Turnout Fault Diagnosis through Dynamic Time Warping and Signal Normalization

Turnout is one key fundamental infrastructure in the railway signal system, which has great influence on the safety of railway systems. Currently, turnout fault diagnoses are conducted manually in China; engineers are obliged to observe the signals and make problem solving decisions. Thus, the accuracies of fault diagnoses totally depend on the engineers’ experience although massive data are produced in real time by the turnout microcomputer-based monitoring systems. This paper aims to develop an intelligent diagnosis method for railway turnout through Dynamic TimeWarping (DTW).We firstly extract the features of normal turnout operation current curve and normalize the collected turnout current curves. Then, five typical fault reference curves are ascertained through the microcomputer-based monitoring system, and DTW is used to identify the turnout current curve fault through test data.The analysis results based on the similarity data indicate that the analyzed five turnout fault types can be diagnosed automatically with 100% accuracy. Finally, the benefits of the proposed method and future research directions were discussed.


Introduction
Recently, the railway system has experienced rapid development all over the world [1] with both the freight and passenger traffic demands increasing.According to a report from the National Railway Administration of the People's Republic of China, the railway passenger and cargo transportation volume in China were 2.535 billion and 3.358 billion tons, respectively, in 2015 [2].Therefore, due to the rapid development, the maintenance of the railway system has become a critical issue.Problems such as lack of relevant experienced professionals and heavy workloads to monitor the railway safety are emerging.
Turnout (shown in Figure 1), with high operation frequency, is the core component of the railway infrastructure since it is an essential device which moves the train from one track to another [3].Turnout failures have caused several major railway accidents recently [4].According to one report, more than 100 turnout failure events occurred in Changsha communication and signal division each year, and this accounted for 17.5% of all faults of the signaling equipment including turnout in the past five years [5].
Given the importance of the turnout system, microcomputer-based monitoring system (MMS) has been introduced to monitor the turnout state in real time in China.The MMS collects turnout operation current and voltage levels data; then engineers perform the failure diagnose analysis based on the displayed curves.The current manual diagnosis system has not only caused low diagnosis efficiency, but also increased the manpower and resources requirements [6].In addition, the accuracy of the diagnosis mainly depends on the engineers' subjective experiences, where any misinterpretation of the data could lead to potential safety issues.Therefore, automatic turnout pattern detection methods are needed to identify the railway faults or failures.
Given the emerging issue, different methods have been utilized in the turnout failure diagnosis area.Zhao and Lu studied the turnout fault diagnosis system based on gray correlation analysis [7].Roberts et al. used single throw mechanical equipment (STME) to detect the fault [8].Neural network and fuzzy theory ( [9][10][11][12][13]), Support Vector Machine (SVM), and improved SVM ( [6,14]) have also been applied for turnout fault diagnoses.Atamuradov et al. utilized Dynamic Time Warping (DTW) and expert systems to recognize three states including one healthy state and two failure states for turnout [15].Ardakani et al. established the health assessment of the turnout by Principal Component Analysis (PCA) [16].
However, the abovementioned methods have several limitations for the targeted problem.For example, gray correlation analysis needs to choose suitable feature vectors or parameters that are hard to ascertain for large scale of fault types since a continuous search of large space is needed until the matching feature vector or parameter is identified.Besides, SVM-based methods cannot efficiently handle the large sample size.And expert system needs to have much a priori knowledge, which requires much manpower from experienced people to summarize the rules and knowledge based on years of experience.However, it is difficult to build a complete knowledge base due to more microcomputer-based monitoring systems and various environments and lack of rich experience.Besides, neural network and fuzzy theory models were developed based on large size of historical fault data, which is difficult to collect, and the model training process would be time consuming for tuning the model parameters.
Recently, Dynamic Time Warping (DTW) [17], a Dynamic Programming (DP) method which has originally been used in isolated word recognition area, has become popular.DTW calculates the distance between reference data and test data which has never been trained, and the smallest distance indicates the greatest similarity [18][19][20].And it holds the benefits of requiring small amount of fault reference data, no need of selecting feature vectors, and requiring limited historical data and a priori knowledge.In this study, DTW was introduced to conduct the turnout fault diagnosis.
In addition to the analysis method, various data have been used in the literature, such as turnout operation current [21][22][23], turnout operation power [24], and data from sensors [25][26][27].Most researchers choose the first two kinds of data since they can be obtained directly by the MMS.Moreover, the data from sensors requires the installation of extra sensors, which may be cost-prohibitive.Therefore, in this study, we utilize turnout operation current data to analyze the fault because current data can intuitively reflect the turnout fault while power data cannot.
Conventional methods to diagnose the fault based on turnout operation current data do not deal with current curve images.The turnout comes from different manufacturers, disparate MMSs, and diverse railway bureaus.Thus, the current curves images have much disturbance such as noise and grids which may seriously affect the accuracy of diagnosis results.Through the normalization of the current curve [28], we can effectively remove the noise in the image and diagnose the types of turnout faults.
Furthermore, in this study the single-action ZD6 turnout will be targeted given its wide range of applications.The operation current curves were collected from the MMS.We will first normalize the original turnout operation current curve and ascertain turnout reference templates.Then, the similarities between the reference templates and the test samples will be calculated through DTW.Each turnout diagnosis result would be recorded and finally the classification accuracy of the proposed approach will be evaluated.
The rest of this paper is organized as follows.Section 2 provides the meaning of the normal turnout operation current curve for each stage and introduces the method of image normalization method for turnout current curves.The method and principles of DTW and reference templates are explained in Section 3. Section 4 presents numerical experiments and result for the real turnout operation current curves to diagnose the faults, followed by the conclusions and discussions in Section 5.

Data Preparation
2.1.Turnout Operation Current Curve.Nowadays, MMS is the main approach to monitor the state of the turnout in China.Turnout operation current curve can intuitively reflect current changes of the switch machine.The operation process of the turnout can be divided into four stages: unlocking, conversion, locking, and slow release [12].We can see the characteristics of the turnout operation current curve of different stages from Figure 2: (1) Stage 1 (unlocking) ( 0 - 1 ): the motor starts with a large starting current which makes the curve rise suddenly.With the operation of the turnout system, the curve shows sharp decline and the turnout enters the unlocking stage.
(3) Stage 3 (locking) ( 2 - 3 ): the point moves to the other side of the rail and the current curve reduces to zero when the locking stage is finished.
(4) Stage 4 (slow release) ( 3 - 4 ): 1DQJ relay slows release and the current stays at zero continuously at slow release stage.turnout operation current curves and improve the accuracy of fault diagnosis.We need to remove the noise such that the disturbance in the images does not affect the experimental results.We have used the following procedure to remove the noise:

Image
(1) Gray-scale transformation: it is a method of producing the gray value for each pixel of the original image according to a target condition.The gray value of the original image pixel is assumed as where  is the gray value of the original image pixel andxis abscissa andyis ordinate.Gray-scale enhancement is expressed as and (, ) is the function of the image threshold.
(2) Binarization: the gray-scale image is converted into binary image which is a two-dimensional array ( × ).The value of the pixel is set to 0 if the gray value is less than the threshold ; otherwise the value of the pixel is set to 1.The function of the image threshold is (3) Noise removal: the first step is to find out the target region surrounded by the axis through the sum of the rows and columns in the two-dimensional array.Then isolated pixels are removed from the object region with open operations including erosion and dilation.
(4) Refinement: the denoising image may have multiple zero pixels in one column.And this situation would lead to the phenomenon that a moment corresponds to a number of current values in the coordinate transformation.It is assumed that the value of  pixel is 0, and the value  in that row is and  1 is the value of 1 pixel.
The number of rows in the th column with the value of pixel of 0 is and others are set to 1.
(5) Coordinate transformation: the purpose of the coordinate transformation is to convert the coordinate of the curve from the RO'C coordinate system to the tO'I coordinate system.We assume that the point  with coordinate (  ,   ) in the RO'C coordinate system has (  ,   ) in the tO'I coordinate system.Thus Point coordinate is 1 ,  2 ,  1 , and  2 are the edge line of location for the target area and  0 ,  0 ,   ,   ,   , and   are set by engineers.
(6) Normalization: the normalization aims to zoom the data proportionally and place it in a specific interval in order to make the algorithm universal because the data comes from disparate systems and manufacturers.It eliminates the influence of different coordinates.For example, the range of t in one image is 0 to 5 and in another it is 0 to 8. The same type of switch machine made in different factories has different current values but the turnout operation current curves have the same tendency.The original data is transformed linearly and the results are mapped to [0, 1].The transformation function is as follows: And   is a point after normalization. is original data. max is the maximum value. min is the minimum value.
The normalization process is shown in Figure 3.All of reference templates and test samples must be processed.

Dynamic Time Warping (DTW).
In this study, DTW was used to calculate the similarities between test samples and reference templates in order to diagnose the fault types.DTW was firstly introduced in the field of speech recognition to recognize the distortion of similar sounds, and it was a flexible distance-based curve comparison model which has been successfully applied to a wide range of time series data [17].DTW is based on the idea of Dynamic Programming (DP), which can match different lengths of the time series and avoid mismatch between the peaks on curves even if the abscissa of peak is different.In Figure 4, the peak of the top line is point .It is incorrect to simply consider that point  would correspond to point   in the bottom curve.DTW can identify point  which corresponds to point  and then calculate the distance of two lines.It compares the similarity of two series by calculating a similarity matrix and searching for an optimal path with the minimum cumulative distance.
The path is not casually selected.The selected path must start from the lower left corner and end in the upper right corner, as shown in Figure 5.It is assumed that the grid points in the path are ( 1 ,  1 ), . . ., (  ,   ), . . ., (  ,   ).If the path has passed the point (, ), the next passing point can only be one of the following three cases: (1) (, ) = ( + 1, ).
The path starts from the point (0, 0) to match the two sequences  and .All the distances of points calculated before will be accumulated when it is reaching each point.After reaching the end point (, ), the optimal path is formed and the cumulative distance is the last total distance.The lower the cumulative distance between the time sequences is, the higher the similarity is [14].Thus the minimum cumulative distance of the time sequences  and  is  [(, )] =  [ () ,  ()] + min { ( + 1, ) ,  ( + 1,  + 1) ,  (,  + 1)} . ( DTW algorithm compares the test samples with the reference templates directly without the need of knowing plenty of prior knowledge and historical data.It is a kind of practical method for fault diagnosis because this algorithm only needs several reference templates to diagnose the fault.

Confirming Reference Templates.
A total of 5 ZD6 turnout fault type current curves were selected as the reference templates, which are shown in Figure 6.
Figure 6(a) presents the fault that the turnout suddenly stops running after starting.It can be seen that conversion time is short and locking line disappears.There are several potential reasons for that: (1) The switch machine action current is too small.(2) The switch machine has a poor performance.(3) The self-closing circuit of 1DQJ has faults that 1DQJ cannot continue after sucking up.The end of turnout jamming curve in Figure 6(b) is stable at a relatively high value.The reason for this is that there is foreign matter between the switch rail and the stock rail or the fault current of the switch machine is too small.Figure 6(c) shows the current curve when start-up circuit disconnects.The trend of this curve is almost straight without parabolic downward trend and the value of current stays at a high value.This failure is due to the mixed line of stator and rotor.Exceeding locking current curve (shown in Figure 6(d)) suddenly rises in the locking section.The reasons for this could be tight turnout adjustment, dirty slide plate, lack of oil in slide plate, the stock rail with a lateral move, and so on.Figure 6(e) shows the fault that automatic actuator is not flexible.This locking line of fault current curve is longer than normal curve and has a larger current value.The reason is that the turnout has been already and several shafts of the automatic actuator action are not flexible due to lack of oil.It causes the following: the automatic actuator operation becomes abnormal and the start connections cannot be disconnected.

Performance Analysis
In this study, a total of 260 current curves have been tested to verify the accuracy of proposed analysis method.Each test sample was compared with the 6 preselected reference templates including one normal template and 5 fault templates through DTW algorithm.Smaller distances between the test sample and reference templates indicate larger similarity.Shown as an example, the distances for 10 current curves between test samples and 6 kinds of reference templates are shown in Table 1.The system fault identification results are shown in Table 2.
From Table 1, we can find the smallest values in each row.That means that this curve has great similarity to the reference template where the smallest numbers are located.Thus, curves (1) and ( 9) indicate turnout jamming.Curves (2) and ( 6) represent that the turnout suddenly stops running after starting.Start-up circuit disconnection corresponds to curves (3) and (10).Exceeding locking current is for curves (5) and (8).Curves (4) and (7) indicate that automatic actuator is not flexible.We have verified in real experiments that all of our fault identification results are correct.
Table 2 shows the diagnosis results for the system.Each type of faults has 52 curves and the normal curve is 0 for total of 260 test samples.
Table 3 presents the accuracy of the system diagnosis result for 5 kinds of faults, and a 100% accuracy rate was achieved, while, in [15], the classification accuracy of fault detection decreased along with the increase of noise level.In this study, image normalization procedure is used to avoid the impact of noise and other interference on the result, and   DTW scheme can greatly improve the diagnosis efficiency and accuracy without using other algorithms, rules, or a priori knowledge.

Discussion and Conclusions
In this study, we have developed an automatic fault diagnostic method based on DTW scheme for the railway turnouts.Firstly, all the turnout current curves captured from the MMS were normalized, and both normal and fault reference templates were identified.Then, a total of 260 turnout current curves which have never been trained were compared with 6 reference templates through DTW scheme.By seeking the minimum cumulative distance between test samples and reference templates, various fault types were identified.The analysis results indicated that the turnout faults could be diagnosed through the proposed method automatically with 100% accuracy for 5 typical fault current curves.Our scheme could avoid accidents caused by new-joined or less experienced technicians' errors and saves much manpower and material resources to improve the railway safety.
In a previous study [15], DTW was also used to analyze a similar problem; however, the analysis results showed that noises in the curves have significantly impacted the accuracy of the result, and the accuracy of system decreases as the noise level increases.This further reduces the system reliability and could even cause accidents because of fault misclassification.Unlike the previous study, in this study, a curve normalization procedure was used to eliminate the impacts of noises.In addition, Atamuradov et al. [15] used some rules to diagnose the turnout failure.However, rule-based system lacks flexibility, since the data from different MMSs may have builtin heterogeneity.In this study, the results indicate that the developed method can diagnose multiple kinds of turnout faults even if the values of current curve fluctuate much and the operation time of turnout is different.Besides, DTW does not require feature vector selection, historical data, or a priori knowledge, which is beneficial for real-time diagnosis.
Our next work would be investigating other algorithms (such as cluster algorithm or manifold learning or deep learning) to recognize the undefined type of turnout fault.Besides, big data learning issues will also be investigated since the collected current or voltage curves are extremely large in data size.

Figure 2 :
Figure 2: The turnout operation current curve.

Figure 4 :
Figure 4: Match the peaks by using DTW.

Table 1 :
The distance between test samples and reference templates.

Table 2 :
The number of different kinds of curves.

Table 3 :
Accuracy of fault diagnosis.