Research Article Vibration-Based Adaptive Novelty ...downloads.hindawi.com/journals/sv/2016/2417856.pdf · applied under a continuous learning framework. e structure of the method

Research ArticleVibration-Based Adaptive Novelty Detection Method forMonitoring Faults in a Kinematic Chain

Jesus Adolfo Carintildeo-Corrales1 Juan Jose Saucedo-Dorantes2

Daniel Zurita-Millaacuten1 Miguel Delgado-Prieto1 Juan Antonio Ortega-Redondo1

Roque Alfredo Osornio-Rios2 and Rene de Jesus Romero-Troncoso3

1MCIA Research Center Department of Electronic Engineering Technical University of Catalonia (UPC) Rbla San Nebridi 22Gaia Research Building Terrassa 08222 Barcelona Spain2CA Mecatronica Facultad de Ingenieria Campus San Juan del Rio Universidad Autonoma de Queretaro Rio Moctezuma 249Col San Cayetano 76807 San Juan del Rio QRO Mexico3CATelematica DICIS Universidad deGuanajuato Carr Salamanca-Valle km 35 + 18 Palo Blanco 36885 Salamanca GTOMexico

Correspondence should be addressed to Rene de Jesus Romero-Troncoso troncosohspdigitalorg

Received 26 April 2016 Accepted 30 August 2016

Academic Editor Mariano Artes

Copyright copy 2016 Jesus Adolfo Carino-Corrales et al This is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

This paper presents an adaptive novelty detection methodology applied to a kinematic chain for the monitoring of faults Theproposed approach has the premise that only information of the healthy operation of the machine is initially available and faultscenarios will eventually develop This approach aims to cover some of the challenges presented when condition monitoring isapplied under a continuous learning framework The structure of the method is divided into two recursive stages first an offlinestage for initialization and retraining of the feature reduction and novelty detection modules and second an online monitoringstage to continuously assess the condition of the machine Contrary to classical static feature reduction approaches the proposedmethod reformulates the features by employing first a Laplacian Score ranking and then the Fisher Score ranking for retrainingThe proposed methodology is validated experimentally by monitoring the vibration measurements of a kinematic chain driven byan induction motor Two faults are induced in the motor to validate the method performance to detect anomalies and adapt thefeature reduction and novelty detection modules to the new information The obtained results show the advantages of employingan adaptive approach for novelty detection and feature reduction making the proposed method suitable for industrial machinerydiagnosis applications

1 Introduction

Condition-based maintenance has become accepted by theindustrial manufacturing sector as a key factor to achievethe currently required high-competitiveness levels In thefield of electromechanical systems monitoring an increasednumber of approaches of health monitoring schemes havebeen proposed during the last decades [1ndash3] These method-ologies have been demonstrated to be a reliable option asfault detection techniques when data of the possible faults isavailable

Nevertheless in order to overcome the practical require-ments of industrial implementation some considerationsmust be taken into account There is a necessity to developmethodologies capable of monitoring complicated industrialmachinery that is diagnosis approaches capable of consider-ing complete electromechanical kinematic chains includingelectric motors gearboxes and additional mechanical ele-ments It is important to mention that information regardingthe possible faults that could occur on complicated machinesis not usually available due to the high cost and obtaining thisinformation requires a great amount of time

Hindawi Publishing CorporationShock and VibrationVolume 2016 Article ID 2417856 12 pageshttpdxdoiorg10115520162417856

2 Shock and Vibration

Unexpected events in the form of faulty scenarios notinitially considered or deviations over the normal operationtake place during the useful life of the machinery These situ-ations must be identified in order to avoid misclassificationsand the consequent incorrect maintenance action To copewith these challenging scenarios the condition-based main-tenance methods must be able to work under the assumptionthat only information of the healthy condition of themachineis available [4] Another relevant characteristic is the capacityto incorporate new knowledge Since new faults could appearduring monitoring the condition-based maintenance meth-ods must also be able to adapt and include new scenarios totheir initial knowledge [5 6]

These challenging scenarios are not always taken intoconsideration in conditionmonitoringmethodologies there-fore the application of classical methods on an industrialenvironment is limited or nonviable

In the field of electromechanical system monitoring agreat deal of health monitoring schemes based on vibrationshave been proposed during the last years [7ndash10] Mainlythese proposals can be divided into two categories regardingthe available information of the monitored machine and theclassification objective

The first category multifault detection consists in thedevelopment of a monitoring system capable of identifyingmultiple faults previously known [1] which implies that theinformation available to train themonitoring system includesmeasurements of the healthy state of the machine but alsomeasurements of the machine working under different faultyconditions Over the last years instead of looking for highlyspecific features for a component fault a trend towardscalculation of more general statistical features and the fusionof information to enhance the performance of the diagnosissystem has been carried out [11] Yet this approach is mostlycarried out at the laboratory scale with successful resultswhere controlled faults are provoked on the machine

However the information of the possible faults that couldappear on a machine is not always available consequentlya different approach must be considered if only informationof healthy condition of the machine is available The secondcategory corresponds to novelty detection also known asanomaly detection and it represents a solution to cope withthe lack of information available of the monitored machineThe objective of novelty detection is to detect novel eventsthat differ in some manner from the information of whichthemodelwas trainedNovelty detection framework has beensuccessfully applied to network intrusion medical diagnosisimage segmentation and handwritten digit recognition faultdetection and condition monitoring [4 6 12ndash14] Specialeffort has been made to improve condition monitoringnovelty detection applications based on vibrations Ma et alproposed a novelty detection approach to rotating machin-ery by means of monitoring the thresholds obtained froman extreme value distribution [15] McBain and Timusk pro-posed a feature extractionmethodology for novelty detectionapplied to a rotating mechanical system [16]

Most of the methodologies for novelty detection arelimited to a static analysis and the incorporation of the novelinformation to the novelty detection system is not usually

considered An approach to include adaptability to thenovelty framework based on vibrations is proposed byWanget al [17] The proposed monitoring scheme includes testingdata on the boundary of the novelty model based on SupportVector Data Description and retrains the model with thisinformation to gain robustness Nevertheless this approachdoes not take into consideration the possibility to includenovel scenarios during the monitored process

Other approaches to develop an adapting conditionmon-itoring schemewere presented by Filev et al where a practicalframework for autonomous monitoring of industrial equip-ment based on novelty detection is analyzed [18] and byCostaet al where a two-stage algorithm for real-time fault detectionand identification is presented [19] Both approaches providethe opportunity to incorporate novel detected faults tothe monitoring system nevertheless in both methods theincorporation is limited to updating the known data basebut an adaptation of the numerical features analyzed is notconsidered

The performance of a novelty detection system is highlydependent on the numerical features consideredWhen thereis no previous information of the possible faults that canoccur the application of a suitable numerical features analysisstrategy represents a critical challenge [4] Considering acontinuous monitoring framework where the initial infor-mation available is the healthy operating condition andlater on different faults are identified progressively when themachine condition deteriorates all the approaches previouslydiscussed do not modify the initial set of features when newinformation of faults is incorporated This static approachhas the advantage of providing a most adequate situation foronline adaptation nevertheless analyzing the informationof the faults detected during the monitoring process couldimprove the identification of amost adequate set of features todiscriminate the possible upcoming or already detected faultscenarios

To improve the performance of the novelty detection taska dynamic approach for reselection of the features includinginformation of new faults detected is proposed on this workTo the extent of the authorsrsquo knowledge the study regardingthe dynamic change of the feature space by considering newoperating scenarios is novel in the applications to electrome-chanical systems for fault diagnosis

Thereby the contribution of this study is to provide anadaptive methodology for novelty detection where the infor-mation of identified faults during the monitoring process isexploited to improve the anomaly detection task The meth-odology is divided into two recursive stages first an offlinestage for initialization and retraining of the feature reductionand novelty detectionmodules second an onlinemonitoringstage to continuously assess the condition of the machineThe novelty model employed is a one-class Support Vectormachine (OC-SVM) [20] which is a domain based modelsuccessfully employed on vibration signals monitoring [4]Contrary to classical static feature reduction approaches theproposed method reformulates the features by employingfirst a Laplacian Score ranking and then the Fisher Scoreranking for retraining

Shock and Vibration 3

The proposed methodology is tested using real datafrom a laboratory based kinematic chain where faults areinduced and the vibration signals from two accelerometersare monitored to assess the system condition To highlightthe benefits of the methodology proposed a comparison isperformed of classical feature reduction approaches to theadaptive feature reduction scheme proposed

This paper is organized as follows In Section 2 the-oretical aspects of the proposed method are describedSection 3 describes the method presented The experimentalsetup employed for assessment and the results obtained arepresented anddiscussed in Section 4Andfinally conclusionsand future work are summarized in Section 5

2 Theoretical Background

21 One-Class Support Vector Machine As mentioned theobjective of novelty detection is to detect abnormal eventsthat differ in some manner from the information of whichthemodel was trainedTheone-class SupportVectormachine(OC-SVM) which is a domain based model successfullyemployed in condition monitoring applications encloses thetraining data by means of a boundary or threshold whichrepresents the limit of normality The OC-SVM presents thepossibility to include a kernel formulation providing it withenough flexibility to adapt to the distribution of the trainingdata To assess the condition of the machine under test newmeasurements are analyzed regarding their position to theboundary created in the training phase The test measure-ments underlying inside the boundary are considered ldquonor-malrdquo or ldquoknownrdquo while the measurements underlying outsidethe boundary are considered ldquooutlierrdquo or ldquonovelrdquo Comparedto other novelty detection techniques the OC-SVM is robustto outliers during the training phase but has the disadvantageof presenting several configuration parameters which aredescribed later in this section

The OC-SVM was proposed by Scholkopf et al for esti-mating the support of a high-dimensional distribution [20]The OC-SVM classification objective is to separate one classof target samples from all other class samples In this typeof problem one class is characterized properly called targetclass while for the other class usually no measurements areavailable Considering 119883 = [1199091 119909119873]119879 isin 119877119873times119872 whichdenotes the normal data set and 119909119894 119894 = 1 119873 whichdenotes training samples (available measurements) charac-terized by119872 numerical features then in order to obtain theboundary an optimization model is considered as follows

min w22 + 1119873V

119873sum119894=1

120585119894 minus 120588Subject to w sdot Φ (119909) ge 120588 minus 120585119894 120585119894 ge 0

(1)

where V is a regularization parameter and 120585119894 is the slackvariable for the point119909119894The constantsw and 120588 are the normalvector and offset of the hyperplane respectively Thus thedecision boundary can be formulated as

119891 (119909) = w sdot Φ (119909) minus 120588 (2)

where 119909 isin 119877119872 andΦ is a higher dimensional projection vec-tor For the classification problem of two categories the datasets are not always linearly separable in the original spacethen Φ projects the original data sets into a higher dimen-sional space the so-called feature space where the data setscan be linearly separable However Φ is inexplicit in thepractical application and only the dot product from Φ(119909119894) sdotΦ(119909119895) is necessary to be known 119870 represents the kernelfunction Φ(119909119894) sdot Φ(119909119895) The most commonly used kernelfunction is the Gaussian

119870(119909119894 119909119895) = exp(minus10038171003817100381710038171003817119909119894 minus 11990911989510038171003817100381710038171003817221205902 ) (3)

In order to solve optimization problem (1) Lagrangemul-tipliers 119886119894 ge 0 and 120573119894 ge 0 (119894 = 1 119873) are introduced andthe Lagrange equation is formed as

119871 (w 120585 120588 a120573) = w22 + 1119873V

119873sum119894=1

120585119894 minus 120588minus 119873sum119894=1

119886119894 (w sdot Φ (119909119894) minus 120588 + 120585119894)minus 119873sum119894=1

120573119894120585119894(4)

The partial derivatives of the Lagrangian equation withrespect to w 120585 and 120588 are set to zero Then w and 119886119894 can beformulated as

w = 119873sum119894=1

119886119894Φ(119909119894) 119886119894 = 1119873V

minus 120573119894119873sum119894=1

119886119894 = 1(5)

Substitute (5) into Lagrangian equation (4) and its dualform is presented as

min a119879119867a

subject to 0 le 119886119894 le 1119873V119873sum119894=1

119886119894 = 1(6)

where a = [119886119894 119886119873]119879 and 119867 is the kernel matrix and thefactor of119867 that is119867119894119895 which can be expressed as

119867119894119895 = 119870 (119909119894 119909119895) = Φ (119909119894) sdot Φ (119909119895) (7)

Solve optimization problem (6) to get a and then 120588 can begiven as

120588 = 1119899119904119899119904sum119894=1

119873sum119895=1

119886119895119870(119909119894 119909119895) a (8)

where 119899119904 is the number of Support Vectors


Tax and Duin proposed another form of OC-SVMwhichis called Support Vector Data Description [21]The basic ideaof SVDD is to construct aminimum-volume hypersphere in ahigh-dimensional feature space to enclose as much as normaldata points Both of these two forms of SVM have equivalentsolution if the diagonal entries of kernel matrix 119867 are equalto a constant

22 Dimensionality Reduction Techniques Working withhigh-dimensional data sets could complicate the learningpart of novelty detection methods not only because of possi-ble presence of noise and redundancy in the data but for otherreasons as well [22] The empty space phenomenon statesthat to cover the whole space a number of samples that growexponentially with dimensionality are needed The curse ofdimensionality implies that in order to learn successfully anumber of training examples that also grow exponentiallywith the dimensionality are needed The ldquoconcentration ofmeasurerdquo phenomenon seems to render distance measuresnot relevant to whatever concept is to be learnt as thedimension of the data increased For these reasons amongothers there is a necessity to apply dimensionality reductiontechniques in novelty detection applications Dimensionalityreduction strategies differ in the question of whether thelearning process is supervised or unsupervised The differ-ence between both learning processes is the availability oflabels to distinguish the different classes

Principal Component Analysis (PCA) is one of the mostcommonly used techniques for unsupervised dimensionalityreduction [23] It aims to find the linear projections that bestcapture the variability of the data

Another well-known dimensionality reduction techniqueis the Laplacian Score (LS) where the merit of each feature ismeasured according to its locality preservation power [24]A nearest neighbor based graph is constructed from thetraining set and analyzed to rank each feature individuallyaccording to a weighting approach selected for the graphedges If the labels of the classes of the analyzed data set aregiven the method changes to a supervised approach and thegraph is created based on these labels instead of the nearestneighbor approach Two variations for weighting the edgescan be applied Heat Kernel and Simpleminded To rank eachfeature its Laplacian Score is computed which is a measureof which extent the analyzed feature preserves the structurepresent in the graph divided by the variance of the feature Fora feature to be selected it must have a low LS which implieshigh variance and locality

Linear discriminant analysis (LDA) is one of the mostwell-known supervised techniques for linear dimensionalityreduction inmulticlass problems [25] LDA searches attemptsto maximize the linear separation between data pointsbelonging to different classes In contrast to most of theother dimensionality reduction techniques LDA as a featureextraction technique finds a linear mapping that maximizesthe linear class separation in the low-dimensional representa-tion of the data The criteria that are used to formulate linearclass separation in LDA are the within-class scatter and thebetween-class scatter

Another variant based on the Fisher coefficient is a featuredimensionality reduction approach called Fisher Score (FS)where the objective is to find a subset of features whichmaximize the Fisher coefficient [26] The main differencebetween LDA and Fisher Score is that LDA is a featureextraction approach and Fisher Score is a feature selectionapproach then instead of extracting new features fromthe subspace obtained FS ranks the features and selectthe reduced subspace of features selected which maximizesthe dimensionality reduction criterion in this case the Fishercriterion

The aforementioned linear feature reduction techniqueshave a different combination of objective (variance preserva-tion topology preservation and discrimination) andmethodof employment (unsupervised or supervised and extractionor selection) and have been widely used in the literature withsuccessful results [16 22] yet they have been applied under astatic framework where the set of features is reduced initiallyand no possible adaptation during the monitoring phase isconsidered

3 Methodology

Given that the data initially available is usually related to thehealthy condition of the machine under analysis condition-based monitoring schemes must be designed to overcometwo main challenges

(i) The identification of significant features to deal withthe characterization of the known conditions of the machineunder the consideration that the occurrence of additionalunknown faults must be detected

(ii) The adaptation of the condition-based monitoringscheme to update the considered data base of the machineonce unknown fault scenarios have been detected

Such requirements are addressed in this work by meansof the proposed adaptive novelty detection method shownin Figure 1 This methodology represents an important stepto the introduction of adaptive novelty detection schemes tothe development of electromechanical system diagnosis pro-cedures

The proposed method is composed of two stages anoffline stage and an online monitoring stageThemain objec-tives of the offline stage are first the analysis of the infor-mation available of the monitored machine to find a reducedset of numerical features to characterize the known machineconditions and second the design of the novelty model bymeans of the selection of the configuration parameters andtraining

Once a reduced set of characteristic features is obtainedand the novelty model is designed the online stage is carriedout During the online stage new measurements are contin-uously compared with the normality threshold 119879119899 definedduring the novelty model training in the offline stageThus ifa novel scenario is detected the supervision of an expert useris proposed in order to confirm and label the new conditionof the machine consequently the monitoring system isretrained to include the characteristics of the novel scenarioDetailed information of each stage and the retraining processis described in Sections 31ndash33


Database Feature calculation

Feature reduction

Novelty model design

New measurement

Calculation of reducedfeatures

NoveltyUser

supervisionKnown

operation False alarm

Offl

ine

Onl

ine

RetrainingIncluding

information of new fault

No

Yes

Noveltymodel

Figure 1 Proposedmethodology for the novelty detection approachThemonitoringmethod is composed of an offline stage for initializationand retraining and an online stage for continuous monitoring

31 Offline Stage During the initialization it is assumed thatonly information of the machine operating under healthycondition is available in the database The first step is thecalculation of numerical features from the vibration mea-surements obtained during the machine operation Since theinformation of the possible faults of the monitored machineis not available yet during this initialization a generic set ofstatistical time-based numerical features is proposed to beextracted from each available vibration axis measurementThe proposed set of potential statistical time-frequency fea-tures is shown as follows

Max Value

119894 = max (119909) (9)

Root Mean Square (RMS)

RMS = radic 1119899 sdot119899sum119896=1

(119909119896)2 (10)

Crest Factor (CF)

CF = 119909119896RMS

(11)

Kurtosis

119896 = 119864 [(119909119896 minus 119909)4]1205904 (12)

Shape Factor

SF = RMS1119899 sdot sum119899119896=1 10038161003816100381610038161199091198961003816100381610038161003816 (13)

These features have been successfully employed for faultdetection in the last years [1 27]

The resulting number of numerical features is propor-tional to the number of available vibration axes collectedduring the acquisition However in order to allow the com-pression and visualization of the data a feature reductionmodule is implemented During such offline stage initializa-tion an unsupervised feature reduction approach must be

used a Laplacian Score ranking is proposed in this work asa good trade-off between simplicity and performance [24]to rank the features according to the topology preservationcapabilities The two or three first ranked features in terms ofLaplacian Score are selected

Next the novelty model is designedThere is a significantnumber of noveltymodels proposed in the literature [4] witheach one demonstrated to be a capable option under certaincircumstances An increasing amount of works implies thatdomain based novelty detection models present promisingresults [19 28] In this work a standard OC-SVMwith Gaus-sian kernel is used The design of the novelty model includesthe selection of the parameters for configuration and trainingemploying the known scenarios stored at the database Thenthe initialization of the offline stage finalizes with the designof the OC-SVM

32 Online Stage This stage continuously monitors thecondition of the machine to detect if an anomaly is presentTo accomplish this each of the new measurements of themachine acquired certain amount of predefined time Eachmeasurement is segmented and a set of features is calculatedfrom each segment The set of numerical features calculatedin this stage are reduced in the offline stage by means ofLaplacian Score

Thus each new measurement characterized by thenumerical features is analyzed by the novelty model In caseof no novelty detection it is assumed that the machine isworking under known conditions However if the analyzedmeasurement is detected as a novelty an alarm is triggeredin order to consider the user assessment Then if theoccurrence of a new scenario is confirmed the correspondingmeasurements are stored at the database and a retrainingprocedure is performed

33 Retraining Once the retraining is trigged the featurereduction and the novelty model design modules are modi-fied at the offline stage A diagram of the retraining procedureis presented in Figure 2

It must be noticed that two important contributiveaspects are proposed in this retraining approach first thereconsideration of the feature reductionmodule and secondthe incorporation of fault scenarios to the novelty model As


New Fault Detected

Continuos Monitoring

Add Information to DDBB

No

Yes

Perform New Feature

Reduction

Novelty model redesign

A supervised method will be

used

Measurementsof the new fault

Data of all stored scenarios

included

Retr

aini

ng

New faultdetected

Continuousmonitoring

Add information to DDBB

Perform new feature

reduction

Novelty modelredesign

Figure 2 Proposed retraining approach First measurements char-acterizing the fault are stored and then the feature reductionmoduleand the novelty design module are modified to incorporate the newscenario encountered

it has been explained the feature reduction module duringits initialization is supported by an unsupervised LaplacianScore ranking due to the lack of additional scenarios yet oncethe information of a fault is available the feature reductionmodule could improve its performance by employing super-vised methods

Then the Laplacian approach is replaced by a FisherScore ranking approach in this work where the features aresorted according to the Fisher coefficient calculated fromeach feature It is important to mention that Laplacian Scorecan be configured towork also under a supervised frameworkand could obtain similar results than employing a Fisherapproach similarities between the two approaches are dis-cussed in [24] Nevertheless Fisher Score is used in thismethodology to search for a possible discriminative scenarioto increase the detection of novel scenarios A comparison ofresults obtained from both features reduction approaches ispresented in the results of this work

The consideration of a faulty scenario in the noveltymodel may contradict the principle of anomaly detectionwhere the objective is to detect healthy behaviors from therest Nevertheless the aim of an adaptive condition mon-itoring system should be to learnt from all the identifiedconditions to subsequently detect them if they are presentedagain by a fault detection module Indeed some workspresent a parallel structure for fault detection and noveltydetection modules [18 19] where the novelty module learnsthe known faults along with the healthy operation because itsobjective is to detect only new scenarios while the objectiveof the fault detection is to identify the condition of the

Gearbox

Load

VFD

DAS

Inductionmotor

AccelerometerEncoder

PC

z

yx

Figure 3 Test bench used for experimentation

machine including known faults This work is based on suchparallel structure approach where the fault scenariosmust betaken into consideration to be included in the novelty modeland then they could be identified by a complementary faultdetection module with a high confidence level

4 Experimental Results

41 Description of the Experimental Platform The test benchused for testing the kinematic chain with different faults andthe acquisition system used to capture the vibration signalsare shown in Figure 3 The test bench consists in a 1492-W-three-phase induction motor (WEG 00236ET3E145T-W22)with its rotational speed controlled through a variable fre-quency drive-VFD (WEG CFW08) the operating speed isfixed to 60Hz for all experiments It also consists of a 4 1 ratiogearbox (BALDORGCF4X01AA) and a DC generator (BAL-DOR CDP3604) which is used as the mechanical load com-prising around 20of the nominal loadThe vibration signalsare acquired using a triaxial accelerometer (LIS3L02AS4)mounted on a board with the signal conditioning andantialias filtering Two 12-bit 4-channel serial output sam-pling analog-to-digital converters (ADS7841) are used in thedata acquisition system (DAS) board Vibration signal fromthe perpendicular plane (119910-119911) to the motor shaft is acquiredusing a triaxial accelerometer LIS3L02AS4 mounted on aboard with the signal conditioning and antialiasing filteringSampling frequency is set to 3 kHz for vibration acquisitionThe data retrieved by the DAS is stored in a regular computer(PC)

Three scenarios are considered to verify the performanceof the proposed method the first one 119867 is the kinematicchain working under healthy condition and the other two 1198651and 1198652 represent the kinematic chain working under faultyconditions For 1198651 the motor is working with a half brokenbar and for 1198652 the motor is working with a fully brokenbar The detail of the failures is shown in Figure 4 The halfbroken bar failure is artificially produced by drilling a 6mmhole with a depth of 3mm that corresponds mostly to the


(a) (b)

Figure 4 Detail of the faults produced in the test bench (a) corresponds to the 12 broken rotor bar and (b) to one broken rotor bar

22 of the section of the rotor bar and the full broken baris produced by a through-hole with a diameter of 6mm anda depth of 14mm which corresponds to the complete sectionof the rotor bar

The information stored from the kinematic chain consistsof an acquisition of 60 seconds of themachine working underthe three scenarios mentioned each acquisition is segmentedin 30 parts of 2 seconds and a set of features is calculatedfrom the 30 segmented measurements Since two axes aretaken into consideration a total of ten features are calculatedfrom each segmented acquisition of the machine workingunder the different scenarios mentioned The first step ofthe methodology is the offline stage where a reduced set offeatures is obtained and the novelty model is designed

42 Parameter Selection Regarding the Laplacian Score con-figuration a simple approach is followed for parameter tun-ing that is a value of 119896 = 3 is used for constructing the adja-cency graph and a ldquosimple mindedrdquo weighting approach isfollowed Since the proposed approach is compared to differ-ent feature reduction modules selecting generic parameterssettings is useful for the purpose of evaluation but ignores thatthere may be dependencies between the feature reductionmodel and the novelty model Regarding the design of thenovelty model the kernel used is the Gaussian and the valueof the configuration parameter is tuned tominimize the errorin the validation In all experiments 80 of the samplesare used for training and 20 for validation To train andadjust the parameter of the novelty model a fivefold cross-validation is used

43 Results and Discussions In order to highlight the contri-butions andmotivation of this work the outline of the resultsis presented as follows first a test is performed by a classicalapproach then the proposedmethodology is applied and theresults are compared The classical static approach impliesconserving the reduced features set obtained at the offline ini-tialization stage meanwhile the proposed dynamic approachimplies a possible reformulation of the reduced features setduring the retraining stage

Healthy

0

005

01

015

02

minus2 minus1minus3 1 2 30LS feature 1 RMS Y

minus3

minus2

minus1

0

1

2

3LS

feat

ure 2

kur

tosis

Y

Figure 5 Initial novelty model representation Limit of the noveltythreshold sdot sdot sdot and measurements used to train the model lowast

During the initialization stage the reduced set of featuresselected bymeans of the Laplacian Score ranking is composedof the RMS of the 119910-axis and the Kurtosis of the 119910-axisThe OC-SVM model is trained employing healthy conditiondata The resulting OC-SVM during training is shown inFigure 5 The marks lowast in the figure represent measurementsof the machine used to train the model in this case andin the subsequence figures only 1-fold of the fivefold cross-validation is displayedThe dotted line represents the noveltythreshold value all data inside the boundaries of the dottedline is considered normal The contour plot represents thenovelty score evaluation over different regions of the featuremap

Once the offline initialization stage is finished the onlinestage follows that is new measurements are obtained toassess the condition of the monitored machine To giverobustness to the novelty detection and avoid false alarms abatch consisting of 30 measurements is evaluated if 75 of


0

005

01

015

02

minus4

minus3

minus2

minus1

0

1

2

3

4

5

60 2 4 8 10minus2minus4LS feature 1 RMS Y

KnownF1

LS fe

atur

e 2 k

urto

sis Y

Figure 6 Evaluation of the fault scenario 1198651 The novelty model istrained employing data from healthy operation condition

minus2 0 2 4 6 8 10minus4LS feature 1 RMS Y

minus4

minus3

minus2

minus1

0

1

2

3

4

5

minus001

0

001

002

003

004

005

006

007

008

Known

LS fe

atur

e 2 k

urto

sis Y

Figure 7 Contour plot of the novelty model after including 1198651

the analyzed measurements are evaluated as novelty then thealarm is triggered Next the 1198651 scenario is presented to testthe performance of the model The plot of the scenario andthe novelty threshold obtained during training is presentedin Figure 6

As can be appreciated the new scenario lays outside thenovelty threshold so it is successfully detected as noveltyOnce a novel scenario is detected and identified as a fault bythe user it is incorporated to the database to consider it as partof the known scenarios and the novelty model is retrainedwithout changing the features to include this informationFigure 7 shows the feature space after the novelty modelis trained using healthy and 1198651 data as part of the knownscenarios

Once the model is retrained the third scenario 1198652 isintroduced The visual representation of the test is presentedin Figure 8 this scenario is not detected as novel because only50 of the measurements are labeled as novel by the model

minus001

0

001

002

003

004

005

006

007

008


minus4

minus3

minus2

minus1

0

1

2

3

4

5

KnownF2

LS fe

atur

e 2 k

urto

sis Y

Figure 8 Evaluation of the fault scenario 1198652 The novelty model istrained employing data from healthy and 1198651 scenarios

A similar result is obtained when the novelty model istrained using healthy and 1198652 data and is tested with data of1198651 A summary of the results of novelty detectionmaintainingthe same features obtained during initialization is shown inTable 1

As it can be seen in Table 1 using the reduced set obtainedduring initialization is easy to detect the novel scenarioswhen there is only information of the healthy condition butwhen a novel scenario is included in the database and thenoveltymodel is retrained the reduced set of features initiallyobtained does not necessarily provide a good representationto detect a new scenario during test

To improve these results the methodology presented inthis work proposes to evaluate again the feature reductionmodule each time a retraining is applied Following theoutline presented for results and parting from the first retrainwhere the scenario 1198651 is included in the database the featurereduction module is applied again but this time includinginformation from healthy condition and 1198651 scenario Sincetwo scenarios are taken into consideration and the labels areknown a supervised approach can be applied in this case aFisher Score ranking for feature reduction is employed Thenovelty model using the initial set of features and the noveltymodel after retraining with the new reduced set of featuresare shown in Figure 9

The new reduced set of features is composed of the RMSof the 119910-axis and the Kurtosis of the 119911-axis as can be appre-ciated the new set of features present a more discriminativedistribution of the three scenarios considered At the laststep (Figure 9(d)) when 1198652 scenario is included the FisherScore still ranked the same features of the last retraining(Figure 9(b)) as the highest yet it would be possible thata different set of features would be obtained The resultsobtained demonstrate the advantages of including the FisherScore reduction module to the retraining procedure the newdistribution of the scenarios avoided an overlapping of the



minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus2 0 2 4 6 8 10minus4FS feature 1 RMS Y

minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)

Figure 9 Process of evaluation and retraining employing the methodology proposed (a) Evaluation of the fault scenario 1198651 (b) Retrainingof the novelty model and reformulation of the reduced set of features including 1198651 (c) Evaluation of the fault scenario 1198652 (d) Retraining ofthe novelty model and reformulation of the reduced set of features including 1198652Table 1 Performance of the novelty detection using only healthy class data to reduce the number of features where DR stands fordimensionality reduction Different scenarios are included according to the information available to train and test the novelty model

DRPerformance using OC-SVM ()

Known Test Known Test Known Test119867 1198651 + 1198652 119867 + 1198651 1198652 119867 + 1198652 1198651LS 954 (plusmn11) 576 (plusmn67) 435 (plusmn64)

new scenario under test therefore the noveltymodel success-fully detected as novelty all measurements corresponding tothe 1198652 scenario

It is worth mentioning that if initially the scenario 1198652is used for training and the scenario 1198651 is used for testingthe new set of features obtained could be different from theaforementionedThe results achieved fromboth scenarios are

shown in Table 2 which also includes a comparison of theresults obtained employing PCA LDA and Laplacian Scoredimensionality reduction techniques in the retraining stepinstead of the Fisher Score proposed

Regarding the classical feature extraction techniquesPCA and LDA the test scenario is not identified as novelin both cases The Fisher Score and the Laplacian Score


Table 2 Performance of the novelty detection employing a reduc-tion of features during retraining Different scenarios are includedaccording to the information available to train and test the noveltymodel


Known Test Known Test119867 + 1198651 1198652 119867 + 1198652 1198651PCA 590 (plusmn52) 642 (plusmn51)LDA 474 (plusmn61) 512 (plusmn61)LS 843 (plusmn24) 100 (plusmn00)Fisher S 100 (plusmn00) 100 (plusmn00)successfully identified the new scenarios as novel since thepercentage obtained in both cases is higher than the 75 pre-defined threshold in the methodology to activate the alarmBoth techniques achieved high scores but still the FisherScore provided a more appropriate selection of features

As mentioned in Section 2 LDA is a feature extractiontechnique based on the Fisher discriminant coefficient sosimilar results between Fisher Score ranking and LDA areexpected however employing Fisher Score ranking scoreachieved a better result this can be caused because the testconsists of novel scenarios and LDA finds the directions onthe feature space specialized for the two supervised scenariosemployed during training meanwhile Fisher Score rankingprovides a more general approach by selecting features

To test the robustness of the Laplacian Score and FisherScore approaches a comparative test is performed where theset of features is increased from 10 to 15 and by varyingthe number of reduced features selected The five featuresincluded in the original set are obtained from the119909-axis of theaccelerometer monitored where the features calculated arepresented in the aforementioned statistical time-frequencyfeatures these features were discarded initially because theyare not part of the perpendicular plane of the motor anddo not contribute significantly to the monitoring in factincluding these features could affect the performance of thefeature reduction and novelty detection modules The resultsare presented in Table 3

The features obtained from the Fisher Score still present abetter distribution to detect the new scenariosThe LaplacianScore performance is affected when irrelevant features areincluded in the feature set but increases when more dimen-sions are taken into consideration Since the objective of theFisher Score is not topology preservation contrary to LS itis capable of discarding all the irrelevant features that wereincluded and the performance is not affected

5 Conclusions

This work proposes an adaptive novelty detection methodol-ogy based on vibration analysis for the condition monitoringand diagnosis of the components of a kinematic chain Themethodology is based on the acquisition of vibration signalsthat are generated in the kinematic chain along with anadequate signal processing to extract features to characterizethe components and an adaptive novelty detection model to

Table 3 Performance of the novelty detection increasing the num-ber of initial features from 10 to 15 and varying the number of thereduced set of features

DRPerformance varying the number of features

reduced ()2 3 4

LS 590 (plusmn41) 643 (plusmn33) 886 (plusmn25)Fisher S 100 (plusmn00) 100 (plusmn00) 100 (plusmn00)detect anomalies The method is composed of two sequentialstages an offline stage to initialize and retrain the modulesand an online stage to continuously assess the conditionof the machine During initialization the model is trainedemploying only information from the machine workingunder healthy condition and two additional faulty scenariosare introduced to test the performance of the method underunknown operations

The adaptive novelty detection approach successfullydetected both novel scenarios and the model incorporatedthe information to avoid generating alarms if the same faultis detected A comparison between the proposed methodand classical dimensionality reduction approaches highlightsthe limitations of maintaining a static set of features duringmonitoring instead of reformulating the feature reductionmodule once new information is available In this particularstudy employing features reduced by Laplacian Score andFisher Score obtained similar results nevertheless it doesnot imply that a similar outcome will be present during theanalysis of other faults Fisher Score is encouraged to beemployed in this methodology rather than maintaining theLaplacian Score approach due to the similarity between themethod objective and the objective function of Fisher Scorein both the ideal case is to find the features to maximizethe distance between scenarios while maintaining compactclusters A specific comparison of performances between LSand Fisher Score was also included in which the Fisher Scoreobtained better results when irrelevant features are includedin the original set of features and when the dimensionalityof the reduced set is increased this highlights the advantagesand robustness of the feature selection approach by FisherScore compared to the LS

From an industrial perspective two aspects should bementioned First dealing with an identical operating sce-nario (ie healthy) the variation of the working conditions(speed and torque patterns) will affect the resulting valueof the estimated numerical features and consequently itsrepresentation in the feature space used for novelty modeldesign Thus the introduction of the working condition asan additional degree of freedom in the novelty detectionscheme implies the increase of the dimensional space thatthe novelty detectionmodel is managingThis fact implies anincrease of the risk of data overlapping (different scenariosunder different operating conditions showing similar featurevalues) and also an increase of complexity for boundariesdefinition during the novelty modelling (data spread) Thusthe proposed method is not limited to stationary conditionsbut to the repetition of the pattern of working condition used


to train the model Indeed in practical industrial applicationwhere cyclostationary processes take place the proposedmethod should be executed during the same period (the sameworking conditions) in order to reduce data variability andincrease novelty performance In case of multiple workingconditions consideration parallel novelty model structurescould be generated in order to activate each one dependingon the corresponding working condition

Second the integration of additional physicalmagnitudesas stator currents temperatures and others represents atrendy approach in electromechanical fault diagnosis inorder to improve fault diagnosis resolution and perfor-mance The corresponding data fusion scheme implies ahigh-dimensional feature space that is multiple numericalfeatures estimated from multiple physical magnitudes con-sidered In this point it must be noted that the presentedmethodology proposes the reevaluation of the feature reduc-tionmodule each time a retraining is appliedThus the largerthe available pool of features is the higher the potentiality ofthe proposedmethod to increase the performance during thedata characterization is

Also the proposed method can be extended andimproved for further development this improvement couldinclude a diagnosis method to not only detect anomalies inthe kinematic chain but also identify the fault causing theabnormal behavior In this sense future work will include aspecific comparison of broken bar fault detection capabilitiesbetween the proposedmethodology and classical approaches

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

This research was partially supported by CONACyT schol-arships Grants 278033 and 313604 and also by the SpanishMinistry of Education Culture and Sport under GrantFPU1300589 The authors would also like to thank a partialfinancial support by the Generalitat de Catalunya (GRCMCIA Grant no SGR 2014-101)

References

[1] H Henao G-A Capolino M Fernandez-Cabanas et alldquoTrends in fault diagnosis for electrical machines a review ofdiagnostic techniquesrdquo IEEE Industrial Electronics Magazinevol 8 no 2 pp 31ndash42 2014

[2] N Bayar S Darmoul S Hajri-Gabouj and H Pierreval ldquoFaultdetection diagnosis and recovery using Artificial Immune Sys-tems a reviewrdquoEngineeringApplications of Artificial Intelligencevol 46 pp 43ndash57 2015

[3] SWang JWan D Zhang D Li and C Zhang ldquoTowards smartfactory for Industry 40 a self-organized multi-agent systemwith big data based feedback and coordinationrdquo ComputerNetworks vol 101 pp 158ndash168 2016

[4] M A F Pimentel D A Clifton L Clifton and L TarassenkoldquoA review of novelty detectionrdquo Signal Processing vol 99 pp215ndash249 2014

[5] R Langone O Mauricio Agudelo B De Moor and J AK Suykens ldquoIncremental kernel spectral clustering for online

learning of non-stationary datardquo Neurocomputing vol 139 pp246ndash260 2014

[6] M-H Chen P-C Chang and J-L Wu ldquoA population-basedincremental learning approach with artificial immune systemfor network intrusion detectionrdquo Engineering Applications ofArtificial Intelligence vol 51 pp 171ndash181 2016

[7] S Khomfoi and L M Tolbert ldquoFault diagnosis and reconfigu-ration for multilevel inverter drive using AI-based techniquesrdquoIEEE Transactions on Industrial Electronics vol 54 no 6 pp2954ndash2968 2007

[8] A Bellini F Filippetti C Tassoni and G-A CapolinoldquoAdvances in diagnostic techniques for induction machinesrdquoIEEE Transactions on Industrial Electronics vol 55 no 12 pp4109ndash4126 2008

[9] V T Tran and B-S Yang ldquoAn intelligent condition-basedmaintenance platform for rotating machineryrdquo Expert Systemswith Applications vol 39 no 3 pp 2977ndash2988 2012

[10] P A Delgado-Arredondo A Garcia-Perez D Morinigo-Soteloet al ldquoComparative study of time-frequency decompositiontechniques for fault detection in induction motors using vibra-tion analysis during startup transientrdquo Shock and Vibration vol2015 Article ID 708034 14 pages 2015

[11] M D Prieto G Cirrincione A G Espinosa J A Ortegaand H Henao ldquoBearing fault detection by a novel condition-monitoring scheme based on statistical-time features and neu-ral networksrdquo IEEE Transactions on Industrial Electronics vol60 no 8 pp 3398ndash3407 2013

[12] P Perner ldquoConcepts for novelty detection and handling basedon a case-based reasoning process schemerdquo Engineering Appli-cations of Artificial Intelligence vol 22 no 1 pp 86ndash91 2009

[13] C Englund and A Verikas ldquoA hybrid approach to outlier detec-tion in the offset lithographic printing processrdquo EngineeringApplications of Artificial Intelligence vol 18 no 6 pp 759ndash7682005

[14] H Garces and D Sbarbaro ldquoOutliers detection in environmen-tal monitoring databasesrdquo Engineering Applications of ArtificialIntelligence vol 24 no 2 pp 341ndash349 2011

[15] S Ma S M Li and Y P Xiong ldquoUncertainty reduced noveltydetection approach applied to rotatingmachinery for conditionmonitoringrdquo Shock and Vibration vol 2015 Article ID 73721310 pages 2015

[16] J McBain and M Timusk ldquoFeature extraction for noveltydetection as applied to fault detection in machineryrdquo PatternRecognition Letters vol 32 no 7 pp 1054ndash1061 2011

[17] S Wang J Yub E Lapirac and J Lee ldquoA modified supportvector data description based novelty detection approach formachinery componentsrdquo Applied Soft Computing vol 13 no 2pp 1193ndash1205 2013

[18] D P Filev R B Chinnam F Tseng and P Baruah ldquoAn indus-trial strength novelty detection framework for autonomousequipment monitoring and diagnosticsrdquo IEEE Transactions onIndustrial Informatics vol 6 no 4 pp 767ndash779 2010

[19] B S J Costa P P Angelov and L A Guedes ldquoFully unsu-pervised fault detection and identification based on recursivedensity estimation and self-evolving cloud-based classifierrdquoNeurocomputing vol 150 pp 289ndash303 2015

[20] B Scholkopf R C Williamson A J Smola J Shawe-Taylorand J C Platt ldquoSupport vector method for novelty detectionrdquoin Neural Information Processing Systems vol 12 pp 582ndash5881999


[21] D M J Tax and R P W Duin ldquoSupport vector domaindescriptionrdquo Pattern Recognition Letters vol 20 no 11ndash13 pp1191ndash1199 1999

[22] S D Villalba and P Cunningham ldquoAn evaluation of dimensionreduction techniques for one-class classificationrdquo ArtificialIntelligence Review vol 27 no 4 pp 273ndash294 2007

[23] I T Jolliffe Principal Component Analysis Springer New YorkNY USA 2nd edition 1986

[24] X He D Cai and P Niyogi ldquoLaplacian score for featureselectionrdquo inAdvances in Neural Information Processing Systems18 pp 507ndash514 MIT Press 2005

[25] A J Izenman ldquoLinear discriminant analysisrdquo in Modern Mul-tivariate Statistical Techniques A J Izenman Ed pp 237ndash280Springer New York NY USA 2008

[26] Q Gu Z Li and J Han ldquoGeneralized fisher score for featureselectionrdquo httpsarxivorgabs12023725

[27] Z Gao C Cecati and S X Ding ldquoA survey of fault diagnosisand fault-tolerant techniquesmdashpart II fault diagnosis withknowledge-based and hybridactive approachesrdquo IEEE Trans-actions on Industrial Electronics vol 62 no 6 pp 3768ndash37742015

[28] A E Lazzaretti D M J Tax H Vieira Neto and V HFerreira ldquoNovelty detection and multi-class classification inpower distribution voltage waveformsrdquo Expert Systems withApplications vol 45 pp 322ndash330 2016

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of



RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



Unexpected events in the form of faulty scenarios notinitially considered or deviations over the normal operationtake place during the useful life of the machinery These situ-ations must be identified in order to avoid misclassificationsand the consequent incorrect maintenance action To copewith these challenging scenarios the condition-based main-tenance methods must be able to work under the assumptionthat only information of the healthy condition of themachineis available [4] Another relevant characteristic is the capacityto incorporate new knowledge Since new faults could appearduring monitoring the condition-based maintenance meth-ods must also be able to adapt and include new scenarios totheir initial knowledge [5 6]

These challenging scenarios are not always taken intoconsideration in conditionmonitoringmethodologies there-fore the application of classical methods on an industrialenvironment is limited or nonviable

In the field of electromechanical system monitoring agreat deal of health monitoring schemes based on vibrationshave been proposed during the last years [7ndash10] Mainlythese proposals can be divided into two categories regardingthe available information of the monitored machine and theclassification objective

The first category multifault detection consists in thedevelopment of a monitoring system capable of identifyingmultiple faults previously known [1] which implies that theinformation available to train themonitoring system includesmeasurements of the healthy state of the machine but alsomeasurements of the machine working under different faultyconditions Over the last years instead of looking for highlyspecific features for a component fault a trend towardscalculation of more general statistical features and the fusionof information to enhance the performance of the diagnosissystem has been carried out [11] Yet this approach is mostlycarried out at the laboratory scale with successful resultswhere controlled faults are provoked on the machine

However the information of the possible faults that couldappear on a machine is not always available consequentlya different approach must be considered if only informationof healthy condition of the machine is available The secondcategory corresponds to novelty detection also known asanomaly detection and it represents a solution to cope withthe lack of information available of the monitored machineThe objective of novelty detection is to detect novel eventsthat differ in some manner from the information of whichthemodelwas trainedNovelty detection framework has beensuccessfully applied to network intrusion medical diagnosisimage segmentation and handwritten digit recognition faultdetection and condition monitoring [4 6 12ndash14] Specialeffort has been made to improve condition monitoringnovelty detection applications based on vibrations Ma et alproposed a novelty detection approach to rotating machin-ery by means of monitoring the thresholds obtained froman extreme value distribution [15] McBain and Timusk pro-posed a feature extractionmethodology for novelty detectionapplied to a rotating mechanical system [16]

Most of the methodologies for novelty detection arelimited to a static analysis and the incorporation of the novelinformation to the novelty detection system is not usually

considered An approach to include adaptability to thenovelty framework based on vibrations is proposed byWanget al [17] The proposed monitoring scheme includes testingdata on the boundary of the novelty model based on SupportVector Data Description and retrains the model with thisinformation to gain robustness Nevertheless this approachdoes not take into consideration the possibility to includenovel scenarios during the monitored process

Other approaches to develop an adapting conditionmon-itoring schemewere presented by Filev et al where a practicalframework for autonomous monitoring of industrial equip-ment based on novelty detection is analyzed [18] and byCostaet al where a two-stage algorithm for real-time fault detectionand identification is presented [19] Both approaches providethe opportunity to incorporate novel detected faults tothe monitoring system nevertheless in both methods theincorporation is limited to updating the known data basebut an adaptation of the numerical features analyzed is notconsidered

The performance of a novelty detection system is highlydependent on the numerical features consideredWhen thereis no previous information of the possible faults that canoccur the application of a suitable numerical features analysisstrategy represents a critical challenge [4] Considering acontinuous monitoring framework where the initial infor-mation available is the healthy operating condition andlater on different faults are identified progressively when themachine condition deteriorates all the approaches previouslydiscussed do not modify the initial set of features when newinformation of faults is incorporated This static approachhas the advantage of providing a most adequate situation foronline adaptation nevertheless analyzing the informationof the faults detected during the monitoring process couldimprove the identification of amost adequate set of features todiscriminate the possible upcoming or already detected faultscenarios

To improve the performance of the novelty detection taska dynamic approach for reselection of the features includinginformation of new faults detected is proposed on this workTo the extent of the authorsrsquo knowledge the study regardingthe dynamic change of the feature space by considering newoperating scenarios is novel in the applications to electrome-chanical systems for fault diagnosis

Thereby the contribution of this study is to provide anadaptive methodology for novelty detection where the infor-mation of identified faults during the monitoring process isexploited to improve the anomaly detection task The meth-odology is divided into two recursive stages first an offlinestage for initialization and retraining of the feature reductionand novelty detectionmodules second an onlinemonitoringstage to continuously assess the condition of the machineThe novelty model employed is a one-class Support Vectormachine (OC-SVM) [20] which is a domain based modelsuccessfully employed on vibration signals monitoring [4]Contrary to classical static feature reduction approaches theproposed method reformulates the features by employingfirst a Laplacian Score ranking and then the Fisher Scoreranking for retraining







min w22 + 1119873V

119873sum119894=1


(1)


119891 (119909) = w sdot Φ (119909) minus 120588 (2)


119870(119909119894 119909119895) = exp(minus10038171003817100381710038171003817119909119894 minus 11990911989510038171003817100381710038171003817221205902 ) (3)


119871 (w 120585 120588 a120573) = w22 + 1119873V

119873sum119894=1

120585119894 minus 120588minus 119873sum119894=1


120573119894120585119894(4)


w = 119873sum119894=1

119886119894Φ(119909119894) 119886119894 = 1119873V

minus 120573119894119873sum119894=1

119886119894 = 1(5)


min a119879119867a


119886119894 = 1(6)


119867119894119895 = 119870 (119909119894 119909119895) = Φ (119909119894) sdot Φ (119909119895) (7)


120588 = 1119899119904119899119904sum119894=1

119873sum119895=1

119886119895119870(119909119894 119909119895) a (8)










3 Methodology









Feature reduction


New measurement


NoveltyUser

supervisionKnown


Offl

ine

Onl

ine

RetrainingIncluding


No

Yes

Noveltymodel



Max Value

119894 = max (119909) (9)



(119909119896)2 (10)

Crest Factor (CF)

CF = 119909119896RMS

(11)

Kurtosis

119896 = 119864 [(119909119896 minus 119909)4]1205904 (12)

Shape Factor

SF = RMS1119899 sdot sum119899119896=1 10038161003816100381610038161199091198961003816100381610038161003816 (13)










New Fault Detected



No

Yes

Perform New Feature

Reduction



used



included

Retr

aini

ng

New faultdetected



Perform new feature

reduction






Gearbox

Load

VFD

DAS

Inductionmotor


PC

z

yx







(a) (b)






Healthy

0

005

01

015

02


minus3

minus2

minus1

0

1

2

3LS

feat

ure 2

kur

tosis

Y





0

005

01

015

02

minus4

minus3

minus2

minus1

0

1

2

3

4

5


KnownF1

LS fe

atur

e 2 k

urto

sis Y



minus4

minus3

minus2

minus1

0

1

2

3

4

5

minus001

0

001

002

003

004

005

006

007

008

Known

LS fe

atur

e 2 k

urto

sis Y





minus001

0

001

002

003

004

005

006

007

008


minus4

minus3

minus2

minus1

0

1

2

3

4

5

KnownF2

LS fe

atur

e 2 k

urto

sis Y








minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3


minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















min w22 + 1119873V

119873sum119894=1


(1)


119891 (119909) = w sdot Φ (119909) minus 120588 (2)


119870(119909119894 119909119895) = exp(minus10038171003817100381710038171003817119909119894 minus 11990911989510038171003817100381710038171003817221205902 ) (3)


119871 (w 120585 120588 a120573) = w22 + 1119873V

119873sum119894=1

120585119894 minus 120588minus 119873sum119894=1


120573119894120585119894(4)


w = 119873sum119894=1

119886119894Φ(119909119894) 119886119894 = 1119873V

minus 120573119894119873sum119894=1

119886119894 = 1(5)


min a119879119867a


119886119894 = 1(6)


119867119894119895 = 119870 (119909119894 119909119895) = Φ (119909119894) sdot Φ (119909119895) (7)


120588 = 1119899119904119899119904sum119894=1

119873sum119895=1

119886119895119870(119909119894 119909119895) a (8)










3 Methodology









Feature reduction


New measurement


NoveltyUser

supervisionKnown


Offl

ine

Onl

ine

RetrainingIncluding


No

Yes

Noveltymodel



Max Value

119894 = max (119909) (9)



(119909119896)2 (10)

Crest Factor (CF)

CF = 119909119896RMS

(11)

Kurtosis

119896 = 119864 [(119909119896 minus 119909)4]1205904 (12)

Shape Factor

SF = RMS1119899 sdot sum119899119896=1 10038161003816100381610038161199091198961003816100381610038161003816 (13)










New Fault Detected



No

Yes

Perform New Feature

Reduction



used



included

Retr

aini

ng

New faultdetected



Perform new feature

reduction






Gearbox

Load

VFD

DAS

Inductionmotor


PC

z

yx







(a) (b)






Healthy

0

005

01

015

02


minus3

minus2

minus1

0

1

2

3LS

feat

ure 2

kur

tosis

Y





0

005

01

015

02

minus4

minus3

minus2

minus1

0

1

2

3

4

5


KnownF1

LS fe

atur

e 2 k

urto

sis Y



minus4

minus3

minus2

minus1

0

1

2

3

4

5

minus001

0

001

002

003

004

005

006

007

008

Known

LS fe

atur

e 2 k

urto

sis Y





minus001

0

001

002

003

004

005

006

007

008


minus4

minus3

minus2

minus1

0

1

2

3

4

5

KnownF2

LS fe

atur

e 2 k

urto

sis Y








minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3


minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation

















3 Methodology









Feature reduction


New measurement


NoveltyUser

supervisionKnown


Offl

ine

Onl

ine

RetrainingIncluding


No

Yes

Noveltymodel



Max Value

119894 = max (119909) (9)



(119909119896)2 (10)

Crest Factor (CF)

CF = 119909119896RMS

(11)

Kurtosis

119896 = 119864 [(119909119896 minus 119909)4]1205904 (12)

Shape Factor

SF = RMS1119899 sdot sum119899119896=1 10038161003816100381610038161199091198961003816100381610038161003816 (13)










New Fault Detected



No

Yes

Perform New Feature

Reduction



used



included

Retr

aini

ng

New faultdetected



Perform new feature

reduction






Gearbox

Load

VFD

DAS

Inductionmotor


PC

z

yx







(a) (b)






Healthy

0

005

01

015

02


minus3

minus2

minus1

0

1

2

3LS

feat

ure 2

kur

tosis

Y





0

005

01

015

02

minus4

minus3

minus2

minus1

0

1

2

3

4

5


KnownF1

LS fe

atur

e 2 k

urto

sis Y



minus4

minus3

minus2

minus1

0

1

2

3

4

5

minus001

0

001

002

003

004

005

006

007

008

Known

LS fe

atur

e 2 k

urto

sis Y





minus001

0

001

002

003

004

005

006

007

008


minus4

minus3

minus2

minus1

0

1

2

3

4

5

KnownF2

LS fe

atur

e 2 k

urto

sis Y








minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3


minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Feature reduction


New measurement


NoveltyUser

supervisionKnown


Offl

ine

Onl

ine

RetrainingIncluding


No

Yes

Noveltymodel



Max Value

119894 = max (119909) (9)



(119909119896)2 (10)

Crest Factor (CF)

CF = 119909119896RMS

(11)

Kurtosis

119896 = 119864 [(119909119896 minus 119909)4]1205904 (12)

Shape Factor

SF = RMS1119899 sdot sum119899119896=1 10038161003816100381610038161199091198961003816100381610038161003816 (13)










New Fault Detected



No

Yes

Perform New Feature

Reduction



used



included

Retr

aini

ng

New faultdetected



Perform new feature

reduction






Gearbox

Load

VFD

DAS

Inductionmotor


PC

z

yx







(a) (b)






Healthy

0

005

01

015

02


minus3

minus2

minus1

0

1

2

3LS

feat

ure 2

kur

tosis

Y





0

005

01

015

02

minus4

minus3

minus2

minus1

0

1

2

3

4

5


KnownF1

LS fe

atur

e 2 k

urto

sis Y



minus4

minus3

minus2

minus1

0

1

2

3

4

5

minus001

0

001

002

003

004

005

006

007

008

Known

LS fe

atur

e 2 k

urto

sis Y





minus001

0

001

002

003

004

005

006

007

008


minus4

minus3

minus2

minus1

0

1

2

3

4

5

KnownF2

LS fe

atur

e 2 k

urto

sis Y








minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3


minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










New Fault Detected



No

Yes

Perform New Feature

Reduction



used



included

Retr

aini

ng

New faultdetected



Perform new feature

reduction






Gearbox

Load

VFD

DAS

Inductionmotor


PC

z

yx







(a) (b)






Healthy

0

005

01

015

02


minus3

minus2

minus1

0

1

2

3LS

feat

ure 2

kur

tosis

Y





0

005

01

015

02

minus4

minus3

minus2

minus1

0

1

2

3

4

5


KnownF1

LS fe

atur

e 2 k

urto

sis Y



minus4

minus3

minus2

minus1

0

1

2

3

4

5

minus001

0

001

002

003

004

005

006

007

008

Known

LS fe

atur

e 2 k

urto

sis Y





minus001

0

001

002

003

004

005

006

007

008


minus4

minus3

minus2

minus1

0

1

2

3

4

5

KnownF2

LS fe

atur

e 2 k

urto

sis Y








minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3


minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










(a) (b)






Healthy

0

005

01

015

02


minus3

minus2

minus1

0

1

2

3LS

feat

ure 2

kur

tosis

Y





0

005

01

015

02

minus4

minus3

minus2

minus1

0

1

2

3

4

5


KnownF1

LS fe

atur

e 2 k

urto

sis Y



minus4

minus3

minus2

minus1

0

1

2

3

4

5

minus001

0

001

002

003

004

005

006

007

008

Known

LS fe

atur

e 2 k

urto

sis Y





minus001

0

001

002

003

004

005

006

007

008


minus4

minus3

minus2

minus1

0

1

2

3

4

5

KnownF2

LS fe

atur

e 2 k

urto

sis Y








minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3


minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










0

005

01

015

02

minus4

minus3

minus2

minus1

0

1

2

3

4

5


KnownF1

LS fe

atur

e 2 k

urto

sis Y



minus4

minus3

minus2

minus1

0

1

2

3

4

5

minus001

0

001

002

003

004

005

006

007

008

Known

LS fe

atur

e 2 k

urto

sis Y





minus001

0

001

002

003

004

005

006

007

008


minus4

minus3

minus2

minus1

0

1

2

3

4

5

KnownF2

LS fe

atur

e 2 k

urto

sis Y








minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3


minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











minus4

minus3

minus2

minus1

0

1

2

3

4

5

0

005

01

015

02

KnownF1

LS fe

atur

e 2 k

urto

sis Y

(a)

minus5

minus4

minus3

minus2

minus1

0

1

2

3


minus001

0

001

002

003

004

005

006

007

008

Known

FS fe

atur

e 2 k

urto

sis Z

(b)

minus001

0

001

002

003

004

005

006

007

008


FS fe

atur

e 2 k

urto

sis Z

minus5

minus4

minus3

minus2

minus1

0

1

2

3

KnownF2

(c)


minus5

minus4

minus3

minus2

minus1

0

1

2

3

minus0005

0

0005

001

0015

002

0025

003

0035

Known

FS fe

atur

e 2 k

urto

sis Z

(d)















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation
















5 Conclusions




reduced ()2 3 4








Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation













Competing Interests


Acknowledgments


References

































RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation




















RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











RoboticsJournal of





Journal of



RotatingMachinery





VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









Documents

Research Article Vibration-Based Adaptive Novelty ...downloads.hindawi.com/journals/sv/2016/2417856.pdf · applied under a continuous learning framework. e structure of the method