proopter: production dynamics analysis and optimization tool · 2017-10-27 · proopter, production...

ProOpter, production dynamics analysis and optimization tool

Gasper Music(a), Miha Glavan(b), Dejan Gradisar(b), Stanko Strmcnik(b)

(a)University of Ljubljana, Faculty of Electrical Engineering, Ljubljana, Slovenia(b)Jozef Stefan Institute, Ljubljana, Slovenia

(a)[email protected]

ABSTRACTThe paper deals with analysis and optimization of produc-tion performance dynamics expressed by key performanceindicators (KPI). The architecture of a supporting softwaretool that integrates into production information systems ispresented. The tool is able to support efficiency indicatorscomposition, determination of performance influentialvariables and building of prediction model that enables ashort-term prediction of the expected performance. Themodel can be used within predictive control algorithms,composing a production management decision support,which assists production operators at on-line productionmanagement. The tool is built modularly and employsindustrial standards that facilitate the tool integration intoproduction information systems of different vendors. Thefunctionality of the tool is demonstrated by simulationbenchmark as well as by real industrial production data.

Keywords: Production control, production efficiency,key performance indicators, modelling, simulation,optimization

1. INTRODUCTIONManufacturing companies are faced with several chal-lenges in their struggle to gain a competitive advantage.Among others, energy and resource efficiency are becom-ing increasingly decisive factors for competitiveness, com-panies are forced to shorten innovation cycles, and marketsare becoming more volatile.

These challenges can be addressed by the extensiveuse of contemporary information technology within all or-ganization levels of a company. A careful orchestration ofthe related information system improvements throughoutthe company has to be maintained in order to avoid bottle-necks on the one hand, and oversized capacity on the other.In any case, production itself is a base of any manufactur-ing company operation. Effective production managementis one of the fundamental operational activities that has tobe carefully designed and integrated into the overall man-agement structure in order to meet the given requirements.

A number of initiatives emerged with a goal to estab-lish a continual production efficiency improvement frame-work by the use of the latest technological advancements.In Europe, the German Industry 4.0 initiative (Industrie 4.0

2013) is perhaps the most known. It proposes to employthe Internet of Things paradigm on the factory floor andderive intelligent, intercommunicating autonomously op-erating production units, co called cyber-physical systems.This should result in flexible and efficient Smart Factory,with consideration of ergonomics and customer needs, andintegration of supply-chain partners along the value chain(Brettel et al. 2014).

A similar North American initiative is Smart Man-ufacturing Leadership Coalition (SMLC 2011), a coali-tion of companies, manufacturing consortia and consul-tants working on Smart Manufacturing. This has been de-fined as the dramatically intensified application of ’man-ufacturing intelligence’ throughout the manufacturing andsupply chain enterprise (Davis et al. 2012).

In terms of organizational advancements Japanesemanufacturing has long tradition dating back in 1980’swith Just-in-time manufacturing and lean production (Hol-weg 2007). Recently, also other Asian manufacturers areforced to adopt these advancements as they are faced withrising consumer sophistication that makes traditional lowcost mass production strategies inadequate.

While these initiatives improve the organization of theproduction process and establish a decision-support frame-work with improved insight into the current state of pro-duction and its current performance, an obvious furtheradvancement is to employ sophisticated data analysis andsystem modelling methods in order to predict the effectof production control measures on the future performance.This way a decision-making process at the production con-trol level would be substantially improved by the possibil-ity of beforehand evaluation of the various decisions.

To this end a holistic production control concept isproposed in Zorzut et al. (2009) and further elaborated inGlavan et al. (2013a) and Glavan et al. (2013b). The pro-posed modelling approach uses historical production per-formance data jointly with history of most influential de-cision variables to derive a black-box model by contempo-rary soft computing techniques, e.g. neural networks. Sev-eral additional steps are provided, such as data preparation,influential variable selection, etc. (Glavan et al. 2013a).

This paper describes the design and prototype imple-mentation of a software tool that supports the applicationof modelling methodology in the context of productionperformance prediction and optimization. This is one of

Proceedings of the European Modeling and Simulation Symposium, 2014 978-88-97999-38-6; Affenzeller, Bruzzone, Jiménez, Longo, Merkuryev, Zhang Eds.

392

the fields where information technology has an immedi-ate and considerable impact on the efficiency and qualityof production control and related manufacturing processes.The tool modular architecture is described and the used in-dustrial standards are presented, which facilitate the toolintegration into production information systems of differ-ent vendors. The functionality of the tool is demonstratedthrough a selection of results from case studies, which in-clude simulation benchmark as well as historical data fromthe actual industrial production.

2. HOLISTIC PRODUCTION CONTROL

The presented tool is based on the holistic production con-trol concept, which attempts to bring classical feedbackcontrol approach, extensively used on the factory floor,also to the higher production control levels.

The concept of holistic production control can be bestexplained by the scheme depicted in Figure 1. The processwe would like to control is indicated by the blockProduc-tion process. Note that this block also covers the low-levelprocess control. Different inputs (u) are available to ma-nipulate the production process. These inputs are mostlythe reference values for the process control loops and/orother manipulating variables not used within the stabilisa-tion loops. On the other hand, there are many measurabledisturbances (d) and outputs (y), which are used to cal-culate on-line the key performance indicators (K) – KPIcalculationblock. The key performance indicators (KPIs)are the production variables that are used by productionmanager to determine the appropriate input values (u) inorder to optimise the production process. The demandsfrom the business control level are given as reference KPIs(K∗). The attempt of the concept described here is to helpthe production manager with the decision-making process,which would close the loop through the introduction of theProduction controller & Optimiser.

One of the possible solutions with this approach isto apply model-based control concept on the productionmanagement level. To enable this, an appropriate modeldescribing the behaviour of the process projected on KPIsis required (theKPI modelM ). The model can be updatedonline and can provide the controller with the predictedoutputsK|M . The model can also consider the measur-able disturbances (d). Based on current valuesK, the pre-dicted outputsK|M and the reference valuesK∗ (i.e., theplanned business goals) theProduction controller & Opti-miserdetermines the appropriate input valuesu and in thisway supports (or substitutes) the production manager (Eq.(1)).

u = argminu∈R

C(K,K∗,K|M) (1)

This concept has been introduced in Zorzut et al.(2009), and further elaborated and extended in Glavan et al.(2013a) and Glavan et al. (2013b).

How to derive an appropriate KPI model representsthe main challenge of the HPC approach.

K

K

u

u

Figure 1: Holistic production control

2.1. Production modelling for holistic production con-trolModern manufacturing systems are in many cases and forvarious reasons too complex to be accurately described an-alytically from first principles. Instead, one can assumethat the relationship between the inputs and the outputs canbe described by a stochastic, high-dimensional model froma class of generally nonlinear model structures.

The production model has to include enough detailsof the production process to reflect the dynamics for pro-duction control. This model should be relatively simplein comparison to the models used for the process controllevel, yet because of the overall complexity and the limita-tions of testing the process, this task is extremely complex.Production control usually requires that the model is easyto adapt online as well. Therefore, the main objective isthe development of the concept of identifying a relativelysimple input–output model of the production.

The main steps of the production modelling are shownin Figure 2. More detailed discussion about each of thesesteps and a short overview of the potential methodologiesare given in the following subsections.

Data preprocessing

Special attention is needed when data from a historicalproduction database are used. From the vast amount ofdata, the informative portions need to be identified. Thesedata segments should cover the interesting dynamics of theKPIs, for which we would like to determine the future be-haviour. Furthermore, any outliers or missing data need tobe properly substituted, and to cover all the process oper-ating conditions an uneven data distribution is needed.

Data-cleaning procedures can be applied to detect andremove any outliers present in the data. As pointed outin Pearson (2006), nonlinear data-cleaning procedures arerecommended. We can find many filters in the literatureproposed for this task: the Martin–Thomson filter, the FIR-median hybrid (FMH) filter, the Hampel filter, etc.


393

Historicalprocess

data

Datapreprocessing

Definition ofproduction KPI

Inputvariableselection

Black-boxprocess

modelling

KPI predictionmodel

Figure 2: Holistic production control design steps

Definition of production performance indicators – KPIs

The research field of performance measurement systems(PMSs) is becoming increasingly important for industryand academics (Neely et al. 1995). Performance indica-tors (PIs) are commonly used by organisations to evalu-ate their overall economic success and many recommen-dations have been presented on how to specify such indi-cators (Folan and Browne 2005). These indicators shouldcover the relevant business aspects of the specific process,where the status of the process should be evaluated withrelatively short-term indicators, since longer-scale busi-ness estimations (e.g., an annual profit report) are uselessfor quick process adjustments (Gerry and Buckbee 2005).

The production objectives are usually aggregated inkey production performance indicators (KPIs). The selec-tion of these KPIs should be performed manually, with ex-tensive consideration of the production expert’s knowledgeof the process.

Input variables selection

To simplify the model and to enhance model’s accuracyonly the most relevant manipulative variables need to beidentified. On the basis of the historical process data, anextensive analysis is needed to evaluate which inputs havethe most significant impact on the selected KPIs.

Variable selection is already a widely applied method-ology in the field of data mining. But, as noted by someauthors, like Smits et al. (2006), in modelling projects it ismostly assumed that true inputs are a-priori known or allthe available inputs are used in a model. To avoid the so-called curse-of-dimensionality, which essentially limits therobustness of a data-based model, only the most relevantinputs need to be selected. This represents an especiallyimportant step for HPC design, as in real-world productionprocesses many potential variables are available. Further-more, as aggregated KPIs are connected with many pro-cess variables it is often found that some a-priori excludedinputs are later identified as significant, and vice versa.

In the literature, three major principles for variableselection are used (Guyon and Elisseeff 2003):

• feature construction,

• variable ranking,

• variable subset selection.

For detailed discussion and evaluation of the relatedmethods, see Glavan et al. (2013b).

As we are dealing with dynamical systems, the cur-rent values of production performance indicators are notdependent only on the current input values, but also ontheir time-delayed values. The input–selection problem istherefore augmented by the selection of lagged inputs andoutputs that are used as regressors.

Black-box process modelling

The HPC efficiency is closely related to the productionmodel, which should describe the main features of the pro-duction process with an acceptable level of approximation.The production process is typically a highly complex pro-cess, with nonlinear relationships among the vast quantityof process variables. Since our model should be simpleenough and the development time needs to be short, black-box modelling techniques are preferred. Furthermore, theproduction model should be extracted mainly from the his-torical process data, since extensive experimentation on thereal process is often too expensive or restricted. If the pro-cess characteristics were to change during the use of theproduction model, new process data should be analysedand a better model extracted. The cyclical generation andvalidation of new models will enable a rather conservativeadaptation of the model-in-use to long-term changes in theproduction.

The main idea of the parametric black-box modellingtechniques is to trim some universal input–output func-tions, with a fixed number of parameters, to represent thetrue process dynamics (2).

y(t) = g(ϕ(t), ϑ) + e(t) (2)

The goal is to minimise the mismatche(t) betweenthe true process responsey(t) and the model predic-tion g(·), where the trimming is performed solely onthe basis of the process input–output data pairsZN ={u(t), y(t)}Nt=1

Black-box models can be seen as the connection oftwo mappings (Sjoberg et al. 1995). The first mappingconstructs the regression vectorϕ(t) from past inputs andoutputs, which enables a representation of the dynamicalmodel behaviour. Another mapping predicts the future be-haviour of the systemy(t), with the nonlinear mappingfrom the regressor space to the system output. This non-linear mapping is found within a family of functions (3),parameterised by the parameter vector (ϑ), wheregk refersto the basis function, usually derived from a single, motherbasis function.

y(t) = g(ϕ(t), ϑ) =∑N

k=1αkgk(ϕ(t), βk, γk)

ϑ = [α1 . . . αn, β1 . . . βn, γ1 . . . γn]T (3)

From such a flexible structure, popular nonlinearmappings can be derived, like Neural Networks, Wavelets,Kernel Estimators, Nearest Neighbors, B-splines, Fuzzymodels, etc. (Ljung 1999).

2.2. Optimization and holistic production controlOptimization is a vital component of holistic productioncontrol concept. Once a production performance dynamicmodel is derived, a natural way of incorporating the modelin a production control scheme is to apply the model pre-dictive control (MPC) approach (Maciejowski 2002, Gruneand Pannek 2011).

The MPC approach is based on repetitive solution ofan optimal control problem, taking measured system stateas the initial state and using system model to evaluate the


394

effect of possible system input sequences. A discrete timerepresentation of system dynamics is used and only the firstsample of the calculated optimal input sequence is appliedto the system. In the next sample time the calculation isrepeated with the newly acquired system state.

The optimization is performed over a finite movinghorizon, which always starts at the current sampling in-stant. Depending on the method, prediction and controlhorizons can be of different lengths. The approach canbe applied to feedback control of nonlinear systems, themethod is then denoted NMPC.

For simplicity, we assume prediction and control hori-zon of lengthN . The determination of the optimal produc-tion manipulative valuesu can therefore be formulated asan optimal control problem

minimize JN (x0, u(·)) :=

N−1∑

k=0

l(xu(k, x0), u(k))

with respect to u(·) ∈ UN (x0) (4)

subject to xu(0, x0) = x0,

xu(k + 1, x0) = f(xu(k, x0), u(k))

Herex0 = x(n) ∈ X is the current state of the system atthe sampling instantn = 0, 1, 2 . . ., u(·) ∈ U

N(x0) is acontrol sequence whereUN (x0) ⊆ UN is a set of admis-sible control sequences over which we optimize,l(x, u) isa distance (deviation) measure,J(x, u) the cost function,andf(x, u) is the nonlinear state transition map represent-ing discrete time dynamics.

When the solution is obtained,u = u(0) is chosen.For a non-constant reference and in the presence of termi-nal constraints the formulation is slightly modified, but thegeneral idea remains the same.

In any case, the NMPC problem can be reformulatedto match the standard problem in nonlinear optimization(NLP) (Grune and Pannek 2011)

minimize F (z)with respect to z ∈ R

nz

subject to G(z) = 0 and H(z) ≥ 0(5)

with mapsF : Rnz → R, G : R

nz → Rrg andH :

Rnz → R

rh . Hererg andrh denote the resulting numberof equality and inequality constraints, respectively.

This reformulation, called discretization, can be donein different ways. Among them, the so-called recursive dis-cretization recursively computesxu(k, x0) from the open-loop system dynamics outside of the optimization problem,thus maintaining low dimension of the optimization vari-ablez and the number of constraintsp = rg + rh. The op-timization variablez reduces toz := (u(0)T , . . . , u(N −1)T ).

Utilizing this form, common solution methods forNLP problems can be applied, such as SequentialQuadratic Programming (SQP) or Interior Point Methods(IPM).

The recursive discretization is optimal regarding thenumber of optimization variables and constraints. Still,

the method has some drawbacks compared to the full dis-cretization. In particular, it is difficult to use the informa-tion on current control sequence solution in search for goodinitial guess in the next step, and the solutionxu(k, x0)may depend very sensitively on the control sequenceu(·),in particular whenN is large (Grune and Pannek 2011).

The advantage of the method is that it separates theNLP solver from the solver of the dynamics. This is par-ticularly well suited to models derived by soft-computingtechniques, because an arbitrary numeric implementationof the model can be used. In the case of HPC, this enablesto use the above mentioned black-box models in the formof nonlinear mappings in a straightforward way.

2.3. Software supportTo assist the modeller in deriving production performancemodel for the HPC, and facilitate the application of theNMPC methods within HPC, a user-friendly software toolis needed. Using such a support tool, the system integratorand the production manager would have the possibility toidentify a production model based on the historical opera-tional data of the process, and integrate it in a model-basedHPC solution.

The main purpose of the developed tool is to facil-itate implementation of performance indicators, analyzethe main performance influences, automate the model-development procedure and to support the manipulationand maintenance of already existing models, as well astheir use in production optimization. As the potential usersof such a tool are non-modelling experts (e.g., productionmanagers), the program tends to simplify the model iden-tification and control optimization procedures, where theuser would not need to understand a detailed identificationand control theory.

3. PROOPTER - PRODUCTION DYNAMICS ANA-LYZER AND OPTIMIZERProOpter extends the functionality of classical Manufac-turing Execution Systems (MES) with embedded intelli-gence. It enables the analysis of production dynamics us-ing complex analytical functions and application of ad-vanced production control concepts that are based on em-bedded models. The Production dynamics analyser andoptimizer represents an upgrade of classical MES systems,and thus increases the functionality and efficiency of pro-duction information systems.

3.1. ArchitectureProOpter enables the analysis of the production dynamicsusing advanced methods like data mining, data reduction,determination of relevant manipulated variables and pro-duction performance indicators model identification, as de-scribed in Section 2. The obtained models enable a short-term prediction of production dynamics, which is the basisfor production optimization. The general architecture isshown in Figure 3.

ProOpter is composed of several modules, wheresome of them are used on-line, and others off-line. Con-nectivity with classical MES systems is enabled by stan-dard IT interfaces, described in the following subsections.


395

ProcesscontrolContinuous controlDiscrete controlBatch control

Bussinesplanning &logisticERP

ProductioncontrolManufacturingOperations & ControlMES

ProOpterISA 95

Performancemeasures

Influentialvariables

Modelling

Performanceprediction

Optimization

off-line

on-line

Datapreparation

Comm-unication

Figure 3: ProOpter architecture

3.2. ProOpter workflowProOpter workflow is depicted in Figure 4. The mainProOpter output is a set of manipulative variable valuesthat should bring the manufacturing process close to theoptimal operating mode.

The performance of the process is estimated througha set of KPI values. The indicators are calculated from thedata acquired by the production management informationsystem (MES) and stored in the production database. Var-ious data irregularities typically appear in a real produc-tion environment, therefore the data is preprocessed andcleaned. During ProOpter development, a particular at-tention has been devoted to simplification of the KPI def-inition, and a corresponding data viewer and KPI formulaeditor have been implemented. When using on-line, theperformance monitoring module supports various viewswith adjustable level of aggregation. Complex KPIs canbe monitored and can also be drilled down to monitor theircomponents.

The detemination of the most influential variables isone of the key steps in deriving applicable performance dy-namic model. To support this a set of standard input vari-able selection (IVS) methods can be called from ProOpterIVS module. A user interface has been designed, which en-ables an aggregate representation and evaluation of resultsobtained by various methods. Due to predominantly non-linear relations within the production process the IVS anal-ysis results are typically non-uniform, and final decision onthe set of variables that will be used in model identificationhas to be performed by the modeller. The developed userinterface assist the modeller by providing graphical repre-sentation of the variable impact in various configurations.

A separate module has been developed for modelidentification support. The modeller can choose the datasegments used for model identification and model valida-tion, and adjust various identification parameters. Twosoft computing methods are currently supported: neural-network based identification and fuzzy model identifica-tion.

The on-line operation of ProOpter provides the in-sight into production performance as described above, aswell as supports the production manager in determining theoptimal production settings. The nonlinear model predic-tive control described in Section 2 is applied here with animportant distinction comparing to the classical MPC ap-plications on the process level: the calculated inputs are notautomatically applied to the production process, the deci-sion on the application of the calculated manipulative val-ues to the process is left to the production operator instead.

Some of the described ProOpter features are furtherillustrated by case studies in Section 4.

3.3. Data integration and communication standardsProOpter was designed as an add-on to existing productionmanagement information systems. These typically consistof ERP and MES systems on the higher levels, which areconnected to SCADA systems, PLCs and other automationequipment on the factory floor.

One of the main requirements for such an add-on is to support established industrial standards to beable to connect and integrate into existing information-communication infrastructure.

This was achieved by supporting standardized datainterchange message format, which is typically used forcommunication among business planning and productioncontrol applications. Equally important is to use standard-ized communication messages among tool modules. Thisenables simple addition or removal of the functional mod-ules as well as their upgrade without compromising thefunctionality of the rest of the system.

Therefore the following standards were consideredand used in the software tool development:

IEC/ISO 62264 (ANSI/ISA 95) standard defines modelsand terminology that are used for marking informa-tion, which has to be exchanged between businesslevel and production control level. The information isstructured by UML models and is a basis of produc-tion control integration into overall business informa-tion systems. The standard is used by IT developersas a guideline for specifying user requirements and asa basis of systems and databases development. Thestandard can be used in any type of production: con-tinuous, batch or discrete.

Within the ProOpter tool the standard was used fordesigning internal communication messages. Themessages were designed based on XML implemen-tation of the standard that was developed by WBF(The Organization for Production Technology). Eachdata model of the standard has an XML equiva-lent within the B2MML (Business to ManufacturingMarkup Language), and the model structure is definedby XML schemes (XSD).

Among the standard data models ProOpter tool usesthe model that is intended for production response re-porting (Production Performance).

Open Math is a standard for representation of mathemat-ical objects that has been developed by OpenMath


396

DB

ProOpter off-line analysis ProOpter on-line optimization

Productionprocess

publishdataget

data(p

ull)

Datapreprocessing

KPIdefinition

Modelidentification

Influentialvariablesdetermination

manipulativevalues

Datapreprocessing

Humanevaluationof results

Optimization(search foroptimal )u

KPI monitoring- drill down

Define KPI targetsModel selection

Production managementinformation system

Figure 4: ProOpter workflow

Society. Two equation representations are supported,XML and binary format. The standard is based on theuse of extendable content dictionaries, which struc-turally document mathematical definitions and unam-biguously define mathematical symbols. Their open-ness enables definition of new functions and symbols.Custom content dictionaries can be added to commonOpenMath repository and used by all interested par-ties. The standard does not require that all of thenewly developed content dictionaries are public, buthave to be accessible to all users of the particular con-cept.

By using OpenMath the KPIs can be described in astandard way that is recognised by various ProOptermodules, and can be exported to other tools.

PMML is used for transfer of identified KPI performancemodels among the ProOpter modules. PMML (Pre-dictive Model Markup Language) is a standardizedXML scheme for model representation in the areaof predictive analytics and data mining. The PMMLstandard has evolved during the years and the recentversion includes a rich set of model types and is sup-ported by large number of statistical software tools.

The use of the described standards for inter-modulecommunication within ProOpter is illustrated in Figure 5.

optimizationmodule

Modelling

KPI definition

DB

Influentialvariables

identification

Da

ta p

rep

ara

tio

n m

od

ule

(filte

rin

g,

KP

I ca

lcu

latio

n,

va

ria

ble

se

lectio

n)

Co

mm

un

ica

tio

n m

od

ule

rea

l-tim

e m

on

ito

rin

gm

od

ule

data module

KPIpredictionmodule

Production information system(MES system)

...

ProOpter

B2MML PMML OpenMath

off-line modules

on-line modules

Communication messages

Figure 5: ProOpter standard messages

3.4. Communication

The use of unified communication protocols within pro-duction information system is of significant importance.


397

Figure 6: ProOpter data module

Use of communication standards guarantees the long termstability of developed applications as well as easier integra-tion with the existing information infrastructure. In com-pliance with IEC 62264 two standard ways of data trans-mission are used within ProOpter:

• Pull model - data user requests the data from the datasupplier, which sends the data upon request. This is apoint-to-point communication.

• Publish model – data supplier is sending data to therecipients that are subscribed to specific data. Thecommunication is carried on according to publish-subscribe principle.

The use of MOM (Message Oriented Middleware)systems is best suited for such an asynchronous XML mes-sage exchange among several clients. It supports bothpoint-to-point and publish-subscribe communication prin-ciples. MOM enables distributed communication amongloosely coupled clients, meaning that communicating ap-plications do not necessarily all have to be active withinthe network, nor do they all have to be aware of each other.

Within ProOpter, MOM communication was imple-mented by JMS (Java Message Service) specification (Fig-ure 6). Even if JMS originates from Java environment it isgeneral and extensible enough to be useful also for clientsthat are designed within other development environmentsand platforms.

4. CASE STUDIESThe performance of the tool was tested through a set ofproduction monitoring and control case studies of variousproduction types. The continuous production was coveredby a simulation benchmark, while in discrete and batchproduction the tool was tested on a real production data.

4.1. Batch productionActual batch production data were considered in connec-tion to performance measurement and influential variableanalysis in water based paints and coatings production.The production is a typical batch process consisting of dos-ing and mixing stage, milling stage, production stage, and

packing stage. In between production and packing stagesthe quality control is performed, where the acceptability ofproduct for packing is determined. In the case of negativeresult, the product can be scrapped, or can go to re-workwhere the quality is improved.

Performance indicators

The performance monitoring in the considered batch pro-duction was mainly oriented toward determination of suit-able batch parameter settings. Therefore batch related in-dicators were of primary concern. In cooperation with pro-duction managerial staff the following relevant indicatorswere identified:

• Quality - calculated per batch as a result of laboratoryanalysis. Different products have different measuredparameter sets.

• Raw materials consumption ratio - ratio of the actualraw material consumption per work order to the nor-mative consumption.

• Timeliness - difference between the actual finish timeand the planned finish time.

• Scrap rate - ration of the actual scrap quantity to theplanned produced quantity.

Among the chosen indicators, Quality is the mostcomplex and needs additional explanation. As the mea-sured parameter sets differ among different products theyhave to be normalized and aggregated to derive indicatorvalues that are comparable among several batches. This isillustrated in Figure 7. For every product (or product fam-ily) the relevant set of laboratory measured quality param-eters is identified, and these are normalized and agregatedinto a standard valued quality indicator (common estimatein Figure 7).

Good Bad

Om

O1

O2

wm

.

.

.

.

.

.

.

.

.

w2

w1

Σ

Q1

Q2

Qm

Qn1Qn1

Qn2

Qnm

normalization aggregation

common

estimate

Figure 7: Aggregated KPI composition

The normalization and aggregation is supported byProOpter KPI definition module. Two of the normalizationscreens are shown in Figures 8 and 9. For every measure-ment, the user can chose one of the predefined normaliza-tion types and then adjust the relevant parameters. Finally,the normalized measurements are aggregated into a KPI.


398

Figure 8: ProOpter normalization

Figure 9: Predefined normalization types

Influential variables

In the considered case of coatings production the qual-ity indicator is most relevant for the production efficiency.Bad quality products go either to re-work, which decreasesproductivity, or go to scrap, which is a direct cost. Theanalysis of influential production variables that determinethe quality level is therefore the most important. Unfor-tunately, the analysis can not be generalized but has to beperformed for every product family due to various produc-tion recipes.

The ProOpter tool supports the analysis by embed-ding several standard and advanced variable selectionmethods that are used in data mining: Linear Correla-tion, Partial correlation (with forward selection approach),PLS (variable importance in projection – PLS VIP), Non-Negative Garrote, LASSO, DMS search (pareto search ofminimum error of linear model as objective function), andothers. The problem of variable selection is elaborated indetail in (Glavan et al. 2013b).

The results for one of the product families are shownin figure 10. The aggregated results of a subset of avail-able variable selection methods are shown as a box plot.In the given case, Max. RPMs of the mixer dominate thebatch quality, and interestingly also the temperature in theproduction hall points out to be important as well.

Figure 10 points out another specific of the given pro-duction case. As several production variables are continu-ously sampled we typically have a range of correspondingvariable values for a batch. Most of the used variable selec-

0 0.2 0.4 0.6 0.8 1

CURRENT_mean(k−0)

timeGross(k−0)

TEMPERATURE_mean(k−0)

Temp_shop_floor 7(k−0)

RPM_max(k−0)

Figure 10: Selection of Quality indicator influental vari-ables for a product family

tion methods require for such values to be aggregated into asingle data point when considering the variable influence tothe batch quality. For this purpose a set of standard data ag-gregation functions are provided within the ProOpter Datamodule, e.g. average, min, max.

4.2. Tennessee Eastman benchmarkThe Tennessee Eastman (TE) benchmark process was in-troduced by Downs and Vogel (1993) as a model of areal chemical production process. The model representsa test problem for researchers to experiment with differ-ent control-related solutions. The process consists of fivemajor units: a chemical reactor, a product condenser, avapour–liquid separator, a product stripper and a recyclecompressor. Four reactants (A, C, D, E) and an inert com-ponent (B) are entering the process, where four exother-mic, irreversible reactions result in two products (G, H)and one byproduct (F). The process products leave the pro-cess through an output stream, and are later separated in adownstream refining section. The production process has41 measured variables (y) and 12 different manipulativevariables (u).

A specific combination of the production rate and/orthe product mix are usually demanded by the market orsome capacity limitations. Therefore, six typical opera-tional modes are defined in the benchmark proposal. Themodel also provides 20 different process disturbances,which imitate the disturbances typical of real TE produc-tion.

As the research literature provides several results onthe control of TE process by various methods, the proceswas used as a primary test bench for verifying the opera-tion of the various ProOpter modules and validation of theobtained results.

Production modelling

The historical data records needed for production perfor-mance evaluation and modelling were generated by sim-ulation. To make the problem more realistic the processmeasurements have some intensionally added noise, typi-cal for the specific measurement.

Production performance is monitored through threeKPIs: Costs, Production, and Quality. The definition ofProduction KPI is quite straightforward, as the quantity ofproduct leaving the process is directly measured. Directlyfrom the process objectives an indicator for the process


399

Figure 11: Modelling with ProOpter

quality is also derived, since the product quality can beviewed as a desired mass ratio between the two final prod-ucts, G and H. For details on the Cost KPI definition, thereader is referred to (Downs and Vogel 1993) or (Glavanet al. 2013a).

Next, the influential variable analysis is performedconsidering available manipulative variables and the de-fined KPIs. As described in Glavan et al. (2013b), five ma-nipulative variables are selected, and then a neural-networkbased model is identified that enables to calculate a short-term prediction of the defined KPIs, provided chosen ma-nipulative values are specified. The modelling is supportedby the ProOpter modelling module, shown in Figure 11.

On-line production efficiency monitoring and optimiza-tion

The use of the derived performance models for real-timemonitoring and performance prediction is illustrated inFigure 12. Observation variables can be chosen in theupper part of the window, while the chosen variable his-tory can be seen in the lower part. The right sections ofthe plots indicate the estimated future KPI evolution basedon the identified performance model and the manipulativevalues. A detailed investigation of the variables that con-tribute to the performance measure as well the insight intothe manipulative values is possible by choosing thedrill-downoption.

Additional optimization module supports the adjust-ment of the manipulative values in a model based predic-tive control manner. The identified performance model isused to determine optimal manipulative variable settings,while the user has to specify appropriate KPI targets. E.g.,Figure 13 illustrates how Production and Quality can bemaintained at certain levels while the production Cost isminimized.

Figure 12: Efficiency monitoring and prediction

Figure 13: Efficiency optimization

6. CONCLUSIONS

Production information systems currently used in produc-tion companies are very efficient in data collection and datapresentation, but are quite limited in the support of deci-sion making and production optimization. The ProOpter -Production dynamics analyser and optimizer enables effec-tive control of various types of manufacturing processes.Its main advantage is in use of simplified dynamic KPImodels that are identified from historical data.

ProOpter will unburden the production manager andhelp him take better decisions in order to improve the pro-duction process. With the introduction of Production dy-namics analyser and optimizer we can expect savings invarious segments in the production through better productquality, increasing the efficiency, reduction of waste, andproduction cost reduction.


400

ACKNOWLEDGMENTSThe presented work has been partially performed withinthe Competence Centre for Advanced Control Technolo-gies, which is partly financed by the Republic of Slovenia,Ministry of Education, Science and Sport, and EuropeanUnion, European Regional Development Fund. The sup-port is gratefully acknowledged.

REFERENCESBrettel, M., Friederichsen, N., Keller, M. and Rosen-

berg, M., 2014. How virtualization, decentraliza-tion and network building change the manufacturinglandscape: An industry 4.0 perspective,InternationalJournal of Mechanical, Industrial Science and Engi-neering, 8 (1), 37 – 44.

Davis, J., Edgar, T., Porter, J., Bernaden, J. and Sarli,M., 2012. Smart manufacturing, manufacturing in-telligence and demand-dynamic performance,Com-puters & Chemical Engineering, 47 (0), 145 – 156.{FOCAPO} 2012.

Downs, J. and Vogel, E., 1993. A plant-wide industrialprocess control problem,Computers & Chemical En-gineering, 17, 245–255.

Folan, P. and Browne, J., 2005. A review of perfor-mance measurement: towards performance manage-ment,Computers in Industry, 56, 663–680.

Gerry, J. and Buckbee, G., 2005.The link between au-tomation and enterprise KPIs, ExpertTune, Plant Per-formance Supervision & PID Tuning Software, WhitePaper.

Glavan, M., Gradisar, D., Atanasijevic-Kunc, M.,Strmcnik, S. and Music, G., 2013b. Input variableselection for model-based production control and op-timisation, The International Journal of AdvancedManufacturing Technology, 68 (9-12), 2743–2759.

Glavan, M., Gradisar, D., Strmcnik, S. and Music, G.,2013a. Production modelling for holistic productioncontrol, Simulation Modelling Practice and Theory,30, 1–20.

Grune, L. and Pannek, J., 2011.Nonlinear Model Pre-dictive Control: Theory and Algorithms, Communi-cations and Control Engineering, 1st edn, Springer.

Guyon, I. and Elisseeff, A., 2003. An introduction to vari-able and feature selection,Journal of Machine Learn-ing Research, 3, 1157–1182.

Holweg, M., 2007. The genealogy of lean production,Journal of Operations Management, 25 (2), 420 –437. Special Issue Evolution of the Field of Op-erations Management SI/ Special Issue OrganisationTheory and Supply Chain Management.

Industrie 4.0, W. G., (2013). Recommendations for im-plementing the strategic initiative INDUSTRIE 4.0,Final report, Industrie 4.0 Working Group.

Ljung, L., 1999. System Identification: Theory for theUser, 2nd edn, Prentice Hall.

Maciejowski, J., 2002. Predictive Control with Con-straints, Prentice-Hall.

Neely, A., Gregory, M. and Platts, K., 1995. Performancemeasurement system design - a literature review andresearch agenda,International Journal of Operations& Production Management, 15 (4), 80–116.

Pearson, R., 2006. Nonlinear empirical modelingtechniques,Computers & Chemical Engineering,30, 1514–1528.

Sjoberg, J., Zhang, Q., Ljung, L., Benveniste, A., De-lyon, B., Glorennec, P.-Y., Hjalmarsson, H. and Ju-ditsky, A., 1995. Nonlinear black-box modeling insystem identification: a unified overview,Automatica,31, 1691–1724.

Smits, G., Kordon, A., Vladislavleva, K., Jordaan, E. andKotanchek, M., 2006. Variable selection in industrialdatasets using pareto genetic programming,in T. Yu,R. Riolo and B. Worzel (eds),Genetic ProgrammingTheory and Practice III, Vol. 9 of Genetic Program-ming, Springer US, pp. 79–92.

SMLC, (2011). Implementing 21st century smart man-ufacturing, Workshop report, Smart ManufacturingLeadership Coalition and U.S. DOE.

Zorzut, S., Gradisar, D., Jovan, V. and Music, G., 2009.Use of a procedural model in the design of productioncontrol for a polymerization plant,The InternationalJournal of Advanced Manufacturing Technology, 44(11-12), 1051–1062.

AUTHORS’ BIOGRAPHIESGASPER MUSIC received B.Sc., M.Sc. and Ph.D. de-grees in electrical engineering from the University ofLjubljana, Slovenia in 1992, 1995, and 1998, respec-tively. He is Associate Professor at the Faculty ofElectrical Engineering, University of Ljubljana. His re-search interests are in discrete event and hybrid dynam-ical systems, supervisory control, planning, scheduling,and industrial informatics. His Web page can be foundat http://msc.fe.uni-lj.si/Staff.asp.

MIHA GLAVAN received B.Sc. in electrical engineer-ing from the University of Ljubljana, Slovenia in 2010.He is Ph.D. student at Department of Systems and Con-trol, Jozef Stefan Institute. His research interests arein modelling of dynamical systems, neural networks andmodel predictive control. His Web page can be foundat http://dsc.ijs.si/en/people/miha_glavan/.

DEJAN GRADI SAR received B.Sc. and Ph.D. de-grees in electrical engineering from the University ofLjubljana, Slovenia in 2001 and 2006, respectively. Heworks at the Jozef Stefan institute as a postdoctoralassociate. His research interests are in production con-trol, scheduling and control systems. His Web page is athttp://dsc.ijs.si/en/people/dejan_gradisar/.

STANKO STRM CNIK received his B.Sc., M.Sc., andPh.D. from the Faculty of Electrical Engineering, Univer-sity of Ljubljana in 1972, 1975 and 1979 respectively. In1973 he joined J. Stefan Institute in Ljubljana. In years1986-2011, he has been the head of the Department ofSystems and Control at the same institute. His researchinterests concern mathematical modelling, identification,process control and non-technical aspects of automation.He is also an associated professor at the Faculty of Electri-cal Engineering, University of Ljubljana and full professorat University of Nova Gorica. His Web page is athttp://dsc.ijs.si/en/people/stanko_strmcnik/.


401

proopter: production dynamics analysis and optimization tool · 2017-10-27 · proopter, production...

Documents