Fault Detection and Diagnosis of Railway Point Machines by ... - MDPI

sensors Article

Fault Detection and Diagnosis of Railway Point Machines by Sound Analysis Jonguk Lee 1 , Heesu Choi 1 , Daihee Park 1, *, Yongwha Chung 1 , Hee-Young Kim 2 and Sukhan Yoon 3 1

2 3

*

Department of Computer and Information Science, Korea University, Sejong Campus, Sejong City 30019, Korea; [email protected] (J.L.); [email protected] (H.C.); [email protected] (Y.C.) Department of Applied Statistics, Korea University, Sejong Campus, Sejong City 30019, Korea; [email protected] Sehwa R&D Center, Techno 2-ro, Yuseong-gu, Daejeon 34026, Korea; [email protected] Correspondence: [email protected]; Tel.: +82-44-860-1344; Fax: +82-44-860-1584

Academic Editor: Vittorio M. N. Passaro Received: 12 February 2016; Accepted: 12 April 2016; Published: 16 April 2016

Abstract: Railway point devices act as actuators that provide different routes to trains by driving switchblades from the current position to the opposite one. Point failure can significantly affect railway operations, with potentially disastrous consequences. Therefore, early detection of anomalies is critical for monitoring and managing the condition of rail infrastructure. We present a data mining solution that utilizes audio data to efficiently detect and diagnose faults in railway condition monitoring systems. The system enables extracting mel-frequency cepstrum coefficients (MFCCs) from audio data with reduced feature dimensions using attribute subset selection, and employs support vector machines (SVMs) for early detection and classification of anomalies. Experimental results show that the system enables cost-effective detection and diagnosis of faults using a cheap microphone, with accuracy exceeding 94.1% whether used alone or in combination with other known methods. Keywords: railway point machine; railway condition monitoring system; audio data; support vector machine

1. Introduction Railway points provide different routes to trains, by driving switchblades between various predetermined positions. The failure of railway points can significantly affect train operations [1]. Consequently, early detection of anomalies is critical for managing railway condition monitoring systems. Technologies for collecting and analyzing data from railway point machinery should be developed in order to minimize detrimental effects of point failure. Many railway condition monitoring systems are equipped with alarms that apply thresholds to electrical sensor readings. However, the application of thresholds does not ensure early detection of faults [2]. In addition to techniques based on electrical thresholds, the literature includes a wide variety of methods for detecting faults in railway points, including statistical analysis, classification, and model-based methods [3–7]. In particular, classification methods are widely used for detecting faults in a variety of point machinery [2]. Several recent studies reported on SVM-based classification methods [8–10] using electrical signals. Eker et al. [11] detected faults using principal component analysis together with SVM based on the measurements of a linear ruler, and classified 20 railway point system operations as either fault-free or indicative of drive misalignment. Asada et al. [1] showed that current and voltage sensors can be used

Sensors 2016, 16, 549; doi:10.3390/s16040549

www.mdpi.com/journal/sensors

Sensors 2016, 16, 549

2 of 12

Sensors 2016, 2016, 16, 16, 549 549 Sensors

of 11 11 22 of

to collect electrical active power data for railway condition monitoring systems. They reported that reported that use the of combined use of of wavelet wavelet transforms andquite SVMs enabled quite accurate accurate detection the combined wavelet use transforms and transforms SVMs enabled accurate detection and diagnosis of reported that the combined and SVMs enabled quite detection and diagnosis of misalignment faults in electrical railway point machinery. Vileiniskis et al. [2] misalignment in electrical railway machinery. Vileiniskis al. [2] presented a methodology and diagnosisfaults of misalignment faults point in electrical railway point etmachinery. Vileiniskis et al. [2] presented methodology for early warning of possible possible point failure failure through early detection of for early warning of possiblefor point failure through early detection of changes in theearly current drawn by presented aa methodology early warning of point through detection of changes in the current current drawn byaccurate the point point motor, which used was more more accurate than than commonly used the pointin motor, which was more than commonly threshold-based methods. Although changes the drawn by the motor, which was accurate commonly used threshold-based methods. Although there has been some recent progress in monitoring railway point there has been some recentAlthough progress in monitoring pointprogress systems in using electricalrailway signalspoint such threshold-based methods. there has beenrailway some recent monitoring systems using electricaltosignals signals such as current current and voltage, voltage, to the the best of of ouremployed knowledge no previous previous as current and voltage, the best of our knowledge no previous studies have audio sensors systems using electrical such as and to best our knowledge no studies have employed employed audio sensors for automated automated investigation investigation of of anomalies. anomalies. for automated investigation ofsensors anomalies. studies have audio for Unlike other current research approaches, this paper puts forth data mining mining solution solution that that Unlike other current research approaches, this paper puts forth aa data mining employs audio audio data data to to detect detect faults faults in in railway railwaycondition conditionmonitoring monitoringsystems. systems. Firstly, Firstly, in in the the employs data to detect faults in railway condition monitoring systems. data-preprocessing phase, MFCC is extracted and feature dimensions are reduced. Two SVMs are data-preprocessing phase, MFCC is extracted and feature dimensions are reduced. Two SVMs are used to detect and diagnose fault sounds, respectively. Experimental results show that this method used to detect and diagnose fault sounds, respectively. respectively. Experimental results show that this method enables cost-effective cost-effective detection detection and and diagnosis diagnosis of of faults faults achieving achieving high high accuracy accuracy levels levels of of 94.1% 94.1% for for enables detection, and 97.0% for diagnosis using a cheap microphone. This is the first study on the detection detection, and 97.0% for diagnosis using a cheap microphone. This is the first study on the detection and diagnosis diagnosis of of faults faults in in railway railway condition condition monitoring monitoring systems systems via via audio audiodata. data. The The results results indicate indicate and that acoustic analysis of railway sounds can be a reliable method for understanding the condition of that acoustic analysis of railway sounds can be a reliable method for understanding the condition of railway point machinery. The remainder of this article is structure as follows: Section 2 describes the railway point machinery. machinery. The remainder remainder of this article is structure structure as as follows: follows: Section 2 describes the proposed fault fault detection detection and and diagnosis diagnosis of of railway railway point point machines, machines, Section Section 33 presents presents the the results results of of proposed simulations, and Section 4 draws the conclusions. simulations, and Section 4 draws the conclusions. conclusions.

2. Fault Fault Detection and Diagnosis of Railway Point Machines by Audio Audio Analysis Analysis 2. FaultDetection Detectionand andDiagnosis Diagnosisof ofRailway RailwayPoint PointMachines Machinesby 2. The proposed proposedreal-time real-timesystem systemconsists consistsof offour fourmodules: modules: two two online online process process modules modules consisting consisting The process The proposed real-time system consists of four modules: two online modules consisting ofaaafeature feature extraction module and fault detector detector module, and twoprocess offlinemodules processconsisting modules of extraction module andand a fault detector module, and twoand offline of feature extraction module aa fault module, two offline process modules consisting of an attribute subset-selection module and an SVM training module (refer to Figure 1). of an attribute subset-selection module and an SVM training module (refer to Figure 1). The feature consisting of an attribute subset-selection module and an SVM training module (refer to Figure 1). The feature extraction module is based on the MFCC algorithm [12–17] (refer to Figure 2). extraction is based on the MFCC [12–17] (refer[12–17] to Figure 2). to Figure 2). The featuremodule extraction module is based onalgorithm the MFCC algorithm (refer

Figure 1. Overall structure for proposed Figure 1. 1. Overall Overall structure structure for for proposed proposed fault fault detection detection and anddiagnosis diagnosissystem systemby byaudio audioanalysis. analysis. Figure fault detection and diagnosis system by audio analysis.

Figure 2. 2. MFCC extraction extraction procedure procedure [17]. [17]. Figure Figure 2. MFCC MFCC extraction procedure [17].

The attribute attribute subset-selection subset-selection module module is is used used to to select select the the optimal optimal feature feature subset subset with with aa view view to to The improving the detection and classification speed of the entire diagnosis system. This study uses improving the detection and classification speed of the entire diagnosis system. This study uses correlation-based feature feature selection selection (CFS), (CFS), which which is is one one of of the the most most popular popular attribute attribute subset-selection subset-selection correlation-based

Sensors 2016, 16, 549

3 of 12

The attribute subset-selection module is used to select the optimal feature subset with a view detection and classification speed of the entire diagnosis system. This study3uses of 11 correlation-based feature selection (CFS), which is one of the most popular attribute subset-selection methods [18–21] (Figure 3). Following Following training, training, the the fault fault detection detection and and classification classification module module detects detects fault sounds by identifying incoming audio signals and classifying classifying them them as subsidiary subsidiary fault-sound fault-sound types such as “ice “ice obstruction”, obstruction”, “ballast obstruction”, or “slackened “slackened nut”. Although Although the the SVM SVM training training module intendedtoto perform training offline based onMFCC the MFCC and the isprocess is not module isisintended perform training offline based on the and CFS, theCFS, process not necessary necessary during the online process. during the online process. to improving the Sensors 2016, 16, 549

Figure Figure 3. 3. Feature Feature dimensionality dimensionality reduction reduction by by CFS CFS [20]. [20].

2.1. 2.1. Mel-frequency Mel-frequency Cepstrum Cepstrum Coefficients Coefficients The obtain the the sequence sequence of feature vectors, vectors, thereby The main main purpose purpose of of feature feature extraction extraction is is to to obtain of feature thereby providing a compact representation from the raw input signal [12]. Sound analysis providing a compact representation from the raw input signal [12]. Sound analysis research research has has investigated various acoustic features for use in signal analysis, such as perceptual linear prediction investigated various acoustic features for use in signal analysis, such as perceptual linear prediction (PLP) features, features,linear linearprediction prediction cepstral coefficients (LPCC), and MFCC. In particular, is (PLP) cepstral coefficients (LPCC), and MFCC. In particular, MFCC MFCC is widely widely used in automatic speech recognition and audio analysis, with its simple processing, used in automatic speech recognition and audio analysis, with its simple processing, outstanding outstanding ability,both containing time and frequency and information, and other[12–17]. advantages [12–17]. ability, containing time andboth frequency information, other advantages Additionally, Additionally, MFCC has been successfully applied to fault diagnosis of engines [22], early MFCC has been successfully applied to fault diagnosis of engines [22], early classifications of bearing classifications bearing faults [23], and quality assurance sound signaling devices [24]. Figure 2 faults [23], andofquality assurance of sound signaling devicesof[24]. Figure 2 shows a structure diagram shows a structure diagram for MFCC extraction. Firstly, pre-emphasis filtering is used to spectrally for MFCC extraction. Firstly, pre-emphasis filtering is used to spectrally flatten the signal. Secondly, flatten the signal. Secondly, the short-time Fourier to transform (STFT) is applied to and extract information the short-time Fourier transform (STFT) is applied extract information on time frequency from on time and frequency from the input signal. In the mel-frequency wrapping step, the frequency is the input signal. In the mel-frequency wrapping step, the frequency is changed from Hz to mel scale. changed from Hz to mel scale. Then, the mel-frequency signal is converted by the logarithmic Then, the mel-frequency signal is converted by the logarithmic mel-spectrum back to the time domain. mel-spectrum to transform the time domain. cosine transform (DCT) is applied to the Finally, discreteback cosine (DCT) isFinally, applieddiscrete to the log-mel-frequency. log-mel-frequency. 2.2. Correlation-Based Feature Selection 2.2. Correlation-Based Feature Selection The literature includes a wide variety of feature-selection methods, including CFS, gain ratio The literature includesanalysis a wide(PCA), varietyetc. of This feature-selection methods, including CFS, gain ratio (GR), principal component study uses CFS, which is one of the most popular (GR), principal component analysis (PCA), etc. This study uses CFS, which is one of the most popular attribute subset-selection methods [18–21]. The main objective of CFS is to obtain the highly relevant attribute [18–21]. The main objective of CFS is toway, obtain highly relevant subset ofsubset-selection features, which methods are uncorrelated to each other [18–21]. In this thethe dimensionality of subset of features, which are uncorrelated to each other [18–21]. In this way, the dimensionality of data sets can be drastically reduced and the performance of learning algorithms can be maintained or data sets can beemploys drastically reduced and the performance can be maintained improved. CFS heuristic evaluation of the worth of or learning merit of aalgorithms subset of features. The merit or improved. CFS employs heuristic evaluation of thefor worth or merit ofclass a subset ofalong features. function considers the usability of individual features predicting the label, withThe themerit level function considers the usability of individual features for predicting the class label, along with of inter-correlation among them [18–21] (refer to Equation (1)). Figure 3 shows a structure diagramthe of level of inter-correlation among them [18–21] (referfeature–class to Equationand (1)).feature–feature Figure 3 shows structure CFS processing. Firstly, feature correlations between areacalculated diagram of CFS processing. Firstly, feature correlations between feature–class and feature–feature are calculated using symmetrical uncertainty, and then search the feature subset space. After estimating symmetrical uncertainty, the merit of a subset is calculated to find the subset with the highest-ranked merit value:

Sensors 2016, 16, 549

4 of 12

using symmetrical uncertainty, and then search the feature subset space. After estimating symmetrical uncertainty, Sensors 2016, 16,the 549merit of a subset is calculated to find the subset with the highest-ranked merit value: 4 of 11 nrc f Merit F “ b = n ` n pn ´ 1q r f f + 1

(1) (1)

Equation(1), (1), a feature subset F contains n features; and represent average In Equation a feature subset F contains n features; rc f and r f f represent average feature–class feature–class correlation and average feature–feature correlation, respectively. correlation and average feature–feature correlation, respectively. 2.3. Support Vector Vector Machine 2.3. Support Machine This This section section presents presents aa brief brief review review of of SVMs SVMs [8–11]. [8–11]. Figure Figure 44 shows shows the the approach approach to to identifying identifying t x ` b “ 0q with maximum margin for linearly separable classifier in a the optimal hyperplane (w the optimal hyperplane ( + = 0 with maximum margin for linearly separable classifier in a geometrical geometrical view view of of SVM. SVM.

Figure 4. Main concept of SVM in a linearly separable case [25]. Figure 4. Main concept of SVM in a linearly separable case [25].

be the training set and let In the linearly separable case, let , ,…, ∈ +1, 1 be the class tx u In the linearly separable case, let , x , . . . , x be the training set and let yi P to t`1, ´1u be z 2 1 label of a D-dimensional feature vector . The margin maximization problem corresponds [8–11]: the class label of a D-dimensional feature vector xi . The margin maximization problem corresponds 1 to [8–11]:  +„ min z 2 min 1 w T w ` C ř ξ (2) i 2 (2) i “ 1 . .` T +˘ 1 ; 0; = 1, … , s.t.yi w xi ` b ě 1 ´ ξ i ; ξ i ě 0; i “ 1, . . . , z Here, is a penalty for misclassification or classification within the margin, and 0 is a Here, ξ i is a penalty for misclassification or classification within the margin, and CpC ą 0q is a tradeoff parameter between error term and margin. The approach described here for a linear SVM tradeoff parameter between error term and margin. The approach described here for a linear SVM can can be extended to the creation of a nonlinear SVM in order to classify linearly inseparable data. be extended to the creation of a nonlinear SVM in order to classify linearly inseparable data. Figure 5 illustrates a solution to the non-linearly separable problem to obtain linear separation by mapping the input training data into the higher-dimensional feature space [8–11]. In the general ` ˘ ` ˘ mathematical formulation, the kernel function, K, is defined as K xi , x j ” φ pxi qT φ x j . In particular, the commonly used kernel function is a radial basis function (RBF) as follows: ´ ¯ ` ˘ K xi , x j “ exp ´γ ||xi ´ x j ||2 , γ ą 0

Figure 5. Graphical view of the SVM in the non-linearly separable case [25].

(3)

2

(2)

. .

+

1

;

0; = 1, … ,

Here, is a penalty for misclassification or classification within the margin, and 0 is a tradeoff parameter between error term and margin. The approach described here for a linear SVM 2016, 16, 549 5 of 12 canSensors be extended to the creation of a nonlinear SVM in order to classify linearly inseparable data.

Sensors 2016, 16, 549

5 of 11

Figure 5 illustrates a solution to the non-linearly separable problem to obtain linear separation by mapping the input training data into the higher-dimensional feature space [8–11]. In the general mathematical formulation, the kernel function, , is defined as , ≡ . In particular, the commonly used kernel function is a radial basis function (RBF) as follows: ,

=

‖

,

0

Figure Graphicalview viewofofthe theSVM SVMin in the the non-linearly non-linearly separable Figure 5. 5. Graphical separablecase case[25]. [25].

Here,

(3)

is a standard deviation parameter [25,26].

Here, γ is a standard deviation parameter [25,26].

3. Results

3. Results

3.1. Data Collection 3.1. Data Collection Audio datadata were collected railwaypoint point machine at Sehwa Company Audio were collectedfrom froman anNS-AM-type NS-AM-type railway machine at Sehwa Company in in Daejeon, South Korea, on 1 January 2016. Figure 6 shows a picture of an NS-AM-type railway Daejeon, South Korea, on 1 January 2016. Figure 6 shows a picture of an NS-AM-type railway pointpoint machine, which isisinstalled withanan audio sensor for this data collection Figure 7 machine, which installed with audio sensor for this data collection experiment.experiment. Figure 7 provides a schematic of the type of type NS-AM railway points used in Korea. typesseveral of fault types can of provides a schematic of the of NS-AM railway points usedIningeneral, Korea. several In general, point faultlead can to lead to failure. point failure.

Figure 6. 6. Data NS-AM-typerailway railway points. Figure Datacollection collection from from NS-AM-type points.

Sensors 2016, 16, 549

Figure 6. Data collection from NS-AM-type railway points.

6 of 12

Figure 7. Schematic for NS-AM-type railway points [1].

Figure 7. Schematic for NS-AM-type railway points [1]. Sensors 2016, 16, 549

6 of 11

Figure 8 shows a fishbone diagram for point failure [7]. As shown in Figure 8, we collected Figure 8 shows a fishbone diagram for point failure [7]. As shown in Figure 8, we collected audio audio data while simulating three fault conditions that include normal data: “ice obstruction”, “ballast data while simulating three fault include normal “ice obstructions obstruction”, “ballast obstruction”, and “slackened nut”conditions (see Figurethat 9). The first two casesdata: concern between obstruction”, and “slackened nut” (see Figure 9). The first two cases concern obstructions between the stock rail and switchblade of the track points. The “slackened nut” scenario may occur when nuts the stock rail and switchblade of the track points. The “slackened nut” scenario may occur when nuts become throughtrain trainvibration, vibration,orormaintenance maintenance misalignment. Apart becomeloose loosedue duetotoaanatural natural process, process, through misalignment. Apart from the faults simulated during data collection, to avoid significant faults, a maintenance task was from the faults simulated during data collection, to avoid significant faults, a maintenance task was performed performedbefore beforecollecting collecting the the data. data.

Figure8.8.Fishbone Fishbone diagram of [7].[7]. Figure of faults faultsfor forrailway railwaypoint pointmachines machines

Sensors 2016, 16, 549

Figure 8. Fishbone diagram of faults for railway point machines [7].

7 of 12

9. Data collection pictures fromsimulating simulating three three fault (a) (a) ice obstruction; (b) ballast FigureFigure 9. Data collection pictures from faultconditions: conditions: ice obstruction; (b) ballast obstruction; andslackened (c) slackened nut. obstruction; and (c) nut.

The sounds emitted by the railway points were recorded using a SHURE SM137 microphone

The sounds emitted by the railway within pointsone were recorded using(see a SHURE SM137 microphone (Shure Inc., Niles, IL, USA) positioned meter of the points Figure 6), and recorded (Shureonto Inc.,aNiles, IL, USA) positioned within one meter the points 6), and recorded Samsung NT-SF310 notebook computer. AdobeofAudition 3.0 (see and Figure R package “tuneR” [17] onto a Samsung NT-SF310 notebook Adobe 3.0 and R package [17] software software were used to digitizecomputer. the recorded signalsAudition in a personal computer with an“tuneR” AC97 soundcard (Realtek, Hsinchu, at 16 bits/44.1 sampling rates. Empirical analysis of the sound were used to digitize theTaiwan) recorded signals in akHz personal computer with an AC97 soundcard (Realtek, spectrogram revealed that ambient (or background) noise existed mainly between 0 and 300 Hz, and Hsinchu, Taiwan) at 16 bits/44.1 kHz sampling rates. Empirical analysis of the sound spectrogram thatthat the operational noise of the point machine occurredmainly from 300 to 13,0000Hz. Thus, revealed ambient (or background) noise existed between and 300noise Hz, filtering and that the process was performed in a passband of 300–13,000 Hz. Figure 10 illustrates the spectrograms and operational noise of the point machine occurred from 300 to 13,000 Hz. Thus, noise filtering process waveforms of standard and various fault sound models using Praat software (Ver. 6.0.05) [27]. was performed in a passband of 300–13,000 Hz. Figure 10 illustrates the spectrograms and waveforms of standard and16,various fault sound models using Praat software (Ver. 6.0.05) [27]. Sensors 2016, 549 7 of 11

10. Waveforms spectrogramsfor forfault fault and and normal samples (the(the rectangle indicates FigureFigure 10. Waveforms andand spectrograms normalsound sound samples rectangle indicates the sound pattern changes at the point of contact between the rod and the obstruction): (a) normal; the sound pattern changes at the point of contact between the rod and the obstruction): (a) normal; (b) ice obstruction; (c) ballast obstruction; and (d) slackened nut. (b) ice obstruction; (c) ballast obstruction; and (d) slackened nut.

3.2. Fault Sound Detection and Classification Results Two different experiments were performed (i.e., one for fault detection with the whole data set, the other for fault classification only using the data labelled as faulty). Figures 11 and 12 show the overall architecture of the SVM-based fault sound detection (a binary-class SVM) and classification system (a multi-class SVM), respectively. The proposed system was implemented using a PC (Intel

Sensors 2016, 16, 549

8 of 12

3.2. Fault Sound Detection and Classification Results Two different experiments were performed (i.e., one for fault detection with the whole data set, the other for fault classification only using the data labelled as faulty). Figures 11 and 12 show the overall architecture of the SVM-based fault sound detection (a binary-class SVM) and classification system (a multi-class SVM), respectively. The proposed system was implemented using a PC (Intel i7-3770K, 16 GB memory), and the experiments used the Weka [28]. In addition, ten-fold cross-validation with ten repetitions was used. The experiment used 430 fault sound data (140 for “ice obstruction”, 140 for “ballast obstruction”, and 150 for “slackened nut”) and 150 normal sound data. The data set is divided into a training set consisting of half of the original set (randomly chosen), with the other half used as a Sensors 2016, set. 16, 549 8 of 11 validation Sensors 2016, 16, 549 8 of 11

Figure 11. Architecture for fault-sound detection based on a binary SVM. Figure 11. Architecture for Figure 11. Architecture for fault-sound fault-sound detection detection based based on on aa binary binary SVM. SVM.

Figure 12. Architecture for fault-sound classification based on a multi-class SVM. Figure 12. Architecture for fault-sound classification based on a multi-class SVM. SVM.

First, an identification test of the proposed mechanism was conducted, to distinguish between First,the an MFCC identification test60 of frames the proposed mechanism conducted, to distinguish between features, perperformance sound and of 12was cepstral coefficients were used, and faultFor and normal sounds (see Figure 11). The the proposed system was evaluated via fault and normal sounds (see Figure 11). The performance of the proposed system was evaluated via 720-dimensional features (12 ˆ 60 = 720) were yielded by using tuneR. The lowest and highest fault detection rate (FDR), false positive rate (FPR), and false negative rate (FNR) [29,30]: True fault frequencies detection rate (FDR), positive (FPR), and whereas false negative rate (FNR) [29,30]: True band set correctly to false 300 and 13,000 rate Hzasrespectively, other parameters were set to positive (TP: faultwere sound identified fault), False positivethe (FP: normal sound incorrectly positive (TP: fault sound correctly identified as fault), False positive (FP: normal sound incorrectly default values. In the True case of CFS, the(TN: dimension the selected optimal-attribute was reduced identified as fault), negative normalofsound correctly identified as subsets normal), and False identified as“CfsSubsetEval” fault), True negative (TN: normal sound correctly identified as normal), and False to 133 using in Weka. negative (FN: fault sound incorrectly identified as normal): negative fault sound incorrectly normal): was conducted, to distinguish between First,(FN: an identification test of the identified proposed as mechanism fault and normal sounds (see Figure via 11). The performance = of the proposed × 100 system was evaluated(4) + (4) = × 100 fault detection rate (FDR), false positive rate (FPR), and false negative rate (FNR) [29,30]: True positive + (TP: fault sound correctly identified as fault), False positive sound incorrectly identified(5)as × 100 = (FP: normal + (5) × 100 = + (6) × 100 = + (6) × 100 = + A summary of detection results for fault sounds is shown in Table 1. According to the A summary of detection results for fault sounds is shown in Table 1. According to the experimental results, when using 720 feature vectors, the fault detection accuracy of the proposed experimental results, when using 720 feature vectors, the fault detection accuracy of the proposed system is 94.1%, and FPR and FNR are 0.6% and 5.9% respectively. Even when only 133 attributes are system is 94.1%, and FPR and FNR are 0.6% and 5.9% respectively. Even when only 133 attributes are used, the accuracy is confirmed as satisfactory. We used the corrected resampled t-test provided by used, the accuracy is confirmed as satisfactory. We used the corrected resampled t-test provided by Weka, with a 95% confidence level, to compare the methods based on the FDR. The results show no Weka, with a 95% confidence level, to compare the methods based on the FDR. The results show no

Sensors 2016, 16, 549

9 of 12

fault), True negative (TN: normal sound correctly identified as normal), and False negative (FN: fault sound incorrectly identified as normal): Fault Detection Rate pFDRq “

TP ˆ 100 TP ` FN

(4)

FP ˆ 100 FP ` TN

False Positive Rate pFPRq “ False Negative Rate pFNRq “

(5)

FN ˆ 100 TP ` FN

(6)

A summary of detection results for fault sounds is shown in Table 1. According to the experimental results, when using 720 feature vectors, the fault detection accuracy of the proposed system is 94.1%, and FPR and FNR are 0.6% and 5.9% respectively. Even when only 133 attributes are used, the accuracy is confirmed as satisfactory. We used the corrected resampled t-test provided by Weka, with a 95% confidence level, to compare the methods based on the FDR. The results show no significant difference between the 133 and the 720 features. Figure 11 indicates that the detector used in this experiment is a binary SVM. In case of using the entire feature-set, an RBF kernel with 0.0275 gamma was used and C was set at 1.7 for this cross-validation experiment. For CFS, an RBF kernel with 0.061 gamma was used and C was set at 1.91. These values were independently chosen by a GridSearch method in a training phase [28]. Our review of the literature did not identify any previous attempts to detect and classify fault sounds, thus a performance comparison cannot be made. Table 1. Performance of proposed system in fault sound detection. CFS (133 Features Used) FDR 94.3%

FPR 2.7%

All Features (720 Used)

FNR 5.6%

FDR 94.1%

FPR 0.6%

FNR 5.9%

Secondly, we classified fault sound data into three types: “ice obstruction”, “ballast obstruction”, and “slackened nut” (see Figure 12). In order to measure the classification accuracy of the proposed system, the precision and recall are used as the performance measurements [29,30]: Precision “ Recall “

TP ˆ 100 TP ` FP

TP ˆ 100 TP ` FN

(7) (8)

A summary of the classification results for the studied fault sounds is shown in Table 2. The experimental results show that the precision and recall of the proposed system approach 97.0% when using 720 feature vectors, compared with 93.1% and 93.0% when only 133 features are used. The corrected resampled t-test provided by Weka (95% confidence level) was used to compare the methods, showing that precision and recall were significantly better when using the entire features-set than with the 133 features used by CFS. The classifiers were branded as a multi-class SVM (Figure 12). When using the entire feature-set, an RBF kernel with 0.0157 gamma was used and C was set at 1.42. For CFS, an RBF kernel with 0.1073 gamma was used and C was set at 5.45. These values were also independently chosen by a GridSearch method in a training phase.

Sensors 2016, 16, 549

10 of 12

Table 2. Performance measurement for fault-sound classification.

Faults Ice obstruction Ballast obstruction Slackened nut Average

CFS(133 Features Used)

All Features (720 Used)

Precision

Recall

Precision

Recall

90.8% 91.1% 97.4% 93.1%

88.1% 91.1% 99.9% 93.0%

94.2% 97.2% 99.7% 97.0% V

97.3% 93.7% 100% 97.0% V

V

The method is significantly better.

4. Conclusions The early discovery of anomalies is critical for systems that monitor the condition of railway infrastructure. Failure to uncover faults in a timely and precise manner can become a critical limiting factor in efficiently managing such systems. This work thus presents a timesaving data mining solution for identifying faults through the use of audio data. The railway sound-acquisition process was performed first, while MFCC was isolated from the data-preprocessing segment. Two SVMs were used in the detection and classification of fault sounds, respectively. The experimental results demonstrated cost-effective, automatic detection and diagnosis of railway faults through the analysis of audio data. The combination of MFCC and SVM identified and classified the sounds of railway faults with accuracies of 94.1% and 97.0% respectively. The results confirm that the proposed method provides a credible means of investigating railway sounds for understanding the condition of rail points, whether used alone or in combination with other known methods. Broader testing of the proposed system in commercial production conditions is a purposeful avenue. A complete real-time system is part of our ongoing research. Acknowledgments: The study and contribution were supported by the project: Small & Medium Business Administration under Project S2312692 “Technological Innovation Development Business” for the innovative company in the year 2015. Author Contributions: Jonguk Lee, Daihee Park, Yongwha Chung, Hee-Young Kim, and Sukhan Yoon designed the overall structure of the proposed system for fault-detection and diagnosis via audio analysis. In addition, they wrote and revised the paper. Jonguk Lee, Daihee Park, and Heesu Choi collected sound data, implemented the proposed method, and helped with the experiments. Conflicts of Interest: The authors declare no conflicts of interest.

References 1. 2. 3.

4.

5. 6.

7.

Asada, T.; Roberts, C.; Koseki, T. An algorithm for improved performance of railway condition monitoring equipment: Alternating-current point machine case study. Transp. Res. C Emerg. Technol. 2013, 30, 81–92. Vileiniskis, M.; Remenyte-Prescott, R.; Rama, D. A fault detection method for railway point systems. Proc. Inst. Mech. Eng. F J. Rail Rapid Transit 2015. [CrossRef] Ardakani, H.D.; Lucas, C.; Siegel, D.; Chang, S.; Dersin, P.; Bonnet, B.; Lee, J. PHM for railway system—A case study on the health assessment of the point machines. In Proceedings of the 2012 IEEE Conference on Prognostics and Health Management (PHM), Denver, CO, USA, 18–21 June 2012; pp. 1–5. Bolbolamiri, N.; Sanai, M.S.; Mirabadi, A. Time-domain stator current condition monitoring: Analyzing point failures detection by Kolmogorov-Smirnov (K-S) test. Int. J. Electr. Comput. Energ. Electron. Commun. Eng. 2012, 6, 587–592. Silmon, J.; Roberts, C. Improving railway switch system reliability with innovative condition monitoring algorithms. Proc. Inst. Mech. Eng. F J. Rail Rapid Transit 2010, 224, 293–302. [CrossRef] Asada, T.; Roberts, C. Development of an effective condition monitoring system for AC point machines. In Proceedings of the 5th IET Conference on Railway Condition Monitoring and Non-Destructive Testing (RCM 2011), Derby, UK, 29–30 November 2011; pp. 1–6. Asada, T.; Roberts, C. Improving the dependability of DC point machines with a novel condition monitoring system. Proc. Inst. Mech. Eng. F J. Rail Rapid Transit 2013, 227, 322–332. [CrossRef]

Sensors 2016, 16, 549

8. 9. 10. 11. 12. 13.

14. 15. 16. 17. 18. 19. 20.

21.

22.

23.

24. 25. 26. 27. 28.

11 of 12

Cristianini, N.; Shawe-Taylor, J. An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods; Cambridge University Press: Cambridge, UK, 2000. Lee, J.; Noh, B.; Jang, S.; Park, D.; Chung, Y.; Chang, H. Stress detection and classification of laying hens by sound analysis. Asian Australas. J. Anim. Sci. 2015, 28, 592–598. [CrossRef] [PubMed] Santos, P.; Villa, L.F.; Reñones, A.; Bustillo, A.; Maudes, J. An SVM-based solution for fault detection in wind turbines. Sensors 2015, 15, 5627–5648. [CrossRef] [PubMed] Eker, O.; Camci, F.; Kumar, U. SVM-based diagnostics on railway turnouts. Int. J. Perform. Eng. 2012, 8, 289–398. Shanthi, T.S.; Lingam, C. Review of feature extraction techniques in automatic speech recognition. Int. J. Sci. Eng. Technol. 2013, 2, 479–484. Shen, P.; Changjun, Z.; Chen, X. Automatic speech emotion recognition using support vector machine. In Proceedings of the International Conference on Electronic & Mechanical Engineering and Information Technology, Harbin, China, 12–14 August 2011; pp. 621–625. Wang, K.C. Time-frequency feature representation using multi-resolution texture analysis and acoustic activity detector for real-life speech emotion recognition. Sensors 2015, 15, 1458–1478. [CrossRef] [PubMed] Chen, C.-H.; Huang, W.-T.; Tan, T.-H.; Chang, C.-C.; Chang, Y.-J. Using k-nearest neighbor classification to diagnose abnormal lung sounds. Sensors 2015, 15, 13132–13158. [CrossRef] [PubMed] Chung, Y.; Oh, S.; Lee, J.; Park, D.; Chang, H.; Kim, S. Automatic detection and recognition of pig wasting diseases using sound data in audio surveillance systems. Sensors 2013, 13, 12929–12942. [CrossRef] [PubMed] Ligges, U.; Preusser, A.; Thieler, A.; Weihs, C. R Package “tuneR” Computer Program. Available online: http://www.inside-r.org/packages/cran/tuneR/docs/melfcc (accessed on 14 April 2016). Hall, M. Correlation-Based Feature Selection for Machine Learning. Ph.D. Thesis, Waikato University, Hamilton, New Zealand, April 1999. Chung, Y.; Lee, J.; Oh, S.; Park, D.; Chang, H.; Kim, S. Automatic detection of cow’s oestrus in audio surveillance system. Asian Australas. J. Anim. Sci. 2013, 26, 1030–1037. [CrossRef] [PubMed] Hall, M.A.; Smith, L.A. Feature selection for machine learning: Comparing a correlation-based filter approach to the wrapper. In Proceedings of the 12th International Conference on Florida Artificial Intelligence Research Society, Orlando, FL, USA, 1–5 May 1999; pp. 235–239. Kaya, H.; Ozkaptan, T.; Salah, A.A.; Gurgen, S.F. Canonical correlation analysis and local fisher discriminant analysis based multi-view acoustic feature reduction for physical load prediction. In Proceedings of the 15th Annual Conference of the International Speech Communication Association, Singapore, 14–18 September 2014; pp. 442–446. Singh, V.; Meena, N. Engine fault diagnosis using DTW, MFCC and FFT. In Proceedings of the First International Conference on Intelligent Human Computer Interaction, Allahabad, India, 20–23 January 2009; pp. 83–94. Nelwamondo, F.V.; Marwala, T.; Mahola, U. Early classifications of bearing faults using hidden Markov models, Gaussian mixture models, mel-frequency cepstral coefficients and fractals. Int. J. Innov. Comput. Inf. Control 2006, 2, 1281–1299. Maniak, T.; Jayne, C.; Iqbal, R.; Doctor, F. Automated intelligent system for sound signalling device quality assurance. Inf. Sci. 2015, 294, 600–611. [CrossRef] Kavzoglu, T.; Colkesen, I. A kernel functions analysis for support vector machines for land cover classification. Int. J. Appl. Earth Obs. Geoinf. 2009, 11, 352–359. [CrossRef] Renukadevi, N.T.; Thangaraj, P. Performance evaluation of SVM–RBF kernel for medical image classification. Glob. J. Comput. Sci. Technol. 2013, 13, 15–19. Boersma, P.; Weenink, D. Praat: Doing Phonetics by Computer; Version 6.0.05; Computer Program; University of Amsterdam: Amsterdam, The Netherland, 2015. Hall, M.; Frank, E.; Holmes, G.; Pfahringer, B.; Reutemann, P.; Witten, I. The WEKA data mining software: An update. SIGKDD Explor. 2009, 11, 10–18. [CrossRef]

Sensors 2016, 16, 549

29. 30.

12 of 12

Han, J.; Kamber, M.; Pei, J. Data Mining: Concepts and Techniques, 3rd ed.; Morgan Kaufman: San Francisco, CA, USA, 2012. Theodoridis, S.; Koutroumbas, K. Pattern Recognition, 4th ed.; Academic Press: Kidlington, Oxford, UK, 2009. © 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).