Research Article Fault Diagnosis for Compensating Capacitors...
Transcript of Research Article Fault Diagnosis for Compensating Capacitors...
-
Research ArticleFault Diagnosis for Compensating Capacitors of Jointless TrackCircuit Based on Dynamic Time Warping
Wei Dong1,2
1 Tsinghua National Laboratory for Information Science and Technology, Beijing 100084, China2Department of Automation, Tsinghua University, Beijing 100084, China
Correspondence should be addressed to Wei Dong; [email protected]
Received 31 January 2014; Accepted 17 February 2014; Published 25 March 2014
Academic Editor: Hamid R. Karimi
Copyright Β© 2014 Wei Dong. This is an open access article distributed under the Creative Commons Attribution License, whichpermits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Aiming at the problem of online fault diagnosis for compensating capacitors of jointless track circuit, a dynamic time warping(DTW) based diagnosis method is proposed in this paper. Different from the existing related works, this method only uses theground indoor monitoring signals of track circuit to locate the faulty compensating capacitor, not depending on the shunt currentof inspection train, which is an indispensable condition for existingmethods. So, it can be used for online diagnosis of compensatingcapacitor, which has not yet been realized by existing methods. To overcome the key problem that track circuit cannot obtain theprecise position of the train, the DTWmethod is used for the first time in this situation to recover the function relationship betweenreceiverβs peak voltage and shunt position.The necessity, thinking, and procedure of the method are described in detail. Besides theclassical DTWbasedmethod, two improvedmethods for improving classification quality and reducing computation complexity areproposed. Finally, the diagnosis experiments based on the simulation model of track circuit show the effectiveness of the proposedmethods.
1. Introduction
Track circuit is one of the most important basic equipmentin train control system. It is used for detection of trainβsoccupancy on the track, train-ground communication, anddetection of broken rails, and its functions are crucial inensuring safe operation of trains. In track circuit, the rails aretaken as parts of the working circuit and the current flowsfrom transmitter to receiver through the rails (Figure 1). Ifsome train is located between the transmitter and the receiver(this interval is called βblock sectionβ), then the circuit willbe shorted by the wheels and axles of the train (Figure 2).So a remarkable drop in the receiverβs voltage will indicatea trainβs occupancy. In this way, the train control system canobtain the informationwhich block sections are occupied andwhich block sections are empty. Based on this information,the train control system can release movement authorities toeach train in order to avoid collisions between trains. If thetrack circuits have some faults, the most serious consequenceis to cause collisions of trains, resulting in catastrophes. Evenif the fail-safety mechanism plays its role and no catastrophic
accident happens, the transportation efficiency of railway willbe affected seriously, causing considerable economic loss. Sothe timely fault diagnosis for track circuit is significant to bothsafety and efficiency of the railway transportation system.
In the practical application, as track circuits must belaid along the rails and the length of one track circuit isabout 1βΌ2 kilometers, the application scope of track circuitsis quite large. And due to the special structure and complexworking environment, track circuits are easily affected bytemperature, humidity, ballast resistance, electromagneticinterference, and mechanical vibration, which results in ahigh fault rate of track circuit [1β4]. As a result of the abovetwo aspects, the loss caused by faults of track circuits is veryhuge. For example, according to the report from the officialwebsite of the Ministry of Railways of China, up to the end of2012, the number of track circuits laid along Chinese railwaysreaches about 600 thousand, and the equipment assets oftrack circuits reach about 60 billion RMB. And according tothe statistical data released by Chinese railway department,there were about 8 thousand faults happening in the signaland communication systems of Chinese railways per year.
Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2014, Article ID 324743, 13 pageshttp://dx.doi.org/10.1155/2014/324743
-
2 Mathematical Problems in Engineering
Current direction
Electricalinsulation joint
Electricalinsulation joint
Receiver
Rails
Transmitter
Figure 1: Basic structure of track circuit (empty).
Current direction
Electricalinsulation joint
Electricalinsulation joint
RailsShunt
current
Wheels and axle
Receiver Transmitter
Figure 2: Basic structure of track circuit (occupied).
The majority of these faults were caused directly or indirectlyby track circuits, resulting in a total traffic delay time ofabout 4 hundred thousandminutes. In otherwords, each faultcaused an average traffic delay of about 50 minutes [5].
Among different fault types of track circuit, compen-sating capacitorβs fault is one of the most important faulttypes. Compensating capacitors are used to compensate theinductive impedance of the rails, so that the current signalcan be transmitted along the rails far enough. To this end,compensating capacitors are installed at equal intervals alongthe rails, and the installation interval is about 80βΌ100 meters(Figure 3). Due to influence of bad weather and mechanicalvibration and so on, compensating capacitors are prone tobe faulty. Once some compensating capacitors are faulty, thesignal transmitted along the rails will be attenuated rapidly,which may result in red light faults. When a track circuithas this kind of faults, the track circuit will declare that itis occupied (i.e., in red light state) when there is no trainon the track in fact. In this case, to find the faulty compen-sating capacitors, maintenance personnel have to check thecompensating capacitors one by one along the very long rails,which not only consumes lots of human resources but alsocauses a long delay in trainβs operation. For example, in thespring of 2009, the large-range faults of the compensatingcapacitors in Chinese Beijing-Tianjin intercity high-speedrailway were caused by snowstorm. In order to repair thefaults as soon as possible, on the condition of lacking relateddiagnosis technologies, the railway department was forced
to replace all the compensating capacitors within 2 days,resulting in a huge consumption of human and materialresources.
Despite the importance, the problem of fault diagnosisfor track circuit has not been solved very well and there arestill many difficulties. In industrial community, up to now, themost commonly adoptedmethods for fault diagnosis of trackcircuit are periodic maintenance and field inspection [4],resulting in low working efficiency, high working strength,and long repair time. In academic community, a lot ofwork has been done. Chen et al. [3, 4] and Huang et al.[6] gave fault diagnosis methods for track circuit based onneurofuzzy systems, which can detect and diagnose the mostcommon faults of track circuit. However, due to lack ofdynamic diagnostic information, this method cannot be usedto diagnose compensating capacitorβs faults. Aiming at thefeatures of track circuitβs multiple fault modes and uncertainand nonaccurate fault symptom information,Oukhellou et al.[7] gave an information fusion method for fault diagnosis ofcompensating capacitor based on Dempster-Shafer classifier.However, this method can only be used by inspection train,resulting in that it cannot be used for online diagnosis.
Similar to [7], most existing methods for fault diagnosisof compensating capacitors in track circuit are based onthe shunt current (βshunt currentβ is the current flowingthrough the wheels and axle of a train when the trackcircuit is occupied (see Figure 2)) of inspection train. Forexample, CoΜme et al. [8, 9] proposed noiseless independentfactor analysis methods for fault diagnosis of compensatingcapacitors in track circuit. Zhao et al. [10] designed a kind ofonboard autotest system to detect the faults of track circuitβscompensating capacitors in real time. Zhao et al. [11] usedthe induced voltage recorded by cab signal as fault featureand gave a fault diagnosis method for track circuit basedon discrete binary wavelet transform (DBWT) and waveletridge (WR), finally detecting the changes of instantaneousfrequency to diagnose the disconnection fault of compen-sating capacitor. Zhao et al. [12] used the transmission linetheory to construct a track circuit model with cascade formof multiple four terminal networks. Based on the constructedmodel, the typical fault modes of compensating capacitorwere analyzed through the fault features of track voltage,shunt current, and input impedance of main track circuit.Linhai et al. [13] proposed a wavelet analysis based methodto detect the integrity of the compensating capacitors ofUM71 track circuit, by making use of the recorded data inthe cab signal recorder. Zhao and Mu [14] analyzed inducedvoltage envelope of cab signal (IVECS) of track circuitby simulation model under different faulty conditions andproposed a compensating capacitor diagnosis method basedon adaptive optimal kernel time-frequency representation(AOK-TFR). Zhao et al. [15] adopted B-spline DBWT andimproved Hilbert-Huang transformation for noise reductionand fault feature retrieve and proposed a diagnosis methodfor compensating capacitor based on cab signal recorderinformation. Zhao et al. [16] proposed a diagnosismethod forcompensating capacitor based on the regression model of theshunt current. This method adopted Levenberg-Marquardt
-
Mathematical Problems in Engineering 3
(L-M) algorithm to verify the model and generalized S-transform (GST) to compute the instantaneous frequency soas to locate the faulty capacitors. Zhao et al. [17] analyzedthe influence law of compensating capacitorβs fault on theinduced voltage envelope of the cab signal and proposeda comprehensive fault diagnosis method based on geneticalgorithm to locate faulty compensating capacitors. Lin-Haiet al. [18] presented a collaborative fault diagnosis system forcompensating capacitors in track circuit using AOK-TFR andadaptive genetic algorithm (AGA) based on the cab signal.Sun et al. [19] presented a diagnosis approach for compen-sating capacitors based on empirical mode decomposition(EMD) and Teager energy operator (TEO) theory, which candetect multiple capacitor faults.
The common feature of the above research on faultdiagnosis of compensating capacitor is that theymust dependon the shunt current of inspection train. The problems ofthese methods can be summarized as follows.
(1) For a specific railway line, the inspection train has aspecific running period and can only be used whenthe railway line is idle, so these methods cannotbe used for online diagnosis. Online diagnosis ofcompensating capacitor is important, because it canhelp to find the faulty compensating capacitors intime so as to avoid traffic delay. Usually, one faultycompensating capacitor may not result in failure oftrack circuit, butmore faulty compensating capacitorsmay result in failure of track circuit. So finding faultycompensating capacitors in time is important foravoiding failure of track circuit.
(2) If there is no inspection train, to use these methods,some shunt current detecting and fault diagnosisdevices must be installed on common trains, whichwill result in huge cost input and huge workloadfor modifications of equipment. Besides, consideringsafety, modifications of on-board equipment are rig-orously restricted by the management department, soit is hard to install these devices.
(3) Since the shunt current must be detected throughelectromagnetic induction, the detected results areeasily interfered by noise.
(4) These methods take the function relationshipbetween shunt current and shunt position as thediagnostic basis, so the position of the trainβs firstwheel-set must be obtained accurately. But inengineering practice, the precise position of the trainis not easy to obtain by track circuit, resulting in thatthe diagnosis results are easily affected by positionerror.
To overcome the above problems, a dynamic time warp-ing (DTW) based fault diagnosis method for compensatingcapacitors in track circuit is proposed in this paper. In order torealize online diagnosis, this method does not depend on theshunt current of inspection train and only collects the voltagesignals in receiver of track circuit. However, static voltageinformation is insufficient to locate the faulty capacitor. Inorder to promote the diagnosis resolution, this method uses
the dynamic voltage information which is varying whena train is passing the track circuit (in this situation, thestructure of track circuit is changing (see Figure 2), so moreinformation can be obtained). Different from the existingmethods, this method only needs the function relationshipbetween receiverβs peak voltage and shunt time (insteadof shunt position) as the diagnostic basis, and the DTWmethod is used in this situation to recover the originalfunction relationship between receiverβs peak voltage andshunt position. Finally, by DTW, the characteristic curve iscompared to the sample curves of different faults to locatethe faulty capacitor. The contributions and advantages of themethod proposed in this paper can be listed as follows.
(1) This method only uses the groundmonitoring signalsof track circuit to locate the faulty compensatingcapacitor, not depending on inspection train. So,this method can be used for online diagnosis ofcompensating capacitor, which is not yet realized byexisting methods.
(2) This method does not need the outdoor monitoringsignals (see Figure 3) or the on-board monitoringsignals, so the existing system does not need to bemodified greatly, which means a good engineeringfeasibility.
(3) Since the precise train position is very hard to obtainbased on the existing monitoring system of trackcircuit (the position information of a train is easilyobtained by on-board devices instead of grounddevices. However, on-board devices are not suitablefor online diagnosis of track circuit), thismethoddoesnot need the position of the train and adopts DTW torecover the function relationship between receiverβspeak voltage and shunt position. So, this methodovercomes the problem that the existing methodsmust depend on the precise position of the train.
This paper is organized as follows. Section 1 gives anoverall background and introduces the problem. Section 2gives a description of track circuit structure and track circuitmodel. Section 3 gives the fault diagnosis method for trackcircuit based on dynamic time warping. Section 4 gives theresults and discussion. Finally, Section 5 gives the conclusion.
2. Track Circuit and Track Circuit Model
2.1. Track Circuit Structure. The structure of a track circuitis shown in Figure 3. A track circuit consists of four parts:transmitter part, receiver part, main track circuit, and smalltrack circuit. Transmitter part, which is used for signaltransmitting, consists of transmitter, lightning protector,cable simulator, SPT (βSPTβ is just a code for digital signalcable in Chinese railway. βSPβ means βdigital signal cableβand βTβ means βrailwayβ) cable, and matching transformer.Transmitter is used to generate a frequency-shift modulatedsignal with high stability and precision. Lightning protectoris used to protect against lightning impulses which areintroduced indoors by the cable. SPT cable is used to transmitsignals from indoor transmitter to outdoor steel rails, and
-
4 Mathematical Problems in Engineering
This track circuit
Lightning protector
Lightning protector
Matchingtransformer
Matchingtransformer
Cablesimulation
network
Cablesimulation
network
Compensatingcapacitors
Small track circuit
Adjacent track circuit
SPT cableSPT cable
Transmitter
Receiver
Indoor
(XGJ, XGHJ)
GJ
Attenuator
Outdoor
Rails
V3
V4
V5
V1
V2
Equivalent to 10km SPT cable
1600mm 3600mm
Tune
uni
t
Tune
uni
t
Air-
core
coil
Tune
uni
t
Tune
uni
t
Air-
core
coil
Main track circuit
Figure 3: Structure of track circuit.
cable simulator is used to compensate the actual SPT cableso that the total equivalent length of SPT cable is equal to10 km.Matching transformer is used to couple the frequency-shiftmodulated signal into steel rails or collect the frequency-shift modulated signal from steel rails. Receiver part, whichis used for signal receiving, consists of receiver, attenuator,lightning protector, cable simulator, SPT cable, and matchingtransformer. Except for attenuator, which is used to attenuatethe signal for receiving, the others are the same as thetransmitter part. Main track circuit, which is the main bodyof the current loop, consists of two steel rails and somecompensating capacitors connected between the two rails.Small track circuit (also called βelectrical insulation jointβ)consists of tune units, air-core coil, and 29 meters long steelrails. These circuit components constitute different kinds ofresonant circuits and are used to separate two signals ofdifferent frequencies in different track sections.
Different track circuits have different working frequen-cies. There are totally four working frequencies for trackcircuits: 1700Hz, 2000Hz, 2300Hz, and 2600Hz. The distri-bution of these fourworking frequencies is shown in Figure 4.By this distribution, the interference between two adjoiningtrack circuits can be reduced effectively.
2000Hz 2000Hz
1700Hz 1700Hz
2600Hz
2300Hz
Travelling Electricalinsulation joint
Electricalinsulation joint
Electricalinsulation joint
Electricalinsulation joint
direction oftrain
Travellingdirection of
train
Rails
Rails
Figure 4: Frequency distribution of track circuits.
2.2. Track Circuit Model. The track circuit model is usedto generate sufficient data for research, so as to make upfor the deficiency of data from actual devices. The βSim-PowerSystemsβ tool box in Matlab is used to construct thetrack circuit model, and receiverβs dynamic peak voltagesare calculated by adjust the structure and parameters of themodel automatically.Themain parameters of the track circuitmodel are shown in Table 1.
For reasonable simplification of the model, the followingassumptions are made.
(1) Due to the frequency-shift modulation mechanism,the signal on track circuit is not a strict sine wave.However, since the frequency offset (11 Hz) is very
-
Mathematical Problems in Engineering 5
Table 1: Main parameters of track circuit model.
Parameter name Parameter valueWorking frequency 2000HzPeak voltage in transmitter 142VCapacitance of compensating capacitor 5 Γ 10β5 FShunt resistance 0.15Ξ©Length 1301mThe number of compensating capacitors 16Contact resistance of compensatingcapacitor 1 Γ 10
β4
Ξ©
0 200 400 600 800 1000 1200 1400
Shunt position (m)
0.115
0.11
0.105
0.1
0.095
0.09
0.085
Peak
vol
tage
of r
ecei
ver (
V)
Figure 5: Peak voltages of receiver under different shunt positions(when track circuit has no faults, without noise).
small relative to theworking frequency (2000Hz), thetrack circuit can be considered as a sine steady circuitwithin an acceptable error range.
(2) Because the transient signal component decays veryquickly when the train is shunting on the rails [20],this paper only considers steady states of track circuit.And the peak voltages of receiver (π
5
, shown inFigure 3) under different shunt positions are pickedas fault features (Figure 5).
(3) Because the shunt resistance is very small (0.04β0.15Ξ©), the electrical effect of multiple wheel-setsshunting on the rails is almost the same as thesituation of one wheel set [21]. So, this paper onlyconsiders one wheel setβs shunt on the rails andthe shunt resistance is supposed to be 0.15Ξ© (forsmaller shunt resistance, transmitterβs current insteadof receiverβs voltage can be selected as fault feature,and the diagnosis process is all the same).
2.3. Verification and Validation of Track Circuit Model. Inorder to verify and validate the track circuit model, aphysical model of track circuit has been set up (shown inFigure 6). This physical model consists of completely real
1
2
3
4
5
76
8
9
Figure 6: The physical model of ZPW-2000A track circuit.A Transmitter; B receiver; C attenuator; D lightning protectorand cable simulator; E small track simulator; F tune unit; G aircore coil;H compensating capacitor;0 main track simulator.
Table 2: The verification and validation results.
π
1
π
2
π
3
π
4
Simulation model 142.31 (V) 41.58 (V) 13.46 (V) 1.09 (V)Physical model 142.56 (V) 41.44 (V) 13.39 (V) 1.06 (V)Absolute error β0.25 (V) 0.14 (V) 0.07 (V) 0.03 (V)Relative error β0.18% 0.34% 0.52% 2.83%
track circuit devices except for the steel rails, which cannotbe installed indoors due to its huge size. The steel rails inmain track circuit and small track circuit are replaced by thecorresponding track simulators, which have been carefullycalibrated according to real steel rails.
The verification and validation scheme is as follows. Fourtypical peak voltages (π
1
, π
2
, π
3
, π
4
, shown in Figure 3) oftrack circuit are selected as features for comparison. One setof voltages are obtained from the simulation model, and theother set of voltages are obtained by averaging 100 sets ofmeasured values from the physical model. The comparativeresults are shown in Table 2, which shows that the deviationbetween the simulationmodel and the physical model is verysmall.
3. Fault Diagnosis Method for Track CircuitBased on Dynamic Time Warping
3.1. Problem Formulation
3.1.1. Targets and Constraints. According to the introduc-tion section, the following targets or constraints should beachieved or met.
(1) Online fault diagnosis for compensating capacitorsof track circuit should be realized. Here, only sin-gle disconnection fault of compensating capacitor is
-
6 Mathematical Problems in Engineering
Table 3: Fault state list of track circuit.
State symbol MeaningπΉ
0
No fault
πΉ
π
(π = 1, 2, 3, . . . , 16)Number π (numbered from transmitter to
receiver) compensating capacitordisconnects from the rails
Table 4: Peak voltage of receiver under different fault states.
Fault state Peak voltage of receiver (V)πΉ
0
0.5917πΉ
1
0.5343πΉ
2
0.5257πΉ
3
0.5246πΉ
4
0.5313πΉ
5
0.5276πΉ
6
0.5258πΉ
7
0.5299πΉ
8
0.5275πΉ
9
0.5272πΉ
10
0.5292πΉ
11
0.5263πΉ
12
0.5275πΉ
13
0.5309πΉ
14
0.5251πΉ
15
0.5258πΉ
16
0.5341
considered, because disconnection fault is the mostpossible fault type of compensating capacitor and twocompensating capacitors are almost impossible to befaulty synchronously.
(2) Only the indoor voltages or currents of track circuit(see Figure 3) can be obtained for fault diagnosis.Because the outdoor voltages or currents are hardto be measured under existing conditions, and theshunt current cannot be used for online diagnosis, asdiscussed in Section 1.
(3) Under existing conditions, a trainβs exact positioninformation cannot be obtained by track circuit or itsmonitoring system. So the shunt position informationcannot be used for fault diagnosis.
Based on the above targets and constraints, the problemto be solved is to distinguish 17 kinds of states (listed inTable 3) of track circuit by specific fault features.Without lossof generality, the peak voltage of receiver is picked as faultfeature.
3.1.2. Insufficiency of Static Voltage Information. (Here,βStaticβ means there is no train passing the track circuit, sothe peak voltage is invariable.) Under different fault states, thepeak voltages of receiver (π
5
in Figure 3) are shown inTable 4.
0 200 400 600 800 1000 1200 1400
Shunt position (m)
0.14
0.13
0.12
0.11
0.1
0.09
0.08
0.07
0.06
0.05
0.04
Peak
vol
tage
of r
ecei
ver (
V)
F0
F1 F2 F10
F0
F1
F2
F10
Figure 7: Peak voltages of receiver under different shunt positions(for fault states πΉ
0
, πΉ1
, πΉ2
, and πΉ10
, without noise).
It can be seen from Table 4 that if only static peak voltageis obtained, πΉ
8
and πΉ12
, as well as πΉ6
and πΉ15
, cannot be distin-guished at all. Besides, if a measurement noise of Β±0.005V isconsidered, there are 120 pairs of indistinguishable faults (i.e.,π
πΉπβ π
πΉπβ€ 0.01(π)), occupying 88% of all fault pairs. In this
case, besides πΉ0
, all other fault states cannot be distinguished,which means the fault can be detected but cannot be located.Even if a measurement noise of Β±0.002V is considered, thereare still 76 pairs of indistinguishable faults.
Besides the peak voltage of receiver, other static voltagesor currents can also be selected as fault features, but the resultsare similar. Moreover, in theory, the phase information ofvoltages or currents can also be used as fault features. But inengineering practice, the accurate phase information is hardto obtain when the working frequency reaches 2000Hz, sothe phase information is not considered in this paper.
From the above analysis, it can be seen that, if onlystatic voltages or currents are used, it is very hard or evenimpossible to locate the compensating capacitorβs fault.
3.1.3. Advantage of Adopting Dynamic Voltage Information.(Here, βDynamicβ means there is a train passing the trackcircuit, so the peak voltage will vary according to the trainβsposition.) Since static information is not enough for locatingthe compensating capacitorβs fault, some other methodsshould be considered according to track circuitβs character-istics. It should be noticed that there are trains periodicallypassing through the track circuit. When a train is passingthrough the track circuit, the circuit structure of the trackcircuit will change, so more information about faults can beobtained. Based on this thinking, the peak voltages of receiverunder different shunt positions can be taken as fault features(Figure 7).
From Figure 7, it can be seen that different curves underdifferent fault states have remarkable differences. In fact, the
-
Mathematical Problems in Engineering 7
Table 5: Description of the diagnosis (classification) experimentbased on SVM.
Item DescriptionVector dimension(the number of samplingpoints in a curve)
156
The number of classes 17Training data set 1 vector per class (without noise)
Testing data set
50 vectors per class (with Β±0.005Vuniformly distributed measurementnoise and smoothing filtering with 3points)
SVM tool LIBSVM [29]
Scaling scheme [30]
Vectors in training data set are linearlyconsistently scaled to the range [0, 1],and testing data set has the samescaling proportion as the training dataset
SVM type C-SVM [29]Kernel type for SVM Linear kernelPenalty factor of theerror term in C-SVM πΆ = 1
Testing result Classification accuracy = 100%
minimal Euclidean distance (π = ββππ₯=1
(π
π
(π₯) β π
π
(π₯))
2
/π.Here, β1/π is taken as a constant factor for comparabilitywith the static situation) between these curves is 0.0113 V,which is remarkably larger than the static situation.
To verify the diagnosis effect based on this kind offault features, a diagnosis (classification) experiment basedon support vector machine (SVM) has been done. Thisexperiment is described by Table 5.
From the testing result in Table 5, it can be seen that,based on this kind of fault features, the diagnosis (classifica-tion) effect is very good.
3.1.4. Necessity of Adopting Dynamic Time Warping (DTW).Though the dynamic voltage information is promising togive a good diagnosis effect, there is still a big problem:the trainβs exact position information cannot be obtained bytrack circuit or its monitoring system. So, in Figure 7, thehorizontal axis cannot be βshunt position,β but βshunt time.βIn this case, if the train travels in a constant speed, the new(actual) curve will have the same shape with the originalcurve, so there is no problem. But if the train does not travelin a constant speed, except for start point and end point (atthe start point, there will be a fall edge, because the trainenters the track circuit. At the end point, there will be a risingedge, because the train leaves the track circuit), other pointsin new curve cannot be alignedwith the corresponding pointsin the original curve, resulting in a nonlinear stretchingdeformation of the curve in horizontal direction (shown inFigure 8).
In this case, the diagnosis (classification) effect based onthe actual deformed curves will become bad. To verify thispoint, the SVMclassification experiment has been performed
0 200 400 600 800 1000 1200 1400
Shunt position (m)
0.15
0.14
0.13
0.12
0.11
0.1
0.09
0.08
0.07
0.06
Peak
vol
tage
of r
ecei
ver (
V)
Original curve Deformed curve
Original curve
Deformedcurve
Figure 8: Original curve and deformed curve of fault state F3 (withnoise and smoothing filtering).
once again with the deformed testing data set (see theappendix for details). This time, the classification accuracydrops to 81.29%, which means there are a lot of incorrectdiagnosis results. So, it is necessary to adopt dynamic timewarping (DTW)method to overcome the deformation effect.
3.2. DTWMethod for Fault Diagnosis ofCompensating Capacitors
3.2.1. Introduction of DTW. Dynamic time warping (DTW)is a well-known dynamic-programming-based method tofind an optimal alignment between two given sequencesunder certain constraints. Intuitively, the sequences arewarped in a nonlinear fashion tomatch each other [22]. DTWwas originally used in speech recognition for comparingdifferent speech patterns [23]. DTW was also successfullyapplied in the fields of data mining and information retrieval,to cope with time deformations of time-dependent data [24,25].
Among main DTW methods, the classical DTW [22]aims at computing the DTW distance and optimal warpingpath using the dynamic programming method, which canbe realized by calculating cost matrix and accumulated costmatrix. Besides the classical DTW, some variations of DTW[22] aim at avoiding abnormal alignments and promotingcomputational efficiency by adding step size condition, localweights, and global constraints. The approximate DTW andmultiscale DTW [26, 27] were proposed to speed up DTWcomputations based on the idea to perform the alignment oncoarsened versions of the sequences π and π.
The related concepts and algorithms of the classical DTWare simply introduced as follows [22].
Objective of DTW. The objective of DTW is to findan optimal alignment between two given sequences:
-
8 Mathematical Problems in Engineering
π := (π₯
1
, π₯
2
, . . . , π₯
π
) of length π and π := (π¦1
, π¦
2
, . . . , π¦
π
)
of lengthπ.
Local Cost Measure and Cost Matrix. To measure the align-ment level of two different features π₯
π
, π¦
π
β πΉ, a local costmeasure π : πΉ Γ πΉ β π
β₯0
is needed. Based on the localcost measure, the cost matrix πΆ β π πΓπ can be defined byπΆ(π, π) := π(π₯
π
, π¦
π
).
Warping Path. An (π,π)-warping path (βwarping pathβ forshort) is a sequence π = (π
1
, π
2
, . . . , π
πΏ
) with ππ
= (π
π
, π
π
) β
[1 : π] Γ [1 : π], (π β [1 : πΏ]) satisfying the followingconditions:
(1) boundary condition: π1
= (1, 1) and ππΏ
= (π,π);
(2) monotonicity condition: π1
β€ π
2
β€ β β β β€ π
πΏ
andπ1
β€
π
2
β€ β β β β€ π
πΏ
;
(3) step size condition: ππ+1
β π
π
β {(1, 0), (0, 1), (1, 1)},(π β [1 : πΏ β 1]).
A warping path defines an alignment between sequenceπ and sequence π.
Total Cost. The total cost of a warping path is defined asπ
π
(π, π) := β
πΏ
π=1
π(π₯
ππ, π¦
ππ), (π
π
, π
π
) β π. The total costrepresents the alignment level of an alignment defined by thewarping path π.
DTW Distance. The minimum total cost is defined as DTWdistance:
πDTW (π, π) := min {ππ (π, π) | π is an
(π,π) -warping path} .(1)
Optimal Warping Path. The warping paths corresponding tothe DTW distance are defined as optimal warping paths,which represent the optimal alignments. An optimal warpingpath defines a path through the cost matrix, which has theminimum total cost.
The objective of DTW can be summarized as finding theDTW distance and the optimal warping paths.
Accumulated Cost Matrix. The accumulated cost matrix isused to calculate the DTW distance and the optimal warpingpaths. It is defined as follows:
π· (π,π) := πDTW (π (1 : π) , π (1 : π))
π (1 : π) := (π₯
1
, π₯
2
, . . . , π₯
π
) , π β [1 : π]
π (1 : π) := (π¦
1
, π¦
2
, . . . , π¦
π
) , π β [1 : π] .
(2)
Algorithm for DTW Distance. Based on the accumulated costmatrix, the algorithm for calculating DTW distance is asfollows:
π· (π, 1) =
π
β
π=1
π (π₯
π
, π¦
1
) , π β [1 : π]
π· (1,π) =
π
β
π=1
π (π₯
1
, π¦
π
) , π β [1 : π]
π· (π,π) = min {π· (π β 1,π β 1) , π· (π β 1,π) ,
π· (π,π β 1)} + π (π₯
π
, π¦
π
) ,
for π β [2 : π] , π β [2 : π]
πDTW (π, π) = π· (π,π) .
(3)
The computation complexity of this algorithm isπ(ππ).
Algorithm for Optimal Warping Path. Based on the accu-mulated cost matrix, the algorithm for calculating optimalwarping paths is as follows:
π
πΏ
= (π,π)
Suppose ππ
= (π,π)
if (π,π) = (1, 1) then π must be 1 and algorithm finishes
else
π
πβ1
=
{
{
{
{
{
{
{
{
{
(1,π β 1) , if π = 1(π β 1, 1) , if π = 1arg min {π· (π β 1,π β 1) ,
π· (π β 1,π) , π· (π,π β 1)} , otherwise.(4)
3.2.2. Basic DTWMethod for Fault Diagnosis of CompensatingCapacitors. Since the problem is to classify the deformedcurves into 17 classes (17 fault states) according to the 17training samples (see Table 5), instead of directly using SVM,the problem can be solved by using DTW between thedeformed curve and the 17 training sample curves. That is,the DTW distances between the deformed curve and the 17training sample curves can be calculated, and the trainingsample curve corresponding to the minimumDTW distancewill be the target class.
This algorithm can be formally described as follows:π
π
:= (π₯
π
1
, π₯
π
2
, . . . , π₯
π
π
) of length π is the training samplefor class π and π := (π¦
1
, π¦
2
, . . . , π¦
π
) of length π is thedeformed curve to be classified. Based on the above classicalDTW, the classification (diagnosis) result can be calculated asfollows:
π‘ = arg min {πDTW (ππ, π) | π β [1 : 17]} . (5)
According to this algorithm, the deformed curve inFigure 8 can be restored to approach the original curve(shown in Figure 9).
-
Mathematical Problems in Engineering 9
0.15
0.14
0.13
0.12
0.11
0.1
0.09
0.08
0.07
0.060 200 400 600 800 1000 1200 1400
Shunt position (m)
Original curve
Original curve
Restored curve
Restored curve
Peak
vol
tage
of r
ecei
ver (
V)
Figure 9: Original curve and restored curve of fault state F3 (withnoise and smoothing filtering).
Table 6: Description of the diagnosis (classification) experimentbased on DTW.
Item DescriptionVector dimension(the number of samplingpoints in a curve)
156
The number of classes 17Local cost measure π (π₯, π¦) =
π₯ β π¦
Training data set 1 vector per class (without noise)
Testing data set
50 vectors per class with:(i) Β±0.005V uniformly distributedmeasurement noise(ii) Smoothing filtering with 3 points(iii) Deformation (see Appendix fordetails)
Testing result Classification accuracy = 99.88%Computation complexity O (156 Γ 156 Γ 17 Γ (50 Γ 17))Classification stability(minimum)
β0.0429 (0.1795 for correctclassifications)
Classification stability(average) 0.9260
To verify the diagnosis effect of the DTW based method,the deformed testing data set in Section 3.1.4 is again used forthe verification experiment. The experiment is described inTable 6.
From the testing result in Table 6, it can be seen that, inspite of the deformation of testing data set, by theDTWbasedmethod, the classification accuracy can still reach 99.88%.In Section 3.1.4, the classification accuracy can only reach81.29% by SVMmethod.
Besides classification accuracy, in Table 6, the βclassi-fication stabilityβ is also used as an index for judging
Table 7: Description of the diagnosis (classification) experimentbased on DTW (with improved local cost measure).
Item Description
Local cost measure π (π₯, π¦) =
π₯ β π¦
2,π
(π) = 2
Testing result Classification accuracy =100%Classification stability (minimum) 0.4165Classification stability (average) 5.0881
the classification quality. The stability π for one classificationis defined as follows:
Suppose π belongs to class π‘, π‘ β [1 : 17]
π
1
= πDTW (ππ‘, π)
π
2
= min {πDTW (ππ, π) | π β [1 : 17] , π ΜΈ= π‘}
π =
(π
2
β π
1
)
min (π1
, π
2
)
.
(6)
If the classification result is incorrect, then π < 0;otherwise, π > 0. Regardless of π < 0 or π > 0, biggerπ implies better classification quality. π = 0 is a specialsituation, which means the minimum DTW distance is notunique.
3.2.3. Improvement onDTWMethod for ClassificationQuality.The classification quality of the above DTW method canstill be improved by simple modification of the local costmeasure. Since local cost measure is a cost measure forone step in a warping path, it can be considered to modifythe cost structure to encourage the correct alignment. Themodification scheme is as follows:
suppose π =
π₯ β π¦
,
π (π₯, π¦) = π (π) , s.t. π (π) > 0.(7)
The basic idea here is that π(π) should increase with πβsincrement, which means bigger π should have a higher costrate. This is because
(1) small π is tolerable in correct alignment, because itmay be caused by noise;
(2) big π is intolerable in correct alignment, because it isunlikely to be caused by noise.
A simple local cost measure satisfying the conditionπ
(π) > 0 is π(π₯, π¦) = |π₯ β π¦|2. The diagnosis (classification)experiment based on this local cost measure is described inTable 7 (other items are the same as Table 6).
From the testing result and classification stability inTable 7, it can be seen that, with the improved local cost mea-sure, there is a remarkable improvement on the classificationquality.
-
10 Mathematical Problems in Engineering
160
140
120
100
80
60
40
20
0
160140120100806040200
Steps in deformed curve (time)
The global constraint region
Step
s in
orig
inal
curv
e (di
spla
cem
ent)
Figure 10: The global constraint region based on the dynamiccharacteristics of CRH3.
Table 8: Related symbols for analysis of the boundary conditions ofthe global constraint region.
Symbol Descriptionπ₯ Abscissa in Figure 10π¦ Ordinate in Figure 10π Displacement (m)V Velocity (m/s)π‘ Time (s)π Acceleration (m/s2)V1
Start velocity on the track circuit (m/s)V2
End velocity on the track circuit (m/s)π Total time on the track circuit (s)πΏ Length of the track circuit (m)π Slope of curve in Figure 10π
1
Start slope of curve in Figure 10π
2
End slope of curve in Figure 10
3.2.4. Improvement on DTW Method for Computation Com-plexity. In order to reduce the computation complexity, theglobal constraints in variations of DTW [22] can be used.However, neither the Sakoe-Chiba band region [22] northe Itakura parallelogram region [22] is adopted for theglobal constraint. According to the characteristic of the trainβsdynamics (see the appendix for details), the following globalconstraint region (Figure 10), which is generated by boundarygears (maximum gear or minimum gear) and boundaryvelocities (V
1
= 0 or V2
= 0.), is adopted for the DTWmethodapplied in fault diagnosis of track circuit.
The boundary conditions of the global constraint regionin Figure 10 can be analyzed as follows (the related symbolsare defined in Table 8).
First, the lower boundary curve is considered. Accordingto the physical relationships between the above variables, thefollowing formulas can be given:
π‘ =
π₯π
160
,
π =
π¦πΏ
160
,
V =ππ
ππ‘
=
ππ¦
ππ₯
β
πΏ
π
,
β΄ π =
ππ¦
ππ₯
= V β π
πΏ
.
(8)
For reasonable simplification, the acceleration is assumedto be constant; then the following formulas can be given:
V22
β V21
= 2ππΏ
π =
2πΏ
V2
+ V1
V2
=β2ππΏ + V2
1
(9)
β΄ Ξπ = π
2
β π
1
= (V2
β V1
) β
π
πΏ
=
2ππΏ
V2
+ V1
β
2
V2
+ V1
=
4ππΏ
(V2
+ V1
)
2
=
4ππΏ
2ππΏ + 2V21
+ 2β2ππΏ + V21
β V1
=
4πΏ
2πΏ + 2V21
/π + 2β2πΏ/π + V21
/π
2
β V1
.
(10)
So, Ξπ will decrease when V1
is increasing and willincrease when π is increasing. Thus, when V
1
is minimumand π is maximum, the maximum Ξπ can be gotten, whichcorresponds to the maximum deformation of the lowerboundary curve.
So, the boundary conditions of the lower boundary curveare V1
= 0 and π = max. The upper boundary curve issymmetrical with the lower boundary curve. So the upperboundary conditions are V
2
= 0 and π = βmax.In the case of the global constraints, the elements beyond
the global constraint region in accumulated cost matrix doesnot need to be calculated and can be directly set as infinity, sothe computation complexity can be reduced to πregion/πsquare(area ratio) of the original computation complexity. Thediagnosis (classification) experiment based on this globalconstraint region is described in Table 9 (other items are thesame as Table 6).
From the results in Table 9, it can be seen that, withthe global constraint, there is a remarkable reduction inthe computation complexity, at the cost of one incorrectclassification.
-
Mathematical Problems in Engineering 11
Table 9: Description of the diagnosis (classification) experimentbased on DTW (with the global constraint).
Item DescriptionLocal cost measure π (π₯, π¦) =
π₯ β π¦
2, π(π) = 2Testing result Classification accuracy = 99.88%
Computation complexity O (29.26% Γ (156 Γ 156) Γ 17 Γ (50 Γ17))Classification stability(minimum)
β0.0468 (0.1299 for correctclassifications)
Classification stability(average) 6.5735
4. Results and Discussion
The results of the above experiments can be summarized inTable 10.
From Table 10, the results of all the experiments can beanalyzed as follows.
(1) The method of comparison of static voltages has verypoor classification quality and can hardly be used forfault location of compensating capacitors.
(2) The method of SVM without DTW has a generallyacceptable classification quality, but there are stillmany incorrect classifications, which will cause trou-ble in practical applications.
(3) The classical DTW has good classification accuracy,but the classification stability is poor, which meansincorrect classifications are prone to appear undersevere conditions such as big noise. Besides, the com-putation complexity of this method is great, whichmeans it is not suitable for online diagnosis.
(4) The DTWmethod with improved local cost measurehas very good classification accuracy and good classi-fication stability, but the computation complexity hasnot been improved.
(5) The DTW method with global constraint has goodclassification quality, great classification stability, andreduced computation complexity.The very few incor-rect classifications, which are not general situations,may be caused by the nonconservative global con-straint region. As a whole, this method is the bestone among these methods for online fault diagnosisof compensating capacitors in track circuit.
5. Conclusions
Aiming at the problem of online fault diagnosis for com-pensating capacitors of jointless track circuit, a dynamictime warping (DTW) based diagnosis method is proposedin this paper. Different from the existing related works,this method only uses the ground monitoring signals oftrack circuit to locate the faulty compensating capacitor, notdepending on the shunt current of inspection train, which isan indispensable condition for existing methods. So, it can be
used for online diagnosis of compensating capacitor, whichhas not yet been realized by existing methods. Besides, thismethod does not need the outdoor monitoring signals or theon-board monitoring signals, so the existing system does notneed to bemodified greatly, whichmeans a good engineeringfeasibility. To overcome the key problem that track circuitcannot obtain the precise position of the train, the DTWmethod is used for the first time in this situation to recoverthe function relationship between receiverβs peak voltage andshunt position. Besides the classical DTW based method,two improved methods for improving classification qualityand reducing computation complexity are proposed. Finally,the diagnosis experiments based on the simulation model oftrack circuit show the effectiveness of the proposed methods.Fault diagnosis for compensating capacitors under uncertainballast resistance and shunt resistance will be the future work.
Appendix
Generation of Deformed Testing Data Set
In order to generate the almost real deformed testing data set,an effective dynamic model of train must be first established.Referring to [28], the dynamic model (state equation) ofCRH3 high-speed train is established as follows:
πVπVππ₯
= πΉ
π
(V) β π (V)
ππ‘
ππ₯
=
1
V.
(A.1)
Here, π₯ instead of π‘ is taken as independent variablefor convenience of integral calculation at given displacementinterval [0, πΏ] (πΏ is the length of the track circuit).
The related symbols are described in Table 11.The traction characteristic of CRH3 high-speed train is
described as follows [28]:
πΉ
π
(V) =
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
{
π
π
(300 β 0.284V) kN,for 0 β€ V β€ 119.7 km/h, π > 0
π
π
(266 Γ 119.7/V) kN,for 119.7 < V β€ 300 km/h, π > 0
0 kN,for π = 0
π
π
(β300 + 0.281V) kN,for 0 β€ V β€ 106.7 km/h, π < 0
π
π
(β270 Γ 106.7/V) kN,for 106.7 < V β€ 300 km/h, π < 0,
(A.2)
where ππ
is the gear coefficient for CRH3, which is defined inTable 12.
The resistance characteristic of CRH3 high-speed train isdescribed as follows [28]:
π (V) = 6.4π + 130π + 0.14πV
+ [0.046 + 0.0065 (π β 1)] π΄V2.(A.3)
-
12 Mathematical Problems in Engineering
Table 10: Summary of results of all the experiments.
Method Classification accuracy Classification stability (average) Computation complexity(for one classification)
Comparison of static voltages Very low (not able to locate thefaults) NAβ NA
SVM (without DTW) 81.29% NA NAClassical DTW 99.88% 0.9260 O (156 Γ 156 Γ 17)DTW with improved local costmeasure 100% 5.0881 O (156 Γ 156 Γ 17)
DTW with global constraint 99.88% 6.5735 O (29.26% Γ 156 Γ 156 Γ 17)ββNAβ means not applicable.
Table 11: Related symbols in the dynamic model of train.
Symbol Descriptionπ₯ Displacement (m)V Velocity (m/s)π‘ Time (s)π Mass (kg, 408t for CRH3)π The number of cars (8 for CRH3)πΉ Tractive force or braking force (N)π Gearπ Resistance (N)π΄ Frontal area (m2, 9m2 for CRH3)π The number of axles (32 for CRH3)
Table 12: Gear coefficients for CRH3.
Gear Tractive gear Neutral gear Braking gear4 3 2 1 0 β1 β2 β3 β4
π
π
1 0.7 0.5 0.3 0 0.3 0.5 0.7 1
The differential equation solver βode45β in Matlab is usedfor solving the state equation with the initial condition β(V
0
,π‘
0
= 0)β and the integral interval [0, πΏ = 1301]. Forembodying randomness of samples, the gear π is randomlyselected in {β4, β3, β2, β1, 0, 1, 2, 3, 4}, and V
0
is randomlyselected in [Vmin, Vmax]. Vmin and Vmax are selected to ensureV β (0, 300 km/h] when the train is running on the trackcircuit.
The typical velocity-time curve and displacement-timecurve generated from the dynamic model are shown inFigures 11 and 12, which are corresponding to Figure 8.
Based on the displacement-time curves generated fromthe dynamic model, the deformed testing data set can begenerated from the original testing data set.
Conflict of Interests
The author declares that there is no conflict of interestsregarding the publication of this paper.
0 5 10 15 20 25 30 35 4020
25
30
35
40
45
Velo
city
(m/s
)
Time (s)
Figure 11: Typical velocity-time curve.
Time (s)0 5 10 15 20 25 30 35 400
200
400
600
800
1000
1200
1400
Disp
lace
men
t (m
)
Figure 12: Typical displacement-time curve.
Acknowledgments
Thiswork was supported by theNSFC under Grants 61104019and 61374123 and the Tsinghua University Initiative ScientificResearch Program.
-
Mathematical Problems in Engineering 13
References
[1] C. J. Baker, L. Chapman, A. Quinn, and K. Dobney, βClimatechange and the railway industry: a review,β Proceedings of theInstitution of Mechanical Engineers C: Journal of MechanicalEngineering Science, vol. 224, no. 3, pp. 519β528, 2010.
[2] F. P. GarcΜΔ±a MaΜrquez, R. W. Lewis, A. M. Tobias, and C.Roberts, βLife cycle costs for railway condition monitoring,βTransportation Research E: Logistics and Transportation Review,vol. 44, no. 6, pp. 1175β1187, 2008.
[3] J. Chen, C. Roberts, and P.Weston, βNeuro-fuzzy fault detectionand diagnosis for railway track circuits,β inProceedings of the 6thIFAC Symposium on Fault Detection, Supervision and Safety ofTechnical Processes (SAFEPROCESS β06), no. 1, pp. 1366β1371,September 2006.
[4] J. Chen, C. Roberts, and P. Weston, βFault detection anddiagnosis for railway track circuits using neuro-fuzzy systems,βControl Engineering Practice, vol. 16, no. 5, pp. 585β596, 2008.
[5] Y. Cheng, Reliability and Safety of Railway Signal System, ChinaRailway, 2010.
[6] Z.-W. Huang, X.-Y. Wei, and Z. Liu, βFault diagnosis of railwaytrack circuits using fuzzy neural network,β Journal of the ChinaRailway Society, vol. 34, no. 11, pp. 54β59, 2012.
[7] L. Oukhellou, A. Debiolles, T. DenΕux, and P. Aknin, βFaultdiagnosis in railway track circuits using Dempster-Shafer clas-sifier fusion,β Engineering Applications of Artificial Intelligence,vol. 23, no. 1, pp. 117β128, 2010.
[8] E. CoΜme, L. Oukhellou, T. DenΕux, and P. Aknin, βNoiselessindependent factor analysis with mixing constraints in a semi-supervised framework. application to railway device fault diag-nosis,β Lecture Notes in Computer Science (including subseriesLecture Notes in Artificial Intelligence and Lecture Notes inBioinformatics), vol. 5769, no. 2, pp. 416β425, 2009.
[9] Z. L. Cherfi, L. Oukhellou, E. CoΜme, T. DenΕux, and P. Aknin,βPartially supervised Independent Factor Analysis using softlabels elicited frommultiple experts: application to railway trackcircuit diagnosis,β Soft Computing, vol. 16, no. 5, pp. 741β754,2012.
[10] H. Zhao, Y. Liu, and B. Liu, βResearch on the onboard auto-testsystem for track circuit compensating capacitors,β in Proceed-ings of the 10th International Conference on Computer SystemDesign and Operation in the Railway and Other Transit Systems,COMPRAIL (CR β06), pp. 965β971, July 2006.
[11] L. Zhao, Z. Li, and W. Liu, βThe compensation capacitors faultdetection method of jointless track circuit based on DBWT andWR,β in Proceedings of the IEEE International Conference onIntelligent Computing and Intelligent Systems (ICIS β09), pp. 875β879, Shanghai, China, November 2009.
[12] L. Zhao, H. Li, J. Guo, and W. Liu, βThe simulation analysisof influence on jointless track circuit signal transmission fromcompensation capacitor based on transmission-line theory,βin Proceedings of the 3rd IEEE International Symposium onMicrowave, Antenna, Propagation and EMC Technologies forWireless Communications (MAPE β09), pp. 1113β1118, Beijing,China, October 2009.
[13] Z. Linhai, L. Baosheng, and X. Xun, βUsing the wavelet analysisbased noise canceling technique to detect the integrity of thecompensating capacitors of UM71 track circuit,β in Proceedingsof the IEEE International Symposium on Microwave, Antenna,Propagation and EMC Technologies for Wireless Communica-tions (MAPE β05), pp. 1268β1274, Beijing, China, August 2005.
[14] L. Zhao and J. Mu, βFault diagnosis method for jointless trackcircuit based on AOK-TFR,β Journal of Southwest JiaotongUniversity, vol. 46, no. 1, pp. 84β91, 2011.
[15] L.-H. Zhao, B.-G. Cai, K.-M. Qiu, and Z.-Q. Lu, βThe methodof diagnosis of compensation capacitor failures with jointlesstrack circuits based on HHT and DBWT,β Journal of the ChinaRailway Society, vol. 33, no. 3, pp. 49β54, 2011.
[16] L.-H. Zhao, J.-J. Xu, W.-N. Liu, and B.-G. Cai, βCompensationcapacitor fault detection method in jointless track circuitbased on Levenberg-Marquardt algorithm and generalized S-transform,β Control Theory and Applications, vol. 27, no. 12, pp.1612β1622, 2010.
[17] L. Zhao, Y. Ran, and J. Mu, βA comprehensive fault diagnosismethod for jointless track circuit based on genetic algorithm,βChina Railway Science, vol. 31, no. 3, pp. 107β114, 2010.
[18] Z. Lin-Hai, W. Jian-Ping, and R. Yi-Kui, βFault diagnosis fortrack circuit using AOK-TFRs and AGA,β Control EngineeringPractice, vol. 20, no. 12, pp. 1270β1280, 2012.
[19] S. P. Sun, H. B. Zhao, and G. Zhou, βA fault diagnosis approachfor railway track circuits trimming capacitors using EMDand teager energy operator,β WIT Transactions on the BuiltEnvironment, vol. 127, pp. 167β176, 2012.
[20] M. Guang-zhi and X. Xue-shu, βThe modeling and simulationof track circuit,β Electric Drive For Locomotives, no. 1, pp. 41β43,2004.
[21] Z. Yong-xian, X. Xue-song, and Y. Jiang-song, βModeling andsimulation of ZPW-2000A jointless track circuit,β Journal ofEast China Jiaotong University, no. 3, pp. 64β68, 2009.
[22] M. MuΜller, Information Retrieval For Music and Motion,Springer Press, 2007.
[23] L. R. Rabiner and B. H. Juang, Fundamentals of Speech Recogni-tion, Prentice Hall Signal Processing Series, 1993.
[24] E. Keogh, βExact indexing of dynamic time warping,β inProceedings of the 28th International Conference on Very LargeData Bases, pp. 406β417, Hong Kong, China, August 2002.
[25] S. Dick, A. Meeks, M. Last, H. Bunke, and A. Kandel, βDatamining in software metrics databases,β Fuzzy Sets and Systems,vol. 145, no. 1, pp. 81β110, 2004.
[26] S. Chu, E. Keogh, D. Hart, andM. Pazzani, βIterative deepeningdynamic time warping for time series,β in Proceedings of the 2ndSIAM International Conference on Data Mining, pp. 195β212,Chicago, Ill, USA, April 2002.
[27] S. . Salvador and P. Chan, βFastDTW: toward accurate dynamictime warping in linear time and space,β in Proceedings of KDDWorkshop on Mining Temporal and Sequential Data, Seattle,Wash, USA, August 2004.
[28] L. Liang, D.Wei, J. Yindong, Z. Zengke, and T. Lang, βMinimal-energy driving strategy for high-speed electric train with hybridsystem model,β IEEE Transactions on Intelligent TransportationSystems, vol. 14, no. 4, pp. 1642β1653, 2013.
[29] C.-C. Chang and C.-J. Lin, βLIBSVM: a library for supportvector machines,β ACM Transactions on Intelligent Systems andTechnology, vol. 2, no. 3, article 27, 2011.
[30] C.-W. Hsu, C.-C. Chang, and C.-J. Lin, βA practical guide tosupport vector classification,β http://www.csie.ntu.edu.tw.
-
Submit your manuscripts athttp://www.hindawi.com
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttp://www.hindawi.com
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
CombinatoricsHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
International Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
The Scientific World JournalHindawi Publishing Corporation http://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com
Volume 2014 Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Stochastic AnalysisInternational Journal of