componential coding in the condition monitoring of electrical...

Componential coding in the condition monitoring ofelectrical machinesPart 2: application to a conventional machine and anovel machine

B S Payne1, F Gu1, C J S Webber2 and A D Ball1*1Maintenance Engineering Research Group, Manchester School of Engineering, University of Manchester, UK2QinetiQ, Malvern, UK

Abstract: This paper (Part 2) presents the practical application of componential coding, theprinciples of which were described in the accompanying Part 1 paper. Four major issues areaddressed, including optimization of the neural network, assessment of the anomaly detection results,development of diagnostic approaches (based on the reconstruction error) and also benchmarking ofcomponential coding with other techniques (including waveform measures, Fourier-based signalreconstruction and principal component analysis). This is achieved by applying componential codingto the data monitored from both a conventional induction motor and from a novel transverse uxmotor. The results reveal that machine condition monitoring using componential coding is not onlycapable of detecting and then diagnosing anomalies but it also outperforms other conventionaltechniques in that it is able to separate very small and localized anomalies.

Keywords: neural network, componential coding, auto-encoder, condition monitoring, inductionmotor, transverse ux motor, fault detection, fault diagnosis

NOTATION

ADI average discrimination indexICAN Independent Channel Architecture

NetworkIR inner root (of transverse ux motor stator

core)IT inner tip (of transverse ux motor stator

core)JCAN Joint Channel Architecture NetworkOR outer root (of transverse ux motor stator

core)OT outer tip (of transverse ux motor stator

core)VDI variance discrimination index

1 INTRODUCTION

Electric motors are at the core of most engineeringprocesses. In general they are inherently reliable(because of their mechanical simplicity) and requirevery little attention, except at infrequent intervals whenthe plant is shut down for inspection. Nevertheless, likeany electromechanical system, electric motors do fail.Unanticipated failure can lead to very costly downtimeand consequential damage [1, 2]. Hence early andaccurate fault detection and diagnosis through conditionmonitoring is of key importance. Within the last decade,much effort has been placed on research and develop-ment in this eld.

The interpretation of vibration and per-phase currentdata is by far the most widely developed means ofdetermining incipient motor faults. In particular, casingvibration monitoring is well established and is capableof detecting a wide range of mechanical motor problems(such as shaft eccentricity/misalignment [1, 3, 4], bearingfaults [1, 5], looseness [5, 6] and rotor imbalance [7]) inaddition to broken rotor bars (which, strictly speaking,are considered as an electrical fault) [1, 2, 8, 9]. Per-phase current measurement and analysis is an approach

The MS was received on 19 November 2002 and was accepted afterrevision for publication on 22 M ay 2003.* Corresponding author: M aintenance Engineering Research Group,M anchester School of Engineering, University of M anchester, OxfordRoad, M anchester M 13 9PL, UK.

901

C15302 Proc. Instn Mech. Engrs Vol. 217 Part C: J. Mechanical Engineering Science

that has more recently been capitalized for motormonitoring. Several researchers and industrial practi-tioners have used per-phase current to detect anddiagnose eccentricity [3, 10], broken rotor bars [11–13],phase imbalance [14] and bearing problems [15, 16].

Analysis of airborne acoustic data is a particularlynew method that has been investigated to detect anddiagnose loose stator coils [17], phase imbalance [14],misalignment [14], broken rotor bars [14] and bearingfaults [18]. The references detail that in some casesacoustic monitoring has the potential to outperformvibration-based approaches in the early detection ofincipient faults. In addition, transducer set-up is easierfor acoustic monitoring than for vibration monitoring,and transducers (microphones) are also non-contactingand thus reduce concerns related to overheating andattachment [14].

Another method used for motor condition monitoringis ux measurement. F lux in the air gap between therotor and stator is directly related to the operationalcharacteristics of an electric machine and hence is likelyto produce accurate indications of electrically relatedfault conditions. Albright [19] placed search coils tomeasure changes in the air gap ux and hence to detectbroken rotor bars. A discussion is also made by Tavnerand Penman [2] and Penman et al. [20] on thepossibilities of using axial ux measurement to detectconditions such as eccentricity and phase imbalance.However, from a practical perspective these approachescan be dif cult to implement [21] and, therefore, havenot widely been applied.

Clearly a range of measurement parameters exists forthe condition monitoring of electrical machines. How-ever, measurement of data alone does not ful lcondition monitoring objectives; the crux of conditionmonitoring is the extraction of useful condition-indicat-ing information from the measured data.

Spectral analysis of data signal frequency componentsis one of the most commonly used methods forinterpretation of the health of a machine [22]. F requencyspectra can be obtained from hand-held data collectors,dedicated signal analysers or ready-written routines insoftware packages such as MATLAB1 . Interpretationof many spectral features (such as fundamentalfrequencies, sidebands and harmonics) can be relativelystraightforward, permitting close correlation with phy-sical machine characteristics.

With the rapid increase in computational capabilitythat has been experienced in the last decade, moreadvanced data processing techniques have becomeviable in the application to machinery conditionmonitoring. As cited in section 1 of Part 1 of this paper[23], these include time–frequency analysis, timescaleanalysis, higher-order statistical analysis and indepen-dent component analysis.

Arti cial neural networks have also been widelystudied for fault detection and diagnosis purposes.

They are often applied as a post-processing tool, withthe input variables having been extracted by moreconventional signal processing techniques. For example,Chow and Yee [24] present several applications ofneural networks to incipient fault detection in rotatingmachines (including induction motors) by using para-meters derived from per-phase power supply current asinput variables. F ilippetti [25] proposed a method thatadopts both induction motor slip and phase current asthe network input variables. Paya et al. [26] describe amethod for detecting motor bearing faults usingvibration data that have been pre-processed by a wavelettransformation. Clearly, the performance of the networkdepends on the choice of input variables and making theappropriate selection can be a less-than-straightforwardtask. A potential solution to this problem is to applyneural networks directly to raw data, thus makingbetter use of the non-linearity bene ts of neural model-ling [27].

In general, even given the advances that have beenmade, the majority of current monitoring techniquesand software often give inaccurate results and are alsopoor at:

1. Handling data collected simultaneously from severalchannels, because most extract fault features on achannel-by-channel basis. This is a disadvantagebecause most machines give a change in multiplemeasurement channels when a fault occurs andcorrelations between these channels can providecrucial detection and diagnosis information.

2. Monitoring novel machines. The reason for this isthat a highly bene cial aid in condition monitoring isprior knowledge, as previous experience often playsan important role in fault detection and diagnosis. Byde nition, for the condition monitoring of a novelmotor, this prior knowledge does not exist.

As explained in Part 1 [23], componential coding modelsdata in an adaptive way and derives signal featuresautomatically using a computationally ef cient algo-rithm. It therefore has the potential to performaccurate fault detection and is applicable to monitor-ing novel machines. In addition, componential codingalso enables simultaneous analysis of multiple datachannels.

To investigate the capabilities of componential codingin fault detection and diagnosis, two lines of investiga-tion were followed. Firstly, an application to noise-contaminated vibration and airborne acoustic data froma conventional induction motor was made. This enabledthe use of componential coding in the conditionmonitoring of a familiar machine, for which conven-tional means of monitoring are well understood, whichhence aided the interpretation of the results. Secondly,componential coding was applied to multichannelmagnetic ux data from a novel electric motor. Thisprovided an opportunity to investigate the application

B S PAYNE, F GU, C J S WEBBER AND A D BALL902

Proc. Instn Mech. Engrs Vol. 217 Part C: J. Mechanical Engineering Science C15302

of componential coding to a previously unseen machine,and for which limited failure mode information wasavailable.

2 APPLICATION TO A CONVENTIONALINDUCTION MOTOR

2.1 Introduction

Componential coding was rst applied to data recordedfrom a 3 kW, three-phase, direct-on-line inductionmotor. Induction machines are the single most commonindustrial prime movers and because they often operatein critical applications, this increases the need foreffective condition monitoring. It is well known thatstator asymmetry faults are of particular concern [28–30] and, therefore, a motor test facility was used to seedvoltage imbalance faults in one phase of the powersupply [14]. Routine vibration data were recorded fromthe motor casing and acoustic measurements were alsotaken (as previous work has demonstrated, this haspotential in this application [14, 17, 31]).

2.2 Seeding of power supply phase imbalance faults

Two main types of imbalance may be found in three-phase supplies: structural and functional. Structuralimbalance occurs because, for economic reasons, thephysical structure of transformers, and especially lines,is not balanced [30]. This leads to imbalanced voltageseven if the generator voltages and load currents areperfectly balanced. The amount of structural imbalancemay also be made worse by poor power management bythe consumer. For example, the single-phase lighting ina large building may be powered unevenly from onephase or power demands may not have been consideredappropriately in the extension of a building.

The second type of imbalance, functional imbalance,is due to uctuations of consumer demand on the three-phase supply. Large machines, such as a large motor oran arc welder [21], are particular causes of this type ofimbalance because of their high current requirements.Alternatively, a bad electrical contact (due to dirt,oxidation or a loose connection) will result in increasedvoltage drop across that contact and thus reduce theproportion of voltage across the motor windings.Current leakage to earth, caused, for example, byinsulation breakdown, may also lead to phase imbal-ance.

Imbalance problems are also compounded by theuse of three-phase variable-speed controllers, as someof the more common controllers on the market onlyhave typical imbalance accuracies of 3 per cent (7.2 Vin the 240 V UK supply). The motor test rig allowed

for controlled imbalances of both 20 V (8.3 per cent)and 40 V (16.7 per cent) to be induced by droppingthe voltage on one phase with a centre-tappedtransformer.

Power supply imbalance to a motor results inadditional losses and overheating—as a rule of thumbthe winding temperature rises by 25 per cent for every3.5 per cent voltage imbalance [28]. F rom the authors’experience, phase imbalance is thus sometimes the causeof apparently random overheating of electric machinesin industry. Temperature rises cause thermal ageing andmake the winding insulation, in particular, morevulnerable to electric, mechanical and environmentalstresses [28]. Voltage imbalance is thus an importantcause to consider in motor failure and there are clearbene ts in detecting its occurrence from measuredmotor parameters.

2.3 Vibration and acoustic data-sets

Componential coding was applied to two channels ofrecorded data. Surface vibration was recorded in aroutine and standard way using an accelerometermounted close to the bearing on the drive-end end-cap(as this is a major vibration transmission path betweenthe rotor and stator and, therefore, provides informa-tion on both machine parts). Airborne acoustics wasrecorded using a microphone pointed towards the motorand placed at 200 mm perpendicularly from the motorcentre height. In contrast to routinely measured vibra-tion data, it is unusual to measure acoustic data,although it is a technique that is receiving increasinginterest.

Data were collected at 75 per cent of the rated motorload, under healthy operation (which will be referred toas case 0) and with the seeding of a 20 V imbalance fault(case 1) and a 40 V imbalance fault (case 2). Thevibration and acoustic data were sampled simulta-neously at 32.5 kHz, giving a usable bandwidth of 0–15 kHz. For each fault case, three data records werecollected—each of 65 536 points and covering more than30 shaft revolutions. This large data record size was usedto ensure that componential coding was trainedsuf ciently and that a reliable model was obtained torepresent the underlying data features. F igure 1 shows aportion of each data record and illustrates that no cleardifferences can be seen between the raw data for thedifferent operational cases.

2.4 Componential coding con guration and optimization

Section 2.1 of Part 1 of the paper [23] describes howcomponential coding may be implemented througheither of two network variants. In summary, the JointChannel Architecture Network (JCAN ) implementation

COMPONENTIAL CODING IN THE CONDITION MONITORING OF ELECTRICAL MACHINES. PART 2 903


can detect anomalous* correlations (including phaserelationships) between simultaneously applied datachannels. Alternatively, the Individual Channel Archi-tecture Network (ICAN ) may be employed in order tomodel the characteristics of individual channels in orderto de-emphasize interchannel correlations. Both theJCAN and the ICAN were studied for the separation ofthe two degrees of severity of voltage imbalance fromthe baseline condition.

Section 1.2 of Part 1 [23] discusses how componentialcoding modelling is achieved by deriving basis vectorsthrough the adaptive training process. The dimension ofthe basis vectors (i.e. the number nsnt of coordinatescomprising a basis vector, using the notation of Part 1[23]) is chosen to cover at least two periods of theobserved principal frequency component in the mon-itored data. This is because at least one complete periodis required to model any cyclic information. Withinreason, the actual dimension of a basis vector is acompromise between computational speed and model-ling accuracy. Previous work has found a basis vectordimension of 64 data points to be a good compromisefor the induction motor data [32]. Signal resampling isnecessary to ensure that each basis vector of dimension64 points covers two periods of the principal frequency(which may be different for different data sources). For

the vibration and acoustic data this was achieved bydesampling by a factor of 2 and 10 respectively.

Componential coding also requires the setting ofits two most in uential non-adaptive parameters:threshold (which determines how many of the neuronsactivate) and the number of basis vectors. Theseparameters were optimized to maximize anomalydiscrimination using an automated genetic algorithm.The nal con gurations for the two network variants,following optimization, are summarized in Table 1. TheICAN, with its inherent independent nature, has aseparate con guration for each type of data, whereasthe JCAN has one con guration for both vibration andacoustics.

The values given in the last two rows are averagediscrimination indices (ADIs) which, as explained insection 2.7 of Part 1 [23], indicate the difference in newlyapplied data-sets from baseline healthy motor operationdata (from case 0). These rows show the differencebetween the model reconstruction and previously unseenhealthy data (from case 0) and the difference betweenthe model reconstruction and data recorded with theintroduction of a 20 V imbalance fault (case 1)respectively.

As expected, the ADI values are close to zero whencomponential coding is applied to previously unseenhealthy data from case 0 (section 2.7 of Part 1 [23]),because it is similar to the data on which the model wastrained. However, when componential coding is appliedto data recorded under faulty operation, the ADI valuesfor each con guration are clearly much larger. Thisindicates that both the ICAN and JCAN are capableof separating the anomalous case 1 from the baseline

Fig. 1 Vibration and acoustic data (normalized amplitudes)

* As described in section 1.1 of Part 1 of this paper [23], an anomaly isany characteristic of an unseen data-set that is different from thecharacteristics of the training data-set. A fault can be thought of as aspecial case of an anomaly in which the anomalous characteristichappens to arise as the result of a real fault in the engineering systembeing monitored.



case 0. In addition, the higher ADI values for the twoseparate ICAN con gurations show that this networkproduces better detection results than the JCAN in thiscontext. This is particularly true for the acoustic dataand, hence, reveals that componential coding is moresensitive to the seeded fault condition when applied tothis form of data. This may be because experience hasshown that a single channel of acoustic data often givesa more comprehensive picture of condition than a singlechannel of surface vibration [33].

It can also be seen in Table 1 that the optimizedthreshold for the ICAN is much higher for the acousticdata than that for the vibration data. The higherthreshold for the acoustic data re ects the fact thatthere is more similarity between input patterns, so thethreshold may be set so high that neurons only activatewhen basis vectors are very similar to the measuredacoustic data. Conversely, there is less similaritybetween input patterns for the vibration data, so asmall threshold is needed to achieve a small modelresidual with baseline control data. In this case, neuronswill activate even if the match between a basis vectorand the vibration data is less good. From an anomalydetection perspective, a higher threshold allows greatersensitivity to smaller changes between the baseline andmonitored data.

2.5 Anomaly detection

So far this paper has shown the ability of both theICAN and JCAN to detect seeded voltage imbalancefaults on the basis of the AD I measure. An alternativemeasure is the variance discrimination index (VDI). Asdiscussed in section 2.7 of Part 1 [23], the ADI is based

on the averaged reconstruction errors over the mon-itored data-set and is, consequently, most useful when arelatively large proportion of the monitored data-set isanomalous. The VDI is based on the variance of thereconstruction errors and is, therefore, more capable ofhighlighting anomalies that manifest themselves onlyover a relatively small portion of data.

Table 2 shows that both the ADI and VDI increasewith severity of fault condition. Therefore both theICAN and the JCAN provide clear discrimination offault severity.

Both the ICAN and the JCAN demonstrate goodfault detection with both the AD I and VDI (althoughthe ICAN provides better detection results in thiscontext). Signi cantly, the acoustic ICAN outperformsthe vibration ICAN. This latter observation suggeststhat the acoustic signal contains more (useful) globalinformation related to the motor operation and thevoltage imbalance problem. The acoustic signal isgenerated not only from the end-cap vibration measuredbut also from sources including the main casing body,the cooling fan and the mount base. Any changes ofthese sources, due to the introduction and increasingseverity of the fault condition, would be incorporated inthe acoustic signal. In contrast, the vibration ismeasured at a speci c location that does not fullyrepresent the change in the motor characteristics [33].

2.6 Anomaly diagnosis using the residual spectrum

Fault diagnosis of many induction motor failure modesis well understood using conventional frequency domainmethods [22]. Componential coding can also diagnosefaults through frequency domain approaches. Since the

Table 1 Optimized componential coding con guration for induction motor data

Componential coding variantICAN JCAN

Data-set Vibration Acoustics Vibration and acoustics

Basis vector dimension 64 64 128 (i.e. 64 ‡ 64)Threshold value 0.101 0.854 0.113Number of basis vectors 23 23 26ADI (case 0) 0.006 0.009 0.056ADI (case 1) 0.417 0.931 0.289

Table 2 Discrimination index values with the introduction of phase imbalance

Componential coding variant ICAN JCAN

Data-set Vibration Acoustics Vibration and acoustics

D iscrimination index ADI VDI ADI VDI ADI VDI

Case 0 (healthy operation) 0.006 0.001 0.009 0.002 0.056 0.001Case 1 (20 V phase imbalance) 0.417 0.847 0.931 2.037 0.289 1.091Case 2 (40 V phase imbalance) 0.653 0.988 1.546 0.087 0.513 1.139



reconstruction errors represent the uniqueness of a newmeasurement, the spectrum of the residual signalcorrespondingly characterizes the changes in the fre-quency structure. (The residual signal, called theanomaly vector in Part 1 [23], is de ned as the vectordifference between the current input data vector and thenetwork’s model-based reconstruction of that input datavector.) Changes in its frequency components cantherefore be used for the identi cation of the occurrenceof a particular fault.

As described in section 2.5, although vibration andacoustics share similar generation mechanisms, theirsignal characteristics differ. In this study, the residualsignals from the ICAN were used for a subsequentspectral analysis, because separation of residual signalsprovided by the ICAN allows easier interpretation thanthe single residual signal produced by the JCAN.

Figure 2 compares the spectra of the measured datafrom the induction motor with those of the networkresiduals. F igures 2a and c show the spectra ofvibration data and acoustic data respectively andF igs 2b and d show the corresponding residual spectra.The spectra of the raw signals show a gradual increasein the amplitude of the 100Hz component. Thisparticular change allows a unique distinction to bemade of the phase imbalance faults from other possiblefaults (such as shaft misalignment and bearing damage[14, 22]). In the corresponding residual spectra, ampli-tude changes at 100 Hz are also clear. The change ofthis component in the vibration residual spectrum is 32per cent higher than in the raw spectrum for case 1and in the acoustic residual spectrum it is 200 per centhigher than in the raw spectrum. By comparison withconventional spectra, the residual spectra therefore

Fig. 2 Spectra of raw signals and residual signals from the ICAN



allow the characteristic fault feature (of the change inthe 100 Hz frequency component) to be more clearlyde ned.

3 APPLICATION TO A NOVEL ELECTRICMOTOR

3.1 Introduction

The second demonstration of componential coding ismade through its application to ux data from a novelhigh-power electric motor prototype. This motor, atransverse ux motor, is currently being developed byRolls-Royce plc and will potentially provide propulsionpower for a range of future marine vessels.

From a theoretical perspective [34, 35], changes of themagnetic ux pro les at different locations within themachine are likely to yield useful information onincipient fault conditions. F lux-based approaches tomonitoring are currently being investigated for thetransverse ux motor, and ux is routinely measuredduring operation. Although the effect of some failuremodes on ux pro les may be predicted, the detection ofparticular pro le changes is currently a labour-intensiveprocess that requires visual examination and interpreta-tion of individual waveforms. For a transverse uxmotor in a working marine propulsion application thishands-on approach is not desirable. The application ofcomponential coding to ux data from this machinetherefore not only provides a demonstration of itsability to monitor novel machines, for which onlylimited information is available, but also providesautomation of waveform analysis and basic leveldiagnosis (by separation of the types of pro le changesfrom a healthy baseline case).

3.2 The transverse ux motor

Several features distinguish the transverse ux motorfrom other more conventional permanent magnetsynchronous machines [36]:

1. The overall ux path is three-dimensional.2. The stationary armature coils are simple solenoids.3. The phases are entirely separated, with no electro-

magnetic coupling, providing fault tolerance andreversionary modes.

4. Substantial freedom of design is achieved by theability to change the magnetic ux geometry and coilsection.

5. The magnetic circuit does not interfere with theelectric circuit.

The simplest transverse ux motor con guration(F ig. 3a) consists of a single rotor disc to which twoopposite rotor rims are attached. The active rotor

(F ig. 3b) comprises alternate soft iron laminated polepieces and permanent magnets. The magnetization ofthe magnets is in the circumferential direction andaligned so that the pole pieces form alternate north (N)and south (S) poles. The stator consists of a number ofC-shaped stator cores and a preformed coil that passeswithin them. The overall output shaft torque is achievedby the combination of the ux produced by thepermanent magnets and that produced by the currentin the solenoid.

3.3 Magnetic ux monitoring

Magnetic ux is an important operational characteristicin any electric machine. Unlike conventional electricmotors, where physical access and design characteristicsmake ux monitoring dif cult [2, 3], the transverse uxmotor is well suited to the monitoring of this parameter.In practice, this is achieved by placement of coppersearch coils around the machine. The induced voltage inthese coils is directly proportional to the normal rate ofchange of ux density through the iron. The two maingroupings of coil are those wound around the C-shapedstator cores and those embedded in the rotor at the inter-face between adjacent magnets and pole pieces. Thoseon the stator cores are located as illustrated in F ig. 4.

Previous work [35] has proven that visual analysis ofthe ux waveforms from the transverse ux motor in thetime domain is particularly useful for detecting changesthat can be directly attributable to machine character-istics. For example, manufacturing tolerances havepreviously been examined by an analysis of the phaserelationship between waveforms [35], incipient sensorfailures identi ed and a partial rotor demagnetizationfault detected, diagnosed and located [35]. During themachine development stages only a small amount of uxdata has been recorded and it has therefore beenpossible to analyse each waveform individually by eyeand, consequently, make interpretations. However, twomain problems exist in this approach. The rst is in thedetection of very small waveform changes from a datachannel (that may only occur over a very small portionof the data covering one shaft revolution). The secondproblem is in performing this (presently visual) wave-form inspection for a motor in service, where measure-ment of data will be continuous. These problems arecompounded by the fact there are multiple sensors, all ofwhich require examination ideally in a synchronousmanner. The work described in this section illustrateshow componential coding is able to address theseproblems.

3.4 Monitored magnetic ux data

To illustrate the potential of componential coding inapplication to the novel motor, six cases were de ned



based upon the type of search coil location from whichdata were collected and upon the date of acquisition (seeTable 3, which de nes cases 0 to 5). Separation wasmade on this basis because:

1. Different (stator) search coil locations are likely toyield slightly different ux pro les as a result ofleakage ux path effects [34].

2. A stator component failure was experienced (andsubsequently repaired) between the two acquisitiondates (and it was possible some machine character-istics could have permanently changed).

Based on the available data, each case consisted of fourchannels of data recorded from four similarly locatedsearch coils.

Fig. 3 Transverse ux motor con guration and ux paths

Fig. 4 Search coil positions on the stator core



Figure 5 provides an example segment from one of thefour channels in each case. Initial visual examination ofthe data-sets revealed that three separate anomalousconditions were contained within them with respect tothe data provided by all channels in cases 0 and 1 (thebaseline pro le expected from theory [34]). The baselinepro les have 38 uniform periods for each complete shaftrevolution. This corresponds to the passing of 38 polepiece pairs on the rotor of this prototype. The rstanomalous condition (channel 3 in case 2 and channel 1in case 3) represents small and localized anomalies. Thesecond anomalous condition (given by channel 1 in case4) relates to severe distortion of most of the uxwaveform. All channels in case 5 represent a thirdanomalous condition. Each of these channels provides ux information from the rotor (rather than stator).Therefore (as expected by theory [34]), the pro les arevery different from the baseline ux for the stator searchcoils. As this difference is across the complete pro le thisis referred to as a global change compared with thebaseline. Componential coding was applied to investi-gate its ability to detect and assess the severity of thethree anomalous conditions described (particularlythose small changes provided in cases 2 and 3).

3.5 Network training and optimization

Both the JCAN and ICAN were studied for theirseparation of the anomalies with respect to the baselinecondition. The JCAN was employed to produce an

overall anomaly detection result for each case (based onthe four channels within each case and the phaserelationships between them). Once the output from theJCAN provided an indication of change from thebaseline, the ICAN was used to identify the anomalousindividual data signals and identify/extract the anomalyfeatures.

The dimension of basis vectors was set to 64 datapoints, as for the application to the induction motordata. For the basis vectors to cover two periods of theprincipal frequency component, the raw data weredesampled. This was achieved using a Fourier trans-form-based approach, where a forward Fourier trans-form was implemented on data covering one completerevolution of the rotor (i.e. 38 periods) and an inverseFourier transform was then applied to the rst 1216 (i.e.32 6 38) data points in the frequency domain. Desam-pling in this way was more suitable than the approachtaken for the vibration and acoustic data because the ux data have little noise and are strongly periodic.

As with the application to induction motor data, boththe threshold and the number of basis vectors wereoptimized using a genetic algorithm. Optimization wasachieved by obtaining maximum discrimination (basedon the ADI) between cases 0 and 1. The resultingcon gurations for the JCAN and ICAN are summarizedin Table 4. Both the JCAN and ICAN have high valuesfor the threshold. (For data vectors normalized to unitEuclidean norm, the theoretical maximum for usefulthreshold values is 1.) As discussed in section 2.4, thisindicates that there is a strong similarity between input

Table 3 Outline of separated data cases

Cases

0 1 2 3 4 5

Location type Inner tip Outer tip Inner root Inner tip and inner root Outer tip Rotor

Collection date March 2000 August 2000

Fig. 5 Example segments of normalized ux data from different search coil locations



patterns from the individual data-sets (whereas in theinduction motor study, differences in the acoustic andvibration data-sets led to differences in optimizedthreshold values).

The optimized number of basis vectors for the ICANand JCAN is much smaller than that in the inductionmotor study. The reason for this is that the ux datastructures are much simpler in comparison to either theacoustic or vibration data.

3.6 Anomaly detection

F igure 6 illustrates the anomaly detection resultsobtained by the JCAN variant of componential coding.It produces an overall value of the AD I and VDI for

each case presented to it. As the network is trained withrespect to case 0, the reconstruction of newly monitoredand previously unseen data, also from case 0, is verygood. Hence the reconstruction error is small (close to 0)and so are the ADI and VDI values for this case.

For those cases (1, 2 and 3) that have small visualdifferences from the baseline set (see F ig. 5), their ADIand VDI values are greater than the thresholds. Therelatively small ADI and VD I values for case 3 relate tothe fact that two of the four channels contained within itwere very similar to the baseline case 0. The higher ADIvalue for case 2 re ects the waveform anomaly ofchannel 3 and the high value of the VDI highlights thefact that the anomaly occurs locally.

For case 1, although no anomalous features wereobserved in the raw data (Fig. 5b), relatively high valuesof the AD I and VDI were obtained. There are twopossible reasons for this: differences in waveformpro les or phase differences between channels. TheICAN can be used to distinguish between these twohypotheses because the ICAN captures the statistics ofindividual channels. However, unlike the JCAN it is notcapable of measuring phase relationships.

For cases 4 and 5, where there is a large variationfrom the baseline, the ADI values are much highercompared with those for the other cases with smallervariations. The VDI value for case 4 is higher than thatfor case 5. Because of the different ways the ADI andVDI interpret the reconstruction error, this implies thatcases 4 and 5 contain local and global anomaliesrespectively. (The formulations of both the ADI andVDI are discussed in section 2.7 of Part 1 [23].)

The discrimination index values from componentialcoding allow separation of the ve cases from thebaseline one (case 0), discrimination between small andlarge differences and identi cation of local and globalvariations.

3.7 Anomaly diagnosis

The diagnosis of globally occurring anomalies can beachieved through residual spectrum analysis (demon-strated in section 2.6). However, residual spectra are notcapable of identifying the locally occurring anomaliesexisting in transverse ux motor ux data. This isbecause the Fourier transform is inadequate at resolving

Table 4 Optimized componential coding con guration for transverse ux motor ux data

Componential coding variant ICAN JCANData-set Channel 1, case 0 All channels, case 0

Basis vector dimension 64 256 (i.e. 64 ‡ 64 ‡ 64 ‡ 64)Threshold value 0.885 0.797Number of basis vectors 4 5ADI (case 0) 0.29 0.053ADI (case 1) 0.807 1.022

Fig. 6 Discrimination indices from the JCAN ( representsthe training case, * represents the optimization case)



waveform changes that are small in amplitude, localizedand positioned randomly. To overcome this problem,both ADI and VD I values can be used for the diagnosisof the small local anomalies.

Having detected the anomalies in each case by theJCAN, the ICAN was used to identify automatically theindividual channels that cause these anomalies. F igure 7presents the VDI values produced by the ICAN, as theVDI is more sensitive to the local small anomalies (seesection 2.7 of Part 1 [23]). Based upon the difference inamplitude from the values in case 0, the VDI clearlycaptures nine anomalous channels (as indicated by * inF ig. 7). Seven of these were initially found by visualinspection of the raw data.

Two new anomalies (channel 3 of case 3 and channel 4of case 4) were revealed by the high amplitudes of theVDI. These were not captured by the initial visualinspection procedure because they are very small wave-form changes, occurring locally, and only once everyshaft rotor revolution.

A better understanding of the localized anomalies canbe achieved by calculation of individual ADI values foreach input data pattern and plotting them against theangular position of the rotor. The obtained ADI is thusreferred to as an instantaneous ADI. F igure 8 shows theADI variation over two complete shaft revolutions. Thelocalized changes in waveform produce large jumps inthe ADI (well above the detection threshold, which wasset at the ADI for case 0 plus three times its standarddeviation). These high jumps not only allow thedetection of the anomalies but also identify the angularposition at which they occur. The characteristics of the

ADI jumps also enable the identi cation of themechanism causing the anomalies. For example, a singlejump per revolution with a constant angular position (asin case 3) indicates a possible fault in one of thepermanent magnets in the rotor. Several jumps perrevolution spaced at a random angular position (as incases 2 and 4) indicate a potentially faulty ux searchcoil.

By de nition, the localized waveform changes cannotbe measured accurately by the ADI and therefore theVDI was also used (see section 2.7 of Part 1 [23]). Thecombination of ADI and VDI produces comprehensivedetection and diagnosis for both globally and locallyoccurring anomalies. This is illustrated in the scatterplot of these two indices. As shown in F ig. 9, those uxchannels with local anomalies have a larger VDI thanADI. However, global changes (as in case 5) lead tosimilarly large amplitudes in both the VDI and ADI.Thus, local changes in a waveform can easily bedistinguished from global ones, providing useful infor-mation in identifying the cause of the anomalies.

3.8 Benchmarking the capability of componential codingin condition monitoring

To benchmark the performance of machine monitoringusing componential coding, its anomaly detectioncapability was compared with several establishedtechniques currently in use. These techniques were:waveform measures (including root-mean-square andkurtosis), spectrum difference, cepstrum difference,

Fig. 7 VDI from ICAN channels and cases



Fig. 8 Instantaneous ADI

Fig. 9 Scatter plot of the ADI against the VDI



Fourier-based signal reconstruction and principal com-ponent analysis. Where possible, these techniques wereoptimized so that each of the anomalies could be mostclearly separated from the baseline healthy data-set (case0). For example, in the Fourier-based signal reconstruc-tion method the number of Fourier components wasoptimized. The calculation of the waveform root-mean-square and kurtosis used the same size of input pattern(basis vector dimension) as componential coding. Thecalculation of the spectrum and cepstrum used longerinput sizes to achieve better resolution. The Fourier-based reconstruction method and principal componentanalysis were used in a similar way to componentialcoding, in that a model was initially developed based onbaseline data and then this was used to examine itsability to reconstruct the other cases. Following agenetic optimization (using a similar approach to thatused for componential coding), the number of compo-nents for the Fourier-based reconstruction and thenumber of eigenvectors for principal component analy-sis was set to ve and four respectively.

Table 5 shows the number of individual channels thatcould be separated as containing anomalies with respectto the baseline data. Componential coding enables allnine anomalous channels to be separated, but the verysmall and localized anomalies could not be detected byany of the conventional techniques. This demonstratesthat, compared with the other methods assessed,monitoring using componential coding is the mostcomprehensive approach to detection of machine faultsusing motor ux search coil data.

4 SUMMARY AND CONCLUSIONS

Advanced data processing has the potential to enableearlier anomaly detection and more accurate diagnosisfor electrical machine faults. This paper has presentedthe use of a novel unsupervised adaptive neuralnetwork, componential coding, for the condition mon-itoring of a conventional motor and a novel transverse ux motor. Componential coding is implemented bymodelling characteristic baseline data and then using themodel for automatic monitoring. The results show thatthe use of the single-valued average discrimination index(ADI) and variance discrimination index (VDI) provides

a straightforward basis on which faults may beautomatically detected and diagnosed.

Componential coding may be implemented througheither of two network derivatives (JCAN or ICAN). Tocombine information from multiple sensor channels(yielding similar data structures) and perform speedcritical on-line monitoring, the JCAN is the mostef cient. However, it has been shown that the ICANmay be employed when either the information contentfrom sensor channels is different (as in the application toinduction motor vibration and acoustic data) orchannel-level anomaly detection is required (as in theapplication to magnetic ux data).

Two approaches have been outlined for diagnosingthe nature of anomalies. In can be concluded that theresidual spectrum approach enables the classi cation ofglobally distributed changes. For example, voltageimbalance in the induction motor three-phase powersupply was identi ed by the change in the 100 Hzcomponent in the residual spectra. For small anomaliesthat occur over short segments of the data-sets (asexperienced in the transverse ux motor ux data), thecombination of the ADI and VDI enabled reliableclassi cation. Furthermore, the application of aninstantaneous ADI allowed identi cation of the positionof the fault features.

A benchmark of componential coding against otherconventional techniques (including waveform statisticalmeasures, Fourier analysis and principal componentanalysis) has also been outlined in this paper. It can beconcluded that, when applied to motor ux data,componential coding outperforms these techniques inthat it can separate small, infrequently occurring andlocally distributed anomalies not achievable to the sameextent by the other approaches examined.

ACKNOWLEDGEMENTS

This work was carried out as part of Technology Group10 of the MoD Corporate Research Programme. TheComponential Coding algorithm is patented intellectualproperty of QinetiQ. Rolls-Royce plc is thanked for itssupport in application of this novel network architectureto electrical machines.

Table 5 Summary of comparisons of different detection performances

Technique Anomalous channels detected Detection performance (%)

Componential coding (ICAN implementation) 9 100Fourier-based reconstruction 7 78Principal component analysis 7 78Waveform statist ics 5 56Spectrum difference 6 67Cepstrum difference 7 78



# QinetiQ limited 2003. Published under licence withthe permission of QinetiQ.

REFERENCES

1 Starr, A. and Wynne, R. A review of condition basedmaintenance for electrical machines. In Handbook ofCondition M onitoring, 1996 (Elsevier Science, Oxford).

2 Tavner, P. J. and Penman, J. Condition M onitoring ofElectrical M achines, 1987 (Research Studies Press,Baldock, Hertfordshire).

3 Dorrel, D. G., Thomson, W. T. and Roach, S. Analysis ofairgap ux, current, and vibration signals as a function ofthe combination of static and dynamic airgap eccentricityin 3-phase induction motors. IEEE Trans. Ind. Applic.,1997, 33(1), 24–34.

4 Seker, S. Determination of air-gap eccentricity in electricmotors using coherence analysis. IEEE Power Engng Rev.,2000, 20(7), 48–50.

5 Wowk, V. M achinery Vibration—M easurement and Analy-sis, 1991 (McGraw-Hill, New York).

6 Finely, W. R., Hodowanec, M. M. and Holter, W. G. Ananalytical approach to solving motor vibration problems.IEEE Trans. Ind. Applic., 2000, 36(5), 1467–1480.

7 Eisenmann, R. C. and Eisenmann Jr, R. C. M achineryM alfunction—Diagnosis and Correction, 2000 (Prentice-Hall, Englewood Cliffs, New Jersey).

8 Hargis, P. K., Mahrt, R. and Doon, H. H. The detection ofrotor defects in induction motors. In Proceedings of IEEEInternational Conference on Electrical M achines, Designand Applications, 1982, pp. 216–220.

9 Arthur, N. and Penman, J. Induction machine conditionmonitoring with higher order spectra. IEEE Trans. Ind.Electron, 2000, 47(5), 1031–1041.

10 Thomson W. T., Rankin, D. and Dorrell, D. G. On-linecurrent monitoring to diagnose airgap eccentricity in largethree-phase induction motors—industrial case historiesverify the predictions. IEEE Trans. Energy Conversion,1999, 14(4), 1372–1378.

11 Walliser, R. F. and Landy, C. F. Determination of interbarcurrent effects in the detection of broken rotor bars insquirrel cage induction motors. IEEE Trans. EnergyConversion, 1994, 9(1), 152–158.

12 Rankin, D. R. The industrial application of phase currentanalysis to detect rotor winding faults in squirrel cageinduction motors. Power Engng J., 1995, 9(2), 77–84.

13 Liang, B., Payne, B. S. and Ball, A. D. Detection anddiagnosis of faults in induction motors using vibration andphase current analysis. In Proceedings of 1st InternationalConference on The Integration of Dynamics, M onitoringand Control, 2000, pp. 337–341.

14 Payne, B. S., Ball, A. D., Gu, F. and Li, W. A head-to-headassessment of the relative fault detection and diagnosiscapabilities of conventional vibration and airborne acous-tic monitoring. In Proceedings of the 13th InternationalCongress on Condition M onitoring and Diagnostic Engi-neering M anagement, Houston, Texas, 2000, pp. 233–242.

15 Schoen, R. R., Habetler, T. G., Kamran, F. and Barthheld,R. G. Motor bearing damage detection using stator current

monitoring. IEEE Trans. Ind. Applic., 1995, 31(6), 1274–1279.

16 Eren, L. and Devaney, M. J. Motor bearing damagedetection via wavelet analysis of the starting currenttransient. IEEE Instrum. M easmt M ag., 2001, 3, 1797–1800.

17 Gaylard, A., Meyer, A. and Landy, C. Acoustic evaluationof faults in electrical machines. In IEE Proceedings of 7thInternational Conference on Electrical M achines andDrives, Durham, 1995, No. 412, pp. 147–150.

18 Shibata, K., Takabashi, A. and Shirai, T. Fault diagnosis ofrotating machinery through visualisation of sound signals.M ech. Systems and S ignal Processing, 2000, 14(2), 229–241.

19 Albright, D. R. Method and apparatus for detecting rotor ux variations in the air-gap of a dynamo-electric machine.United States Patent Of ce, US Pat. 3506914, PatentedApril 1970.

20 Penman, J., Dye, M. N., Tait, A. J. and Bryan, W. E.Condition monitoring of electrical drives. Proc. IEE, PartB, 1986, 133(3), 142–148.

21 Payne, B. S., Ball, A. D. and Gu, F. Detection anddiagnosis of induction motor faults using statisticalmeasures. Int. J. Condition M onitoring and DiagnosticEngng M anagmt, 2002, 5(2), 5–19.

22 Lyon, R. H. M achinery Noise and Diagnosis, 1987 (Butter-worth Publishers, London).

23 Webber, C. J. S., Payne, B. S., Gu, F. and Ball, A. D.Componential coding in the condition monitoring ofelectrical machines. Part 1: principles and illustrationsusing simulated typical faults. Proc. Instn Mech. Engrs,Part C: J. M echanical Engineering Science, 2003, 217(C8),883–899.

24 Chow, M. and Yee, S.O. Application of neural networks toincipient fault detection in induction motors. J. NeuralNetwork Computing, 1991, 2(3), 26–32.

25 Filippetti, F. Neural networks aided on-line diagnostics ofinduction motor rotor faults. IEEE Trans. Ind. Applic.,1995, 31(4), 892–898.

26 Paya, B. A., Esat, I. I. and Badi, M. N. M. Arti cial neuralnetwork based fault diagnosis of rotating machinery usingwavelet transforms as a pre-processor. M ech. Systems andSignal Processing, 1997, 11(5), 751–765.

27 Luo, F. and Unbehauen, R. Applied Neural Networks forSignal Processing, 1997 (Cambridge University Press,Cambridge).

28 Bonnett, A. H. and Soukup, G. C. Cause and analysis ofstator and rotor failures in three-phase squirrel-cageinduction motors. IEEE Trans. Ind. Applic., 1992, 28(4),921–937.

29 Arkan, M., Perovic, D. K. and Unsworth, P. Online statorfault diagnosis in induction motors. Proc. Electl PowerApplic., 2001, 148(6), 537–547.

30 Pierrat, L. and Morrison, R. E. Probabilistic modelling ofvoltage asymmetry. IEEE Trans. Power Delivery, 1995,10(3), 1614–1620.

31 Ball, A. D, Li, W., Gu, F. and Shi, Z. The analysis of engineacoustic signals using higher order statistics. In Proceed-ings of 4th Annual Maintenance and Reliability Con-ference, Knoxville, Tennessee, 2000, Vol. 2, pp. 33.01–33.09.

32 Gu, F., Payne, B. S. and Ball, A. D. Optimisation of a self-organising autoencoder for the condition monitoring of a



novel electric motor. University of Manchester DocumentMERG-0600, 2000.

33 Elhaj, M., Gu, F., Shi, Z. and Ball, A. D. A comparison ofthe condition monitoring of reciprocating compressorvalves using vibration, acoustic, temperature and pressuremeasurements. In Electronic Proceedings of 6th Main-tenance and Reliability Conference, Knoxville, Tennessee,2002.

34 Payne, B. S., Voyce, J., Husband, M. and Ball, A. D.Development of condition monitoring technology for

electric propulsion systems. In IMarEST Proceedings ofInternational Naval Engineering Conference, Bath, 2002,pp. 238–247.

35 Payne, B. S., Husband, M., Simmers, B., Gu, F. and Ball,A. D. The development of ux monitoring for a novelmotor. In Proceedings of the 14th International Congresson Condition M onitoring and Diagnostic EngineeringM anagement, Manchester, 2001, pp. 353–360.

36 Mitcham, A. J. Transverse ux motors for electricpropulsion of ships. In IEE Colloquium, 1997, pp. 3/1–3/6.



componential coding in the condition monitoring of electrical...

Documents