Research ArticleDecision Fusion System for Bolted Joint Monitoring

Dong Liang1 and Shen-fang Yuan2

1Department of Aeronautics College of Physics and Electromechanics Xiamen University Xiamen 361005 China2The State Key Lab of Mechanics and Control of Mechanical Structures Nanjing University of Aeronautics and AstronauticsNanjing 210016 China

Correspondence should be addressed to Dong Liang ld19821213126com

Received 3 July 2014 Revised 26 October 2014 Accepted 10 November 2014

Academic Editor Anindya Ghoshal

Copyright copy 2015 D Liang and S-f Yuan This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited

Bolted joint is widely used inmechanical and architectural structures such as machine tools industrial robots transport machinespower plants aviation stiffened plate bridges and steel towers The bolt loosening induced by flight load and environment factorcan cause joint failure leading to a disastrous accident Hence structural health monitoring is critical for the bolted joint detectionIn order to realize a real-time and convenient monitoring and satisfy the requirement of advanced maintenance of the structurethis paper proposes an intelligent bolted joint failuremonitoring approach using a developed decision fusion system integrated withLamb wave propagation based actuator-sensor monitoring method Firstly the basic knowledge of decision fusion and classifierselection techniques is briefly introduced Then a developed decision fusion system is presented Finally three fusion algorithmswhich consist of majority voting Bayesian belief and multiagent method are adopted for comparison in a real-world monitoringexperiment for the large aviation aluminum plate Based on the results shown in the experiment a big potential in real-timeapplication is presented that the method can accurately and rapidly identify the bolt loosening by analyzing the acquired strainsignal using proposed decision fusion system

1 Introduction

Bolted joint is widely used in mechanical and architecturalstructures such asmachine tools industrial robots transportmachines power plants aviation stiffened plate bridgesand steel towers The bolt loosening induced by flight loadand environment factor can cause joint failure leading to adisastrous accident for the aircraft In order to keep up theintegrity and operation safety of these structures detectingbolted joint in real time is an important concern in structuralhealth monitoring

Till now for the bolt loosening detection there are someconventional nondestructive inspection techniques whichuse the ultrasonic waves and electromagnetic resonance [12] However these methods are costly labor intensive andtime consuming to perform for a large structure and canonly be performed when the aircraft is out of service beingintermittent condition monitoring Accordingly structuralhealth monitoring (SHM) has been being recently focusedon by many researchers since the new inspection approach

utilizes advanced sensor and actuator devices being inte-grated in the structural material with aim to achieve a widerange of real-time online monitoring There are a number ofsignificant works in the SHM area concerning bolt looseningmonitoring

The problem of detecting bolt loosening has been studiedby different researchers The principle in these techniquesis to seek out the changes in the dynamic properties asindicators of damage in the structure Pai and Hess studythe loosening of threaded fasteners due to shear loads aswell as the effect of fastener placement on a structure as avariable promoting self-loosening [3 4] Caccese et al exhibitthe promise of the transmittance function for bolt load lossdetection in hybrid compositemetal bolted connections [5]Brown and Adams examine the equilibrium point damageprognosis method across the joint [6] Todd et al assessthe effectiveness of structural frequencies and mode shapesfor bolt loosening monitoring [7] Nichols et al use statespace models to detect joint preload loss in a framestructure

Hindawi Publishing CorporationShock and VibrationVolume 2015 Article ID 592043 11 pageshttpdxdoiorg1011552015592043

2 Shock and Vibration

and utilize data-driven phase space models to assess theconditions of a bolted joint in a composite beam [8 9] Monizet al use a multivariate attractor-based approach to detectthe bolt loosening with FBG sensor [10] Rutherford et alutilize nonlinear feature identifications based on self-sensingimpedance measurements for jointed portal frame structurestructural health assessment [11] Ritdumrongkul and Parkpresent the use of a PZT actuator-sensor and the impedance-based monitoring techniques in conjunction with numericalmodel-basedmethodology in structural healthmonitoring toquantitatively detect damage of bolted joint of two aluminumbeams [12ndash15] Yang et al use the Lamb wave propagation tomonitor the loosening of bolts in a space thermal protectionpanel [16] Okugawa employs the subspace state space identi-fication algorithm (4SID) to identify the natural frequency ofthe smart washer for bolt loosening detection [17] Milaneseet al research modeling and detection of joint looseningusing output-only broadband vibration data [18] Doyle et aluse the acoustoelastic and magnetomechanical impedance todetect bolt loosening in satellite bolted joints [19]

Recently the development of artificial intelligence tech-niques has led to their application in the structure healthmonitoring Somemethods such as artificial neural networksand support vector machines have been employed to esti-mate the structure damage [20ndash23]They are capable of mod-eling extremely complex nonlinear relationships betweenknown structure damage and structure output responseHowever for practical applications a single decision methodcan only acquire a limited recognition capability for specialdata Therefore a decision fusion method is introduced tocombine the advantages of different recognition algorithmsand give more reliable result for the complex task

This paper proposes an intelligent bolted joint failuremonitoring approach using a developed decision fusion sys-tem integrated with Lamb wave propagation-based actuator-sensor monitoring method Firstly the basic knowledge ofdecision fusion and classifier selection techniques is brieflyintroduced Then a developed decision fusion system ispresented Finally three fusion algorithms which consistof majority voting Bayesian belief and multiagent methodare adopted for comparison in a real-world monitoringexperiment for the large aviation aluminum plate Based onthe results shown in the experiment a big potential in real-time application is presented that the method can accuratelyand rapidly identify the bolt loosening by analyzing theacquired strain signal using the proposed decision fusionsystem

The rest of this paper is structured in the followingmanner Section 2 introduces the background knowledgeof decision fusion In Section 3 decision fusion system forbonded joint monitoring is presented Section 4 gives exper-imental results and discussion for large aviation aluminumplate structure Finally Section 5 concludes the paper

2 Decision Fusion Method

This section covers a brief introduction of decision fusionmethod which consists of the classifier selection based onentropy and multiagent fusion algorithm

21 Classifier Selection Based on Entropy Studies [24] haveshown that classifer selection can affect the performance ofthe subsequent multiclassifier fusion A proper combinationof classifiers can generate the best recognization performanceand reduce the calculated time However there still exit manyproblems to select the proper classifier team from a largepool of different classifiers Accordingly classifier selectiontechnology has been being recently focused on by manyresearchers The optimal combinations of classifiers shouldhave good individual performances and sufficient level ofdiversity The diversity quality of classifier selection reliesmostly on the goodness of the selection criterion whichincludes correlation coefficent procuct-moment correlationmeasure 119876 statistics and entropy measure [25] In thesemethods entropy-based diversity measure is a new andeffective method for classifier selection based on [26]

ED = minus11987311

119873log2(11987311119873)

minus11987300

119873log2(11987300119873)

minus11987310

119873log2(11987310119873)

minus11987301

119873log2(11987301119873)

(1)

where11987311 means the number of samples which are classifiedcorrectly by two classifiers 119890

1and 119890211987300means those samples

which are misclassified by the two classifiers 11987310 denotesthose samples which are classified correctly by classifier 119890

1

and misclassified by classifier 1198902 11987301 denotes the number of

samples which are misclassified by classifier 1198901and classified

correctly by classifier 1198902 and 119873 is the total number of

experiment samplesGenerally the diversity of classifiers can give more effec-

tive information so smaller correlation degree among theclassifiers can lead to better fusion performance According tothe diversitymeasurement principle it is neccesacy to select ateam of classifiers and the flowchart of classifier selection canbe shown in Procedure 1

22 Multiagent Decision Fusion Generally the classifiersrsquooutput information can be divided into three levels [27]

(1) the abstract level a classifier 119890 only outputs a singleclass label for an input 119909

(2) the rank level a classifier 119890 ranks all classesrsquo labels ina queue with the one at the top being the first choice

(3) the measurement level a classifier 119890 evaluates thedegree that 119909 has for each class using a measurementvalue

Among the levels mentioned above from the abstractlevel to the measurement level the amount of informationof the classifiersrsquo output increases in sequence Accordinglythe classification algorithms of themeasurement informationcan produce the best results However the classifiers that cansupply the abstract information are more available in the realapplication

According to the three levels in the classifiersrsquo outputinformation decision fusion methods can be divided into

Shock and Vibration 3

Define119864 = 119890

1 1198902 119890

119870 is the set of the classifier needed to be selected

1198641015840= 1198901015840

1 1198901015840

2 119890

1015840

119870 is the set of the classifier selected

119894 119895 = 1 2 119870 to index the classifier in set 119864119890119895is the 119895th classifier in set 119864

1198901015840

119895is the 119895th classifier in set 1198641015840

120572(119890119894) is the accuracy rate of the 119894th classifier in set 119864 which is the ratio of number of samples classified correctly to

the total samplesED(119890 119890

119894) is the Entropy-base diversity measure between classifier 119890 and 119890

119894

BeginStep 1 Select the initial evaluation criterion such as 120572(119890

119894)

Step 2119890 larr max

119890119894isin119864

120572(119890119894) 119864 larr 119864 minus 119890 119895 larr 1 119890

1015840

119895larr 119890

Step 3119890 larr max

119890119894isin119864

ED(119890 119890119894) 119864 larr 119864 minus 119890 119895 larr 119895 + 1 119890

1015840

119895larr 119890

Note when a similar low correlation degree appears for more than one classifier the classifier that has the highestaccuracy rate is chosen

Step 4 If 119864 = Φ then go to Step 5 otherwise go to Step 3 endStep 5 Find the optimal classifier sequence 1198641015840 = 119890

1015840

1 1198901015840

2 119890

1015840

119870

Procedure 1 Proposed procedure for classifier selection

three types Multiple classifiersrsquo fusion integrates differentdecisions from multiple classifiers to boost the accuracy ofrecognitionThe decision fusionmethods of the used abstractinformation are widely adopted which include majorityvoting [28] Bayesian belief [27] and multiagent method[29 30]

In the section the multiagent fusion algorithm is intro-duced in detail In recent years multiagent system (MAS)of artificial intelligence (AI) has been a natural model fordeveloping a large-scale complex distributed system whichis loosely coupled and heterogeneous [31] In this way acomplex system is decomposed into some small autonomoussystems which can interact and cooperate with each otherThey can finish the complex mission via communication andnegotiation

In themultiagent fusionmethod each classifier is deemedas a single agent The confusion matrix of the classifierdenotes the recognition ability of the agent For a test sampleBayesian belief decision can be given by each classifier agentA two-order correlation degree for information exchangebetween any two classifiers is introduced to dynamicallymodify each agentrsquos belief decision Once there are no moredifferent decisions for these agents a final combinationdecision ismadeHence Bayesian beliefmethod andmajorityvoting are integrated creatively in the method It considers abehaviour of population decision The flowchart of multia-gent method is shown in Figure 1

Firstly a sample setU consists ofU1U2 andU

3U1is the

training set of each classifier for obtaining the parameter ofthe classifierU

2is the test set of each classifier and is also the

training set of the fusionmethod for acquiring the parameterof the fusion method U

3is the test set of the fusion method

Confusion matrix N(119896) is firstly created on the basisof Bayesian belief method N(119896) is regarded as the prior

Confusion matrix

Majority voting

Belief matrix

matrix

Initial vote rate Vote

Create new vote rate

Create label decision vector

Normalization

Modify belief matrix

Input sample x

No

Label i

Yes

rate gt threshold

Codecision

Figure 1 Flowchart of multiagent decision fusion algorithm

knowledge of each classifier agent It can be calculated easilyfor test samples ofU

1based on the trained classifier agent for

U2Secondly a five-dimensional codecision matrix D =

[1198891198951 119895211989411989611198962

]119872times119872times119872times119870times119870

is required as the training parameterIt stands for decision correlation between any two classifieragents and its element is calculated by

1198891198951 11989521198941198961 1198962

= 119875 (119909 isin 119894 | 1198901198961(119909) = 119895

1 1198901198962(119909) = 119895

2)

=|A|

radic|B| sdot radic|C|

(2)

4 Shock and Vibration

7

4

8

2

9

6

3PZT sensor 1

5

(a) Single-actuator multisensor

7

4

8

2

9

6

31

5

(b) Cycle-actuator multisensor

Figure 2 The active monitoring method for bolt loosening

where 119894 is the expected class of input sample 119909 1198951and 1198952are

respectively the decisions of classifiers 1198901198961and 1198901198962 where 119896

1=

1198962 set ABC to be defined as

A = 119909 | 119909 isin 119894 1198901198961(119909) = 119895

1 1198901198962(119909) = 119895

2 forall119909 isin U

2

B = 119909 | 119909 isin 119894 1198901198961(119909) = 119895

1 forall119909 isin U

2

C = 119909 | 119909 isin 119894 1198901198962(119909) = 119895

2 forall119909 isin U

2

(3)

The element 1198891198951 11989521198941198961 1198962

in the matrix shows the probabil-ity of the sample119909 of the class 119894 assigned as 119895

1class by classifier

1198901198961and classified as 119895

2by classifier 119890

1198962 | sdot | denotes the cardinal

number of setsAfter obtaining the confusion matrix and codecision

matrix the initial belief matrix B(119909) for input sample 119909 canbe calculated B(119909) is regarded as the initial belief probabilityof each classifier agent for test samples ofU

3 Each row in the

belief matrix is corresponding to each classifier agentrsquos beliefprobability of different column classes for the input sample119909 If the class of the maximum probability in the 119896th rowis regarded as the 119896th classifier agentrsquos decision a decisionlabel vector can be directly obtained from the belief matrixAccording to the majority voting strategy the initial vote rateof each class can be calculated for input 119909

Next if the initial maximum vote rate is less than anaccordance threshold there are more different decisions forthe classifier agents Then the agents can interact with eachother andmodify the original belief degrees themselves usingthe codecision matrix The repeated modification scheme isrepresented as

119887119896119894(119909) = 119887

119896119894(119909) + (

1

119870)

119870

sum

119896119899=1119896119899 =119896

119889119895119895119899 119894119896119896119899

sdot radic119887119896119894(119909) sdot 119887

119896119899119894(119909)

(4)

where 119887119896119894(119909) is the element of Bayesian belief matrixB(119909) and

represents belief probability of classifier 119896 for the sample 119909

belonging to class 119894119870 is the number of total fusion classifiersand 119889

119895119895119899 119894119896119896119899is the weight of information exchange between

119896th classifier and 119896119899th classifier The correction term of the

right formula means the information summation of classifier119896 interacting with other classifiers for the sample 119909 belongingto class 119894

Whenever the belief matrix is modified a normalizationprocess is required to ensure the row element of new beliefmatrix being the significant probability value On the basis ofthe new belief matrix a decision vector of the classifier agentsis acquired to generate a new vote rates If the maximumvote rate is still less than the predetermined threshold theclassifier agents have less accordance for the input sampleHence the interaction between the agents will continueand their belief matrix will be modified repeatedly untiltheir decision reaches the accordance criterion Finally themultiagent classifiers use amajority votingmethod to give outthe output of fusion decision

3 Decision Fusion System forBolted Joint Monitoring

The active SHM method is generally adopted to monitorthe joint failure induced by bolt loosening [32] The methodis based on the structural vibration response and uses thepiezoelectric ceramic material (PZT) element as the actu-ator or the sensor Its actuator-sensor scheme includes thesingle-actuator multisensor and cycle-actuator multisensoras shown in Figure 2 [33] In the first scheme the fixed driverPZT element is arranged on the structure to stimulate thesensors surrounding the structure simultaneouslyThe powerof the actuator is finite and accordingly the second schemeis presented for the large structure The PZT element aroundthe boundary acts as actuator in turn Each time the signals oftwo adjacent PZT sensors (left and right or upper and lower)are sampled For instance in Figure 2 when PZT element 1acts as the actuator the signals of PZT elements 2 and 5 asthe sensors are sampled

Shock and Vibration 5

Joint failure

Charge amplifier

Waveform generator

Computer

Data acquisition card

PCI bus

Wiring board

PZT sensor

Figure 3 Sensor layout and joint failure position on the specimen

Output final decision

Classifiers selection

Decision fusion

Classifier 1 Classifier 2

Feature extraction

Feature extraction

Feature extraction

Sensor 1 Sensor 2

Feature combination

Classifier n

Sensor n

middot middot middot

middot middot middot

Figure 4 Framework of the proposed fusion decision system

Generally a sine wave can be excited by the PZT actuatorto the structure at a frequency under which the vibrationresponse of the structure is sensitive to the bolt looseningThe experiment shows that the sensor signal varies before andafter the bolt loosening [32] An active monitoring system forbolt loosening is shown in Figure 3 which includes waveformgenerator charge amplifier and data acquisition card

In this paper a decision fusion system is presentedfor bolted joint monitoring It is based on a self-designedfusion diagnosis toolbox byMATLAB language R2006aThissystem consists of six levels sensor feature extraction featurecombination multiclassifier decision classifier selection anddecision fusion The framework of the proposed system isshown in Figure 4 Firstly the input signal is acquired fromthe sensorwhen the actuator stimulates the structure periodi-cally Secondly the signal feature is extracted and the featuresof different sensors are combined to be a feature vectorThena decision vector is the output of a team of classifiers andthe algorithm of classifier selection is employed to obtainthe optimization classifier combination Finally the decision

fusion method combines the selected classifiersrsquo decisions togive out the final evaluation This paper adopts three fusionalgorithms majority voting Bayesian belief and multiagentmethod to assess decision fusionrsquos performance

31 Experiment Setup In order to verify the effectiveness ofthe presented decision fusion system integrated with Lambwave propagation based actuator-sensormonitoringmethodin this paper the large aviation aluminum plate structureis studied as the experimental object Figure 5 depicts a flatstructure and the sensor distribution diagram The platestructural material is the aviation hard aluminum LY12whose basic dimensions and thickness are 120 cm times 200 cmtimes 025 cm Around the structure there are 64M6-bolts whichare used to fix the plate with bracket and the bolt space is10 cm The structure is divided into eight subregions each ofwhich is 49 cm times 45 cm except its edge The PZT sensors arelaid on the vertices of each subregion

In this study tests are conducted with healthy andunhealthy configuration which includes the full loose state of20 bolts in different locations around the structure Hencetwenty joint failure patterns and one health pattern are con-sidered In the experiment tests are conducted with healthyand damage configuration which includes the completelyloose state of 20 bolts in different locations around thestructure and each time only one bolt is loosening In orderto quantitatively measure the loosening degrees of bolt thetightening condition 119878 is introduced and defined as

119878 =119879119904

1198790

times 100 [] (5)

where 119879119904is axial tension of a tightening condition and 119879

0

is axial tension equivalent to 100 of tightening condition100 tightening condition is defined as the condition ofthat bolted joint being tightened to standard tightening axialtension In our study for uneasy calibration of partiallyloosening bolt and pattern overlapping obvious existencein our large structure since numerous structure joint boltsdistribute densely only the tight 119878 = 100 and completelyloose state 119878 = 0 of bolts are considered

Hence the various cases tested are (i) healthy case thestructure is tested without any bolt loosening from the joint(ii) unhealthy the cases tested are the complete loosening of

6 Shock and Vibration

1 2 3

4 5 6

7 8 9

PZT SA Bolt

10 cm

49 cm

45 cm

120 cm

200 cm

11 cm

10 cm

Subarea 2 Subarea 4

Subarea 6 Subarea 7 Subarea 8

Figure 5 System setup

PZT 3

PZT 13 PZT 1 PZT 7

PZT 15 PZT 9

Bolt 2

Bolt 9

Bolt 22 Bolt 27

Bolt 39

Bolt 49Bolt 54

Bolt 5

Bolt 19

Bolt 33

Bolt 43

Bolt 63middot middot middot

middot middot middot

Figure 6 The loosening bolts monitored location

Actuator

integrated device

Power amplifier

Computer

Sensor

Piezoelectric sensor arrays

Waveform generator

Charge amplifier Data

acquisition

Switch Digital IOMultichannel

Figure 7 The principle structure of the active monitoring system

Bolts 2 5 9 19 22 23 24 25 26 27 33 39 43 49 50 51 5253 54 or 63 as shown in Figure 6

In the experiment twelve PZT sensors around the bound-ary are employed to detect the bolt loosening with thecycle-actuator multisensor method For the measurement

hardware the self-design integrate and program control mul-tichannel piezoelectric scanning system is used in the activemonitoring for bolt loosening [34] As shown in Figure 7the system integrates the waveform generator module dataacquisition module charge amplifier module digital IO

Shock and Vibration 7

0 1 2 3 4 5 6Time (s)

Volta

ge (V

)5

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

times10minus5

(a) The excited sine signal

0 1 2 3 4 5 6

Volta

ge (V

)

Time (s) times10minus5

2

15

1

05

0

minus05

minus1

minus15

minus2

(b) The signal before bolt loosening

0 1 2 3 4 5 6times10

minus5

Volta

ge (V

)

2

15

1

05

0

minus05

minus1

minus15

minus2

Time (s)

(c) The sensor signal after bolt loosening

Figure 8 The sensor signal change before and after Bolt 9 loosening

module multichannel scanning switch board and poweramplifier It can interrogate the large numbers of actuator-sensor channel automatically and efficiently The software isprogrammedwith LabVIEW85 andMATLABR2006a in theindustry control computer

32 Data and Feature The computer controls twelve PZTsensors circularly and periodically to excite and sense thestructure strain signal The excitation signal is the sine wavewith 100KHz Lots of experiments [33] have shown thatthe vibration response of the structure under this excitationfrequency is sensitive to the bolt loosening The number ofsampled data is 6000 and the measured time is 00006 s Thesample frequency is 10MHz Figure 8 gives the signal of PZTsensor 1 as actuator and PZT sensor 4 as sensor before andafter bolt loosening (Bolt 9) There are two reasons for theacquired strain signal changes Firstly bolt loosening couldcause the change of the prestress distribution in the structurewhich makes the structure thickness change Hence Lamb

wave with different modes generated by the PZT actuatorpropagates in the plate structure with a different velocityHence PZT sensor acquires different Lamb wave signalsbefore and after bolt loosening Secondly the bolt is deemedto be a scattering source on the Lamb wave propagation pathbetween the actuator and the sensor When bolt looseningcan partly affect the scattering Lamb wave coupled with boltits previous propagation path changes so the acquired Lambwave signal changes For sine wave excitation signal theexperiment [33] shows that the peak change is obvious beforeand after the bolt loosening So twenty-four acquired signalpeaks of the twelve sensors on the plate border are combinedto be a feature vector For the chosen bolts twenty-onemodesrsquo conditions are measured twenty-five times and tensamples are measured to train parameters of the classifiersand ten ones are used to train the fusion method So finallywe obtain a total of 525 samples which consist of 210samples for training classifiers 210 samples for training fusionalgorithms and the remaining 105 samples for test

8 Shock and Vibration

Table 1 Parameters of individual classifier

Classifier SVM C45 119896-NN IIS LVQ

Parameterssetup

Kernel function119896(x y) = (07xTy + 1)

2

Euclidean distance typepenalty coefficient = 10

Percentage of incorrectlyassigned samplesat a node = 5

119896 = 3 Number ofiterations = 50

Number ofneurons = 50 epochs = 50

33 Classifier Description Six pattern classification methodsare utilized to identify the loosening bolt The utilizedclassifiers are described as follows

(1) Support vector machine (SVM) the method canimplement the good recognition rate derived froma few training samples and it is based on statisticallearning theory [35] Kernel function is a key param-eter for SVM which includes linear polynomialGaussian RBF and sigmoid

(2) C45 the algorithm implements ldquoIf-Thenrdquo rulesderived from the training data set [36] These rulesare used to classify the ldquounseenrdquo data

(3) 119896 nearest neighbor (119896-NN) the classifier is verysimple and effective [37] The 119896 nearest neighbors ofthe unidentified test pattern are searched within ahypersphere of predefined radius in order to deter-mine its true class which is the most class in the 119896

samples If only one nearest neighbor is detected 119896-NN is the minimum-distance classification

(4) Improved iterative scaling (IIS) IIS is one of themajor algorithms for finding the optimal parame-ters for the conditional exponential model [38] Itsunderlying idea is that by approximating the log-likelihood function of the conditional exponentialmodel as some kind of ldquosimplerdquo auxiliary functionit is able to decouple the correlation between theparameters and search for the maximum point alongmany directions simultaneously By carrying out thisprocedure iteratively the approximated optimal pointfound over the ldquosimplifiedrdquo function is guaranteedto converge to the true optimal point due to theconvexity of the objective function

(5) Gaussian mixture model (GMM) the classifier isbased on Gaussian component functions [39] Thelinear combination of Gaussian functions is capableof representing a large class of the sample distribu-tion In principle it is a compromise between theperformance and the complexity Gaussian mixturehas remarkable capability to model the irregular data

(6) Learning vector quantization (LVQ) it is a neuralnetwork classifier proposed by Villmann et al [40] Itcombines the simplicity of competitive learning withthe accuracy of supervision It is a simple and intuitiveprototype-based clustering algorithm

4 Results and Discussion

This section describes the result of an experiment of thebolted joint monitoring using the proposed decision fusion

Table 2 Classification results

Classifier SVM C45 119896-NN IIS GMM LVQAccuracy 08952 05385 08571 01904 07524 03077

Table 3 Result of optimal sequence of classifiers fused

Number ofclassifiersselected

Serial number of classifiersEntropy-based

diversitymeasure

1 1 mdash2 1 4 11273 1 4 6 09784 1 4 6 2 08785 1 4 6 2 3 08266 1 4 6 2 3 5 0596

system Then comparison and discussion are given for eachpart of the presented system

41 Individual Classification Next six classifiers are utilizedto classify the calculated features of the bolt loosening Therelevant parameters setup for these classifiers can be found inTable 1 Table 2 gives the test accuracy of the six classifiers Inthe experiment the classification accuracy is evaluated usinga ratio of the number of the samples classified correctly tothe total sample It can be seen that the best classificationaccuracy is 08952 and 08571 of SVM and 119896-NN agents Asfar as performance of the six classifiers is concerned SVMand 119896-NN produce superior results followed by GMM agentLVQ and IIS are not suitable in this work for joint failure Itimplies that the two types of classifiers do not fit for the scatterof training samples But in practice it is almost impossiblethat all the predetermined classifiers will achieve the bestperformance at the same time Otherwise the fusion of puregood or bad classifiers groupmay not necessarily improve theaccuracy [24]Therefore LVQand IIS are still reserved for thefusion method

42 Selection of Classifiers Based on the individual clas-sification decisions acquired in the first step the entropy-based diversity measure method introduced in Section 21 isused for sequence selection of six classifiers The optimizedselected results for different numbers of classifiers and theentropy-based diversity degrees are shown in Table 3

To evaluate the effect of classifier selection Bayesianfusion method with classifier selection is compared withthe one without classifier selection as shown in Figure 9

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

2 Shock and Vibration

and utilize data-driven phase space models to assess theconditions of a bolted joint in a composite beam [8 9] Monizet al use a multivariate attractor-based approach to detectthe bolt loosening with FBG sensor [10] Rutherford et alutilize nonlinear feature identifications based on self-sensingimpedance measurements for jointed portal frame structurestructural health assessment [11] Ritdumrongkul and Parkpresent the use of a PZT actuator-sensor and the impedance-based monitoring techniques in conjunction with numericalmodel-basedmethodology in structural healthmonitoring toquantitatively detect damage of bolted joint of two aluminumbeams [12ndash15] Yang et al use the Lamb wave propagation tomonitor the loosening of bolts in a space thermal protectionpanel [16] Okugawa employs the subspace state space identi-fication algorithm (4SID) to identify the natural frequency ofthe smart washer for bolt loosening detection [17] Milaneseet al research modeling and detection of joint looseningusing output-only broadband vibration data [18] Doyle et aluse the acoustoelastic and magnetomechanical impedance todetect bolt loosening in satellite bolted joints [19]

Recently the development of artificial intelligence tech-niques has led to their application in the structure healthmonitoring Somemethods such as artificial neural networksand support vector machines have been employed to esti-mate the structure damage [20ndash23]They are capable of mod-eling extremely complex nonlinear relationships betweenknown structure damage and structure output responseHowever for practical applications a single decision methodcan only acquire a limited recognition capability for specialdata Therefore a decision fusion method is introduced tocombine the advantages of different recognition algorithmsand give more reliable result for the complex task

This paper proposes an intelligent bolted joint failuremonitoring approach using a developed decision fusion sys-tem integrated with Lamb wave propagation-based actuator-sensor monitoring method Firstly the basic knowledge ofdecision fusion and classifier selection techniques is brieflyintroduced Then a developed decision fusion system ispresented Finally three fusion algorithms which consistof majority voting Bayesian belief and multiagent methodare adopted for comparison in a real-world monitoringexperiment for the large aviation aluminum plate Based onthe results shown in the experiment a big potential in real-time application is presented that the method can accuratelyand rapidly identify the bolt loosening by analyzing theacquired strain signal using the proposed decision fusionsystem

The rest of this paper is structured in the followingmanner Section 2 introduces the background knowledgeof decision fusion In Section 3 decision fusion system forbonded joint monitoring is presented Section 4 gives exper-imental results and discussion for large aviation aluminumplate structure Finally Section 5 concludes the paper

2 Decision Fusion Method

This section covers a brief introduction of decision fusionmethod which consists of the classifier selection based onentropy and multiagent fusion algorithm

21 Classifier Selection Based on Entropy Studies [24] haveshown that classifer selection can affect the performance ofthe subsequent multiclassifier fusion A proper combinationof classifiers can generate the best recognization performanceand reduce the calculated time However there still exit manyproblems to select the proper classifier team from a largepool of different classifiers Accordingly classifier selectiontechnology has been being recently focused on by manyresearchers The optimal combinations of classifiers shouldhave good individual performances and sufficient level ofdiversity The diversity quality of classifier selection reliesmostly on the goodness of the selection criterion whichincludes correlation coefficent procuct-moment correlationmeasure 119876 statistics and entropy measure [25] In thesemethods entropy-based diversity measure is a new andeffective method for classifier selection based on [26]

ED = minus11987311

119873log2(11987311119873)

minus11987300

119873log2(11987300119873)

minus11987310

119873log2(11987310119873)

minus11987301

119873log2(11987301119873)

(1)

where11987311 means the number of samples which are classifiedcorrectly by two classifiers 119890

1and 119890211987300means those samples

which are misclassified by the two classifiers 11987310 denotesthose samples which are classified correctly by classifier 119890

1

and misclassified by classifier 1198902 11987301 denotes the number of

samples which are misclassified by classifier 1198901and classified

correctly by classifier 1198902 and 119873 is the total number of

experiment samplesGenerally the diversity of classifiers can give more effec-

tive information so smaller correlation degree among theclassifiers can lead to better fusion performance According tothe diversitymeasurement principle it is neccesacy to select ateam of classifiers and the flowchart of classifier selection canbe shown in Procedure 1

22 Multiagent Decision Fusion Generally the classifiersrsquooutput information can be divided into three levels [27]

(1) the abstract level a classifier 119890 only outputs a singleclass label for an input 119909

(2) the rank level a classifier 119890 ranks all classesrsquo labels ina queue with the one at the top being the first choice

(3) the measurement level a classifier 119890 evaluates thedegree that 119909 has for each class using a measurementvalue

Among the levels mentioned above from the abstractlevel to the measurement level the amount of informationof the classifiersrsquo output increases in sequence Accordinglythe classification algorithms of themeasurement informationcan produce the best results However the classifiers that cansupply the abstract information are more available in the realapplication

According to the three levels in the classifiersrsquo outputinformation decision fusion methods can be divided into

Shock and Vibration 3

Define119864 = 119890

1 1198902 119890

119870 is the set of the classifier needed to be selected

1198641015840= 1198901015840

1 1198901015840

2 119890

1015840

119870 is the set of the classifier selected

119894 119895 = 1 2 119870 to index the classifier in set 119864119890119895is the 119895th classifier in set 119864

1198901015840

119895is the 119895th classifier in set 1198641015840

120572(119890119894) is the accuracy rate of the 119894th classifier in set 119864 which is the ratio of number of samples classified correctly to

the total samplesED(119890 119890

119894) is the Entropy-base diversity measure between classifier 119890 and 119890

119894

BeginStep 1 Select the initial evaluation criterion such as 120572(119890

119894)

Step 2119890 larr max

119890119894isin119864

120572(119890119894) 119864 larr 119864 minus 119890 119895 larr 1 119890

1015840

119895larr 119890

Step 3119890 larr max

119890119894isin119864

ED(119890 119890119894) 119864 larr 119864 minus 119890 119895 larr 119895 + 1 119890

1015840

119895larr 119890

Note when a similar low correlation degree appears for more than one classifier the classifier that has the highestaccuracy rate is chosen

Step 4 If 119864 = Φ then go to Step 5 otherwise go to Step 3 endStep 5 Find the optimal classifier sequence 1198641015840 = 119890

1015840

1 1198901015840

2 119890

1015840

119870

Procedure 1 Proposed procedure for classifier selection

three types Multiple classifiersrsquo fusion integrates differentdecisions from multiple classifiers to boost the accuracy ofrecognitionThe decision fusionmethods of the used abstractinformation are widely adopted which include majorityvoting [28] Bayesian belief [27] and multiagent method[29 30]

In the section the multiagent fusion algorithm is intro-duced in detail In recent years multiagent system (MAS)of artificial intelligence (AI) has been a natural model fordeveloping a large-scale complex distributed system whichis loosely coupled and heterogeneous [31] In this way acomplex system is decomposed into some small autonomoussystems which can interact and cooperate with each otherThey can finish the complex mission via communication andnegotiation

In themultiagent fusionmethod each classifier is deemedas a single agent The confusion matrix of the classifierdenotes the recognition ability of the agent For a test sampleBayesian belief decision can be given by each classifier agentA two-order correlation degree for information exchangebetween any two classifiers is introduced to dynamicallymodify each agentrsquos belief decision Once there are no moredifferent decisions for these agents a final combinationdecision ismadeHence Bayesian beliefmethod andmajorityvoting are integrated creatively in the method It considers abehaviour of population decision The flowchart of multia-gent method is shown in Figure 1

Firstly a sample setU consists ofU1U2 andU

3U1is the

training set of each classifier for obtaining the parameter ofthe classifierU

2is the test set of each classifier and is also the

training set of the fusionmethod for acquiring the parameterof the fusion method U

3is the test set of the fusion method

Confusion matrix N(119896) is firstly created on the basisof Bayesian belief method N(119896) is regarded as the prior

Confusion matrix

Majority voting

Belief matrix

matrix

Initial vote rate Vote

Create new vote rate

Create label decision vector

Normalization

Modify belief matrix

Input sample x

No

Label i

Yes

rate gt threshold

Codecision

Figure 1 Flowchart of multiagent decision fusion algorithm

knowledge of each classifier agent It can be calculated easilyfor test samples ofU

1based on the trained classifier agent for

U2Secondly a five-dimensional codecision matrix D =

[1198891198951 119895211989411989611198962

]119872times119872times119872times119870times119870

is required as the training parameterIt stands for decision correlation between any two classifieragents and its element is calculated by

1198891198951 11989521198941198961 1198962

= 119875 (119909 isin 119894 | 1198901198961(119909) = 119895

1 1198901198962(119909) = 119895

2)

=|A|

radic|B| sdot radic|C|

(2)

4 Shock and Vibration

7

4

8

2

9

6

3PZT sensor 1

5

(a) Single-actuator multisensor

7

4

8

2

9

6

31

5

(b) Cycle-actuator multisensor

Figure 2 The active monitoring method for bolt loosening

where 119894 is the expected class of input sample 119909 1198951and 1198952are

respectively the decisions of classifiers 1198901198961and 1198901198962 where 119896

1=

1198962 set ABC to be defined as

A = 119909 | 119909 isin 119894 1198901198961(119909) = 119895

1 1198901198962(119909) = 119895

2 forall119909 isin U

2

B = 119909 | 119909 isin 119894 1198901198961(119909) = 119895

1 forall119909 isin U

2

C = 119909 | 119909 isin 119894 1198901198962(119909) = 119895

2 forall119909 isin U

2

(3)

The element 1198891198951 11989521198941198961 1198962

in the matrix shows the probabil-ity of the sample119909 of the class 119894 assigned as 119895

1class by classifier

1198901198961and classified as 119895

2by classifier 119890

1198962 | sdot | denotes the cardinal

number of setsAfter obtaining the confusion matrix and codecision

matrix the initial belief matrix B(119909) for input sample 119909 canbe calculated B(119909) is regarded as the initial belief probabilityof each classifier agent for test samples ofU

3 Each row in the

belief matrix is corresponding to each classifier agentrsquos beliefprobability of different column classes for the input sample119909 If the class of the maximum probability in the 119896th rowis regarded as the 119896th classifier agentrsquos decision a decisionlabel vector can be directly obtained from the belief matrixAccording to the majority voting strategy the initial vote rateof each class can be calculated for input 119909

Next if the initial maximum vote rate is less than anaccordance threshold there are more different decisions forthe classifier agents Then the agents can interact with eachother andmodify the original belief degrees themselves usingthe codecision matrix The repeated modification scheme isrepresented as

119887119896119894(119909) = 119887

119896119894(119909) + (

1

119870)

119870

sum

119896119899=1119896119899 =119896

119889119895119895119899 119894119896119896119899

sdot radic119887119896119894(119909) sdot 119887

119896119899119894(119909)

(4)

where 119887119896119894(119909) is the element of Bayesian belief matrixB(119909) and

represents belief probability of classifier 119896 for the sample 119909

belonging to class 119894119870 is the number of total fusion classifiersand 119889

119895119895119899 119894119896119896119899is the weight of information exchange between

119896th classifier and 119896119899th classifier The correction term of the

right formula means the information summation of classifier119896 interacting with other classifiers for the sample 119909 belongingto class 119894

Whenever the belief matrix is modified a normalizationprocess is required to ensure the row element of new beliefmatrix being the significant probability value On the basis ofthe new belief matrix a decision vector of the classifier agentsis acquired to generate a new vote rates If the maximumvote rate is still less than the predetermined threshold theclassifier agents have less accordance for the input sampleHence the interaction between the agents will continueand their belief matrix will be modified repeatedly untiltheir decision reaches the accordance criterion Finally themultiagent classifiers use amajority votingmethod to give outthe output of fusion decision

3 Decision Fusion System forBolted Joint Monitoring

The active SHM method is generally adopted to monitorthe joint failure induced by bolt loosening [32] The methodis based on the structural vibration response and uses thepiezoelectric ceramic material (PZT) element as the actu-ator or the sensor Its actuator-sensor scheme includes thesingle-actuator multisensor and cycle-actuator multisensoras shown in Figure 2 [33] In the first scheme the fixed driverPZT element is arranged on the structure to stimulate thesensors surrounding the structure simultaneouslyThe powerof the actuator is finite and accordingly the second schemeis presented for the large structure The PZT element aroundthe boundary acts as actuator in turn Each time the signals oftwo adjacent PZT sensors (left and right or upper and lower)are sampled For instance in Figure 2 when PZT element 1acts as the actuator the signals of PZT elements 2 and 5 asthe sensors are sampled

Shock and Vibration 5

Joint failure

Charge amplifier

Waveform generator

Computer

Data acquisition card

PCI bus

Wiring board

PZT sensor

Figure 3 Sensor layout and joint failure position on the specimen

Output final decision

Classifiers selection

Decision fusion

Classifier 1 Classifier 2

Feature extraction

Feature extraction

Feature extraction

Sensor 1 Sensor 2

Feature combination

Classifier n

Sensor n

middot middot middot

middot middot middot

Figure 4 Framework of the proposed fusion decision system

Generally a sine wave can be excited by the PZT actuatorto the structure at a frequency under which the vibrationresponse of the structure is sensitive to the bolt looseningThe experiment shows that the sensor signal varies before andafter the bolt loosening [32] An active monitoring system forbolt loosening is shown in Figure 3 which includes waveformgenerator charge amplifier and data acquisition card

In this paper a decision fusion system is presentedfor bolted joint monitoring It is based on a self-designedfusion diagnosis toolbox byMATLAB language R2006aThissystem consists of six levels sensor feature extraction featurecombination multiclassifier decision classifier selection anddecision fusion The framework of the proposed system isshown in Figure 4 Firstly the input signal is acquired fromthe sensorwhen the actuator stimulates the structure periodi-cally Secondly the signal feature is extracted and the featuresof different sensors are combined to be a feature vectorThena decision vector is the output of a team of classifiers andthe algorithm of classifier selection is employed to obtainthe optimization classifier combination Finally the decision

fusion method combines the selected classifiersrsquo decisions togive out the final evaluation This paper adopts three fusionalgorithms majority voting Bayesian belief and multiagentmethod to assess decision fusionrsquos performance

31 Experiment Setup In order to verify the effectiveness ofthe presented decision fusion system integrated with Lambwave propagation based actuator-sensormonitoringmethodin this paper the large aviation aluminum plate structureis studied as the experimental object Figure 5 depicts a flatstructure and the sensor distribution diagram The platestructural material is the aviation hard aluminum LY12whose basic dimensions and thickness are 120 cm times 200 cmtimes 025 cm Around the structure there are 64M6-bolts whichare used to fix the plate with bracket and the bolt space is10 cm The structure is divided into eight subregions each ofwhich is 49 cm times 45 cm except its edge The PZT sensors arelaid on the vertices of each subregion

In this study tests are conducted with healthy andunhealthy configuration which includes the full loose state of20 bolts in different locations around the structure Hencetwenty joint failure patterns and one health pattern are con-sidered In the experiment tests are conducted with healthyand damage configuration which includes the completelyloose state of 20 bolts in different locations around thestructure and each time only one bolt is loosening In orderto quantitatively measure the loosening degrees of bolt thetightening condition 119878 is introduced and defined as

119878 =119879119904

1198790

times 100 [] (5)

where 119879119904is axial tension of a tightening condition and 119879

0

is axial tension equivalent to 100 of tightening condition100 tightening condition is defined as the condition ofthat bolted joint being tightened to standard tightening axialtension In our study for uneasy calibration of partiallyloosening bolt and pattern overlapping obvious existencein our large structure since numerous structure joint boltsdistribute densely only the tight 119878 = 100 and completelyloose state 119878 = 0 of bolts are considered

Hence the various cases tested are (i) healthy case thestructure is tested without any bolt loosening from the joint(ii) unhealthy the cases tested are the complete loosening of

6 Shock and Vibration

1 2 3

4 5 6

7 8 9

PZT SA Bolt

10 cm

49 cm

45 cm

120 cm

200 cm

11 cm

10 cm

Subarea 2 Subarea 4

Subarea 6 Subarea 7 Subarea 8

Figure 5 System setup

PZT 3

PZT 13 PZT 1 PZT 7

PZT 15 PZT 9

Bolt 2

Bolt 9

Bolt 22 Bolt 27

Bolt 39

Bolt 49Bolt 54

Bolt 5

Bolt 19

Bolt 33

Bolt 43

Bolt 63middot middot middot

middot middot middot

Figure 6 The loosening bolts monitored location

Actuator

integrated device

Power amplifier

Computer

Sensor

Piezoelectric sensor arrays

Waveform generator

Charge amplifier Data

acquisition

Switch Digital IOMultichannel

Figure 7 The principle structure of the active monitoring system

Bolts 2 5 9 19 22 23 24 25 26 27 33 39 43 49 50 51 5253 54 or 63 as shown in Figure 6

In the experiment twelve PZT sensors around the bound-ary are employed to detect the bolt loosening with thecycle-actuator multisensor method For the measurement

hardware the self-design integrate and program control mul-tichannel piezoelectric scanning system is used in the activemonitoring for bolt loosening [34] As shown in Figure 7the system integrates the waveform generator module dataacquisition module charge amplifier module digital IO

Shock and Vibration 7

0 1 2 3 4 5 6Time (s)

Volta

ge (V

)5

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

times10minus5

(a) The excited sine signal

0 1 2 3 4 5 6

Volta

ge (V

)

Time (s) times10minus5

2

15

1

05

0

minus05

minus1

minus15

minus2

(b) The signal before bolt loosening

0 1 2 3 4 5 6times10

minus5

Volta

ge (V

)

2

15

1

05

0

minus05

minus1

minus15

minus2

Time (s)

(c) The sensor signal after bolt loosening

Figure 8 The sensor signal change before and after Bolt 9 loosening

module multichannel scanning switch board and poweramplifier It can interrogate the large numbers of actuator-sensor channel automatically and efficiently The software isprogrammedwith LabVIEW85 andMATLABR2006a in theindustry control computer

32 Data and Feature The computer controls twelve PZTsensors circularly and periodically to excite and sense thestructure strain signal The excitation signal is the sine wavewith 100KHz Lots of experiments [33] have shown thatthe vibration response of the structure under this excitationfrequency is sensitive to the bolt loosening The number ofsampled data is 6000 and the measured time is 00006 s Thesample frequency is 10MHz Figure 8 gives the signal of PZTsensor 1 as actuator and PZT sensor 4 as sensor before andafter bolt loosening (Bolt 9) There are two reasons for theacquired strain signal changes Firstly bolt loosening couldcause the change of the prestress distribution in the structurewhich makes the structure thickness change Hence Lamb

wave with different modes generated by the PZT actuatorpropagates in the plate structure with a different velocityHence PZT sensor acquires different Lamb wave signalsbefore and after bolt loosening Secondly the bolt is deemedto be a scattering source on the Lamb wave propagation pathbetween the actuator and the sensor When bolt looseningcan partly affect the scattering Lamb wave coupled with boltits previous propagation path changes so the acquired Lambwave signal changes For sine wave excitation signal theexperiment [33] shows that the peak change is obvious beforeand after the bolt loosening So twenty-four acquired signalpeaks of the twelve sensors on the plate border are combinedto be a feature vector For the chosen bolts twenty-onemodesrsquo conditions are measured twenty-five times and tensamples are measured to train parameters of the classifiersand ten ones are used to train the fusion method So finallywe obtain a total of 525 samples which consist of 210samples for training classifiers 210 samples for training fusionalgorithms and the remaining 105 samples for test

8 Shock and Vibration

Table 1 Parameters of individual classifier

Classifier SVM C45 119896-NN IIS LVQ

Parameterssetup

Kernel function119896(x y) = (07xTy + 1)

2

Euclidean distance typepenalty coefficient = 10

Percentage of incorrectlyassigned samplesat a node = 5

119896 = 3 Number ofiterations = 50

Number ofneurons = 50 epochs = 50

33 Classifier Description Six pattern classification methodsare utilized to identify the loosening bolt The utilizedclassifiers are described as follows

(1) Support vector machine (SVM) the method canimplement the good recognition rate derived froma few training samples and it is based on statisticallearning theory [35] Kernel function is a key param-eter for SVM which includes linear polynomialGaussian RBF and sigmoid

(2) C45 the algorithm implements ldquoIf-Thenrdquo rulesderived from the training data set [36] These rulesare used to classify the ldquounseenrdquo data

(3) 119896 nearest neighbor (119896-NN) the classifier is verysimple and effective [37] The 119896 nearest neighbors ofthe unidentified test pattern are searched within ahypersphere of predefined radius in order to deter-mine its true class which is the most class in the 119896

samples If only one nearest neighbor is detected 119896-NN is the minimum-distance classification

(4) Improved iterative scaling (IIS) IIS is one of themajor algorithms for finding the optimal parame-ters for the conditional exponential model [38] Itsunderlying idea is that by approximating the log-likelihood function of the conditional exponentialmodel as some kind of ldquosimplerdquo auxiliary functionit is able to decouple the correlation between theparameters and search for the maximum point alongmany directions simultaneously By carrying out thisprocedure iteratively the approximated optimal pointfound over the ldquosimplifiedrdquo function is guaranteedto converge to the true optimal point due to theconvexity of the objective function

(5) Gaussian mixture model (GMM) the classifier isbased on Gaussian component functions [39] Thelinear combination of Gaussian functions is capableof representing a large class of the sample distribu-tion In principle it is a compromise between theperformance and the complexity Gaussian mixturehas remarkable capability to model the irregular data

(6) Learning vector quantization (LVQ) it is a neuralnetwork classifier proposed by Villmann et al [40] Itcombines the simplicity of competitive learning withthe accuracy of supervision It is a simple and intuitiveprototype-based clustering algorithm

4 Results and Discussion

This section describes the result of an experiment of thebolted joint monitoring using the proposed decision fusion

Table 2 Classification results

Classifier SVM C45 119896-NN IIS GMM LVQAccuracy 08952 05385 08571 01904 07524 03077

Table 3 Result of optimal sequence of classifiers fused

Number ofclassifiersselected

Serial number of classifiersEntropy-based

diversitymeasure

1 1 mdash2 1 4 11273 1 4 6 09784 1 4 6 2 08785 1 4 6 2 3 08266 1 4 6 2 3 5 0596

system Then comparison and discussion are given for eachpart of the presented system

41 Individual Classification Next six classifiers are utilizedto classify the calculated features of the bolt loosening Therelevant parameters setup for these classifiers can be found inTable 1 Table 2 gives the test accuracy of the six classifiers Inthe experiment the classification accuracy is evaluated usinga ratio of the number of the samples classified correctly tothe total sample It can be seen that the best classificationaccuracy is 08952 and 08571 of SVM and 119896-NN agents Asfar as performance of the six classifiers is concerned SVMand 119896-NN produce superior results followed by GMM agentLVQ and IIS are not suitable in this work for joint failure Itimplies that the two types of classifiers do not fit for the scatterof training samples But in practice it is almost impossiblethat all the predetermined classifiers will achieve the bestperformance at the same time Otherwise the fusion of puregood or bad classifiers groupmay not necessarily improve theaccuracy [24]Therefore LVQand IIS are still reserved for thefusion method

42 Selection of Classifiers Based on the individual clas-sification decisions acquired in the first step the entropy-based diversity measure method introduced in Section 21 isused for sequence selection of six classifiers The optimizedselected results for different numbers of classifiers and theentropy-based diversity degrees are shown in Table 3

To evaluate the effect of classifier selection Bayesianfusion method with classifier selection is compared withthe one without classifier selection as shown in Figure 9

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 3

Define119864 = 119890

1 1198902 119890

119870 is the set of the classifier needed to be selected

1198641015840= 1198901015840

1 1198901015840

2 119890

1015840

119870 is the set of the classifier selected

119894 119895 = 1 2 119870 to index the classifier in set 119864119890119895is the 119895th classifier in set 119864

1198901015840

119895is the 119895th classifier in set 1198641015840

120572(119890119894) is the accuracy rate of the 119894th classifier in set 119864 which is the ratio of number of samples classified correctly to

the total samplesED(119890 119890

119894) is the Entropy-base diversity measure between classifier 119890 and 119890

119894

BeginStep 1 Select the initial evaluation criterion such as 120572(119890

119894)

Step 2119890 larr max

119890119894isin119864

120572(119890119894) 119864 larr 119864 minus 119890 119895 larr 1 119890

1015840

119895larr 119890

Step 3119890 larr max

119890119894isin119864

ED(119890 119890119894) 119864 larr 119864 minus 119890 119895 larr 119895 + 1 119890

1015840

119895larr 119890

Note when a similar low correlation degree appears for more than one classifier the classifier that has the highestaccuracy rate is chosen

Step 4 If 119864 = Φ then go to Step 5 otherwise go to Step 3 endStep 5 Find the optimal classifier sequence 1198641015840 = 119890

1015840

1 1198901015840

2 119890

1015840

119870

Procedure 1 Proposed procedure for classifier selection

three types Multiple classifiersrsquo fusion integrates differentdecisions from multiple classifiers to boost the accuracy ofrecognitionThe decision fusionmethods of the used abstractinformation are widely adopted which include majorityvoting [28] Bayesian belief [27] and multiagent method[29 30]

In the section the multiagent fusion algorithm is intro-duced in detail In recent years multiagent system (MAS)of artificial intelligence (AI) has been a natural model fordeveloping a large-scale complex distributed system whichis loosely coupled and heterogeneous [31] In this way acomplex system is decomposed into some small autonomoussystems which can interact and cooperate with each otherThey can finish the complex mission via communication andnegotiation

In themultiagent fusionmethod each classifier is deemedas a single agent The confusion matrix of the classifierdenotes the recognition ability of the agent For a test sampleBayesian belief decision can be given by each classifier agentA two-order correlation degree for information exchangebetween any two classifiers is introduced to dynamicallymodify each agentrsquos belief decision Once there are no moredifferent decisions for these agents a final combinationdecision ismadeHence Bayesian beliefmethod andmajorityvoting are integrated creatively in the method It considers abehaviour of population decision The flowchart of multia-gent method is shown in Figure 1

Firstly a sample setU consists ofU1U2 andU

3U1is the

training set of each classifier for obtaining the parameter ofthe classifierU

2is the test set of each classifier and is also the

training set of the fusionmethod for acquiring the parameterof the fusion method U

3is the test set of the fusion method

Confusion matrix N(119896) is firstly created on the basisof Bayesian belief method N(119896) is regarded as the prior

Confusion matrix

Majority voting

Belief matrix

matrix

Initial vote rate Vote

Create new vote rate

Create label decision vector

Normalization

Modify belief matrix

Input sample x

No

Label i

Yes

rate gt threshold

Codecision

Figure 1 Flowchart of multiagent decision fusion algorithm

knowledge of each classifier agent It can be calculated easilyfor test samples ofU

1based on the trained classifier agent for

U2Secondly a five-dimensional codecision matrix D =

[1198891198951 119895211989411989611198962

]119872times119872times119872times119870times119870

is required as the training parameterIt stands for decision correlation between any two classifieragents and its element is calculated by

1198891198951 11989521198941198961 1198962

= 119875 (119909 isin 119894 | 1198901198961(119909) = 119895

1 1198901198962(119909) = 119895

2)

=|A|

radic|B| sdot radic|C|

(2)

4 Shock and Vibration

7

4

8

2

9

6

3PZT sensor 1

5

(a) Single-actuator multisensor

7

4

8

2

9

6

31

5

(b) Cycle-actuator multisensor

Figure 2 The active monitoring method for bolt loosening

where 119894 is the expected class of input sample 119909 1198951and 1198952are

respectively the decisions of classifiers 1198901198961and 1198901198962 where 119896

1=

1198962 set ABC to be defined as

A = 119909 | 119909 isin 119894 1198901198961(119909) = 119895

1 1198901198962(119909) = 119895

2 forall119909 isin U

2

B = 119909 | 119909 isin 119894 1198901198961(119909) = 119895

1 forall119909 isin U

2

C = 119909 | 119909 isin 119894 1198901198962(119909) = 119895

2 forall119909 isin U

2

(3)

The element 1198891198951 11989521198941198961 1198962

in the matrix shows the probabil-ity of the sample119909 of the class 119894 assigned as 119895

1class by classifier

1198901198961and classified as 119895

2by classifier 119890

1198962 | sdot | denotes the cardinal

number of setsAfter obtaining the confusion matrix and codecision

matrix the initial belief matrix B(119909) for input sample 119909 canbe calculated B(119909) is regarded as the initial belief probabilityof each classifier agent for test samples ofU

3 Each row in the

belief matrix is corresponding to each classifier agentrsquos beliefprobability of different column classes for the input sample119909 If the class of the maximum probability in the 119896th rowis regarded as the 119896th classifier agentrsquos decision a decisionlabel vector can be directly obtained from the belief matrixAccording to the majority voting strategy the initial vote rateof each class can be calculated for input 119909

Next if the initial maximum vote rate is less than anaccordance threshold there are more different decisions forthe classifier agents Then the agents can interact with eachother andmodify the original belief degrees themselves usingthe codecision matrix The repeated modification scheme isrepresented as

119887119896119894(119909) = 119887

119896119894(119909) + (

1

119870)

119870

sum

119896119899=1119896119899 =119896

119889119895119895119899 119894119896119896119899

sdot radic119887119896119894(119909) sdot 119887

119896119899119894(119909)

(4)

where 119887119896119894(119909) is the element of Bayesian belief matrixB(119909) and

represents belief probability of classifier 119896 for the sample 119909

belonging to class 119894119870 is the number of total fusion classifiersand 119889

119895119895119899 119894119896119896119899is the weight of information exchange between

119896th classifier and 119896119899th classifier The correction term of the

right formula means the information summation of classifier119896 interacting with other classifiers for the sample 119909 belongingto class 119894

Whenever the belief matrix is modified a normalizationprocess is required to ensure the row element of new beliefmatrix being the significant probability value On the basis ofthe new belief matrix a decision vector of the classifier agentsis acquired to generate a new vote rates If the maximumvote rate is still less than the predetermined threshold theclassifier agents have less accordance for the input sampleHence the interaction between the agents will continueand their belief matrix will be modified repeatedly untiltheir decision reaches the accordance criterion Finally themultiagent classifiers use amajority votingmethod to give outthe output of fusion decision

3 Decision Fusion System forBolted Joint Monitoring

The active SHM method is generally adopted to monitorthe joint failure induced by bolt loosening [32] The methodis based on the structural vibration response and uses thepiezoelectric ceramic material (PZT) element as the actu-ator or the sensor Its actuator-sensor scheme includes thesingle-actuator multisensor and cycle-actuator multisensoras shown in Figure 2 [33] In the first scheme the fixed driverPZT element is arranged on the structure to stimulate thesensors surrounding the structure simultaneouslyThe powerof the actuator is finite and accordingly the second schemeis presented for the large structure The PZT element aroundthe boundary acts as actuator in turn Each time the signals oftwo adjacent PZT sensors (left and right or upper and lower)are sampled For instance in Figure 2 when PZT element 1acts as the actuator the signals of PZT elements 2 and 5 asthe sensors are sampled

Shock and Vibration 5

Joint failure

Charge amplifier

Waveform generator

Computer

Data acquisition card

PCI bus

Wiring board

PZT sensor

Figure 3 Sensor layout and joint failure position on the specimen

Output final decision

Classifiers selection

Decision fusion

Classifier 1 Classifier 2

Feature extraction

Feature extraction

Feature extraction

Sensor 1 Sensor 2

Feature combination

Classifier n

Sensor n

middot middot middot

middot middot middot

Figure 4 Framework of the proposed fusion decision system

Generally a sine wave can be excited by the PZT actuatorto the structure at a frequency under which the vibrationresponse of the structure is sensitive to the bolt looseningThe experiment shows that the sensor signal varies before andafter the bolt loosening [32] An active monitoring system forbolt loosening is shown in Figure 3 which includes waveformgenerator charge amplifier and data acquisition card

In this paper a decision fusion system is presentedfor bolted joint monitoring It is based on a self-designedfusion diagnosis toolbox byMATLAB language R2006aThissystem consists of six levels sensor feature extraction featurecombination multiclassifier decision classifier selection anddecision fusion The framework of the proposed system isshown in Figure 4 Firstly the input signal is acquired fromthe sensorwhen the actuator stimulates the structure periodi-cally Secondly the signal feature is extracted and the featuresof different sensors are combined to be a feature vectorThena decision vector is the output of a team of classifiers andthe algorithm of classifier selection is employed to obtainthe optimization classifier combination Finally the decision

fusion method combines the selected classifiersrsquo decisions togive out the final evaluation This paper adopts three fusionalgorithms majority voting Bayesian belief and multiagentmethod to assess decision fusionrsquos performance

31 Experiment Setup In order to verify the effectiveness ofthe presented decision fusion system integrated with Lambwave propagation based actuator-sensormonitoringmethodin this paper the large aviation aluminum plate structureis studied as the experimental object Figure 5 depicts a flatstructure and the sensor distribution diagram The platestructural material is the aviation hard aluminum LY12whose basic dimensions and thickness are 120 cm times 200 cmtimes 025 cm Around the structure there are 64M6-bolts whichare used to fix the plate with bracket and the bolt space is10 cm The structure is divided into eight subregions each ofwhich is 49 cm times 45 cm except its edge The PZT sensors arelaid on the vertices of each subregion

In this study tests are conducted with healthy andunhealthy configuration which includes the full loose state of20 bolts in different locations around the structure Hencetwenty joint failure patterns and one health pattern are con-sidered In the experiment tests are conducted with healthyand damage configuration which includes the completelyloose state of 20 bolts in different locations around thestructure and each time only one bolt is loosening In orderto quantitatively measure the loosening degrees of bolt thetightening condition 119878 is introduced and defined as

119878 =119879119904

1198790

times 100 [] (5)

where 119879119904is axial tension of a tightening condition and 119879

0

is axial tension equivalent to 100 of tightening condition100 tightening condition is defined as the condition ofthat bolted joint being tightened to standard tightening axialtension In our study for uneasy calibration of partiallyloosening bolt and pattern overlapping obvious existencein our large structure since numerous structure joint boltsdistribute densely only the tight 119878 = 100 and completelyloose state 119878 = 0 of bolts are considered

Hence the various cases tested are (i) healthy case thestructure is tested without any bolt loosening from the joint(ii) unhealthy the cases tested are the complete loosening of

6 Shock and Vibration

1 2 3

4 5 6

7 8 9

PZT SA Bolt

10 cm

49 cm

45 cm

120 cm

200 cm

11 cm

10 cm

Subarea 2 Subarea 4

Subarea 6 Subarea 7 Subarea 8

Figure 5 System setup

PZT 3

PZT 13 PZT 1 PZT 7

PZT 15 PZT 9

Bolt 2

Bolt 9

Bolt 22 Bolt 27

Bolt 39

Bolt 49Bolt 54

Bolt 5

Bolt 19

Bolt 33

Bolt 43

Bolt 63middot middot middot

middot middot middot

Figure 6 The loosening bolts monitored location

Actuator

integrated device

Power amplifier

Computer

Sensor

Piezoelectric sensor arrays

Waveform generator

Charge amplifier Data

acquisition

Switch Digital IOMultichannel

Figure 7 The principle structure of the active monitoring system

Bolts 2 5 9 19 22 23 24 25 26 27 33 39 43 49 50 51 5253 54 or 63 as shown in Figure 6

In the experiment twelve PZT sensors around the bound-ary are employed to detect the bolt loosening with thecycle-actuator multisensor method For the measurement

hardware the self-design integrate and program control mul-tichannel piezoelectric scanning system is used in the activemonitoring for bolt loosening [34] As shown in Figure 7the system integrates the waveform generator module dataacquisition module charge amplifier module digital IO

Shock and Vibration 7

0 1 2 3 4 5 6Time (s)

Volta

ge (V

)5

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

times10minus5

(a) The excited sine signal

0 1 2 3 4 5 6

Volta

ge (V

)

Time (s) times10minus5

2

15

1

05

0

minus05

minus1

minus15

minus2

(b) The signal before bolt loosening

0 1 2 3 4 5 6times10

minus5

Volta

ge (V

)

2

15

1

05

0

minus05

minus1

minus15

minus2

Time (s)

(c) The sensor signal after bolt loosening

Figure 8 The sensor signal change before and after Bolt 9 loosening

module multichannel scanning switch board and poweramplifier It can interrogate the large numbers of actuator-sensor channel automatically and efficiently The software isprogrammedwith LabVIEW85 andMATLABR2006a in theindustry control computer

32 Data and Feature The computer controls twelve PZTsensors circularly and periodically to excite and sense thestructure strain signal The excitation signal is the sine wavewith 100KHz Lots of experiments [33] have shown thatthe vibration response of the structure under this excitationfrequency is sensitive to the bolt loosening The number ofsampled data is 6000 and the measured time is 00006 s Thesample frequency is 10MHz Figure 8 gives the signal of PZTsensor 1 as actuator and PZT sensor 4 as sensor before andafter bolt loosening (Bolt 9) There are two reasons for theacquired strain signal changes Firstly bolt loosening couldcause the change of the prestress distribution in the structurewhich makes the structure thickness change Hence Lamb

wave with different modes generated by the PZT actuatorpropagates in the plate structure with a different velocityHence PZT sensor acquires different Lamb wave signalsbefore and after bolt loosening Secondly the bolt is deemedto be a scattering source on the Lamb wave propagation pathbetween the actuator and the sensor When bolt looseningcan partly affect the scattering Lamb wave coupled with boltits previous propagation path changes so the acquired Lambwave signal changes For sine wave excitation signal theexperiment [33] shows that the peak change is obvious beforeand after the bolt loosening So twenty-four acquired signalpeaks of the twelve sensors on the plate border are combinedto be a feature vector For the chosen bolts twenty-onemodesrsquo conditions are measured twenty-five times and tensamples are measured to train parameters of the classifiersand ten ones are used to train the fusion method So finallywe obtain a total of 525 samples which consist of 210samples for training classifiers 210 samples for training fusionalgorithms and the remaining 105 samples for test

8 Shock and Vibration

Table 1 Parameters of individual classifier

Classifier SVM C45 119896-NN IIS LVQ

Parameterssetup

Kernel function119896(x y) = (07xTy + 1)

2

Euclidean distance typepenalty coefficient = 10

Percentage of incorrectlyassigned samplesat a node = 5

119896 = 3 Number ofiterations = 50

Number ofneurons = 50 epochs = 50

33 Classifier Description Six pattern classification methodsare utilized to identify the loosening bolt The utilizedclassifiers are described as follows

(1) Support vector machine (SVM) the method canimplement the good recognition rate derived froma few training samples and it is based on statisticallearning theory [35] Kernel function is a key param-eter for SVM which includes linear polynomialGaussian RBF and sigmoid

(2) C45 the algorithm implements ldquoIf-Thenrdquo rulesderived from the training data set [36] These rulesare used to classify the ldquounseenrdquo data

(3) 119896 nearest neighbor (119896-NN) the classifier is verysimple and effective [37] The 119896 nearest neighbors ofthe unidentified test pattern are searched within ahypersphere of predefined radius in order to deter-mine its true class which is the most class in the 119896

samples If only one nearest neighbor is detected 119896-NN is the minimum-distance classification

(4) Improved iterative scaling (IIS) IIS is one of themajor algorithms for finding the optimal parame-ters for the conditional exponential model [38] Itsunderlying idea is that by approximating the log-likelihood function of the conditional exponentialmodel as some kind of ldquosimplerdquo auxiliary functionit is able to decouple the correlation between theparameters and search for the maximum point alongmany directions simultaneously By carrying out thisprocedure iteratively the approximated optimal pointfound over the ldquosimplifiedrdquo function is guaranteedto converge to the true optimal point due to theconvexity of the objective function

(5) Gaussian mixture model (GMM) the classifier isbased on Gaussian component functions [39] Thelinear combination of Gaussian functions is capableof representing a large class of the sample distribu-tion In principle it is a compromise between theperformance and the complexity Gaussian mixturehas remarkable capability to model the irregular data

(6) Learning vector quantization (LVQ) it is a neuralnetwork classifier proposed by Villmann et al [40] Itcombines the simplicity of competitive learning withthe accuracy of supervision It is a simple and intuitiveprototype-based clustering algorithm

4 Results and Discussion

This section describes the result of an experiment of thebolted joint monitoring using the proposed decision fusion

Table 2 Classification results

Classifier SVM C45 119896-NN IIS GMM LVQAccuracy 08952 05385 08571 01904 07524 03077

Table 3 Result of optimal sequence of classifiers fused

Number ofclassifiersselected

Serial number of classifiersEntropy-based

diversitymeasure

1 1 mdash2 1 4 11273 1 4 6 09784 1 4 6 2 08785 1 4 6 2 3 08266 1 4 6 2 3 5 0596

system Then comparison and discussion are given for eachpart of the presented system

41 Individual Classification Next six classifiers are utilizedto classify the calculated features of the bolt loosening Therelevant parameters setup for these classifiers can be found inTable 1 Table 2 gives the test accuracy of the six classifiers Inthe experiment the classification accuracy is evaluated usinga ratio of the number of the samples classified correctly tothe total sample It can be seen that the best classificationaccuracy is 08952 and 08571 of SVM and 119896-NN agents Asfar as performance of the six classifiers is concerned SVMand 119896-NN produce superior results followed by GMM agentLVQ and IIS are not suitable in this work for joint failure Itimplies that the two types of classifiers do not fit for the scatterof training samples But in practice it is almost impossiblethat all the predetermined classifiers will achieve the bestperformance at the same time Otherwise the fusion of puregood or bad classifiers groupmay not necessarily improve theaccuracy [24]Therefore LVQand IIS are still reserved for thefusion method

42 Selection of Classifiers Based on the individual clas-sification decisions acquired in the first step the entropy-based diversity measure method introduced in Section 21 isused for sequence selection of six classifiers The optimizedselected results for different numbers of classifiers and theentropy-based diversity degrees are shown in Table 3

To evaluate the effect of classifier selection Bayesianfusion method with classifier selection is compared withthe one without classifier selection as shown in Figure 9

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

4 Shock and Vibration

7

4

8

2

9

6

3PZT sensor 1

5

(a) Single-actuator multisensor

7

4

8

2

9

6

31

5

(b) Cycle-actuator multisensor

Figure 2 The active monitoring method for bolt loosening

where 119894 is the expected class of input sample 119909 1198951and 1198952are

respectively the decisions of classifiers 1198901198961and 1198901198962 where 119896

1=

1198962 set ABC to be defined as

A = 119909 | 119909 isin 119894 1198901198961(119909) = 119895

1 1198901198962(119909) = 119895

2 forall119909 isin U

2

B = 119909 | 119909 isin 119894 1198901198961(119909) = 119895

1 forall119909 isin U

2

C = 119909 | 119909 isin 119894 1198901198962(119909) = 119895

2 forall119909 isin U

2

(3)

The element 1198891198951 11989521198941198961 1198962

in the matrix shows the probabil-ity of the sample119909 of the class 119894 assigned as 119895

1class by classifier

1198901198961and classified as 119895

2by classifier 119890

1198962 | sdot | denotes the cardinal

number of setsAfter obtaining the confusion matrix and codecision

matrix the initial belief matrix B(119909) for input sample 119909 canbe calculated B(119909) is regarded as the initial belief probabilityof each classifier agent for test samples ofU

3 Each row in the

belief matrix is corresponding to each classifier agentrsquos beliefprobability of different column classes for the input sample119909 If the class of the maximum probability in the 119896th rowis regarded as the 119896th classifier agentrsquos decision a decisionlabel vector can be directly obtained from the belief matrixAccording to the majority voting strategy the initial vote rateof each class can be calculated for input 119909

Next if the initial maximum vote rate is less than anaccordance threshold there are more different decisions forthe classifier agents Then the agents can interact with eachother andmodify the original belief degrees themselves usingthe codecision matrix The repeated modification scheme isrepresented as

119887119896119894(119909) = 119887

119896119894(119909) + (

1

119870)

119870

sum

119896119899=1119896119899 =119896

119889119895119895119899 119894119896119896119899

sdot radic119887119896119894(119909) sdot 119887

119896119899119894(119909)

(4)

where 119887119896119894(119909) is the element of Bayesian belief matrixB(119909) and

represents belief probability of classifier 119896 for the sample 119909

belonging to class 119894119870 is the number of total fusion classifiersand 119889

119895119895119899 119894119896119896119899is the weight of information exchange between

119896th classifier and 119896119899th classifier The correction term of the

right formula means the information summation of classifier119896 interacting with other classifiers for the sample 119909 belongingto class 119894

Whenever the belief matrix is modified a normalizationprocess is required to ensure the row element of new beliefmatrix being the significant probability value On the basis ofthe new belief matrix a decision vector of the classifier agentsis acquired to generate a new vote rates If the maximumvote rate is still less than the predetermined threshold theclassifier agents have less accordance for the input sampleHence the interaction between the agents will continueand their belief matrix will be modified repeatedly untiltheir decision reaches the accordance criterion Finally themultiagent classifiers use amajority votingmethod to give outthe output of fusion decision

3 Decision Fusion System forBolted Joint Monitoring

The active SHM method is generally adopted to monitorthe joint failure induced by bolt loosening [32] The methodis based on the structural vibration response and uses thepiezoelectric ceramic material (PZT) element as the actu-ator or the sensor Its actuator-sensor scheme includes thesingle-actuator multisensor and cycle-actuator multisensoras shown in Figure 2 [33] In the first scheme the fixed driverPZT element is arranged on the structure to stimulate thesensors surrounding the structure simultaneouslyThe powerof the actuator is finite and accordingly the second schemeis presented for the large structure The PZT element aroundthe boundary acts as actuator in turn Each time the signals oftwo adjacent PZT sensors (left and right or upper and lower)are sampled For instance in Figure 2 when PZT element 1acts as the actuator the signals of PZT elements 2 and 5 asthe sensors are sampled

Shock and Vibration 5

Joint failure

Charge amplifier

Waveform generator

Computer

Data acquisition card

PCI bus

Wiring board

PZT sensor

Figure 3 Sensor layout and joint failure position on the specimen

Output final decision

Classifiers selection

Decision fusion

Classifier 1 Classifier 2

Feature extraction

Feature extraction

Feature extraction

Sensor 1 Sensor 2

Feature combination

Classifier n

Sensor n

middot middot middot

middot middot middot

Figure 4 Framework of the proposed fusion decision system

Generally a sine wave can be excited by the PZT actuatorto the structure at a frequency under which the vibrationresponse of the structure is sensitive to the bolt looseningThe experiment shows that the sensor signal varies before andafter the bolt loosening [32] An active monitoring system forbolt loosening is shown in Figure 3 which includes waveformgenerator charge amplifier and data acquisition card

In this paper a decision fusion system is presentedfor bolted joint monitoring It is based on a self-designedfusion diagnosis toolbox byMATLAB language R2006aThissystem consists of six levels sensor feature extraction featurecombination multiclassifier decision classifier selection anddecision fusion The framework of the proposed system isshown in Figure 4 Firstly the input signal is acquired fromthe sensorwhen the actuator stimulates the structure periodi-cally Secondly the signal feature is extracted and the featuresof different sensors are combined to be a feature vectorThena decision vector is the output of a team of classifiers andthe algorithm of classifier selection is employed to obtainthe optimization classifier combination Finally the decision

fusion method combines the selected classifiersrsquo decisions togive out the final evaluation This paper adopts three fusionalgorithms majority voting Bayesian belief and multiagentmethod to assess decision fusionrsquos performance

31 Experiment Setup In order to verify the effectiveness ofthe presented decision fusion system integrated with Lambwave propagation based actuator-sensormonitoringmethodin this paper the large aviation aluminum plate structureis studied as the experimental object Figure 5 depicts a flatstructure and the sensor distribution diagram The platestructural material is the aviation hard aluminum LY12whose basic dimensions and thickness are 120 cm times 200 cmtimes 025 cm Around the structure there are 64M6-bolts whichare used to fix the plate with bracket and the bolt space is10 cm The structure is divided into eight subregions each ofwhich is 49 cm times 45 cm except its edge The PZT sensors arelaid on the vertices of each subregion

In this study tests are conducted with healthy andunhealthy configuration which includes the full loose state of20 bolts in different locations around the structure Hencetwenty joint failure patterns and one health pattern are con-sidered In the experiment tests are conducted with healthyand damage configuration which includes the completelyloose state of 20 bolts in different locations around thestructure and each time only one bolt is loosening In orderto quantitatively measure the loosening degrees of bolt thetightening condition 119878 is introduced and defined as

119878 =119879119904

1198790

times 100 [] (5)

where 119879119904is axial tension of a tightening condition and 119879

0

is axial tension equivalent to 100 of tightening condition100 tightening condition is defined as the condition ofthat bolted joint being tightened to standard tightening axialtension In our study for uneasy calibration of partiallyloosening bolt and pattern overlapping obvious existencein our large structure since numerous structure joint boltsdistribute densely only the tight 119878 = 100 and completelyloose state 119878 = 0 of bolts are considered

Hence the various cases tested are (i) healthy case thestructure is tested without any bolt loosening from the joint(ii) unhealthy the cases tested are the complete loosening of

6 Shock and Vibration

1 2 3

4 5 6

7 8 9

PZT SA Bolt

10 cm

49 cm

45 cm

120 cm

200 cm

11 cm

10 cm

Subarea 2 Subarea 4

Subarea 6 Subarea 7 Subarea 8

Figure 5 System setup

PZT 3

PZT 13 PZT 1 PZT 7

PZT 15 PZT 9

Bolt 2

Bolt 9

Bolt 22 Bolt 27

Bolt 39

Bolt 49Bolt 54

Bolt 5

Bolt 19

Bolt 33

Bolt 43

Bolt 63middot middot middot

middot middot middot

Figure 6 The loosening bolts monitored location

Actuator

integrated device

Power amplifier

Computer

Sensor

Piezoelectric sensor arrays

Waveform generator

Charge amplifier Data

acquisition

Switch Digital IOMultichannel

Figure 7 The principle structure of the active monitoring system

Bolts 2 5 9 19 22 23 24 25 26 27 33 39 43 49 50 51 5253 54 or 63 as shown in Figure 6

In the experiment twelve PZT sensors around the bound-ary are employed to detect the bolt loosening with thecycle-actuator multisensor method For the measurement

hardware the self-design integrate and program control mul-tichannel piezoelectric scanning system is used in the activemonitoring for bolt loosening [34] As shown in Figure 7the system integrates the waveform generator module dataacquisition module charge amplifier module digital IO

Shock and Vibration 7

0 1 2 3 4 5 6Time (s)

Volta

ge (V

)5

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

times10minus5

(a) The excited sine signal

0 1 2 3 4 5 6

Volta

ge (V

)

Time (s) times10minus5

2

15

1

05

0

minus05

minus1

minus15

minus2

(b) The signal before bolt loosening

0 1 2 3 4 5 6times10

minus5

Volta

ge (V

)

2

15

1

05

0

minus05

minus1

minus15

minus2

Time (s)

(c) The sensor signal after bolt loosening

Figure 8 The sensor signal change before and after Bolt 9 loosening

module multichannel scanning switch board and poweramplifier It can interrogate the large numbers of actuator-sensor channel automatically and efficiently The software isprogrammedwith LabVIEW85 andMATLABR2006a in theindustry control computer

32 Data and Feature The computer controls twelve PZTsensors circularly and periodically to excite and sense thestructure strain signal The excitation signal is the sine wavewith 100KHz Lots of experiments [33] have shown thatthe vibration response of the structure under this excitationfrequency is sensitive to the bolt loosening The number ofsampled data is 6000 and the measured time is 00006 s Thesample frequency is 10MHz Figure 8 gives the signal of PZTsensor 1 as actuator and PZT sensor 4 as sensor before andafter bolt loosening (Bolt 9) There are two reasons for theacquired strain signal changes Firstly bolt loosening couldcause the change of the prestress distribution in the structurewhich makes the structure thickness change Hence Lamb

wave with different modes generated by the PZT actuatorpropagates in the plate structure with a different velocityHence PZT sensor acquires different Lamb wave signalsbefore and after bolt loosening Secondly the bolt is deemedto be a scattering source on the Lamb wave propagation pathbetween the actuator and the sensor When bolt looseningcan partly affect the scattering Lamb wave coupled with boltits previous propagation path changes so the acquired Lambwave signal changes For sine wave excitation signal theexperiment [33] shows that the peak change is obvious beforeand after the bolt loosening So twenty-four acquired signalpeaks of the twelve sensors on the plate border are combinedto be a feature vector For the chosen bolts twenty-onemodesrsquo conditions are measured twenty-five times and tensamples are measured to train parameters of the classifiersand ten ones are used to train the fusion method So finallywe obtain a total of 525 samples which consist of 210samples for training classifiers 210 samples for training fusionalgorithms and the remaining 105 samples for test

8 Shock and Vibration

Table 1 Parameters of individual classifier

Classifier SVM C45 119896-NN IIS LVQ

Parameterssetup

Kernel function119896(x y) = (07xTy + 1)

2

Euclidean distance typepenalty coefficient = 10

Percentage of incorrectlyassigned samplesat a node = 5

119896 = 3 Number ofiterations = 50

Number ofneurons = 50 epochs = 50

33 Classifier Description Six pattern classification methodsare utilized to identify the loosening bolt The utilizedclassifiers are described as follows

(1) Support vector machine (SVM) the method canimplement the good recognition rate derived froma few training samples and it is based on statisticallearning theory [35] Kernel function is a key param-eter for SVM which includes linear polynomialGaussian RBF and sigmoid

(2) C45 the algorithm implements ldquoIf-Thenrdquo rulesderived from the training data set [36] These rulesare used to classify the ldquounseenrdquo data

(3) 119896 nearest neighbor (119896-NN) the classifier is verysimple and effective [37] The 119896 nearest neighbors ofthe unidentified test pattern are searched within ahypersphere of predefined radius in order to deter-mine its true class which is the most class in the 119896

samples If only one nearest neighbor is detected 119896-NN is the minimum-distance classification

(4) Improved iterative scaling (IIS) IIS is one of themajor algorithms for finding the optimal parame-ters for the conditional exponential model [38] Itsunderlying idea is that by approximating the log-likelihood function of the conditional exponentialmodel as some kind of ldquosimplerdquo auxiliary functionit is able to decouple the correlation between theparameters and search for the maximum point alongmany directions simultaneously By carrying out thisprocedure iteratively the approximated optimal pointfound over the ldquosimplifiedrdquo function is guaranteedto converge to the true optimal point due to theconvexity of the objective function

(5) Gaussian mixture model (GMM) the classifier isbased on Gaussian component functions [39] Thelinear combination of Gaussian functions is capableof representing a large class of the sample distribu-tion In principle it is a compromise between theperformance and the complexity Gaussian mixturehas remarkable capability to model the irregular data

(6) Learning vector quantization (LVQ) it is a neuralnetwork classifier proposed by Villmann et al [40] Itcombines the simplicity of competitive learning withthe accuracy of supervision It is a simple and intuitiveprototype-based clustering algorithm

4 Results and Discussion

This section describes the result of an experiment of thebolted joint monitoring using the proposed decision fusion

Table 2 Classification results

Classifier SVM C45 119896-NN IIS GMM LVQAccuracy 08952 05385 08571 01904 07524 03077

Table 3 Result of optimal sequence of classifiers fused

Number ofclassifiersselected

Serial number of classifiersEntropy-based

diversitymeasure

1 1 mdash2 1 4 11273 1 4 6 09784 1 4 6 2 08785 1 4 6 2 3 08266 1 4 6 2 3 5 0596

system Then comparison and discussion are given for eachpart of the presented system

41 Individual Classification Next six classifiers are utilizedto classify the calculated features of the bolt loosening Therelevant parameters setup for these classifiers can be found inTable 1 Table 2 gives the test accuracy of the six classifiers Inthe experiment the classification accuracy is evaluated usinga ratio of the number of the samples classified correctly tothe total sample It can be seen that the best classificationaccuracy is 08952 and 08571 of SVM and 119896-NN agents Asfar as performance of the six classifiers is concerned SVMand 119896-NN produce superior results followed by GMM agentLVQ and IIS are not suitable in this work for joint failure Itimplies that the two types of classifiers do not fit for the scatterof training samples But in practice it is almost impossiblethat all the predetermined classifiers will achieve the bestperformance at the same time Otherwise the fusion of puregood or bad classifiers groupmay not necessarily improve theaccuracy [24]Therefore LVQand IIS are still reserved for thefusion method

42 Selection of Classifiers Based on the individual clas-sification decisions acquired in the first step the entropy-based diversity measure method introduced in Section 21 isused for sequence selection of six classifiers The optimizedselected results for different numbers of classifiers and theentropy-based diversity degrees are shown in Table 3

To evaluate the effect of classifier selection Bayesianfusion method with classifier selection is compared withthe one without classifier selection as shown in Figure 9

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 5

Joint failure

Charge amplifier

Waveform generator

Computer

Data acquisition card

PCI bus

Wiring board

PZT sensor

Figure 3 Sensor layout and joint failure position on the specimen

Output final decision

Classifiers selection

Decision fusion

Classifier 1 Classifier 2

Feature extraction

Feature extraction

Feature extraction

Sensor 1 Sensor 2

Feature combination

Classifier n

Sensor n

middot middot middot

middot middot middot

Figure 4 Framework of the proposed fusion decision system

Generally a sine wave can be excited by the PZT actuatorto the structure at a frequency under which the vibrationresponse of the structure is sensitive to the bolt looseningThe experiment shows that the sensor signal varies before andafter the bolt loosening [32] An active monitoring system forbolt loosening is shown in Figure 3 which includes waveformgenerator charge amplifier and data acquisition card

In this paper a decision fusion system is presentedfor bolted joint monitoring It is based on a self-designedfusion diagnosis toolbox byMATLAB language R2006aThissystem consists of six levels sensor feature extraction featurecombination multiclassifier decision classifier selection anddecision fusion The framework of the proposed system isshown in Figure 4 Firstly the input signal is acquired fromthe sensorwhen the actuator stimulates the structure periodi-cally Secondly the signal feature is extracted and the featuresof different sensors are combined to be a feature vectorThena decision vector is the output of a team of classifiers andthe algorithm of classifier selection is employed to obtainthe optimization classifier combination Finally the decision

fusion method combines the selected classifiersrsquo decisions togive out the final evaluation This paper adopts three fusionalgorithms majority voting Bayesian belief and multiagentmethod to assess decision fusionrsquos performance

31 Experiment Setup In order to verify the effectiveness ofthe presented decision fusion system integrated with Lambwave propagation based actuator-sensormonitoringmethodin this paper the large aviation aluminum plate structureis studied as the experimental object Figure 5 depicts a flatstructure and the sensor distribution diagram The platestructural material is the aviation hard aluminum LY12whose basic dimensions and thickness are 120 cm times 200 cmtimes 025 cm Around the structure there are 64M6-bolts whichare used to fix the plate with bracket and the bolt space is10 cm The structure is divided into eight subregions each ofwhich is 49 cm times 45 cm except its edge The PZT sensors arelaid on the vertices of each subregion

In this study tests are conducted with healthy andunhealthy configuration which includes the full loose state of20 bolts in different locations around the structure Hencetwenty joint failure patterns and one health pattern are con-sidered In the experiment tests are conducted with healthyand damage configuration which includes the completelyloose state of 20 bolts in different locations around thestructure and each time only one bolt is loosening In orderto quantitatively measure the loosening degrees of bolt thetightening condition 119878 is introduced and defined as

119878 =119879119904

1198790

times 100 [] (5)

where 119879119904is axial tension of a tightening condition and 119879

0

is axial tension equivalent to 100 of tightening condition100 tightening condition is defined as the condition ofthat bolted joint being tightened to standard tightening axialtension In our study for uneasy calibration of partiallyloosening bolt and pattern overlapping obvious existencein our large structure since numerous structure joint boltsdistribute densely only the tight 119878 = 100 and completelyloose state 119878 = 0 of bolts are considered

Hence the various cases tested are (i) healthy case thestructure is tested without any bolt loosening from the joint(ii) unhealthy the cases tested are the complete loosening of

6 Shock and Vibration

1 2 3

4 5 6

7 8 9

PZT SA Bolt

10 cm

49 cm

45 cm

120 cm

200 cm

11 cm

10 cm

Subarea 2 Subarea 4

Subarea 6 Subarea 7 Subarea 8

Figure 5 System setup

PZT 3

PZT 13 PZT 1 PZT 7

PZT 15 PZT 9

Bolt 2

Bolt 9

Bolt 22 Bolt 27

Bolt 39

Bolt 49Bolt 54

Bolt 5

Bolt 19

Bolt 33

Bolt 43

Bolt 63middot middot middot

middot middot middot

Figure 6 The loosening bolts monitored location

Actuator

integrated device

Power amplifier

Computer

Sensor

Piezoelectric sensor arrays

Waveform generator

Charge amplifier Data

acquisition

Switch Digital IOMultichannel

Figure 7 The principle structure of the active monitoring system

Bolts 2 5 9 19 22 23 24 25 26 27 33 39 43 49 50 51 5253 54 or 63 as shown in Figure 6

In the experiment twelve PZT sensors around the bound-ary are employed to detect the bolt loosening with thecycle-actuator multisensor method For the measurement

hardware the self-design integrate and program control mul-tichannel piezoelectric scanning system is used in the activemonitoring for bolt loosening [34] As shown in Figure 7the system integrates the waveform generator module dataacquisition module charge amplifier module digital IO

Shock and Vibration 7

0 1 2 3 4 5 6Time (s)

Volta

ge (V

)5

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

times10minus5

(a) The excited sine signal

0 1 2 3 4 5 6

Volta

ge (V

)

Time (s) times10minus5

2

15

1

05

0

minus05

minus1

minus15

minus2

(b) The signal before bolt loosening

0 1 2 3 4 5 6times10

minus5

Volta

ge (V

)

2

15

1

05

0

minus05

minus1

minus15

minus2

Time (s)

(c) The sensor signal after bolt loosening

Figure 8 The sensor signal change before and after Bolt 9 loosening

module multichannel scanning switch board and poweramplifier It can interrogate the large numbers of actuator-sensor channel automatically and efficiently The software isprogrammedwith LabVIEW85 andMATLABR2006a in theindustry control computer

32 Data and Feature The computer controls twelve PZTsensors circularly and periodically to excite and sense thestructure strain signal The excitation signal is the sine wavewith 100KHz Lots of experiments [33] have shown thatthe vibration response of the structure under this excitationfrequency is sensitive to the bolt loosening The number ofsampled data is 6000 and the measured time is 00006 s Thesample frequency is 10MHz Figure 8 gives the signal of PZTsensor 1 as actuator and PZT sensor 4 as sensor before andafter bolt loosening (Bolt 9) There are two reasons for theacquired strain signal changes Firstly bolt loosening couldcause the change of the prestress distribution in the structurewhich makes the structure thickness change Hence Lamb

wave with different modes generated by the PZT actuatorpropagates in the plate structure with a different velocityHence PZT sensor acquires different Lamb wave signalsbefore and after bolt loosening Secondly the bolt is deemedto be a scattering source on the Lamb wave propagation pathbetween the actuator and the sensor When bolt looseningcan partly affect the scattering Lamb wave coupled with boltits previous propagation path changes so the acquired Lambwave signal changes For sine wave excitation signal theexperiment [33] shows that the peak change is obvious beforeand after the bolt loosening So twenty-four acquired signalpeaks of the twelve sensors on the plate border are combinedto be a feature vector For the chosen bolts twenty-onemodesrsquo conditions are measured twenty-five times and tensamples are measured to train parameters of the classifiersand ten ones are used to train the fusion method So finallywe obtain a total of 525 samples which consist of 210samples for training classifiers 210 samples for training fusionalgorithms and the remaining 105 samples for test

8 Shock and Vibration

Table 1 Parameters of individual classifier

Classifier SVM C45 119896-NN IIS LVQ

Parameterssetup

Kernel function119896(x y) = (07xTy + 1)

2

Euclidean distance typepenalty coefficient = 10

Percentage of incorrectlyassigned samplesat a node = 5

119896 = 3 Number ofiterations = 50

Number ofneurons = 50 epochs = 50

33 Classifier Description Six pattern classification methodsare utilized to identify the loosening bolt The utilizedclassifiers are described as follows

(1) Support vector machine (SVM) the method canimplement the good recognition rate derived froma few training samples and it is based on statisticallearning theory [35] Kernel function is a key param-eter for SVM which includes linear polynomialGaussian RBF and sigmoid

(2) C45 the algorithm implements ldquoIf-Thenrdquo rulesderived from the training data set [36] These rulesare used to classify the ldquounseenrdquo data

(3) 119896 nearest neighbor (119896-NN) the classifier is verysimple and effective [37] The 119896 nearest neighbors ofthe unidentified test pattern are searched within ahypersphere of predefined radius in order to deter-mine its true class which is the most class in the 119896

samples If only one nearest neighbor is detected 119896-NN is the minimum-distance classification

(4) Improved iterative scaling (IIS) IIS is one of themajor algorithms for finding the optimal parame-ters for the conditional exponential model [38] Itsunderlying idea is that by approximating the log-likelihood function of the conditional exponentialmodel as some kind of ldquosimplerdquo auxiliary functionit is able to decouple the correlation between theparameters and search for the maximum point alongmany directions simultaneously By carrying out thisprocedure iteratively the approximated optimal pointfound over the ldquosimplifiedrdquo function is guaranteedto converge to the true optimal point due to theconvexity of the objective function

(5) Gaussian mixture model (GMM) the classifier isbased on Gaussian component functions [39] Thelinear combination of Gaussian functions is capableof representing a large class of the sample distribu-tion In principle it is a compromise between theperformance and the complexity Gaussian mixturehas remarkable capability to model the irregular data

(6) Learning vector quantization (LVQ) it is a neuralnetwork classifier proposed by Villmann et al [40] Itcombines the simplicity of competitive learning withthe accuracy of supervision It is a simple and intuitiveprototype-based clustering algorithm

4 Results and Discussion

This section describes the result of an experiment of thebolted joint monitoring using the proposed decision fusion

Table 2 Classification results

Classifier SVM C45 119896-NN IIS GMM LVQAccuracy 08952 05385 08571 01904 07524 03077

Table 3 Result of optimal sequence of classifiers fused

Number ofclassifiersselected

Serial number of classifiersEntropy-based

diversitymeasure

1 1 mdash2 1 4 11273 1 4 6 09784 1 4 6 2 08785 1 4 6 2 3 08266 1 4 6 2 3 5 0596

system Then comparison and discussion are given for eachpart of the presented system

41 Individual Classification Next six classifiers are utilizedto classify the calculated features of the bolt loosening Therelevant parameters setup for these classifiers can be found inTable 1 Table 2 gives the test accuracy of the six classifiers Inthe experiment the classification accuracy is evaluated usinga ratio of the number of the samples classified correctly tothe total sample It can be seen that the best classificationaccuracy is 08952 and 08571 of SVM and 119896-NN agents Asfar as performance of the six classifiers is concerned SVMand 119896-NN produce superior results followed by GMM agentLVQ and IIS are not suitable in this work for joint failure Itimplies that the two types of classifiers do not fit for the scatterof training samples But in practice it is almost impossiblethat all the predetermined classifiers will achieve the bestperformance at the same time Otherwise the fusion of puregood or bad classifiers groupmay not necessarily improve theaccuracy [24]Therefore LVQand IIS are still reserved for thefusion method

42 Selection of Classifiers Based on the individual clas-sification decisions acquired in the first step the entropy-based diversity measure method introduced in Section 21 isused for sequence selection of six classifiers The optimizedselected results for different numbers of classifiers and theentropy-based diversity degrees are shown in Table 3

To evaluate the effect of classifier selection Bayesianfusion method with classifier selection is compared withthe one without classifier selection as shown in Figure 9

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

6 Shock and Vibration

1 2 3

4 5 6

7 8 9

PZT SA Bolt

10 cm

49 cm

45 cm

120 cm

200 cm

11 cm

10 cm

Subarea 2 Subarea 4

Subarea 6 Subarea 7 Subarea 8

Figure 5 System setup

PZT 3

PZT 13 PZT 1 PZT 7

PZT 15 PZT 9

Bolt 2

Bolt 9

Bolt 22 Bolt 27

Bolt 39

Bolt 49Bolt 54

Bolt 5

Bolt 19

Bolt 33

Bolt 43

Bolt 63middot middot middot

middot middot middot

Figure 6 The loosening bolts monitored location

Actuator

integrated device

Power amplifier

Computer

Sensor

Piezoelectric sensor arrays

Waveform generator

Charge amplifier Data

acquisition

Switch Digital IOMultichannel

Figure 7 The principle structure of the active monitoring system

Bolts 2 5 9 19 22 23 24 25 26 27 33 39 43 49 50 51 5253 54 or 63 as shown in Figure 6

In the experiment twelve PZT sensors around the bound-ary are employed to detect the bolt loosening with thecycle-actuator multisensor method For the measurement

hardware the self-design integrate and program control mul-tichannel piezoelectric scanning system is used in the activemonitoring for bolt loosening [34] As shown in Figure 7the system integrates the waveform generator module dataacquisition module charge amplifier module digital IO

Shock and Vibration 7

0 1 2 3 4 5 6Time (s)

Volta

ge (V

)5

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

times10minus5

(a) The excited sine signal

0 1 2 3 4 5 6

Volta

ge (V

)

Time (s) times10minus5

2

15

1

05

0

minus05

minus1

minus15

minus2

(b) The signal before bolt loosening

0 1 2 3 4 5 6times10

minus5

Volta

ge (V

)

2

15

1

05

0

minus05

minus1

minus15

minus2

Time (s)

(c) The sensor signal after bolt loosening

Figure 8 The sensor signal change before and after Bolt 9 loosening

module multichannel scanning switch board and poweramplifier It can interrogate the large numbers of actuator-sensor channel automatically and efficiently The software isprogrammedwith LabVIEW85 andMATLABR2006a in theindustry control computer

32 Data and Feature The computer controls twelve PZTsensors circularly and periodically to excite and sense thestructure strain signal The excitation signal is the sine wavewith 100KHz Lots of experiments [33] have shown thatthe vibration response of the structure under this excitationfrequency is sensitive to the bolt loosening The number ofsampled data is 6000 and the measured time is 00006 s Thesample frequency is 10MHz Figure 8 gives the signal of PZTsensor 1 as actuator and PZT sensor 4 as sensor before andafter bolt loosening (Bolt 9) There are two reasons for theacquired strain signal changes Firstly bolt loosening couldcause the change of the prestress distribution in the structurewhich makes the structure thickness change Hence Lamb

wave with different modes generated by the PZT actuatorpropagates in the plate structure with a different velocityHence PZT sensor acquires different Lamb wave signalsbefore and after bolt loosening Secondly the bolt is deemedto be a scattering source on the Lamb wave propagation pathbetween the actuator and the sensor When bolt looseningcan partly affect the scattering Lamb wave coupled with boltits previous propagation path changes so the acquired Lambwave signal changes For sine wave excitation signal theexperiment [33] shows that the peak change is obvious beforeand after the bolt loosening So twenty-four acquired signalpeaks of the twelve sensors on the plate border are combinedto be a feature vector For the chosen bolts twenty-onemodesrsquo conditions are measured twenty-five times and tensamples are measured to train parameters of the classifiersand ten ones are used to train the fusion method So finallywe obtain a total of 525 samples which consist of 210samples for training classifiers 210 samples for training fusionalgorithms and the remaining 105 samples for test

8 Shock and Vibration

Table 1 Parameters of individual classifier

Classifier SVM C45 119896-NN IIS LVQ

Parameterssetup

Kernel function119896(x y) = (07xTy + 1)

2

Euclidean distance typepenalty coefficient = 10

Percentage of incorrectlyassigned samplesat a node = 5

119896 = 3 Number ofiterations = 50

Number ofneurons = 50 epochs = 50

33 Classifier Description Six pattern classification methodsare utilized to identify the loosening bolt The utilizedclassifiers are described as follows

(1) Support vector machine (SVM) the method canimplement the good recognition rate derived froma few training samples and it is based on statisticallearning theory [35] Kernel function is a key param-eter for SVM which includes linear polynomialGaussian RBF and sigmoid

(2) C45 the algorithm implements ldquoIf-Thenrdquo rulesderived from the training data set [36] These rulesare used to classify the ldquounseenrdquo data

(3) 119896 nearest neighbor (119896-NN) the classifier is verysimple and effective [37] The 119896 nearest neighbors ofthe unidentified test pattern are searched within ahypersphere of predefined radius in order to deter-mine its true class which is the most class in the 119896

samples If only one nearest neighbor is detected 119896-NN is the minimum-distance classification

(4) Improved iterative scaling (IIS) IIS is one of themajor algorithms for finding the optimal parame-ters for the conditional exponential model [38] Itsunderlying idea is that by approximating the log-likelihood function of the conditional exponentialmodel as some kind of ldquosimplerdquo auxiliary functionit is able to decouple the correlation between theparameters and search for the maximum point alongmany directions simultaneously By carrying out thisprocedure iteratively the approximated optimal pointfound over the ldquosimplifiedrdquo function is guaranteedto converge to the true optimal point due to theconvexity of the objective function

(5) Gaussian mixture model (GMM) the classifier isbased on Gaussian component functions [39] Thelinear combination of Gaussian functions is capableof representing a large class of the sample distribu-tion In principle it is a compromise between theperformance and the complexity Gaussian mixturehas remarkable capability to model the irregular data

(6) Learning vector quantization (LVQ) it is a neuralnetwork classifier proposed by Villmann et al [40] Itcombines the simplicity of competitive learning withthe accuracy of supervision It is a simple and intuitiveprototype-based clustering algorithm

4 Results and Discussion

This section describes the result of an experiment of thebolted joint monitoring using the proposed decision fusion

Table 2 Classification results

Classifier SVM C45 119896-NN IIS GMM LVQAccuracy 08952 05385 08571 01904 07524 03077

Table 3 Result of optimal sequence of classifiers fused

Number ofclassifiersselected

Serial number of classifiersEntropy-based

diversitymeasure

1 1 mdash2 1 4 11273 1 4 6 09784 1 4 6 2 08785 1 4 6 2 3 08266 1 4 6 2 3 5 0596

system Then comparison and discussion are given for eachpart of the presented system

41 Individual Classification Next six classifiers are utilizedto classify the calculated features of the bolt loosening Therelevant parameters setup for these classifiers can be found inTable 1 Table 2 gives the test accuracy of the six classifiers Inthe experiment the classification accuracy is evaluated usinga ratio of the number of the samples classified correctly tothe total sample It can be seen that the best classificationaccuracy is 08952 and 08571 of SVM and 119896-NN agents Asfar as performance of the six classifiers is concerned SVMand 119896-NN produce superior results followed by GMM agentLVQ and IIS are not suitable in this work for joint failure Itimplies that the two types of classifiers do not fit for the scatterof training samples But in practice it is almost impossiblethat all the predetermined classifiers will achieve the bestperformance at the same time Otherwise the fusion of puregood or bad classifiers groupmay not necessarily improve theaccuracy [24]Therefore LVQand IIS are still reserved for thefusion method

42 Selection of Classifiers Based on the individual clas-sification decisions acquired in the first step the entropy-based diversity measure method introduced in Section 21 isused for sequence selection of six classifiers The optimizedselected results for different numbers of classifiers and theentropy-based diversity degrees are shown in Table 3

To evaluate the effect of classifier selection Bayesianfusion method with classifier selection is compared withthe one without classifier selection as shown in Figure 9

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 7

0 1 2 3 4 5 6Time (s)

Volta

ge (V

)5

4

3

2

1

0

minus1

minus2

minus3

minus4

minus5

times10minus5

(a) The excited sine signal

0 1 2 3 4 5 6

Volta

ge (V

)

Time (s) times10minus5

2

15

1

05

0

minus05

minus1

minus15

minus2

(b) The signal before bolt loosening

0 1 2 3 4 5 6times10

minus5

Volta

ge (V

)

2

15

1

05

0

minus05

minus1

minus15

minus2

Time (s)

(c) The sensor signal after bolt loosening

Figure 8 The sensor signal change before and after Bolt 9 loosening

module multichannel scanning switch board and poweramplifier It can interrogate the large numbers of actuator-sensor channel automatically and efficiently The software isprogrammedwith LabVIEW85 andMATLABR2006a in theindustry control computer

32 Data and Feature The computer controls twelve PZTsensors circularly and periodically to excite and sense thestructure strain signal The excitation signal is the sine wavewith 100KHz Lots of experiments [33] have shown thatthe vibration response of the structure under this excitationfrequency is sensitive to the bolt loosening The number ofsampled data is 6000 and the measured time is 00006 s Thesample frequency is 10MHz Figure 8 gives the signal of PZTsensor 1 as actuator and PZT sensor 4 as sensor before andafter bolt loosening (Bolt 9) There are two reasons for theacquired strain signal changes Firstly bolt loosening couldcause the change of the prestress distribution in the structurewhich makes the structure thickness change Hence Lamb

wave with different modes generated by the PZT actuatorpropagates in the plate structure with a different velocityHence PZT sensor acquires different Lamb wave signalsbefore and after bolt loosening Secondly the bolt is deemedto be a scattering source on the Lamb wave propagation pathbetween the actuator and the sensor When bolt looseningcan partly affect the scattering Lamb wave coupled with boltits previous propagation path changes so the acquired Lambwave signal changes For sine wave excitation signal theexperiment [33] shows that the peak change is obvious beforeand after the bolt loosening So twenty-four acquired signalpeaks of the twelve sensors on the plate border are combinedto be a feature vector For the chosen bolts twenty-onemodesrsquo conditions are measured twenty-five times and tensamples are measured to train parameters of the classifiersand ten ones are used to train the fusion method So finallywe obtain a total of 525 samples which consist of 210samples for training classifiers 210 samples for training fusionalgorithms and the remaining 105 samples for test

8 Shock and Vibration

Table 1 Parameters of individual classifier

Classifier SVM C45 119896-NN IIS LVQ

Parameterssetup

Kernel function119896(x y) = (07xTy + 1)

2

Euclidean distance typepenalty coefficient = 10

Percentage of incorrectlyassigned samplesat a node = 5

119896 = 3 Number ofiterations = 50

Number ofneurons = 50 epochs = 50

33 Classifier Description Six pattern classification methodsare utilized to identify the loosening bolt The utilizedclassifiers are described as follows

(1) Support vector machine (SVM) the method canimplement the good recognition rate derived froma few training samples and it is based on statisticallearning theory [35] Kernel function is a key param-eter for SVM which includes linear polynomialGaussian RBF and sigmoid

(2) C45 the algorithm implements ldquoIf-Thenrdquo rulesderived from the training data set [36] These rulesare used to classify the ldquounseenrdquo data

(3) 119896 nearest neighbor (119896-NN) the classifier is verysimple and effective [37] The 119896 nearest neighbors ofthe unidentified test pattern are searched within ahypersphere of predefined radius in order to deter-mine its true class which is the most class in the 119896

samples If only one nearest neighbor is detected 119896-NN is the minimum-distance classification

(4) Improved iterative scaling (IIS) IIS is one of themajor algorithms for finding the optimal parame-ters for the conditional exponential model [38] Itsunderlying idea is that by approximating the log-likelihood function of the conditional exponentialmodel as some kind of ldquosimplerdquo auxiliary functionit is able to decouple the correlation between theparameters and search for the maximum point alongmany directions simultaneously By carrying out thisprocedure iteratively the approximated optimal pointfound over the ldquosimplifiedrdquo function is guaranteedto converge to the true optimal point due to theconvexity of the objective function

(5) Gaussian mixture model (GMM) the classifier isbased on Gaussian component functions [39] Thelinear combination of Gaussian functions is capableof representing a large class of the sample distribu-tion In principle it is a compromise between theperformance and the complexity Gaussian mixturehas remarkable capability to model the irregular data

(6) Learning vector quantization (LVQ) it is a neuralnetwork classifier proposed by Villmann et al [40] Itcombines the simplicity of competitive learning withthe accuracy of supervision It is a simple and intuitiveprototype-based clustering algorithm

4 Results and Discussion

This section describes the result of an experiment of thebolted joint monitoring using the proposed decision fusion

Table 2 Classification results

Classifier SVM C45 119896-NN IIS GMM LVQAccuracy 08952 05385 08571 01904 07524 03077

Table 3 Result of optimal sequence of classifiers fused

Number ofclassifiersselected

Serial number of classifiersEntropy-based

diversitymeasure

1 1 mdash2 1 4 11273 1 4 6 09784 1 4 6 2 08785 1 4 6 2 3 08266 1 4 6 2 3 5 0596

system Then comparison and discussion are given for eachpart of the presented system

41 Individual Classification Next six classifiers are utilizedto classify the calculated features of the bolt loosening Therelevant parameters setup for these classifiers can be found inTable 1 Table 2 gives the test accuracy of the six classifiers Inthe experiment the classification accuracy is evaluated usinga ratio of the number of the samples classified correctly tothe total sample It can be seen that the best classificationaccuracy is 08952 and 08571 of SVM and 119896-NN agents Asfar as performance of the six classifiers is concerned SVMand 119896-NN produce superior results followed by GMM agentLVQ and IIS are not suitable in this work for joint failure Itimplies that the two types of classifiers do not fit for the scatterof training samples But in practice it is almost impossiblethat all the predetermined classifiers will achieve the bestperformance at the same time Otherwise the fusion of puregood or bad classifiers groupmay not necessarily improve theaccuracy [24]Therefore LVQand IIS are still reserved for thefusion method

42 Selection of Classifiers Based on the individual clas-sification decisions acquired in the first step the entropy-based diversity measure method introduced in Section 21 isused for sequence selection of six classifiers The optimizedselected results for different numbers of classifiers and theentropy-based diversity degrees are shown in Table 3

To evaluate the effect of classifier selection Bayesianfusion method with classifier selection is compared withthe one without classifier selection as shown in Figure 9

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

8 Shock and Vibration

Table 1 Parameters of individual classifier

Classifier SVM C45 119896-NN IIS LVQ

Parameterssetup

Kernel function119896(x y) = (07xTy + 1)

2

Euclidean distance typepenalty coefficient = 10

Percentage of incorrectlyassigned samplesat a node = 5

119896 = 3 Number ofiterations = 50

Number ofneurons = 50 epochs = 50

33 Classifier Description Six pattern classification methodsare utilized to identify the loosening bolt The utilizedclassifiers are described as follows

(1) Support vector machine (SVM) the method canimplement the good recognition rate derived froma few training samples and it is based on statisticallearning theory [35] Kernel function is a key param-eter for SVM which includes linear polynomialGaussian RBF and sigmoid

(2) C45 the algorithm implements ldquoIf-Thenrdquo rulesderived from the training data set [36] These rulesare used to classify the ldquounseenrdquo data

(3) 119896 nearest neighbor (119896-NN) the classifier is verysimple and effective [37] The 119896 nearest neighbors ofthe unidentified test pattern are searched within ahypersphere of predefined radius in order to deter-mine its true class which is the most class in the 119896

samples If only one nearest neighbor is detected 119896-NN is the minimum-distance classification

(4) Improved iterative scaling (IIS) IIS is one of themajor algorithms for finding the optimal parame-ters for the conditional exponential model [38] Itsunderlying idea is that by approximating the log-likelihood function of the conditional exponentialmodel as some kind of ldquosimplerdquo auxiliary functionit is able to decouple the correlation between theparameters and search for the maximum point alongmany directions simultaneously By carrying out thisprocedure iteratively the approximated optimal pointfound over the ldquosimplifiedrdquo function is guaranteedto converge to the true optimal point due to theconvexity of the objective function

(5) Gaussian mixture model (GMM) the classifier isbased on Gaussian component functions [39] Thelinear combination of Gaussian functions is capableof representing a large class of the sample distribu-tion In principle it is a compromise between theperformance and the complexity Gaussian mixturehas remarkable capability to model the irregular data

(6) Learning vector quantization (LVQ) it is a neuralnetwork classifier proposed by Villmann et al [40] Itcombines the simplicity of competitive learning withthe accuracy of supervision It is a simple and intuitiveprototype-based clustering algorithm

4 Results and Discussion

This section describes the result of an experiment of thebolted joint monitoring using the proposed decision fusion

Table 2 Classification results

Classifier SVM C45 119896-NN IIS GMM LVQAccuracy 08952 05385 08571 01904 07524 03077

Table 3 Result of optimal sequence of classifiers fused

Number ofclassifiersselected

Serial number of classifiersEntropy-based

diversitymeasure

1 1 mdash2 1 4 11273 1 4 6 09784 1 4 6 2 08785 1 4 6 2 3 08266 1 4 6 2 3 5 0596

system Then comparison and discussion are given for eachpart of the presented system

41 Individual Classification Next six classifiers are utilizedto classify the calculated features of the bolt loosening Therelevant parameters setup for these classifiers can be found inTable 1 Table 2 gives the test accuracy of the six classifiers Inthe experiment the classification accuracy is evaluated usinga ratio of the number of the samples classified correctly tothe total sample It can be seen that the best classificationaccuracy is 08952 and 08571 of SVM and 119896-NN agents Asfar as performance of the six classifiers is concerned SVMand 119896-NN produce superior results followed by GMM agentLVQ and IIS are not suitable in this work for joint failure Itimplies that the two types of classifiers do not fit for the scatterof training samples But in practice it is almost impossiblethat all the predetermined classifiers will achieve the bestperformance at the same time Otherwise the fusion of puregood or bad classifiers groupmay not necessarily improve theaccuracy [24]Therefore LVQand IIS are still reserved for thefusion method

42 Selection of Classifiers Based on the individual clas-sification decisions acquired in the first step the entropy-based diversity measure method introduced in Section 21 isused for sequence selection of six classifiers The optimizedselected results for different numbers of classifiers and theentropy-based diversity degrees are shown in Table 3

To evaluate the effect of classifier selection Bayesianfusion method with classifier selection is compared withthe one without classifier selection as shown in Figure 9

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 9

Table 4 Relationship of accordance criterion number of classifiers fused and accuracy

Accordance criterion 120588

Number of classifiers fused1 2 3 4 5 6

Accuracy050 0895 0933 0962 0952 0949 0956055 0895 0933 0962 0952 0949 0956060 0895 0933 0962 0952 0949 0956065 0895 0933 0962 0952 0949 0956070 0895 0933 0962 0952 0971 0971075 0895 0933 0962 0952 0971 0971080 0895 0933 0962 0952 0971 0971085 0895 0933 0962 0952 0971 0971090 0895 0933 0962 0952 0971 0971095 0895 0933 0962 0952 0971 0971

1 2 3 4 5 6089

09

091

092

093

094

095

096

097

Number of classifiers fused (Bayesian)

Accu

racy

SelectionNo selection

Figure 9 Effect of classifiers selection (Bayesian method)

The results show that the fusion accuracy rate with theselection process is higher than that of the no selectionprocess Therefore selection of classifiers is proposed asa potential optimization process before the final decisionfusion

43 Decision Fusion After the six classifiers are sequen-tially selected the decision vectors of multiclassifiers arefused using three fusion methods namely majority votingBayesian belief and multiagent method In the multiagentmethod accordance criterion is a vital parameter The largerthe value is configured the longer computation time it takesand the better accuracy rate it produces In order to searchthe optimization value the value is traversed from 05 to 1with a step size of 005 and the corresponding fusion resultsare shown in Table 4 When the value is 07 and the numberof classifiers fused is 5 there is the optimization in the costof time and the accuracy While the accordance criterion is

1 2 3 4 5 6082

084

086

088

09

092

094

096

098

1

Number of classifiers fused

Accu

racy

Majority votingBayesianMultiagent

Figure 10 Fusion performances of three algorithms for currentdata

gradually increased from 07 to 09 the fusion result is notmuch improved

The performance of the three fusion algorithms is com-pared as shown in Figure 10 It can be seen that multiagentmethod is better than Bayesian method when the numberof classifiers fused is more than 3 The maximum fusionaccuracy formultiagentmethod is 0971 while it needs fusing5 classifiers While the maximum accuracy using Bayesianmethod is 0962 it only needs 3 classifiers The minimumaccuracy for the two methods is 0895 Compared with theother two methods the maximum and minimum fusionaccuracy for majority voting method are 0904 and 0838and it gives the worst fusion performance The reason isthat multiagent and Bayesian methods involve soft dynamicfusion and majority voting is only a static fusion processSince multiagent method includes two-order correlation

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

10 Shock and Vibration

degree between the classifiers it provides better performancethan Bayesian method

5 Conclusions

In this paper a decision system for bolted joint monitor-ing is presented which consists of individual classificationclassifier selection and decision fusion The effectivenessof the proposed methodology is tested with examples ofthe large aviation aluminum plate structure In the processclassification accuracy considering the classifier selection issuperior to the oneswithout the step To compare three fusionmethods the multiagent method is the best since the methodnot only considers the character of individual classifiers butalso the information exchange between the classifiers Deci-sion fusion strategy can improve the classification accuracyremarkable

Based on the decision fusion framework further studiesare required concentrating on the following three parts

(1) investigating more joint failure modes including thelevel of the bolt loosening and validating the effec-tiveness of the presented method more studies areneeded with complex structures to fully validate thenew method

(2) comparing other different methods of classifier selec-tion and evaluating these methods

(3) studying deeply the relation among the individualclassifier classifier selection and fusion method

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

Thiswork is supported by theNational Natural Science Foun-dation of China (Grant no 51405409) and the FundamentalResearch Funds for the Central Universities

References

[1] T Sakai et al ldquoBolt clamping force measurement with ultra-sonic wavesrdquo Transactions of the Japan Society of MechanicalEngineers vol 43 pp 723ndash729 1977

[2] Y Gotoh M Tanaka and H Yanoh ldquoStudy on inspectionmethod to measure slack of a bolt using electromagneticvibrationrdquo Journal of the Japanese Society for Non-DestructiveInspection vol 51 pp 24ndash31 2002

[3] N G Pai and D P Hess ldquoExperimental study of looseningof threaded fasteners due to dynamic shear loadsrdquo Journal ofSound and Vibration vol 253 no 3 pp 585ndash602 2002

[4] N G Pai and D P Hess ldquoInfluence of fastener placement onvibration-induced looseningrdquo Journal of Sound and Vibrationvol 268 no 3 pp 617ndash626 2003

[5] V Caccese R Mewer and S S Vel ldquoDetection of bolt load lossin hybrid compositemetal bolted connectionsrdquo EngineeringStructures vol 26 no 7 pp 895ndash906 2004

[6] R L Brown and D E Adams ldquoEquilibrium point damageprognosis models for structural health monitoringrdquo Journal ofSound and Vibration vol 262 no 3 pp 591ndash611 2003

[7] M D Todd J M Nichols C J Nichols and L N Virgin ldquoAnassessment of modal property effectiveness in detecting boltedjoint degradation theory and experimentrdquo Journal of Sound andVibration vol 275 no 3ndash5 pp 1113ndash1126 2004

[8] J M Nichols M D Todd and J R Wait ldquoUsing state spacepredictive modeling with chaotic interrogation in detectingjoint preload loss in a frame structure experimentrdquo SmartMaterials and Structures vol 12 no 4 pp 580ndash601 2003

[9] J M Nichols C J Nichols M D Todd M Seaver S T Trickeyand L N Virgin ldquoUse of data-driven phase space models inassessing the strength of a bolted connection in a compositebeamrdquo Smart Materials and Structures vol 13 no 2 pp 241ndash250 2004

[10] L Moniz J M Nichols C J Nichols et al ldquoA multivariateattractor-based approach to structural healthmonitoringrdquo Jour-nal of Sound and Vibration vol 283 no 1-2 pp 295ndash310 2005

[11] A C Rutherford G Park and C R Farrar ldquoNon-linear featureidentifications based on self-sensing impedance measurementsfor structural health assessmentrdquoMechanical Systems and SignalProcessing vol 21 no 1 pp 322ndash333 2007

[12] S Ritdumrongkul and Y Fujino ldquoIdentification of the locationand level of damage in multiple-bolted-joint structures by PZTactuator-sensorsrdquo Journal of Structural Engineering vol 132 no2 pp 304ndash311 2006

[13] S Ritdumrongkul M Abe Y Fujino and T Miyashita ldquoQuan-titative health monitoring of bolted joints using a piezoceramicactuator-sensorrdquo Smart Materials and Structures vol 13 no 1pp 20ndash29 2004

[14] G Park H Sohn C R Farrar and D J Inman ldquoOverviewof piezoelectric impedance-based health monitoring and pathforwardrdquo Shock and Vibration Digest vol 35 no 6 pp 451ndash4632003

[15] G Park H H Cudney and D J Inman ldquoFeasibility of usingimpedance-based damage assessment for pipeline structuresrdquoEarthquake Engineering and Structural Dynamics vol 30 no10 pp 1463ndash1474 2001

[16] J Yang F-K Chang and M M Derriso ldquoDesign of a hier-archical health monitoring system for detection of multileveldamage in bolted thermal protection panels a preliminarystudyrdquo Structural Health Monitoring vol 2 no 2 pp 115ndash1222003

[17] M Okugawa ldquoBolt loosening detection methods by usingsmart washer adopted 4SIDrdquo in Proceedings of the 45thAIAAASMEASCEAHSASC Structures Structural Dynamicsamp Materials Conference AIAA2004 -1981 Palm Spring CalifUSA April 2004

[18] A Milanese P Marzocca J M Nichols M Seaver and ST Trickey ldquoModeling and detection of joint loosening usingoutput-only broad-band vibration datardquo StructuralHealthMon-itoring vol 7 no 4 pp 309ndash328 2008

[19] D Doyle A Zagrai B Arritt and H Cakan ldquoDamage detec-tion in bolted space structuresrdquo Journal of Intelligent MaterialSystems and Structures vol 21 no 3 pp 251ndash264 2010

[20] L H Yam Y J Yan L Cheng and J S Jiang ldquoIdentificationof complex crack damage for honeycomb sandwich plate usingwavelet analysis and neural networksrdquo Smart Materials andStructures vol 12 no 5 pp 661ndash671 2003

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Shock and Vibration 11

[21] G I Lemoine K W Love and T A Anderson ldquoAn electricpotential-based structural health monitoring technology usingneural networkrdquo in Proceedings of the 4th International Work-shop on Structural HealthMonitoring F-K Chang Ed pp 387ndash395 DEStech Publications Stanford Calif USA 2003

[22] C-B Yun and E Y Bahng ldquoSubstructural identification usingneural networksrdquo Computers and Structures vol 77 no 1 pp41ndash52 2000

[23] X Zhao C Kwan R Xu et al ldquoNon-destructive inspectionof metal matrix composites using guided wavesrdquo in Review ofQuantitative Nondestructive Evaluation D O Thompson andD E Chimenti Eds vol 23 pp 914ndash920 American Institute ofPhysics Melville NY USA 2004

[24] M Petrakos J A Benediktsson and I Kanellopoulos ldquoTheeffect of classifier agreement on the accuracy of the combinedclassifier in decision level fusionrdquo IEEE Transactions on Geo-science and Remote Sensing vol 39 no 11 pp 2539ndash2546 2001

[25] D Ruta and B Gabrys ldquoClassifier selection formajority votingrdquoInformation Fusion vol 6 no 1 pp 63ndash81 2005

[26] W-Y Liu Z-H Wu and G Pa ldquoAn entropy-based diversitymeasure for classifier combining and its application to faceclassifier ensemble thinningrdquo Advances in Biometric PersonAuthentication vol 3338 pp 118ndash124 2005

[27] L Xu A Krzyzak and C Y Suen ldquoMethods of combiningmultiple classifiers and their applications to handwriting recog-nitionrdquo IEEETransactions on SystemsMan andCybernetics vol22 no 3 pp 418ndash435 1992

[28] R Battiti and A M Colla ldquoDemocracy in neural nets votingschemes for classificationrdquo Neural Networks vol 7 no 4 pp691ndash707 1994

[29] J-B Kou and C-S Zhang ldquoMulti-agent based classifier combi-nationrdquo Chinese Journal of Computers vol 26 pp 1ndash5 2003

[30] G Niu T Han B S Yang and A C C Tan ldquoMulti-agentdecision fusion for motor fault diagnosisrdquo Mechanical Systemsand Signal Processing vol 21 no 3 pp 1285ndash1299 2007

[31] W Michael and N R Jennings ldquoIntelligent agents theory andpracticerdquo Knowledge Engineering Review vol 10 pp 115ndash1521995

[32] B-Q Tao Smart Materials and Structures National DefenseIndustry Press Beijing China 1997

[33] X Zhao S-F Yuan H-B Zhou H-B Sun and L Qiu ldquoAnevaluation on the multi-agent system based structural healthmonitoring for large scale structuresrdquo Expert Systems withApplications vol 36 no 3 pp 4900ndash4914 2009

[34] L Qiu and S-F Yuan ldquoOn development of amulti-channel PZTarray scanning system and its evaluating application on UAVwing boxrdquo Sensors and Actuators A Physical vol 151 no 2 pp220ndash230 2009

[35] C Cortes and V Vapnik ldquoSupport-vector networksrdquo MachineLearning vol 20 no 3 pp 273ndash297 1995

[36] J R Quinlan C45 Programs for Machine Learning MorganKaufmann San Mateo Calif USA 1992

[37] S Singh J Haddon and M Markou ldquoNearest-neighbourclassifiers in natural scene analysisrdquo Pattern Recognition vol 34no 8 pp 1601ndash1612 2001

[38] A Berger ldquoThe improved iterative scaling algorithm a gentleintroductionrdquo httpluthulicsuiucedusimdafcoursesOptimi-zationPapersberger-iispdf

[39] Q Y Hong and S Kwong ldquoA genetic classification methodfor speaker recognitionrdquo Engineering Applications of ArtificialIntelligence vol 18 no 1 pp 13ndash19 2005

[40] T Villmann F Schleif and B Hammer ldquoComparison of rele-vance learning vector quantization with other metric adaptiveclassification methodsrdquoNeural Networks vol 19 no 5 pp 610ndash622 2006

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of