed Methodology

31 Correlation Analysis of Variables Firstly there is a qual-itative analysis of the correlative variables of key parametersof MHIs and these correlation variables are composed intoa high-dimensional data set Each variable is divided intoa coordinate axis that is equal in distance and parallel toeach other Each axis represents an attribute dimension andthe value of the variable corresponds to the correspondingposition on the axis In this way every chemical processvariable can be divided into a broken line on the parallel axisof n according to its attribute valueThe implementation stepsare described in the following

The associated variables of production key parameters inthe chemical industry MHIs system are combined into thefollowing scoring matrix S

S =[[[[[[[[

11987811 11987812 sdot sdot sdot 1198781p11987821 11987822 sdot sdot sdot 1198782119901 d1198781198991 1198781198992 sdot sdot sdot 119878119899119901

]]]]]]]] (1)

In the plane with the Descartes coordinate system eachrow in the score matrix S 119878119899 = [1198781198991 1198781198992 sdot sdot sdot 119878119899119901] cor-responds to a fold line connected by a k-1-line segment ina parallel coordinate system These data must be convertedfrom the Descartes coordinate system to the parallel coordi-nate system because each coordinate axis is equal in lengththus the scoring matrix is homogenized 0ndash1The relationshipbetween the processed data s1015840119899p and the original data s119899p iss1015840np = (S119899p minus S119899min)(s119899maxminusS119899min) where S119899min and S119899maxare the maximum and minimum values of the correlationvariable respectively

Parallel axes represent the correlation variables of eachkey parameter in MHIs which are connected by the param-eter values of these correlation variables If the line between

the two variables is crossed to show the X type then the twovariables are negatively correlated If the line between the twovariables is parallel then the two variables indicate positivecorrelation If the lines between variables are randomlycrossed then no unique relationship exists between the twovariables

Then quantitative analyzed the correlation variables ofkey parameters of MHIs [33] Strong and medium correla-tion variables are further divided to verify the visualizationmethod The correlation coefficient is reflected in the symbolldquorrdquo and the Pearson correlation coefficient is as follows

rAB = 1119899 minus 1119899sum119894=1

(119860 119894 minus Avg (A)120575A )(B119894 minus Avg (B)120575B ) (2)

where n is equal to the number of selected data itemsAvg(A) and Avg(B) are the average values of selected dataon attributes A and B respectively and 120575A 120575119861 are theircorresponding standard deviations

32 Feature Analysis and Feature Vector Acquisition Con-cerning the majority of chemical processes the key param-eters have numerous strong correlation variables Thesestrongly correlation variables constitute a high-dimensionaldata which must be reduced in order to improve theefficiency and accuracy of prediction of key parameters [34]n-dimensional features to K dimensions (kltn) The steps areas follows

(i) The strong correlation variables are selected fromvariable correlation analysis constituting the data matrix X

119883 =[[[[[[[[

x11 x12 sdot sdot sdot x1mx21 x22 sdot sdot sdot x2m d

xn1 xn2 sdot sdot sdot xnm

]]]]]]]] (3)

(ii)The dimensionality of each data matrix dimension hasa serious influence on the principal component which maycause a considerable difference between the number and thevalue of the principal component and the actual situationThe strong correlation variables constitute the data matrix Xand each variable is standardized reprocessed to obtained thestandardized data matrix Z znm = (xnmminusxn)Y2mThe overallmean of 119909119898 is xm = (1n) sumn

n=1 xnm and the total varianceof 119909119898 is Y2m = (1(n minus 1))sumn

n=1(xnm minus xn)2 The correlationmatrix C = (1n)ZTZ is calculated in which ZT is thetranspose of the standardized data matrix Z n represents therow of matrix Z m represents the column of matrix Z andthe calculated correlation matrix C comprises the D principalcomponent

(iii) Jacobirsquos method is employed to find the eigenvectorwi i = 1 2 m of the correlation matrix C and eigenvaluematrix Λ = diag(1205821 1205822 120582m) where 120582 is the eigenvalueand diag represents a diagonal matrix

(iv) The eigenvalues are arranged in order from large tosmall 1205821 gt 1205822 gt sdot sdot sdot gt 120582m The order of the eigenvectorcolumns is adjusted correspondingly allowing the maximum

Complexity 5

variance to be the first principal components the sublargevariance as the second principal components and the leastvariance as the D principal component

(v)The K principal component of the maximum varianceis selected to allow the K principal component to containmost of the original data information The cumulative vari-ance contribution of the commonly selected K principalcomponents is 85 larger than that of the total variance Thatis sumk

j=1 120582jsumdj=1 120582j ge 85

(vi) By selecting the eigenvectors wi i = 1 2 k of Kprincipal components k independent linear combination ofnew variables is obtained

1205851 = 119909111990811 + 119909211990821 + sdot sdot sdot + 11990911988911990811988911205852 = 119909111990812 + 119909211990822 + sdot sdot sdot + 1199091198891199081198892120585119896 = 11990911199081119896 + 11990921199082119896 + sdot sdot sdot + 119909119889119908119889119896

(4)

(vii) 1205851 1205852 120585119896 is the rebuilt K feature after the reduc-tion of dimension These features represent the principalcomponents of all strong correlation variables and thematrixT of the K principal components of the strong correlationvariable is as follows

119879 = [[[[[[[

x111990811 119909211990821 sdot sdot sdot 1199091198891199081198891119909111990812 119909211990822 sdot sdot sdot 1199091198891199081198892 d11990911199081119896 11990921199082119896 sdot sdot sdot 119909119889119908119889119896

]]]]]]] (5)

33 Dynamic Recurrent Prediction In the prediction modelin traditional neural network we add a state layer to constructthe dynamic recurrent network As in Figure 2 which showsfour layers namely input layer hidden layer state layer andoutput layer [35]

The mathematical model is as follows

119909 (119905) = 119891 (119908(1)119909119888 (119905) + 119908(2)119906 (119905 minus 1) + 120579(1)) 119909119888 (119905) = 119909 (119905 minus 1) y (t) = w(3)x (t) + 120579(2)

(6)

where y(t) is the output of the output layer u(t) is theexternal input of the input layer the matrix T of the Kprincipal component of the strong correlation variable isinput xc(t) is the output of the state layer x(t) is the outputof the hidden layer w(1) is the connection weight of the statelayer and the hidden layer and w(2) is the connection weightbetween the input layer and the hidden layer w(3) is theconnection weight of the hidden layer and the output layerthe 120579(1) is the hidden layer threshold and the 120579(2) is the outputthreshold value where 119891() represents the activation function

y(t)x(t)

hidden output layer

Input layer

state layer

u(t-1) layer

w(2)

w(3)

w(1)

Ｒ＝(t)

Figure 2 Recurrent neural network structures

of the neural network using the Sigmoid function presentedin formula (7)

f (x) = 11 + exp [minusx] 0 lt 119891 (x) lt 1 (7)

Formula (6) can be introduced as follows

xc (t) = f (w(1)kminus1xc (t minus 1) + w(2)kminus2xc (t minus 2)) (8)

where w(1)kminus1 and w(2)kminus2 represent the connection weight atprevious different times Formula (8) indicates that xc(t) isrelated to the weight of the connection at the previous timeand realizes the characteristic of dynamic recursion [36]

The input and output of input layer are respectively

I(1)120572 (t) = u120572 (t minus 1) (9)

O(1)120572 (t) = I(1)120572 (t) (10)

120572 = [1 2 sdot sdot sdot 1198641] where the input and output of thehidden layer are respectively

I(2)120573 (t) = E1sum120572=1

w(2)120573120572O(1)120572 (t) + E2sum

120573=1

w(1)120573120574O(c)120574 (t) + 120579(1) (11)

O(2)120573 (t) = f (I(2)120573 (t)) = 11 + 119890I(2)120573 (t) (12)

120573 = [1 2 sdot sdot sdot 1198642] The input and output of the statelayer are respectively as follows

I(C)120574 (t) = O(2)120573 (t minus 1) (13)

O(C)120574 (t) = I(C)120574 (t) = x(C) (t) (14)

6 Complexity

Valu

e y(t)

Time

G high alarm value

=

expert evaluating method

Second-level fuzzy comprehensive evaluation

Major hazard

Key parameter

First-level fuzzy comprehensive

evaluation

11 11

Major hazard installations 2 2

installations 1 1

Major hazard installations 3 3

Major hazard installations p p

Key parameter

21 22

Key parameter

31 33

Key parameter

11 pressure

p3 pressure 3

p2 pressure 2 p1 pressure 1

12 temperature

21 temperature31 temperature 1

32 temperature 2

33 temperature 3

13 level

23 level 2

22 level 1

11 pressure

22 pressure 33

pressure

pp pressure

Ｓ(Ｎ)minus

times 100

ppp1

１11

１12

１13

１21

１22

１23１32

１33

１31

11

12

13

21

22

23 31

33

32

＄1

＄1

＄1

＄1

B1

B2

B3

Bp

＄2

＄p

＄2

＄2 ＄2 ＄3

＄3

＄3

＄3

p = Apmiddot Dp

50⩽R lt 100

10⩽ R lt 50

R ⩽ 10

A11

A22

１11

１22

A33

１33

１p3

１p2

１p1

R ⩾ 100

p= p

middot １p

R = 1+2+3+ p

＄p Ap3

＄p Ap2

＄p Ap1

１ppApp

p

I Disastrous

II High risk

III Moderate risk

IV Low risk

Figure 3 Hierarchical chart of dynamic assessment of MHIs

120574 = [1 2 sdot sdot sdot 1198642] The input and output of the outputlayer are respectively

I(3)120583 (t) = s2sumj=1w(3)120583120573O

(2)120573 (t) + 120579(2) (15)

y(120583) (t) = O(3)120583 (t) = g ( I(3)120583 (t)) (16)

where 120583 = [1 2 sdot sdot sdot 1198643] Among them 1198641 1198642 and 1198643are the input hidden and output layers respectively and thenumber of layers of the state layer is the same as that of thehidden layer The final y(t) is the predictive value of the keyparameters

34 Dynamic Risk Assessment and Early Warning Fuzzycomprehensive evaluation is a comprehensive decision-making methodology of a multivariable problem-solvingcomplex decision process [37] It is developed from fuzzy

sets We through the first-level and second-level fuzzy com-prehensive evaluation to determine the risk R value of CIPHierarchical chart is presented in Figure 3 The specific stepsare as follows

Step 1 (construction of second-level fuzzy comprehensiveevaluation) The second-level fuzzy comprehensive eval-uation defines an evaluation matrix of key parametersQ119901120590119901 = [Q11205901 Q21205902 Q119901120590119901 ] 119901 is the number of MHIsand 1205901 1205902 120590119901 is the number of key parameters for eachMHIs they may be different values The percentage of thepredicted value of the key parameter y(t) exceeding the highalarm value G C119901120590119901 = (y(t) minus G)G times 100 multiplies thecorresponding elements between Q119901120590119901 and C119901120590119901 Then thesecond-level fuzzy comprehensive evaluation index matrix isobtained

A119901120590119901 = C119901120590119901 sdot Q119901120590119901 (17)

Complexity 7

Air

Compressor

oxygennitrogen

Choke valve

Air precooling system

Air purification system

Filter

Turboexpander

Air separation tower system

Air

Figure 4 Air separation distillation section

Table 1 Dynamic risk level

comment grade rangeDisastrous I R ge 100High risk II 50 le R lt 100Moderate risk III 10 le R lt 50Low risk IV R lt 10Step 2 (construction of the first-level fuzzy comprehensiveevaluation) According to the risk degree of each key param-eter in MHIs the weight of the first-level fuzzy comprehen-sive evaluation is determined D119901 = [D1D2 D119901] Thesecond-level fuzzy comprehensive evaluation matrix A119901120590119901 =[A11205901 A21205902 A119901120590119901 ] obtained by (6) and the weight of thefirst-level fuzzy comprehensive evaluation are conducted bypoint multiplication and the risk grade index of the MHIs inthe CIP is obtained as follows

B119901 = A119901120590119901 sdot D119901 (18)

where 119901 is the total number of MHIs in CIP

Finally the dynamic level of MHIs in CIP is determinedby the evaluation result The score of dynamic evaluation isdecided by the scoring method and the range of the grade isprovided in Table 1 R which is specified in Identification andControl of MHIs (GB18218) is used as a grading index [38]Here R obtains the value of hazard installations dynamicindex B119901The historical factors ofMHIsmay affect the overallrisk The data of all dynamic risk grades are derived fromthe last time the true value of the historical key parameterexceeds the high alarm value to return to the safety value orthe current value

Step 3 The final dynamic assessment of MHIs and earlywarning results are used to urge security personnel to excludethe potential risks of MHIs for the first time

4 Practical Application

CIP is composed of many MHIs the air separation distil-lation section is one of the typical representatives of MHIsThis practical application relies on the industrial data of alarge CIP in Quzhou City from July 2017 and a series ofmodel establishment correlation analysis prediction andassessment is completed to realize the dynamic early warningfunction of MHIs in the CIP

This practical application selects an industry of air sep-aration distillation section in the CIP as a dynamic earlywarning object of MHIs [39] a simplified graphic sketchof the air separation distillation section in Figure 4 Thissection simplifies and removes some industrial process fromthe air separation distillation section To make it easier tounderstand the industrial process we are based on data-driven method and can neglect to describe the mechanismof the chemical process industry The section can separateair into oxygen and nitrogen Risk degree of the section isaffected by several key links such as air separation towersystem air purification system and air precooling systemThese production lines are strictly dependent on pressureand temperature if pressure and temperature are abnormalthey may result in an explosion and fire Air separationtower system consists of Turboexpander and distillationcolumn etcWe removed several weakly parameters and linksfor ensuring factories data privacy but did not affect ourapplication result We take liquid air level in lower columnLI101 as the key parameter in air separation distillation

8 Complexity

Table 2

item DescriptionPI101 Air intake chamber pressure (MPa)PI102 Nitrogen Outlet Tower Pressure (MPa)PI103 Oxygen-enriched air outlet pressure (MPa)PI1201 Air outlet left adsorption bucket pressure (MPa)PI1202 Air outlet right adsorption barrel pressure(MPa)PI1203 Instrument air pressure (MPa)PI01 Pressure of air outlet main heat exchanger(MPa)PI02 Lower column pressure (MPa)PI03 Evaporator condenser pressure (MPa)PDI01 The resistance of lower column (MPa)PI401 Turboexpander1 Intake pressure (MPa)PI402 Turboexpander1Exhaust pressure (MPa)PI403 Turboexpander2Intake pressure (MPa)PI404 Turboexpander2Exhaust pressure (MPa)PI2001 Turboexpander1Bearing air pressure (MPa)PI2002 Turboexpander2 Bearing air pressure (MPa)PI2003 Turboexpander1Sealing pressure (MPa)PI2004 Turboexpander2Sealing pressure (MPa)TI101 Air temperature before entering tower (∘C)TI102 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI103 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI1201 Air outlet left adsorption bucket temperature (∘C)TI1202 Air outlet right adsorption bucket temperature (∘C)TE1204 Temperature of regenerated gas outlet adsorption bucket (∘C)TI01 Temperature of main air outlet heat exchanger (∘C)TI02 Turboexpander1 Intake temperature (∘C)TI03 Turboexpander2 Intake temperature (∘C)TI04 Turboexpander1 Exhaust temperature (∘C)TI05 Turboexpander2 Exhaust temperature (∘C)TI07 Exit temperature of precooler (∘C)FI1201 Regenerative gas flow rate (Nm3h)FI101 Air flow rate (Nm3h)FI102 Nitrogen flow rate (Nm3h)LI101 liquid air level in lower column (mm)LI01 Water cooled liquid level (mm)LI02 Liquid oxygen level in the main condenser (mm)SI402 Expander speed (RPM)LV02 Liquid-air throttle valve ()

section Since most of the top feed comes from the bottomof the tower the working conditions at the bottom of thetower play a decisive role in the air separation distillationsection The abnormal liquid air level in lower column willresult in the decrease of oxygen purity which poses a seriousthreat to the normal production of the system The sectionhas 38 liquid levels pressure temperature and other sensorvariables as shown in Table 2 in which the liquid air level inlower columnof the key parameter is LI101 In order to predictLI101 we need to find the strong correlation variables in 38variables The change trend of LI101 is predicted by the subtlechanges in these strong correlation variables It should benoted that LI101 is only one of many key parameters analysis

process of other key parameters is similar enough to analysisprocess of LI101 so that they are not repeated here

The 38 variables of air separation distillation section areanalyzed by visual correlation analysis The experimentaldatasets comprise 1000 sets as shown in Figure 5 The valueof LI101 exceeds the high alarm value when the broken linebetween two adjacent variables is red Conversely the blueline indicates that the LI101 is not exceeding the high alarmvalue

The 38 variables are not convenient to observe Soanalyze the correlation between 22 variables on the localgraph As shown in Figure 6 the lines between LI101 andadjacent variables PI404 and PI403 are randomly crossed

Complexity 9

Original dataset Parallel Coordinates

0700

PI1201

PI1202

PI1203PI02PI03

PI01PI2001PI2002

LI01LI02

LV02

TI01TI02TI03TI04TI05

TI101

TI103

PI101

LI101TI102FI102PI102

FI101

TI07SI402PI103

PI2003PI2004

PI401PI402PI403PI404

PDI01

TI1201

FI1201

TI1202

TE1204

1920minus13890 830

0 0 0 0 0 0 0 01200

1490000

001

1490

003

010002

002002

003009008006010290866289118260

15560153601685018550106

000078

07420040

114391068033010

1073055033041038073062063027030

40000109679162507840

126309280

1597010000

24189802501450

0093990073

30527386010003430

136528033

40802109

Figure 5 Visual correlation analysis of 38 variables

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0078010 0015914 002 002 0 0023 0089 0064078830

074 20040 114391 069 033 01 055 033 041 03810 073 062 063 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

LI101

Figure 6 Visual correlation analysis local graph A

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

4 Complexity

(2) The average value and standard deviation of keyparameters and related variables are calculated

(3) The correlation coefficient r value between two vari-ables is obtained and the strong correlation variable (4)The covariance matrix of the strong correlation variable isalso calculated and its eigenvalue and the eigenvector areobtained

(5) The order of the eigenvalue from large to smallis selected and the maximum is chosen according to theestablished rule K eigenvalue constitutes the correspondingeigenvector matrix

(6)The samples of all relevant variables are projected intothe selected K feature vectors

(7) The strong correlation matrix is used for dynamicallyrecursive prediction

(8) The strong correlation variables used dynamic recur-rent prediction and evaluated the risk level using the dynam-ics prediction results of each parameter The assessmentstandard is divided into two levels ultimately determining therisk level of CIP

3 The Proposed Methodology

31 Correlation Analysis of Variables Firstly there is a qual-itative analysis of the correlative variables of key parametersof MHIs and these correlation variables are composed intoa high-dimensional data set Each variable is divided intoa coordinate axis that is equal in distance and parallel toeach other Each axis represents an attribute dimension andthe value of the variable corresponds to the correspondingposition on the axis In this way every chemical processvariable can be divided into a broken line on the parallel axisof n according to its attribute valueThe implementation stepsare described in the following

The associated variables of production key parameters inthe chemical industry MHIs system are combined into thefollowing scoring matrix S

S =[[[[[[[[

11987811 11987812 sdot sdot sdot 1198781p11987821 11987822 sdot sdot sdot 1198782119901 d1198781198991 1198781198992 sdot sdot sdot 119878119899119901

]]]]]]]] (1)

In the plane with the Descartes coordinate system eachrow in the score matrix S 119878119899 = [1198781198991 1198781198992 sdot sdot sdot 119878119899119901] cor-responds to a fold line connected by a k-1-line segment ina parallel coordinate system These data must be convertedfrom the Descartes coordinate system to the parallel coordi-nate system because each coordinate axis is equal in lengththus the scoring matrix is homogenized 0ndash1The relationshipbetween the processed data s1015840119899p and the original data s119899p iss1015840np = (S119899p minus S119899min)(s119899maxminusS119899min) where S119899min and S119899maxare the maximum and minimum values of the correlationvariable respectively

Parallel axes represent the correlation variables of eachkey parameter in MHIs which are connected by the param-eter values of these correlation variables If the line between

the two variables is crossed to show the X type then the twovariables are negatively correlated If the line between the twovariables is parallel then the two variables indicate positivecorrelation If the lines between variables are randomlycrossed then no unique relationship exists between the twovariables

Then quantitative analyzed the correlation variables ofkey parameters of MHIs [33] Strong and medium correla-tion variables are further divided to verify the visualizationmethod The correlation coefficient is reflected in the symbolldquorrdquo and the Pearson correlation coefficient is as follows

rAB = 1119899 minus 1119899sum119894=1

(119860 119894 minus Avg (A)120575A )(B119894 minus Avg (B)120575B ) (2)

where n is equal to the number of selected data itemsAvg(A) and Avg(B) are the average values of selected dataon attributes A and B respectively and 120575A 120575119861 are theircorresponding standard deviations

32 Feature Analysis and Feature Vector Acquisition Con-cerning the majority of chemical processes the key param-eters have numerous strong correlation variables Thesestrongly correlation variables constitute a high-dimensionaldata which must be reduced in order to improve theefficiency and accuracy of prediction of key parameters [34]n-dimensional features to K dimensions (kltn) The steps areas follows

(i) The strong correlation variables are selected fromvariable correlation analysis constituting the data matrix X

119883 =[[[[[[[[

x11 x12 sdot sdot sdot x1mx21 x22 sdot sdot sdot x2m d

xn1 xn2 sdot sdot sdot xnm

]]]]]]]] (3)

(ii)The dimensionality of each data matrix dimension hasa serious influence on the principal component which maycause a considerable difference between the number and thevalue of the principal component and the actual situationThe strong correlation variables constitute the data matrix Xand each variable is standardized reprocessed to obtained thestandardized data matrix Z znm = (xnmminusxn)Y2mThe overallmean of 119909119898 is xm = (1n) sumn

n=1 xnm and the total varianceof 119909119898 is Y2m = (1(n minus 1))sumn

n=1(xnm minus xn)2 The correlationmatrix C = (1n)ZTZ is calculated in which ZT is thetranspose of the standardized data matrix Z n represents therow of matrix Z m represents the column of matrix Z andthe calculated correlation matrix C comprises the D principalcomponent

(iii) Jacobirsquos method is employed to find the eigenvectorwi i = 1 2 m of the correlation matrix C and eigenvaluematrix Λ = diag(1205821 1205822 120582m) where 120582 is the eigenvalueand diag represents a diagonal matrix

(iv) The eigenvalues are arranged in order from large tosmall 1205821 gt 1205822 gt sdot sdot sdot gt 120582m The order of the eigenvectorcolumns is adjusted correspondingly allowing the maximum

Complexity 5

variance to be the first principal components the sublargevariance as the second principal components and the leastvariance as the D principal component

(v)The K principal component of the maximum varianceis selected to allow the K principal component to containmost of the original data information The cumulative vari-ance contribution of the commonly selected K principalcomponents is 85 larger than that of the total variance Thatis sumk

j=1 120582jsumdj=1 120582j ge 85

(vi) By selecting the eigenvectors wi i = 1 2 k of Kprincipal components k independent linear combination ofnew variables is obtained

1205851 = 119909111990811 + 119909211990821 + sdot sdot sdot + 11990911988911990811988911205852 = 119909111990812 + 119909211990822 + sdot sdot sdot + 1199091198891199081198892120585119896 = 11990911199081119896 + 11990921199082119896 + sdot sdot sdot + 119909119889119908119889119896

(4)

(vii) 1205851 1205852 120585119896 is the rebuilt K feature after the reduc-tion of dimension These features represent the principalcomponents of all strong correlation variables and thematrixT of the K principal components of the strong correlationvariable is as follows

119879 = [[[[[[[

x111990811 119909211990821 sdot sdot sdot 1199091198891199081198891119909111990812 119909211990822 sdot sdot sdot 1199091198891199081198892 d11990911199081119896 11990921199082119896 sdot sdot sdot 119909119889119908119889119896

]]]]]]] (5)

33 Dynamic Recurrent Prediction In the prediction modelin traditional neural network we add a state layer to constructthe dynamic recurrent network As in Figure 2 which showsfour layers namely input layer hidden layer state layer andoutput layer [35]

The mathematical model is as follows

119909 (119905) = 119891 (119908(1)119909119888 (119905) + 119908(2)119906 (119905 minus 1) + 120579(1)) 119909119888 (119905) = 119909 (119905 minus 1) y (t) = w(3)x (t) + 120579(2)

(6)

where y(t) is the output of the output layer u(t) is theexternal input of the input layer the matrix T of the Kprincipal component of the strong correlation variable isinput xc(t) is the output of the state layer x(t) is the outputof the hidden layer w(1) is the connection weight of the statelayer and the hidden layer and w(2) is the connection weightbetween the input layer and the hidden layer w(3) is theconnection weight of the hidden layer and the output layerthe 120579(1) is the hidden layer threshold and the 120579(2) is the outputthreshold value where 119891() represents the activation function

y(t)x(t)

hidden output layer

Input layer

state layer

u(t-1) layer

w(2)

w(3)

w(1)

Ｒ＝(t)

Figure 2 Recurrent neural network structures

of the neural network using the Sigmoid function presentedin formula (7)

f (x) = 11 + exp [minusx] 0 lt 119891 (x) lt 1 (7)

Formula (6) can be introduced as follows

xc (t) = f (w(1)kminus1xc (t minus 1) + w(2)kminus2xc (t minus 2)) (8)

where w(1)kminus1 and w(2)kminus2 represent the connection weight atprevious different times Formula (8) indicates that xc(t) isrelated to the weight of the connection at the previous timeand realizes the characteristic of dynamic recursion [36]

The input and output of input layer are respectively

I(1)120572 (t) = u120572 (t minus 1) (9)

O(1)120572 (t) = I(1)120572 (t) (10)

120572 = [1 2 sdot sdot sdot 1198641] where the input and output of thehidden layer are respectively

I(2)120573 (t) = E1sum120572=1

w(2)120573120572O(1)120572 (t) + E2sum

120573=1

w(1)120573120574O(c)120574 (t) + 120579(1) (11)

O(2)120573 (t) = f (I(2)120573 (t)) = 11 + 119890I(2)120573 (t) (12)

120573 = [1 2 sdot sdot sdot 1198642] The input and output of the statelayer are respectively as follows

I(C)120574 (t) = O(2)120573 (t minus 1) (13)

O(C)120574 (t) = I(C)120574 (t) = x(C) (t) (14)

6 Complexity

Valu

e y(t)

Time

G high alarm value

=

expert evaluating method

Second-level fuzzy comprehensive evaluation

Major hazard

Key parameter

First-level fuzzy comprehensive

evaluation

11 11

Major hazard installations 2 2

installations 1 1

Major hazard installations 3 3

Major hazard installations p p

Key parameter

21 22

Key parameter

31 33

Key parameter

11 pressure

p3 pressure 3

p2 pressure 2 p1 pressure 1

12 temperature

21 temperature31 temperature 1

32 temperature 2

33 temperature 3

13 level

23 level 2

22 level 1

11 pressure

22 pressure 33

pressure

pp pressure

Ｓ(Ｎ)minus

times 100

ppp1

１11

１12

１13

１21

１22

１23１32

１33

１31

11

12

13

21

22

23 31

33

32

＄1

＄1

＄1

＄1

B1

B2

B3

Bp

＄2

＄p

＄2

＄2 ＄2 ＄3

＄3

＄3

＄3

p = Apmiddot Dp

50⩽R lt 100

10⩽ R lt 50

R ⩽ 10

A11

A22

１11

１22

A33

１33

１p3

１p2

１p1

R ⩾ 100

p= p

middot １p

R = 1+2+3+ p

＄p Ap3

＄p Ap2

＄p Ap1

１ppApp

p

I Disastrous

II High risk

III Moderate risk

IV Low risk

Figure 3 Hierarchical chart of dynamic assessment of MHIs

120574 = [1 2 sdot sdot sdot 1198642] The input and output of the outputlayer are respectively

I(3)120583 (t) = s2sumj=1w(3)120583120573O

(2)120573 (t) + 120579(2) (15)

y(120583) (t) = O(3)120583 (t) = g ( I(3)120583 (t)) (16)

where 120583 = [1 2 sdot sdot sdot 1198643] Among them 1198641 1198642 and 1198643are the input hidden and output layers respectively and thenumber of layers of the state layer is the same as that of thehidden layer The final y(t) is the predictive value of the keyparameters

34 Dynamic Risk Assessment and Early Warning Fuzzycomprehensive evaluation is a comprehensive decision-making methodology of a multivariable problem-solvingcomplex decision process [37] It is developed from fuzzy

sets We through the first-level and second-level fuzzy com-prehensive evaluation to determine the risk R value of CIPHierarchical chart is presented in Figure 3 The specific stepsare as follows

Step 1 (construction of second-level fuzzy comprehensiveevaluation) The second-level fuzzy comprehensive eval-uation defines an evaluation matrix of key parametersQ119901120590119901 = [Q11205901 Q21205902 Q119901120590119901 ] 119901 is the number of MHIsand 1205901 1205902 120590119901 is the number of key parameters for eachMHIs they may be different values The percentage of thepredicted value of the key parameter y(t) exceeding the highalarm value G C119901120590119901 = (y(t) minus G)G times 100 multiplies thecorresponding elements between Q119901120590119901 and C119901120590119901 Then thesecond-level fuzzy comprehensive evaluation index matrix isobtained

A119901120590119901 = C119901120590119901 sdot Q119901120590119901 (17)

Complexity 7

Air

Compressor

oxygennitrogen

Choke valve

Air precooling system

Air purification system

Filter

Turboexpander

Air separation tower system

Air

Figure 4 Air separation distillation section

Table 1 Dynamic risk level

comment grade rangeDisastrous I R ge 100High risk II 50 le R lt 100Moderate risk III 10 le R lt 50Low risk IV R lt 10Step 2 (construction of the first-level fuzzy comprehensiveevaluation) According to the risk degree of each key param-eter in MHIs the weight of the first-level fuzzy comprehen-sive evaluation is determined D119901 = [D1D2 D119901] Thesecond-level fuzzy comprehensive evaluation matrix A119901120590119901 =[A11205901 A21205902 A119901120590119901 ] obtained by (6) and the weight of thefirst-level fuzzy comprehensive evaluation are conducted bypoint multiplication and the risk grade index of the MHIs inthe CIP is obtained as follows

B119901 = A119901120590119901 sdot D119901 (18)

where 119901 is the total number of MHIs in CIP

Finally the dynamic level of MHIs in CIP is determinedby the evaluation result The score of dynamic evaluation isdecided by the scoring method and the range of the grade isprovided in Table 1 R which is specified in Identification andControl of MHIs (GB18218) is used as a grading index [38]Here R obtains the value of hazard installations dynamicindex B119901The historical factors ofMHIsmay affect the overallrisk The data of all dynamic risk grades are derived fromthe last time the true value of the historical key parameterexceeds the high alarm value to return to the safety value orthe current value

Step 3 The final dynamic assessment of MHIs and earlywarning results are used to urge security personnel to excludethe potential risks of MHIs for the first time

4 Practical Application

CIP is composed of many MHIs the air separation distil-lation section is one of the typical representatives of MHIsThis practical application relies on the industrial data of alarge CIP in Quzhou City from July 2017 and a series ofmodel establishment correlation analysis prediction andassessment is completed to realize the dynamic early warningfunction of MHIs in the CIP

This practical application selects an industry of air sep-aration distillation section in the CIP as a dynamic earlywarning object of MHIs [39] a simplified graphic sketchof the air separation distillation section in Figure 4 Thissection simplifies and removes some industrial process fromthe air separation distillation section To make it easier tounderstand the industrial process we are based on data-driven method and can neglect to describe the mechanismof the chemical process industry The section can separateair into oxygen and nitrogen Risk degree of the section isaffected by several key links such as air separation towersystem air purification system and air precooling systemThese production lines are strictly dependent on pressureand temperature if pressure and temperature are abnormalthey may result in an explosion and fire Air separationtower system consists of Turboexpander and distillationcolumn etcWe removed several weakly parameters and linksfor ensuring factories data privacy but did not affect ourapplication result We take liquid air level in lower columnLI101 as the key parameter in air separation distillation

8 Complexity

Table 2

item DescriptionPI101 Air intake chamber pressure (MPa)PI102 Nitrogen Outlet Tower Pressure (MPa)PI103 Oxygen-enriched air outlet pressure (MPa)PI1201 Air outlet left adsorption bucket pressure (MPa)PI1202 Air outlet right adsorption barrel pressure(MPa)PI1203 Instrument air pressure (MPa)PI01 Pressure of air outlet main heat exchanger(MPa)PI02 Lower column pressure (MPa)PI03 Evaporator condenser pressure (MPa)PDI01 The resistance of lower column (MPa)PI401 Turboexpander1 Intake pressure (MPa)PI402 Turboexpander1Exhaust pressure (MPa)PI403 Turboexpander2Intake pressure (MPa)PI404 Turboexpander2Exhaust pressure (MPa)PI2001 Turboexpander1Bearing air pressure (MPa)PI2002 Turboexpander2 Bearing air pressure (MPa)PI2003 Turboexpander1Sealing pressure (MPa)PI2004 Turboexpander2Sealing pressure (MPa)TI101 Air temperature before entering tower (∘C)TI102 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI103 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI1201 Air outlet left adsorption bucket temperature (∘C)TI1202 Air outlet right adsorption bucket temperature (∘C)TE1204 Temperature of regenerated gas outlet adsorption bucket (∘C)TI01 Temperature of main air outlet heat exchanger (∘C)TI02 Turboexpander1 Intake temperature (∘C)TI03 Turboexpander2 Intake temperature (∘C)TI04 Turboexpander1 Exhaust temperature (∘C)TI05 Turboexpander2 Exhaust temperature (∘C)TI07 Exit temperature of precooler (∘C)FI1201 Regenerative gas flow rate (Nm3h)FI101 Air flow rate (Nm3h)FI102 Nitrogen flow rate (Nm3h)LI101 liquid air level in lower column (mm)LI01 Water cooled liquid level (mm)LI02 Liquid oxygen level in the main condenser (mm)SI402 Expander speed (RPM)LV02 Liquid-air throttle valve ()

section Since most of the top feed comes from the bottomof the tower the working conditions at the bottom of thetower play a decisive role in the air separation distillationsection The abnormal liquid air level in lower column willresult in the decrease of oxygen purity which poses a seriousthreat to the normal production of the system The sectionhas 38 liquid levels pressure temperature and other sensorvariables as shown in Table 2 in which the liquid air level inlower columnof the key parameter is LI101 In order to predictLI101 we need to find the strong correlation variables in 38variables The change trend of LI101 is predicted by the subtlechanges in these strong correlation variables It should benoted that LI101 is only one of many key parameters analysis

process of other key parameters is similar enough to analysisprocess of LI101 so that they are not repeated here

The 38 variables of air separation distillation section areanalyzed by visual correlation analysis The experimentaldatasets comprise 1000 sets as shown in Figure 5 The valueof LI101 exceeds the high alarm value when the broken linebetween two adjacent variables is red Conversely the blueline indicates that the LI101 is not exceeding the high alarmvalue

The 38 variables are not convenient to observe Soanalyze the correlation between 22 variables on the localgraph As shown in Figure 6 the lines between LI101 andadjacent variables PI404 and PI403 are randomly crossed

Complexity 9

Original dataset Parallel Coordinates

0700

PI1201

PI1202

PI1203PI02PI03

PI01PI2001PI2002

LI01LI02

LV02

TI01TI02TI03TI04TI05

TI101

TI103

PI101

LI101TI102FI102PI102

FI101

TI07SI402PI103

PI2003PI2004

PI401PI402PI403PI404

PDI01

TI1201

FI1201

TI1202

TE1204

1920minus13890 830

0 0 0 0 0 0 0 01200

1490000

001

1490

003

010002

002002

003009008006010290866289118260

15560153601685018550106

000078

07420040

114391068033010

1073055033041038073062063027030

40000109679162507840

126309280

1597010000

24189802501450

0093990073

30527386010003430

136528033

40802109

Figure 5 Visual correlation analysis of 38 variables

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0078010 0015914 002 002 0 0023 0089 0064078830

074 20040 114391 069 033 01 055 033 041 03810 073 062 063 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

LI101

Figure 6 Visual correlation analysis local graph A

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Complexity 5

variance to be the first principal components the sublargevariance as the second principal components and the leastvariance as the D principal component

(v)The K principal component of the maximum varianceis selected to allow the K principal component to containmost of the original data information The cumulative vari-ance contribution of the commonly selected K principalcomponents is 85 larger than that of the total variance Thatis sumk

j=1 120582jsumdj=1 120582j ge 85

(vi) By selecting the eigenvectors wi i = 1 2 k of Kprincipal components k independent linear combination ofnew variables is obtained

1205851 = 119909111990811 + 119909211990821 + sdot sdot sdot + 11990911988911990811988911205852 = 119909111990812 + 119909211990822 + sdot sdot sdot + 1199091198891199081198892120585119896 = 11990911199081119896 + 11990921199082119896 + sdot sdot sdot + 119909119889119908119889119896

(4)

(vii) 1205851 1205852 120585119896 is the rebuilt K feature after the reduc-tion of dimension These features represent the principalcomponents of all strong correlation variables and thematrixT of the K principal components of the strong correlationvariable is as follows

119879 = [[[[[[[

x111990811 119909211990821 sdot sdot sdot 1199091198891199081198891119909111990812 119909211990822 sdot sdot sdot 1199091198891199081198892 d11990911199081119896 11990921199082119896 sdot sdot sdot 119909119889119908119889119896

]]]]]]] (5)

33 Dynamic Recurrent Prediction In the prediction modelin traditional neural network we add a state layer to constructthe dynamic recurrent network As in Figure 2 which showsfour layers namely input layer hidden layer state layer andoutput layer [35]

The mathematical model is as follows

119909 (119905) = 119891 (119908(1)119909119888 (119905) + 119908(2)119906 (119905 minus 1) + 120579(1)) 119909119888 (119905) = 119909 (119905 minus 1) y (t) = w(3)x (t) + 120579(2)

(6)

where y(t) is the output of the output layer u(t) is theexternal input of the input layer the matrix T of the Kprincipal component of the strong correlation variable isinput xc(t) is the output of the state layer x(t) is the outputof the hidden layer w(1) is the connection weight of the statelayer and the hidden layer and w(2) is the connection weightbetween the input layer and the hidden layer w(3) is theconnection weight of the hidden layer and the output layerthe 120579(1) is the hidden layer threshold and the 120579(2) is the outputthreshold value where 119891() represents the activation function

y(t)x(t)

hidden output layer

Input layer

state layer

u(t-1) layer

w(2)

w(3)

w(1)

Ｒ＝(t)

Figure 2 Recurrent neural network structures

of the neural network using the Sigmoid function presentedin formula (7)

f (x) = 11 + exp [minusx] 0 lt 119891 (x) lt 1 (7)

Formula (6) can be introduced as follows

xc (t) = f (w(1)kminus1xc (t minus 1) + w(2)kminus2xc (t minus 2)) (8)

where w(1)kminus1 and w(2)kminus2 represent the connection weight atprevious different times Formula (8) indicates that xc(t) isrelated to the weight of the connection at the previous timeand realizes the characteristic of dynamic recursion [36]

The input and output of input layer are respectively

I(1)120572 (t) = u120572 (t minus 1) (9)

O(1)120572 (t) = I(1)120572 (t) (10)

120572 = [1 2 sdot sdot sdot 1198641] where the input and output of thehidden layer are respectively

I(2)120573 (t) = E1sum120572=1

w(2)120573120572O(1)120572 (t) + E2sum

120573=1

w(1)120573120574O(c)120574 (t) + 120579(1) (11)

O(2)120573 (t) = f (I(2)120573 (t)) = 11 + 119890I(2)120573 (t) (12)

120573 = [1 2 sdot sdot sdot 1198642] The input and output of the statelayer are respectively as follows

I(C)120574 (t) = O(2)120573 (t minus 1) (13)

O(C)120574 (t) = I(C)120574 (t) = x(C) (t) (14)

6 Complexity

Valu

e y(t)

Time

G high alarm value

=

expert evaluating method

Second-level fuzzy comprehensive evaluation

Major hazard

Key parameter

First-level fuzzy comprehensive

evaluation

11 11

Major hazard installations 2 2

installations 1 1

Major hazard installations 3 3

Major hazard installations p p

Key parameter

21 22

Key parameter

31 33

Key parameter

11 pressure

p3 pressure 3

p2 pressure 2 p1 pressure 1

12 temperature

21 temperature31 temperature 1

32 temperature 2

33 temperature 3

13 level

23 level 2

22 level 1

11 pressure

22 pressure 33

pressure

pp pressure

Ｓ(Ｎ)minus

times 100

ppp1

１11

１12

１13

１21

１22

１23１32

１33

１31

11

12

13

21

22

23 31

33

32

＄1

＄1

＄1

＄1

B1

B2

B3

Bp

＄2

＄p

＄2

＄2 ＄2 ＄3

＄3

＄3

＄3

p = Apmiddot Dp

50⩽R lt 100

10⩽ R lt 50

R ⩽ 10

A11

A22

１11

１22

A33

１33

１p3

１p2

１p1

R ⩾ 100

p= p

middot １p

R = 1+2+3+ p

＄p Ap3

＄p Ap2

＄p Ap1

１ppApp

p

I Disastrous

II High risk

III Moderate risk

IV Low risk

Figure 3 Hierarchical chart of dynamic assessment of MHIs

120574 = [1 2 sdot sdot sdot 1198642] The input and output of the outputlayer are respectively

I(3)120583 (t) = s2sumj=1w(3)120583120573O

(2)120573 (t) + 120579(2) (15)

y(120583) (t) = O(3)120583 (t) = g ( I(3)120583 (t)) (16)

where 120583 = [1 2 sdot sdot sdot 1198643] Among them 1198641 1198642 and 1198643are the input hidden and output layers respectively and thenumber of layers of the state layer is the same as that of thehidden layer The final y(t) is the predictive value of the keyparameters

34 Dynamic Risk Assessment and Early Warning Fuzzycomprehensive evaluation is a comprehensive decision-making methodology of a multivariable problem-solvingcomplex decision process [37] It is developed from fuzzy

sets We through the first-level and second-level fuzzy com-prehensive evaluation to determine the risk R value of CIPHierarchical chart is presented in Figure 3 The specific stepsare as follows

Step 1 (construction of second-level fuzzy comprehensiveevaluation) The second-level fuzzy comprehensive eval-uation defines an evaluation matrix of key parametersQ119901120590119901 = [Q11205901 Q21205902 Q119901120590119901 ] 119901 is the number of MHIsand 1205901 1205902 120590119901 is the number of key parameters for eachMHIs they may be different values The percentage of thepredicted value of the key parameter y(t) exceeding the highalarm value G C119901120590119901 = (y(t) minus G)G times 100 multiplies thecorresponding elements between Q119901120590119901 and C119901120590119901 Then thesecond-level fuzzy comprehensive evaluation index matrix isobtained

A119901120590119901 = C119901120590119901 sdot Q119901120590119901 (17)

Complexity 7

Air

Compressor

oxygennitrogen

Choke valve

Air precooling system

Air purification system

Filter

Turboexpander

Air separation tower system

Air

Figure 4 Air separation distillation section

Table 1 Dynamic risk level

comment grade rangeDisastrous I R ge 100High risk II 50 le R lt 100Moderate risk III 10 le R lt 50Low risk IV R lt 10Step 2 (construction of the first-level fuzzy comprehensiveevaluation) According to the risk degree of each key param-eter in MHIs the weight of the first-level fuzzy comprehen-sive evaluation is determined D119901 = [D1D2 D119901] Thesecond-level fuzzy comprehensive evaluation matrix A119901120590119901 =[A11205901 A21205902 A119901120590119901 ] obtained by (6) and the weight of thefirst-level fuzzy comprehensive evaluation are conducted bypoint multiplication and the risk grade index of the MHIs inthe CIP is obtained as follows

B119901 = A119901120590119901 sdot D119901 (18)

where 119901 is the total number of MHIs in CIP

Finally the dynamic level of MHIs in CIP is determinedby the evaluation result The score of dynamic evaluation isdecided by the scoring method and the range of the grade isprovided in Table 1 R which is specified in Identification andControl of MHIs (GB18218) is used as a grading index [38]Here R obtains the value of hazard installations dynamicindex B119901The historical factors ofMHIsmay affect the overallrisk The data of all dynamic risk grades are derived fromthe last time the true value of the historical key parameterexceeds the high alarm value to return to the safety value orthe current value

Step 3 The final dynamic assessment of MHIs and earlywarning results are used to urge security personnel to excludethe potential risks of MHIs for the first time

4 Practical Application

CIP is composed of many MHIs the air separation distil-lation section is one of the typical representatives of MHIsThis practical application relies on the industrial data of alarge CIP in Quzhou City from July 2017 and a series ofmodel establishment correlation analysis prediction andassessment is completed to realize the dynamic early warningfunction of MHIs in the CIP

This practical application selects an industry of air sep-aration distillation section in the CIP as a dynamic earlywarning object of MHIs [39] a simplified graphic sketchof the air separation distillation section in Figure 4 Thissection simplifies and removes some industrial process fromthe air separation distillation section To make it easier tounderstand the industrial process we are based on data-driven method and can neglect to describe the mechanismof the chemical process industry The section can separateair into oxygen and nitrogen Risk degree of the section isaffected by several key links such as air separation towersystem air purification system and air precooling systemThese production lines are strictly dependent on pressureand temperature if pressure and temperature are abnormalthey may result in an explosion and fire Air separationtower system consists of Turboexpander and distillationcolumn etcWe removed several weakly parameters and linksfor ensuring factories data privacy but did not affect ourapplication result We take liquid air level in lower columnLI101 as the key parameter in air separation distillation

8 Complexity

Table 2

item DescriptionPI101 Air intake chamber pressure (MPa)PI102 Nitrogen Outlet Tower Pressure (MPa)PI103 Oxygen-enriched air outlet pressure (MPa)PI1201 Air outlet left adsorption bucket pressure (MPa)PI1202 Air outlet right adsorption barrel pressure(MPa)PI1203 Instrument air pressure (MPa)PI01 Pressure of air outlet main heat exchanger(MPa)PI02 Lower column pressure (MPa)PI03 Evaporator condenser pressure (MPa)PDI01 The resistance of lower column (MPa)PI401 Turboexpander1 Intake pressure (MPa)PI402 Turboexpander1Exhaust pressure (MPa)PI403 Turboexpander2Intake pressure (MPa)PI404 Turboexpander2Exhaust pressure (MPa)PI2001 Turboexpander1Bearing air pressure (MPa)PI2002 Turboexpander2 Bearing air pressure (MPa)PI2003 Turboexpander1Sealing pressure (MPa)PI2004 Turboexpander2Sealing pressure (MPa)TI101 Air temperature before entering tower (∘C)TI102 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI103 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI1201 Air outlet left adsorption bucket temperature (∘C)TI1202 Air outlet right adsorption bucket temperature (∘C)TE1204 Temperature of regenerated gas outlet adsorption bucket (∘C)TI01 Temperature of main air outlet heat exchanger (∘C)TI02 Turboexpander1 Intake temperature (∘C)TI03 Turboexpander2 Intake temperature (∘C)TI04 Turboexpander1 Exhaust temperature (∘C)TI05 Turboexpander2 Exhaust temperature (∘C)TI07 Exit temperature of precooler (∘C)FI1201 Regenerative gas flow rate (Nm3h)FI101 Air flow rate (Nm3h)FI102 Nitrogen flow rate (Nm3h)LI101 liquid air level in lower column (mm)LI01 Water cooled liquid level (mm)LI02 Liquid oxygen level in the main condenser (mm)SI402 Expander speed (RPM)LV02 Liquid-air throttle valve ()

section Since most of the top feed comes from the bottomof the tower the working conditions at the bottom of thetower play a decisive role in the air separation distillationsection The abnormal liquid air level in lower column willresult in the decrease of oxygen purity which poses a seriousthreat to the normal production of the system The sectionhas 38 liquid levels pressure temperature and other sensorvariables as shown in Table 2 in which the liquid air level inlower columnof the key parameter is LI101 In order to predictLI101 we need to find the strong correlation variables in 38variables The change trend of LI101 is predicted by the subtlechanges in these strong correlation variables It should benoted that LI101 is only one of many key parameters analysis

process of other key parameters is similar enough to analysisprocess of LI101 so that they are not repeated here

The 38 variables of air separation distillation section areanalyzed by visual correlation analysis The experimentaldatasets comprise 1000 sets as shown in Figure 5 The valueof LI101 exceeds the high alarm value when the broken linebetween two adjacent variables is red Conversely the blueline indicates that the LI101 is not exceeding the high alarmvalue

The 38 variables are not convenient to observe Soanalyze the correlation between 22 variables on the localgraph As shown in Figure 6 the lines between LI101 andadjacent variables PI404 and PI403 are randomly crossed

Complexity 9

Original dataset Parallel Coordinates

0700

PI1201

PI1202

PI1203PI02PI03

PI01PI2001PI2002

LI01LI02

LV02

TI01TI02TI03TI04TI05

TI101

TI103

PI101

LI101TI102FI102PI102

FI101

TI07SI402PI103

PI2003PI2004

PI401PI402PI403PI404

PDI01

TI1201

FI1201

TI1202

TE1204

1920minus13890 830

0 0 0 0 0 0 0 01200

1490000

001

1490

003

010002

002002

003009008006010290866289118260

15560153601685018550106

000078

07420040

114391068033010

1073055033041038073062063027030

40000109679162507840

126309280

1597010000

24189802501450

0093990073

30527386010003430

136528033

40802109

Figure 5 Visual correlation analysis of 38 variables

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0078010 0015914 002 002 0 0023 0089 0064078830

074 20040 114391 069 033 01 055 033 041 03810 073 062 063 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

LI101

Figure 6 Visual correlation analysis local graph A

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

6 Complexity

Valu

e y(t)

Time

G high alarm value

=

expert evaluating method

Second-level fuzzy comprehensive evaluation

Major hazard

Key parameter

First-level fuzzy comprehensive

evaluation

11 11

Major hazard installations 2 2

installations 1 1

Major hazard installations 3 3

Major hazard installations p p

Key parameter

21 22

Key parameter

31 33

Key parameter

11 pressure

p3 pressure 3

p2 pressure 2 p1 pressure 1

12 temperature

21 temperature31 temperature 1

32 temperature 2

33 temperature 3

13 level

23 level 2

22 level 1

11 pressure

22 pressure 33

pressure

pp pressure

Ｓ(Ｎ)minus

times 100

ppp1

１11

１12

１13

１21

１22

１23１32

１33

１31

11

12

13

21

22

23 31

33

32

＄1

＄1

＄1

＄1

B1

B2

B3

Bp

＄2

＄p

＄2

＄2 ＄2 ＄3

＄3

＄3

＄3

p = Apmiddot Dp

50⩽R lt 100

10⩽ R lt 50

R ⩽ 10

A11

A22

１11

１22

A33

１33

１p3

１p2

１p1

R ⩾ 100

p= p

middot １p

R = 1+2+3+ p

＄p Ap3

＄p Ap2

＄p Ap1

１ppApp

p

I Disastrous

II High risk

III Moderate risk

IV Low risk

Figure 3 Hierarchical chart of dynamic assessment of MHIs

120574 = [1 2 sdot sdot sdot 1198642] The input and output of the outputlayer are respectively

I(3)120583 (t) = s2sumj=1w(3)120583120573O

(2)120573 (t) + 120579(2) (15)

y(120583) (t) = O(3)120583 (t) = g ( I(3)120583 (t)) (16)

where 120583 = [1 2 sdot sdot sdot 1198643] Among them 1198641 1198642 and 1198643are the input hidden and output layers respectively and thenumber of layers of the state layer is the same as that of thehidden layer The final y(t) is the predictive value of the keyparameters

34 Dynamic Risk Assessment and Early Warning Fuzzycomprehensive evaluation is a comprehensive decision-making methodology of a multivariable problem-solvingcomplex decision process [37] It is developed from fuzzy

sets We through the first-level and second-level fuzzy com-prehensive evaluation to determine the risk R value of CIPHierarchical chart is presented in Figure 3 The specific stepsare as follows

Step 1 (construction of second-level fuzzy comprehensiveevaluation) The second-level fuzzy comprehensive eval-uation defines an evaluation matrix of key parametersQ119901120590119901 = [Q11205901 Q21205902 Q119901120590119901 ] 119901 is the number of MHIsand 1205901 1205902 120590119901 is the number of key parameters for eachMHIs they may be different values The percentage of thepredicted value of the key parameter y(t) exceeding the highalarm value G C119901120590119901 = (y(t) minus G)G times 100 multiplies thecorresponding elements between Q119901120590119901 and C119901120590119901 Then thesecond-level fuzzy comprehensive evaluation index matrix isobtained

A119901120590119901 = C119901120590119901 sdot Q119901120590119901 (17)

Complexity 7

Air

Compressor

oxygennitrogen

Choke valve

Air precooling system

Air purification system

Filter

Turboexpander

Air separation tower system

Air

Figure 4 Air separation distillation section

Table 1 Dynamic risk level

comment grade rangeDisastrous I R ge 100High risk II 50 le R lt 100Moderate risk III 10 le R lt 50Low risk IV R lt 10Step 2 (construction of the first-level fuzzy comprehensiveevaluation) According to the risk degree of each key param-eter in MHIs the weight of the first-level fuzzy comprehen-sive evaluation is determined D119901 = [D1D2 D119901] Thesecond-level fuzzy comprehensive evaluation matrix A119901120590119901 =[A11205901 A21205902 A119901120590119901 ] obtained by (6) and the weight of thefirst-level fuzzy comprehensive evaluation are conducted bypoint multiplication and the risk grade index of the MHIs inthe CIP is obtained as follows

B119901 = A119901120590119901 sdot D119901 (18)

where 119901 is the total number of MHIs in CIP

Finally the dynamic level of MHIs in CIP is determinedby the evaluation result The score of dynamic evaluation isdecided by the scoring method and the range of the grade isprovided in Table 1 R which is specified in Identification andControl of MHIs (GB18218) is used as a grading index [38]Here R obtains the value of hazard installations dynamicindex B119901The historical factors ofMHIsmay affect the overallrisk The data of all dynamic risk grades are derived fromthe last time the true value of the historical key parameterexceeds the high alarm value to return to the safety value orthe current value

Step 3 The final dynamic assessment of MHIs and earlywarning results are used to urge security personnel to excludethe potential risks of MHIs for the first time

4 Practical Application

CIP is composed of many MHIs the air separation distil-lation section is one of the typical representatives of MHIsThis practical application relies on the industrial data of alarge CIP in Quzhou City from July 2017 and a series ofmodel establishment correlation analysis prediction andassessment is completed to realize the dynamic early warningfunction of MHIs in the CIP

This practical application selects an industry of air sep-aration distillation section in the CIP as a dynamic earlywarning object of MHIs [39] a simplified graphic sketchof the air separation distillation section in Figure 4 Thissection simplifies and removes some industrial process fromthe air separation distillation section To make it easier tounderstand the industrial process we are based on data-driven method and can neglect to describe the mechanismof the chemical process industry The section can separateair into oxygen and nitrogen Risk degree of the section isaffected by several key links such as air separation towersystem air purification system and air precooling systemThese production lines are strictly dependent on pressureand temperature if pressure and temperature are abnormalthey may result in an explosion and fire Air separationtower system consists of Turboexpander and distillationcolumn etcWe removed several weakly parameters and linksfor ensuring factories data privacy but did not affect ourapplication result We take liquid air level in lower columnLI101 as the key parameter in air separation distillation

8 Complexity

Table 2

item DescriptionPI101 Air intake chamber pressure (MPa)PI102 Nitrogen Outlet Tower Pressure (MPa)PI103 Oxygen-enriched air outlet pressure (MPa)PI1201 Air outlet left adsorption bucket pressure (MPa)PI1202 Air outlet right adsorption barrel pressure(MPa)PI1203 Instrument air pressure (MPa)PI01 Pressure of air outlet main heat exchanger(MPa)PI02 Lower column pressure (MPa)PI03 Evaporator condenser pressure (MPa)PDI01 The resistance of lower column (MPa)PI401 Turboexpander1 Intake pressure (MPa)PI402 Turboexpander1Exhaust pressure (MPa)PI403 Turboexpander2Intake pressure (MPa)PI404 Turboexpander2Exhaust pressure (MPa)PI2001 Turboexpander1Bearing air pressure (MPa)PI2002 Turboexpander2 Bearing air pressure (MPa)PI2003 Turboexpander1Sealing pressure (MPa)PI2004 Turboexpander2Sealing pressure (MPa)TI101 Air temperature before entering tower (∘C)TI102 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI103 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI1201 Air outlet left adsorption bucket temperature (∘C)TI1202 Air outlet right adsorption bucket temperature (∘C)TE1204 Temperature of regenerated gas outlet adsorption bucket (∘C)TI01 Temperature of main air outlet heat exchanger (∘C)TI02 Turboexpander1 Intake temperature (∘C)TI03 Turboexpander2 Intake temperature (∘C)TI04 Turboexpander1 Exhaust temperature (∘C)TI05 Turboexpander2 Exhaust temperature (∘C)TI07 Exit temperature of precooler (∘C)FI1201 Regenerative gas flow rate (Nm3h)FI101 Air flow rate (Nm3h)FI102 Nitrogen flow rate (Nm3h)LI101 liquid air level in lower column (mm)LI01 Water cooled liquid level (mm)LI02 Liquid oxygen level in the main condenser (mm)SI402 Expander speed (RPM)LV02 Liquid-air throttle valve ()

section Since most of the top feed comes from the bottomof the tower the working conditions at the bottom of thetower play a decisive role in the air separation distillationsection The abnormal liquid air level in lower column willresult in the decrease of oxygen purity which poses a seriousthreat to the normal production of the system The sectionhas 38 liquid levels pressure temperature and other sensorvariables as shown in Table 2 in which the liquid air level inlower columnof the key parameter is LI101 In order to predictLI101 we need to find the strong correlation variables in 38variables The change trend of LI101 is predicted by the subtlechanges in these strong correlation variables It should benoted that LI101 is only one of many key parameters analysis

process of other key parameters is similar enough to analysisprocess of LI101 so that they are not repeated here

The 38 variables of air separation distillation section areanalyzed by visual correlation analysis The experimentaldatasets comprise 1000 sets as shown in Figure 5 The valueof LI101 exceeds the high alarm value when the broken linebetween two adjacent variables is red Conversely the blueline indicates that the LI101 is not exceeding the high alarmvalue

The 38 variables are not convenient to observe Soanalyze the correlation between 22 variables on the localgraph As shown in Figure 6 the lines between LI101 andadjacent variables PI404 and PI403 are randomly crossed

Complexity 9

Original dataset Parallel Coordinates

0700

PI1201

PI1202

PI1203PI02PI03

PI01PI2001PI2002

LI01LI02

LV02

TI01TI02TI03TI04TI05

TI101

TI103

PI101

LI101TI102FI102PI102

FI101

TI07SI402PI103

PI2003PI2004

PI401PI402PI403PI404

PDI01

TI1201

FI1201

TI1202

TE1204

1920minus13890 830

0 0 0 0 0 0 0 01200

1490000

001

1490

003

010002

002002

003009008006010290866289118260

15560153601685018550106

000078

07420040

114391068033010

1073055033041038073062063027030

40000109679162507840

126309280

1597010000

24189802501450

0093990073

30527386010003430

136528033

40802109

Figure 5 Visual correlation analysis of 38 variables

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0078010 0015914 002 002 0 0023 0089 0064078830

074 20040 114391 069 033 01 055 033 041 03810 073 062 063 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

LI101

Figure 6 Visual correlation analysis local graph A

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Complexity 7

Air

Compressor

oxygennitrogen

Choke valve

Air precooling system

Air purification system

Filter

Turboexpander

Air separation tower system

Air

Figure 4 Air separation distillation section

Table 1 Dynamic risk level

comment grade rangeDisastrous I R ge 100High risk II 50 le R lt 100Moderate risk III 10 le R lt 50Low risk IV R lt 10Step 2 (construction of the first-level fuzzy comprehensiveevaluation) According to the risk degree of each key param-eter in MHIs the weight of the first-level fuzzy comprehen-sive evaluation is determined D119901 = [D1D2 D119901] Thesecond-level fuzzy comprehensive evaluation matrix A119901120590119901 =[A11205901 A21205902 A119901120590119901 ] obtained by (6) and the weight of thefirst-level fuzzy comprehensive evaluation are conducted bypoint multiplication and the risk grade index of the MHIs inthe CIP is obtained as follows

B119901 = A119901120590119901 sdot D119901 (18)

where 119901 is the total number of MHIs in CIP

Finally the dynamic level of MHIs in CIP is determinedby the evaluation result The score of dynamic evaluation isdecided by the scoring method and the range of the grade isprovided in Table 1 R which is specified in Identification andControl of MHIs (GB18218) is used as a grading index [38]Here R obtains the value of hazard installations dynamicindex B119901The historical factors ofMHIsmay affect the overallrisk The data of all dynamic risk grades are derived fromthe last time the true value of the historical key parameterexceeds the high alarm value to return to the safety value orthe current value

Step 3 The final dynamic assessment of MHIs and earlywarning results are used to urge security personnel to excludethe potential risks of MHIs for the first time

4 Practical Application

CIP is composed of many MHIs the air separation distil-lation section is one of the typical representatives of MHIsThis practical application relies on the industrial data of alarge CIP in Quzhou City from July 2017 and a series ofmodel establishment correlation analysis prediction andassessment is completed to realize the dynamic early warningfunction of MHIs in the CIP

This practical application selects an industry of air sep-aration distillation section in the CIP as a dynamic earlywarning object of MHIs [39] a simplified graphic sketchof the air separation distillation section in Figure 4 Thissection simplifies and removes some industrial process fromthe air separation distillation section To make it easier tounderstand the industrial process we are based on data-driven method and can neglect to describe the mechanismof the chemical process industry The section can separateair into oxygen and nitrogen Risk degree of the section isaffected by several key links such as air separation towersystem air purification system and air precooling systemThese production lines are strictly dependent on pressureand temperature if pressure and temperature are abnormalthey may result in an explosion and fire Air separationtower system consists of Turboexpander and distillationcolumn etcWe removed several weakly parameters and linksfor ensuring factories data privacy but did not affect ourapplication result We take liquid air level in lower columnLI101 as the key parameter in air separation distillation

8 Complexity

Table 2

item DescriptionPI101 Air intake chamber pressure (MPa)PI102 Nitrogen Outlet Tower Pressure (MPa)PI103 Oxygen-enriched air outlet pressure (MPa)PI1201 Air outlet left adsorption bucket pressure (MPa)PI1202 Air outlet right adsorption barrel pressure(MPa)PI1203 Instrument air pressure (MPa)PI01 Pressure of air outlet main heat exchanger(MPa)PI02 Lower column pressure (MPa)PI03 Evaporator condenser pressure (MPa)PDI01 The resistance of lower column (MPa)PI401 Turboexpander1 Intake pressure (MPa)PI402 Turboexpander1Exhaust pressure (MPa)PI403 Turboexpander2Intake pressure (MPa)PI404 Turboexpander2Exhaust pressure (MPa)PI2001 Turboexpander1Bearing air pressure (MPa)PI2002 Turboexpander2 Bearing air pressure (MPa)PI2003 Turboexpander1Sealing pressure (MPa)PI2004 Turboexpander2Sealing pressure (MPa)TI101 Air temperature before entering tower (∘C)TI102 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI103 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI1201 Air outlet left adsorption bucket temperature (∘C)TI1202 Air outlet right adsorption bucket temperature (∘C)TE1204 Temperature of regenerated gas outlet adsorption bucket (∘C)TI01 Temperature of main air outlet heat exchanger (∘C)TI02 Turboexpander1 Intake temperature (∘C)TI03 Turboexpander2 Intake temperature (∘C)TI04 Turboexpander1 Exhaust temperature (∘C)TI05 Turboexpander2 Exhaust temperature (∘C)TI07 Exit temperature of precooler (∘C)FI1201 Regenerative gas flow rate (Nm3h)FI101 Air flow rate (Nm3h)FI102 Nitrogen flow rate (Nm3h)LI101 liquid air level in lower column (mm)LI01 Water cooled liquid level (mm)LI02 Liquid oxygen level in the main condenser (mm)SI402 Expander speed (RPM)LV02 Liquid-air throttle valve ()

section Since most of the top feed comes from the bottomof the tower the working conditions at the bottom of thetower play a decisive role in the air separation distillationsection The abnormal liquid air level in lower column willresult in the decrease of oxygen purity which poses a seriousthreat to the normal production of the system The sectionhas 38 liquid levels pressure temperature and other sensorvariables as shown in Table 2 in which the liquid air level inlower columnof the key parameter is LI101 In order to predictLI101 we need to find the strong correlation variables in 38variables The change trend of LI101 is predicted by the subtlechanges in these strong correlation variables It should benoted that LI101 is only one of many key parameters analysis

process of other key parameters is similar enough to analysisprocess of LI101 so that they are not repeated here

The 38 variables of air separation distillation section areanalyzed by visual correlation analysis The experimentaldatasets comprise 1000 sets as shown in Figure 5 The valueof LI101 exceeds the high alarm value when the broken linebetween two adjacent variables is red Conversely the blueline indicates that the LI101 is not exceeding the high alarmvalue

The 38 variables are not convenient to observe Soanalyze the correlation between 22 variables on the localgraph As shown in Figure 6 the lines between LI101 andadjacent variables PI404 and PI403 are randomly crossed

Complexity 9

Original dataset Parallel Coordinates

0700

PI1201

PI1202

PI1203PI02PI03

PI01PI2001PI2002

LI01LI02

LV02

TI01TI02TI03TI04TI05

TI101

TI103

PI101

LI101TI102FI102PI102

FI101

TI07SI402PI103

PI2003PI2004

PI401PI402PI403PI404

PDI01

TI1201

FI1201

TI1202

TE1204

1920minus13890 830

0 0 0 0 0 0 0 01200

1490000

001

1490

003

010002

002002

003009008006010290866289118260

15560153601685018550106

000078

07420040

114391068033010

1073055033041038073062063027030

40000109679162507840

126309280

1597010000

24189802501450

0093990073

30527386010003430

136528033

40802109

Figure 5 Visual correlation analysis of 38 variables

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0078010 0015914 002 002 0 0023 0089 0064078830

074 20040 114391 069 033 01 055 033 041 03810 073 062 063 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

LI101

Figure 6 Visual correlation analysis local graph A

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

8 Complexity

Table 2

item DescriptionPI101 Air intake chamber pressure (MPa)PI102 Nitrogen Outlet Tower Pressure (MPa)PI103 Oxygen-enriched air outlet pressure (MPa)PI1201 Air outlet left adsorption bucket pressure (MPa)PI1202 Air outlet right adsorption barrel pressure(MPa)PI1203 Instrument air pressure (MPa)PI01 Pressure of air outlet main heat exchanger(MPa)PI02 Lower column pressure (MPa)PI03 Evaporator condenser pressure (MPa)PDI01 The resistance of lower column (MPa)PI401 Turboexpander1 Intake pressure (MPa)PI402 Turboexpander1Exhaust pressure (MPa)PI403 Turboexpander2Intake pressure (MPa)PI404 Turboexpander2Exhaust pressure (MPa)PI2001 Turboexpander1Bearing air pressure (MPa)PI2002 Turboexpander2 Bearing air pressure (MPa)PI2003 Turboexpander1Sealing pressure (MPa)PI2004 Turboexpander2Sealing pressure (MPa)TI101 Air temperature before entering tower (∘C)TI102 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI103 Outlet Tower Temperature of Contaminated Nitrogen (∘C)TI1201 Air outlet left adsorption bucket temperature (∘C)TI1202 Air outlet right adsorption bucket temperature (∘C)TE1204 Temperature of regenerated gas outlet adsorption bucket (∘C)TI01 Temperature of main air outlet heat exchanger (∘C)TI02 Turboexpander1 Intake temperature (∘C)TI03 Turboexpander2 Intake temperature (∘C)TI04 Turboexpander1 Exhaust temperature (∘C)TI05 Turboexpander2 Exhaust temperature (∘C)TI07 Exit temperature of precooler (∘C)FI1201 Regenerative gas flow rate (Nm3h)FI101 Air flow rate (Nm3h)FI102 Nitrogen flow rate (Nm3h)LI101 liquid air level in lower column (mm)LI01 Water cooled liquid level (mm)LI02 Liquid oxygen level in the main condenser (mm)SI402 Expander speed (RPM)LV02 Liquid-air throttle valve ()

section Since most of the top feed comes from the bottomof the tower the working conditions at the bottom of thetower play a decisive role in the air separation distillationsection The abnormal liquid air level in lower column willresult in the decrease of oxygen purity which poses a seriousthreat to the normal production of the system The sectionhas 38 liquid levels pressure temperature and other sensorvariables as shown in Table 2 in which the liquid air level inlower columnof the key parameter is LI101 In order to predictLI101 we need to find the strong correlation variables in 38variables The change trend of LI101 is predicted by the subtlechanges in these strong correlation variables It should benoted that LI101 is only one of many key parameters analysis

process of other key parameters is similar enough to analysisprocess of LI101 so that they are not repeated here

The 38 variables of air separation distillation section areanalyzed by visual correlation analysis The experimentaldatasets comprise 1000 sets as shown in Figure 5 The valueof LI101 exceeds the high alarm value when the broken linebetween two adjacent variables is red Conversely the blueline indicates that the LI101 is not exceeding the high alarmvalue

The 38 variables are not convenient to observe Soanalyze the correlation between 22 variables on the localgraph As shown in Figure 6 the lines between LI101 andadjacent variables PI404 and PI403 are randomly crossed

Complexity 9

Original dataset Parallel Coordinates

0700

PI1201

PI1202

PI1203PI02PI03

PI01PI2001PI2002

LI01LI02

LV02

TI01TI02TI03TI04TI05

TI101

TI103

PI101

LI101TI102FI102PI102

FI101

TI07SI402PI103

PI2003PI2004

PI401PI402PI403PI404

PDI01

TI1201

FI1201

TI1202

TE1204

1920minus13890 830

0 0 0 0 0 0 0 01200

1490000

001

1490

003

010002

002002

003009008006010290866289118260

15560153601685018550106

000078

07420040

114391068033010

1073055033041038073062063027030

40000109679162507840

126309280

1597010000

24189802501450

0093990073

30527386010003430

136528033

40802109

Figure 5 Visual correlation analysis of 38 variables

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0078010 0015914 002 002 0 0023 0089 0064078830

074 20040 114391 069 033 01 055 033 041 03810 073 062 063 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

LI101

Figure 6 Visual correlation analysis local graph A

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Complexity 9

Original dataset Parallel Coordinates

0700

PI1201

PI1202

PI1203PI02PI03

PI01PI2001PI2002

LI01LI02

LV02

TI01TI02TI03TI04TI05

TI101

TI103

PI101

LI101TI102FI102PI102

FI101

TI07SI402PI103

PI2003PI2004

PI401PI402PI403PI404

PDI01

TI1201

FI1201

TI1202

TE1204

1920minus13890 830

0 0 0 0 0 0 0 01200

1490000

001

1490

003

010002

002002

003009008006010290866289118260

15560153601685018550106

000078

07420040

114391068033010

1073055033041038073062063027030

40000109679162507840

126309280

1597010000

24189802501450

0093990073

30527386010003430

136528033

40802109

Figure 5 Visual correlation analysis of 38 variables

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0078010 0015914 002 002 0 0023 0089 0064078830

074 20040 114391 069 033 01 055 033 041 03810 073 062 063 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

LI101

Figure 6 Visual correlation analysis local graph A

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

10 Complexity

Original dataset Parallel Coordinates070

0 1920 minus1389 0 0 0 0 0 0010 002 002 002 003 009 008 006078830

074 20040 114391 068 033 010 055 033 041 038 075 062 063 10 02710734080 2109

PI1201

PI1202

PI1203

PI02

PI03

PI01

PI2001PI2002

LI101

PI2003

PI401

PI402

PI403

PI404

PDI01

TI1201

FI1201

TI1202

TE1204

Figure 7 Visual correlation analysis local graph B

Table 3 Strong correlation variable table

item PI02 PI03 PI401 PI402 PI1201 PI2003 PI102correlation 0674lowastlowast 0661lowastlowast 0645lowastlowast 0621lowastlowast 0670lowastlowast 0669lowastlowast 0674lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000item PI103 FI102 FI101 FI1201 TI03 TI04 TI102correlation 0659lowastlowast 0659lowastlowast 0602lowastlowast 0697lowastlowast 0617lowastlowast -0611lowastlowast -0633lowastlowastconspicuousness 0000 0000 0000 0000 0000 0000 0000N 1000 1000 1000 1000 1000 1000 1000

indicating that no special correlation exists between thesevariables

As shown in Figure 7 lines between LI101 and PI2002LI101 and PI2003 are overall parallel indicating a positivelycorrelated relationship This method can intuitively andrapidly select strong correlation of key parameter

After visual correlation analysis we filtered 27 corre-lation variables and eliminated 11 weakly correlation vari-ables

Next in order to verify the accuracy of visual correlationanalysis and further filter strong correlation variables Pear-son correlation analysis of 27 correlation variables and theresults of the filtered strong correlation variables are shownin Table 3

The table of strong correlation variable has three rowsThefirst row is the correlation value inwhich 0 to 033 isweakcorrelation 033 to 067 is medium correlation and 067 to 1is strong correlation [40] The second row is the significant

test result that is sig bilateral test The asterisk behind thecorrelation row values represents the degree of saliency lowastrepresenting 001 lt sig lt 005 indicates significance lowastlowastrepresenting sig lt 001 denotes a high degree of significanceand the third row is the sample capacity

Take the data of PI02 in Table 3 as an example wherecorrelation coefficient r = 0674 significant P = 0 and samplecapacity N = 1000 The results indicate that PI02 is the lowercolumn pressure and this pressure has a strong correlationwith LI101

The strong correlation variable inTable 3 is used to reducethe dimension

The data matrix X is composed of strong correlationvariables and the matrix is normalized The normalizeddata matrix Z is obtained and the correlation matrix C iscalculated The correlation matrix C comprises 14 principalcomponents The correlation coefficient matrix of solid cor-relation variables is shown in Table 4

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Complexity 11

Table 4 Correlation coefficient matrix

Z-Score PI02 PI03 PI401 PI402 PI1201 PI2003 PI102 PI103 FI102 FI101 FI1201 TI03 TI04 TI102PI02 1000 0354 0744 0376 0720 0805 0519 0884 -0468 0396 0250 0798 0114 0749PI03 0354 1000 0854 0976 0694 0714 0699 0477 -0623 0984 0866 0156 0543 0850PI401 0744 0854 1000 0856 0878 0933 0727 0779 -0675 0871 0750 0505 0463 1000PI402 0376 0976 0856 1000 0637 0695 0659 0462 -0512 0993 0862 0218 0444 0855PI1201 0720 0694 0878 0637 1000 0958 0798 0828 -0902 0693 0650 0379 0677 0868PI2003 0805 0714 0933 0695 0958 1000 0785 0873 -0766 0734 0599 0525 0511 0930PI102 0519 0699 0727 0659 0798 0785 1000 0674 -0742 0696 0670 0233 0531 0716PI103 0884 0477 0779 0462 0828 0873 0674 1000 -0695 0509 0331 0726 0288 0780FI102 -0468 -0623 -0675 -0512 -0902 -0766 -0742 -0695 1000 -0590 -0611 -0144 -0796 -0657FI101 0396 0984 0871 0993 0693 0734 0696 0509 -0590 1000 0866 0211 0515 0869FI1201 0250 0866 0750 0862 0650 0599 0670 0331 -0611 0866 1000 -0038 0672 0738TI03 0798 0156 0505 0218 0379 0525 0233 0726 -0144 0211 -0038 1000 -0313 0519TI04 0114 0543 0463 0444 0677 0511 0531 0288 -0796 0515 0672 -0313 1000 0441TI102 0749 0850 1000 855 0868 0930 0716 0780 -0657 0869 0738 0519 0441 1000

Table 5 Principal component characteristic root and contribution rate

assemblyInitial Eigenvalues Extraction of load square sum

total variance cumulative total variance cumulativepercentage percentage

1 9543 68165 68165 9543 68165 681652 2328 16631 84796 2328 16631 847963 1232 8800 93597 1232 8800 935974 0341 2434 960315 0204 1454 974846 0139 0990 984747 0076 0543 990178 0061 0432 994509 0045 0322 9977110 0021 0149 9992011 0007 0051 9997112 0002 0017 9998813 0002 0012 10000014 5197E-5 0000 100000

The correlation coefficient matrix shows a strong cor-relation between the indexes For example the correlationcoefficient between PI03 and PI402 and FI101 is large as inTable 4This result shows that an overlap exists between theirindex information Next the Jacobi method is used to findthe eigenvector wi i = 1 2 m of the correlation matrixΛ = diag(1205821 1205822 120582m) where 120582 is the eigenvalue and diagis a diagonalmatrixThe eigenvalues are sorted by the order of1205821 gt 1205822 gt sdot sdot sdot gt 120582m and the order of the eigenvector columnis adjusted correspondingly making the maximum variancethe first principal components the second largest variance asthe second principal components and theminimumvarianceas the thirteenth principal components

As shown in Table 5 the principal component character-istic root and contribution rate are characteristic root 1205821 =9543 characteristic root 1205822 = 2328 and characteristic root1205823 = 123 The cumulative variance contribution rate of

the first three principal components reached 93597 whichcovers most of the information This finding suggests thatthe first three principal components can represent the 13principal components therefore the first three indexes canbe extracted The principal component is taken as 1205851 1205852 and1205853 The principal component load vector is calculated by thethree principal components and the result is presented inTable 6

The indexes PI02 PI03 PI401 PI402 PI1201 PI2003PI102 PI103 PI102 FI101 FI1201 and TI102 have high loadon the first principal component and the correlation is highTI03 has high load on the second principal components anda strong correlation TI04 has high load on the third principalcomponents and a strong correlation

The load matrix of the principal component is the eigen-vector of the principal component Principal component loadvector is divided by the arithmetic square root of the principal

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

12 Complexity

Table 6 Principal component load vector

assembly1 2 3

PI02 07 0654 minus006PI03 0874 -0338 0302PI401 097 0093 0154PI402 0851 -0284 0429PI1201 0936 0074 -0314PI2003 0948 0212 -0118PI102 0833 -0075 -0165PI103 0801 0515 -0199PI102 -0805 0135 0512FI101 0884 -0289 034FI1201 079 -05 0155TI03 0419 0841 0224TI04 0596 -0552 -0617TI102 0965 0109 0174

Table 7 Load matrix of the principal component

assembly1 2 3

PI02 023 043 -005PI03 028 -022 027PI401 031 006 014PI402 028 -019 039PI1201 03 005 -028PI2003 031 014 -011PI102 027 -005 -015PI103 026 034 -018PI102 -026 009 046FI101 029 -019 031FI1201 026 -033 014TI03 014 055 02TI04 019 -036 -056TI102 031 007 016

component characteristic root that is the load matrix of theprincipal component

119880i = Airadic120582i (19)

The results are shown in Table 7 The expression of theprincipal component coefficient is

1205851 = 11990911198801 + 11990921198802 + sdot sdot sdot + 119909119889119880119889 (20)

The eigenvector is the load matrix Ui i = 1 2 d andthe eigenvalue is the value of other variables at the same time

Table 7 uses the three principal component to dynamicrecurrent prediction

41 Prediction Model BP neural network is a feedback-freefeedforward neural network which is composed of a set of

interconnected neurons in which each neuron is connectedwith the corresponding weight [41] The extreme learningmachine (ELM) is a single hidden layer feedforward neuralnetwork It is made up by a simple structure with stronglearning and generalization ability This model has beenextensively used in the fields of pattern recognition andregression estimation [42]

Through filtered of key parameter correlation variablesand dimensionality reduction 14 groups of strong correlationvariables were filtered from 38 groups of correlation variablesAfter dimensionality reduction 3 groups of principal compo-nents representing 14 groups of strong correlation variableswere obtained

The following is the prediction of key parameter Themethod is compared with BP neural network ELM anddynamic recurrent prediction model without dimensional-ity reduction For fairness the selected variables are also

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Complexity 13

Table 8 Sample set allocation table

Data sets prorate amountTraining 70 700Validation 15 150Testing 15 150

analyzed by visual qualitative analysis and quantitativecorrelation analysis The predicted experimental data arederived from 14 sets of solid correlation variables whichremove abnormal values data and acquire 1000 sets ofeffective data as experimental samplesThe distribution of thesample set is presented in Table 8

In this study the configuration of hardware involves CPUof Intel Core i5-7300HQ with main clock frequency of 250GHz and GPU is GeForce GTX1050Ti by NVIDIA Thisstudy adopts the following international evaluation index tomeasure the prediction efficiency of the model

(1) Maximum Error (ME)

ME = max (1003816100381610038161003816fi minus yi1003816100381610038161003816) i = 1 2 3 n (21)

Among them n is the sample size fi is the predicted valueand yi is the true value

(2) Mean Absolute Error (MAE) is the average of theabsolute value of the deviation between all observed andpredicted values

MAE = 1n

i=nsumi=1

1003816100381610038161003816fi minus yi1003816100381610038161003816 (22)

(3) Root Mean Square Error (RMSE) represents thediscreteness of the error distribution A small RMSE leads toconcentrated error distribution and high prediction accuracy

RMSE = radic 1n

i=nsumi=1(fi minus yi)2 (23)

(4) Mean Absolute Percentage Error (MAPE) can beused to measure the prediction results of a model and itscalculation formula is as follows

MAPE = sum((1003816100381610038161003816fi minus yi

1003816100381610038161003816 times 100) yi)n

(24)

For the dynamic recurrent predictionmodel the determi-nation of the number of hidden layer neurons is an empiricalproblem that must be constantly tested The number ofhidden layer neurons in the dynamic recurrent networkstructure which is not reduced by dimensionality reductionis set to 2ndash10 The prediction error proportionality to thenumber of neurons in different hidden layers is shown inFigure 8 In the dynamic recurrent network model MAPEgradually decreases when the number of hidden layer neu-rons increases from 2 to 9 When the number of hiddenlayer neurons increases from 9 to 10 MAPE also graduallyincreases indicating that the optimal number of hiddenlayer neurons was 8 For the Dimensionality Reduction-Dynamic Recurrent model the optimal number of hidden

Dynamic Recurrent

Dimensionality Reduction-Dynamic Recurrent

2 3 4 5 6 7 8 9 10 111The number of hidden layer neurons

1

11

12

13

14

15

16

17

18

19

Pred

ictio

n er

ror

Figure 8 MAPE value and the number of neurons

layer neurons is 5When the number of hidden layer neuronsis larger than 5 the increasing trend of MAPE graduallyflattens and eventually stabilizes with the increase in thenumber of hidden layer neurons indicating that the stabilityof the Dimensionality Reduction-Dynamic Recurrent modelis strong

According to the analysis results of the two precedingmodels the network structure of dynamic recurrent modelis 14-9-1 and the network structure of DimensionalityReduction-Dynamic Recurrent model is 3-5-1

Four methods were tested 30 times using the sample datawith a time step of 1 min to visualize their prediction effectWhen the dynamic recurrent network is trained the transferfunction of the hidden layer unit takes the S tangent functionTansig the output unit also uses the S tangent functionTansig and the training function adopts the momentum andadaptive gradient decreasing training function traingdx

An excerpt fragment from the 150 sets of test data isshown in Figure 9 Results show that the four predictionmodeling methods have achieved good prediction results forthe time series of key parameter However there is a highcorrespondence between fitting curve of predicted data andactual real values and have greater prediction accuracy thanthe three other prediction models

Several kinds of evaluation indexes of the four predictionmodels are calculated Figure 10 indicates the prediction errorchart of the four prediction methods and Table 9 shows theperformance indexes of several prediction methods and theresult is presented The MAE and the RMSE of this methodare all excellent among the four methods which show thatthis method has higher precision than the other methods inthe chemical key parameter prediction From the mean valueof themaximum error and the absolute value of the percentilerelative error the method demonstrates more stable than the

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

14 Complexity

Table 9 Performance evaluation of different prediction models

Method ME MAE RMSE MAPE () TIME (s)BP 74641 9982 129060 1023 337414ELM 70414 7916 99137 0793 077Dynamic Recurrent 49381 5768 69117 0577 40341Proposed Method 31360 4879 55958 0498 34732

RealProposed MethodDynamic RecurrentELMBP

10 15 20 25 30 35 40 45 505Time (min)

880

900

920

940

960

980

1000

1020

1040

1060

Pred

ictio

n va

lue (

mm

)

Figure 9 Prediction results of key parameters

other methods in predicting the dynamic time series of thechemical industry

Among the fourmethods of this application ELMhas theshortest prediction time but has weaker prediction accuracyand stability than the proposedmethod in this paper and theBP neural network is poor in all aspects After dimensionalityreduction the performance of this method is superior to thesingle dynamic recurrent model method

42 Evaluation and Discussion In this application the riskdegree of air separation distillation section is assumed to beaffected by several key links such as air separation towersystem air purification system and air precooling systemMoreover the risk degree of air separation tower system isconsiderably influenced by several key links The evaluationindex and evaluation coefficient are presented in Table 10

The second-level fuzzy comprehensive evaluation is usedto define the evaluation matrix Q21205902 = [Q21Q22Q23Q24]of the fine air separation distillation section and the keyparameter y(t) exceeding the high alarm value G

C21205902 = y (t) minus GG

times 100 (25)

5 504540353025201510

6

minus8

minus6

minus4

minus2

0

2

4

Time (min)

Pred

ictio

n er

ror (

)

Proposed MethodDynamic RecurrentELMBP

Figure 10 Prediction error of four prediction models

The following is only for liquid air level in lowercolumnMultiply the elements corresponding to Q21 andC21 The second-level fuzzy comprehensive evaluation indexmatrix of liquid air level in lower column

A21 = C21 sdot Q21 (26)

The 0ndash25 period of real-time prediction of the excerpt keyparameter is presented in Figure 11 The yellow line at 1030mm is LI101 high alarm value and purple line at 1040 mm isHigh-High alarm value The current predictive value of timeis 24 min The corresponding value is 960236 mm

The effect of the fuzzy comprehensive evaluation ofhistorical key parameter becomes smaller with the increasetime This study considers an algorithm for generating anattenuation factor to accurately evaluate the influence of thefuzzy comprehensive evaluation of historical key parameteron MHIs

Assuming that the second-level fuzzy comprehensiveevaluation of key parameter is A allowing A to assign atimeliness weight 119905119901 to the numerical adjustment to eliminatethe effect of time factors on the overall MHIs rating results L

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Complexity 15

Table 10 Second-level fuzzy evaluation index and evaluation coefficient

key production links key parameters evaluation labeling evaluation coefficient

Air precooling system Circulation waterlevel Q11 880

Air separation towersystem (contain Distillationcolumn)

Liquid air level inlower column Q21 1110

Distillation columnpressure Q22 1050

Liquid oxygen level inthe main condenser Q23 1010

Distillation columntop temperature Q24 1040

Air purification system

Adsorption barreltemperature Q31 890

Adsorption barrelpressure Q32 870

RealProposed Method

5 10 15 20 250Time (min)

940

960

980

1000

1020

1040

1060

Pred

ictio

n Va

lue (

mm

)

Figure 11 Real-time prediction and display of key parameters

is the number of abnormal after the last time the real value ofthe key parameter exceeds the high alarm

119871sum119901=1

119905119901 = 1 1 le 119901 le 119871 (27)

Time weight 119905119901 is called time attenuation factor In thispaper the attenuation trend of second-level fuzzy compre-hensive evaluations with time is presented by decreasing theratio of equal ratio where 119905(119901minus1) = 119905119901 = 119905(119901+1)

119905(119901minus1)119905119901 = 119905119901119905(119901+1) 2 le 119901 le 119871 minus 1 (28)

Input L historical key parameter matrix A21 and correspond-ing time attenuation factor 1199051 1199052 119905119871

Table 11 Percentage of key parameter over normal thresholdsecond-level fuzzy comprehensive evaluation and attenuation fac-tor

key parameter C21 A21 1199051199011037760 0753 8283 00381035590 0543 5973 00551038630 0838 9218 00781038190 0795 8745 01121042530 1217 13387 01601044700 1427 15697 02291042530 1217 13387 0327

Output Second-level fuzzy comprehensive evaluation valueof eliminating time factor is A1015840

A1015840 = Lsump=1

A times tp (29)

The interval L is 7 for the last time the segment dataexceeds the high alarm value to the last return to the securityvalue in which the key parameter exceeds the percentage ofthe normal threshold The second-level fuzzy comprehensiveevaluation and the attenuation factor are shown in Table 11

The second-level fuzzy comprehensive evaluation valueexcluding time factor is A101584021 which is calculated as follows

A101584021 = [8283 5973 9218 8745 13387 15697 13387][[[[[[[[[[[[[[[

0038005500780112016002290327

]]]]]]]]]]]]]]]

= 124558 (30)

Then the weight set of the first layer is determined theeffect of each index on the air separation distillation sectionrisk grade that is the weight allocation is shown in Table 12

Multiply the second-level fuzzy comprehensive evalu-ation matrix A10158402 and the weight of the first-level fuzzy

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

16 Complexity

Table 12 Level weight allocation

Hazard installation Key production number link Weight(D)

Air separation distillation section BAir precooling system B1 007

Air separation tower system B2 076Air purification system B3 017

comprehensive evaluation so the risk grade index of the airseparation distillation section in CIP is as follows

B2 = A101584021205902 sdotD2 (31)

We assume that the risk grade index of the air pre-cooling system and the air purification system is B1 =4796 B3 = 6163 the second-level fuzzy comprehensiveevaluation values A101584022 A

101584023 and A101584024 of distillation column

pressure liquid oxygen level in the main condenser anddistillation column top temperature are 10637 958 and 637respectively which are the key parameters of air separationtower system therefore the risk grade index

R = B1 times 007 + B2 times 076 + B3 times 017 = 4796 times 007 +(10637+958+637+124558)times 076+6163times017 = 31056of the air separation distillation section Combined with thedynamic evaluation level in Table 1 the dynamic grade ofthe air separation distillation section in CIP is determinedto be 31056 which is a moderate risk grade The divisionof the dynamic risk level fully considers the influence ofthe historical factors of MHIs on the overall risk and thehistorical data will have decreasing impact on the overall riskover time

5 Conclusion

This study introduces a dynamic early warning methodof MHIs in CIP Data visual qualitative analysis methodand quantitative analysis are conducted to filter the corre-lation variables Feature analysis feature vector acquisitionperforms the dimensionality reduction of key parameterseffectively improving the performance of dynamic recurrentprediction in the chemical process industry parametersHowever there are still some defects in the current study thatwe cannot solve In future research we need to pay moreattention to identification of false alarms and consider humanfactors and social factorsThese research directionswill be thekey points of our future research

Nomenclature

S Composition matrixs1015840119899p Composition matrix is

homogenized 0ndash1rAB Correlation coefficient119883 = xnm Strong correlation variables

data matrixZ = znm Standardized data matrixC Correlation matrixY2m Total variance of 119909119898

xm Overall mean of 119909119898wi i = 1 2 m Eigenvector of the

correlation matrix CΛ = diag(1205821 1205822 120582m) Eigenvalue matrix of thecorrelation matrix C

K Principal component of Cwi i = 1 2 k Eigenvectors of K1205851 1205852 120585119896 Rebuilt K feature after the

reduction of dimension119879 K principal components of

the strong correlationvariable

y(t) Output of the output layeru(t) External input of the input

layerxc(t) Output of the state layerx(t) Output of the hidden layerw(1) Connection weight of the

state layer and the hiddenlayer

w(2) Connection weightbetween the input layerand the hidden layer

w(3) Connection weight of thehidden layer and theoutput layer120579(1) Hidden layer threshold120579(2) Output threshold value

f(x) Represents the activationfunction of the neuralnetwork

I(1)120572 (t) Input of input layer isrespectively

O(1)120572 (t) Output of input layer isrespectively

I(2)120573 (t) Input of the hidden layerO(2)120573(t) Output of the hidden layer

I(C)120574 (t) Input of the state layerO(C)120574 (t) Output of the state layerI(3)120583 (t) Input of the output layerG High alarm value119901 The number of MHIs120590119901 The number of key

parameters for MHIsC119901120590119901 The percentage of the

predicted value of the keyparameter y(t) exceedingthe high alarm value

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Complexity 17

Q119901120590119901 = [Q11205901 Q21205902 Q119901120590119901] The weight of thesecond-level of fuzzycomprehensive evaluation

D119901 = [D1D2 D119901] The weight of the first-levelof fuzzy comprehensiveevaluation

A119901120590119901 = [A11205901 A21205902 A119901120590119901 ] Second-level fuzzycomprehensive evaluationindex matrix

B119901 Risk grade index of theMHIs119905119901 Timeliness weight

A1015840 Second-level fuzzycomprehensive evaluationvalue of eliminating timefactor

L Last time the segment dataexceeds the high alarmvalue to the last return tothe security value

Data Availability

Data source is Quzhou Chemical Industry Park EnterpriseProduction Data Authors do not have permission to publishdata

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by Provincial Key RampD Programof Zhejiang Province (No 2017C03019) National Key RampDProgram of China (No 2016YFC0201400) Zhejiang JointFund for Integrating of Informatization and Industrialization(NoU1509217) International Science and Technology Coop-eration Program of Zhejiang Province for Joint Researchin High-Tech Industry (No 2016C54007) and NationalNatural Science Foundation of China (Nos U1609212 andU61304211)

References

[1] N Khakzad F Khan and P Amyotte ldquoDynamic safety analysisof process systems by mapping bow-tie into Bayesian networkrdquoProcess Safety and Environmental Protection vol 91 no 1-2 pp46ndash53 2013

[2] A S Andersson Development of an Environment-AccidentIndex A Planning Tool to Protect the Environment in Case of aChemical Spill Miljovetenskap 2004

[3] F P Lees Loss Prevention in the Process Industries vol 1ndash3Butterworths London UK 3rd edition 2004

[4] H Zhang H Duan J Zuo et al ldquoCharacterization of post-disaster environmental management for hazardous materialsincidents lessons learnt from the tianjin warehouse explosion

Chinardquo Journal of Environmental Management vol 199 pp 21ndash30 2017

[5] N Khakzad F Khan and P Amyotte ldquoDynamic risk analy-sis using bow-tie approachrdquo Reliability Engineering amp SystemSafety vol 104 pp 36ndash44 2012

[6] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53 pp61ndash73 2018

[7] AMeel andWD Seider ldquoPlant-specific dynamic failure assess-ment using Bayesian theoryrdquo Chemical Engineering Science vol61 no 21 pp 7036ndash7056 2006

[8] C Lv Z Zhang X Ren and S Li ldquoPredicting the frequency ofabnormal events in chemical process with Bayesian theory andvine copulardquo Journal of Loss Prevention in the Process Industriesvol 32 pp 192ndash200 2014

[9] B Abdolhamidzadeh T Abbasi D Rashtchian and S AAbbasi ldquoA newmethod for assessing domino effect in chemicalprocess industryrdquo Journal of Hazardous Materials vol 182 no1-3 pp 416ndash426 2010

[10] A Pariyani W D Seider U G Oktem and M SoroushldquoDynamic risk analysis using alarm databases to improveprocess safety and product quality part iimdashbayesian analysisrdquoAIChE Journal vol 58 no 3 pp 826ndash841 2012

[11] J Zhou G Reniers and L Zhang ldquoA weighted fuzzy Petri-net based approach for security risk assessment in the chemicalindustryrdquo Chemical Engineering Science vol 174 pp 136ndash1452017

[12] P Jain H J Pasman S Waldram E N Pistikopoulos and MS Mannan ldquoProcess resilience analysis framework (PRAF) asystems approach for improved risk and safety managementrdquoJournal of Loss Prevention in the Process Industries vol 53Article ID S0950423017307027 pp 61ndash73 2018

[13] T Aven and M Ylonen ldquoA risk interpretation of sociotechnicalsafety perspectivesrdquoReliability Engineering amp System Safety vol175 pp 13ndash18 2018

[14] G L L Reniers B J M Ale W Dullaert and K SoudanldquoDesigning continuous safety improvement within chemicalindustrial areasrdquo Safety Science vol 47 no 5 pp 578ndash590 2009

[15] R Irani and R Nasimi ldquoEvolving neural network using realcoded genetic algorithm for permeability estimation of thereservoirrdquo Expert Systems with Applications vol 38 no 8 pp9862ndash9866 2011

[16] MKhashei andM Bijari ldquoA new class of hybridmodels for timeseries forecastingrdquo Expert Systems with Applications vol 39 no4 pp 4344ndash4357 2012

[17] X Luo X Zuo and D Du ldquoVarying model based adaptivepredictive control of highly nonlinear chemical processrdquo inProceedings of the International Conference on Control andAutomation vol 1 pp 537ndash540 IEEE 2005

[18] P Goel A Datta andM S Mannan ldquoIndustrial alarm systemschallenges and opportunitiesrdquo Journal of Loss Prevention in theProcess Industries Article ID S0950423017306320 2017

[19] O J Reichman M B Jones andM P Schildhauer ldquoChallengesand opportunities of open data in ecologyrdquo Science vol 331 no6018 pp 703ndash705 2011

[20] P Goel A Datta and M SamMannan ldquoApplication of big dataanalytics in process safety and riskmanagementrdquo in Proceedingsof the 5th IEEE International Conference on Big Data Big Datarsquo17 pp 1143ndash1152 IEEE 2017

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

18 Complexity

[21] F Higuchi I Yamamoto T Takai M Noda and H NishitanildquoUse of event correlation analysis to reduce number of alarmsrdquoComputer Aided Chemical Engineering vol 27 no 9 pp 1521ndash1526 2009

[22] E Saatci ldquoCorrelation analysis of respiratory signals by usingparallel coordinate plotsrdquo Computer Methods and Programs inBiomedicine vol 153 pp 41ndash51 2018

[23] S B Azhar and M J Rissanen ldquoieee 2011 15th internationalconference information visualisation (iv) - london UnitedKingdom (20110713-20110715)rdquo inProceedings of the 15th Inter-national Conference on Information Visualisation - Evaluation ofParallel Coordinates for Interactive Alarm Filtering pp 102ndash109London UK 2011

[24] M R Berthold and F HoppnerOn Clustering Time Series UsingEuclidean Distance and Pearson Correlation 2016

[25] J Zhao X Zhu W Wang and Y Liu ldquoExtended Kalman filter-based Elman networks for industrial time series prediction withGPU accelerationrdquoNeurocomputing vol 118 no 6 pp 215ndash2242013

[26] L G B Ruiz R Rueda M P Cuellar and M C Pegala-jar ldquoEnergy consumption forecasting based on Elman neuralnetworks with evolutive optimizationrdquo Expert Systems withApplications vol 92 pp 380ndash389 2017

[27] B Erkmen and T Yildirim ldquoImproving classification perfor-mance of sonar targets by applying general regression neuralnetwork with PCArdquo Expert Systems with Applications vol 35no 1-2 pp 472ndash475 2008

[28] M G DeGiorgi MMalvoni and PM Congedo ldquoComparisonof strategies for multi-step ahead photovoltaic power forecast-ing models based on hybrid group method of data handlingnetworks and least square support vectormachinerdquoEnergy vol107 pp 360ndash373 2016

[29] F Yu and X Z Xu ldquoA short-term load forecasting model ofnatural gas based on optimized genetic algorithm and improvedBP neural networkrdquo Applied Energy vol 134 pp 102ndash113 2014

[30] A Patil and Z-Q Deng ldquoBayesian approach to estimatingmargin of safety for total maximum daily load developmentrdquoJournal of Environmental Management vol 92 no 3 pp 910ndash918 2011

[31] S Qin J Wang J Wu and G Zhao ldquoA hybrid model based onsmooth transition periodic autoregressive and Elman artificialneural network for wind speed forecasting of the Hebei regionin Chinardquo International Journal of Green Energy vol 13 no 6pp 595ndash607 2016

[32] C A Steed G Shipman P Thornton D Ricciuto D EricksonandM Branstetter ldquoPractical application of parallel coordinatesfor climate model analysisrdquo Procedia Computer Science vol 9no 11 pp 877ndash886 2012

[33] J L Rodgers and W A Nicewander ldquoThirteen ways to look atthe correlation coefficientrdquo The American Statistician vol 42no 1 pp 59ndash66 1988

[34] A Gupta and A Barbu ldquoParameterized principal componentanalysisrdquo Pattern Recognition vol 78 pp 215ndash227 2018

[35] R Dutta J Aryal A Das and J B Kirkpatrick ldquoDeep cognitiveimaging systems enable estimation of continental-scale fireincidence from climate datardquo Scientific Reports vol 3 no 11 p3188 2013

[36] X Tai and LWang ldquoPrediction of short-termpower load basedon elman neural network and fuzzy controlrdquo World Sci-Tech Ramp D 2016

[37] Y Bai H Zhang and Y Hao ldquoThe performance of thebackpropagation algorithm with varying slope of the activationfunctionrdquo Chaos Solitons amp Fractals vol 40 no 1 pp 69ndash772009

[38] L Shifei and W Jiang Identification and Control of MajorHazard Sources Metallurgical Industry Press 2012

[39] C Bo X Qiao G Zhang Y Bai and S Zhang ldquoAn integratedmethod of independent component analysis and support vectormachines for industry distillation process monitoringrdquo Journalof Process Control vol 20 no 10 pp 1133ndash1140 2010

[40] R Taylor ldquoInterpretation of the correlation coefficient a basicreviewrdquo Journal of Diagnostic Medical Sonography vol 6 no 1pp 35ndash39 1990

[41] Z-H Guo J Wu H-Y Lu and J-Z Wang ldquoA case studyon a hybrid wind speed forecasting method using BP neuralnetworkrdquo Knowledge-Based Systems vol 24 no 7 pp 1048ndash1056 2011

[42] S Ding Y Zhang X Xu and L Bao ldquoA novel extremelearning machine based on hybrid kernel functionrdquo Journal ofComputers vol 8 no 8 pp 2110ndash2117 2013

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom

Hindawiwwwhindawicom Volume 2018

MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Applied MathematicsJournal of

Hindawiwwwhindawicom Volume 2018

Probability and StatisticsHindawiwwwhindawicom Volume 2018

Journal of

Hindawiwwwhindawicom Volume 2018

Mathematical PhysicsAdvances in

Complex AnalysisJournal of

Hindawiwwwhindawicom Volume 2018

OptimizationJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

Hindawiwwwhindawicom Volume 2018

Operations ResearchAdvances in

Journal of

Hindawiwwwhindawicom Volume 2018

Function SpacesAbstract and Applied AnalysisHindawiwwwhindawicom Volume 2018

International Journal of Mathematics and Mathematical Sciences

Hindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018Volume 2018

Numerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisNumerical AnalysisAdvances inAdvances in Discrete Dynamics in

Nature and SocietyHindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Dierential EquationsInternational Journal of

Volume 2018

Hindawiwwwhindawicom Volume 2018

Decision SciencesAdvances in

Hindawiwwwhindawicom Volume 2018

AnalysisInternational Journal of

Hindawiwwwhindawicom Volume 2018

Stochastic AnalysisInternational Journal of

Submit your manuscripts atwwwhindawicom