Fault diagnosis method of track circuit based on KPCA-SAE

JIN Zuchen， DONG Yu

(School of Automation and Electrical Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China)

Abstract： At present, ZPW-2000 track circuit fault diagnosis is artificially analyzed and monitored. Its discrimination method not only is low efficient and takes a long period, but also requires highly experienced personnel to analyze the data. Therefore, we introduce kernel principal component analysis and stacked auto-encoder network (KPCA-SAD) into the fault diagnosis of ZPW-2000 track circuit. According to the working principle and fault characteristics of track circuit, a fault diagnosis model of KPCA-SAE network is established. The relevant parameters of key components recorded in the data collected by field staff are used as the fault feature parameters. The KPCA method is used to reduce the dimension and noise of fault document matrix to avoid information redundancy. The SAE network is trained by the processed fault data. The model parameters are optimized overall by using back propagation (BP) algorithm. The KPCA-SAE model is simulated in Matlab platform and is finally proved to be effective and feasible. Compared with the traditional method of artificially analyzing fault data and other intelligent algorithms, the KPCA-SAE based classifier has higher fault identification accuracy.

Key words： ZPW-2000 track circuit; fault diagnosis; stacked auto-encoder (SAE); kernel principal component analysis (KPCA)

0 Introduction

In a railway signal control system, non-insulated track circuit plays a key role. ZPW-2000 non-insulated frequency-shift track circuits are often used in railways. Since the fault handling of the track circuit is still based on the experience of the staff at present, which is not only intensive and low efficient, but also greatly affected by human factors. Therefore, it becomes urgent to use intelligent diagnosis methods to assist the staff and judge the faults timely and accurately.

Scholars have proposed many intelligent algorithms to judge the faults of track circuits. In Ref.[1], fuzzy neural network (FNN) fault diagnosis model is established according to the working principle and fault characteristics of track circuit. In Ref.[2], fuzzy cognition map and rough set theory are introduced into ZPW-2000 track circuit, and a classifier based on attribute reduction and fuzzy cognition map is established. Through verification, accuracy of the classifier has been improved and the time needed is shortened. In Ref.[3], the theory for processing big data, namely deep learning theory, has been proposed for the first time, and then it has been widely used in various fields of research and practical application. In Ref.[4], a fault diagnosis method for induction motor based on kernel principal component analysis (KPCA) and relevance vector machine (RVM) is proposed. In Ref.[5], stacked denoising auto-encoder is used to extract the features of the faults of asynchronous motor, and the diagnostic efficiency is relatively high. In Ref.[6], the automatic encoder (AE) model and its extended model are stacked into a deep structure to complete fault diagnosis of mechanical equipment.

In this work, we analyze fault category and causes of ZPW-2000 track circuit by collecting the common fault data actually generated at the site. A kernel principal component analysis and stacked auto-encoder (KPCA-SAE) based track circuit fault diagnosis model is established by using the data recorded in the microcomputer monitoring system and the ZPW-2000 track circuit test record table, so as to locate and analyze track circuit fault.

1 Track circuit principle

As shown in Fig.1, the ZPW-2000 track circuit consists of a transmitter, 29 m electrical insulation section, some receivers, lightning protections inside the station, transmission cables and matching transformers[7]. The small track is the continuation of the main track of ZPW-2000 track circuit[8]. The main track and the small track are two important components of the track circuit. The track circuit supplies power first, and then sends control signals to through power transformer, service parallel thermoplastic (SPT) cable, cable analog network, the lightning protection inside the station, and the receiver. The receiver not only receives the circuit signal of the track segment of the line, but also receives the small track circuit signal of another track segment, so as to judge whether to occupy or clear[9].

The schematic diagram of the ZPW-2000 track circuit is shown in Fig.1.

Fig.1 Schematic diagram of ZPW-2000 track circuit

2 SAE and KPCA

2.1 SAE

Auto-encoder (AE) is a basic building block of SAE. As shown in Fig.2, AE network is a 3-layer unsupervised network, including input layer, hidden layer, and out putlayer[10].

Fig.2 AE network structure

AE is composed of two parts: encoding network and decoding network. When data are input, the encoding network converts the input data into a feature vector, and then the feature vector is reconstructed by the decoding network to obtain output data[11]. In an ideal state, the input data and the output data are the same, but in actual cases, the closer the input data to the output data, the smaller the error of the AE network.

Assuming a set of network training data setsxm, the encoding network converts the original dataxminto a feature vector by encoding, and propagates it from the input layer to the hidden layer through activation functions such as Sigmoid, Tanh, etc. The activation functions used in this model is Sigmoid function[12]

(1)

and its propagation process is expressed as

(2)

wherexis the data sample input;xiis its component;f(·) is the activation function;θ={w,b} is the parameter of the encoding network, wherewis the weight,bis the bias, andwijandbjare the corresponding components

The decoding process is to propagate the feature vector converted from encoding from the hidden layer to the output layer through the activation function to realize the reconstruction of the input. The reconstruction equation is expressed as

(3)

The goal of training the AE network is to find a set of optimal encoding network parameters and decoding reconstruction parameters, so that the error between the input data and the output data is minimized,which is consequently to realize the purpose of minimizing the loss function equation[13]. The loss function equation is expressed as

(4)

(5)

The gradient reduction method and BP algorithm can be used to iteratively update the AE network, so that the error functionL(w,w′,b,b′) tends to be minimized.

The SAE network is composed ofseveral multiple auto-encoders, and the hidden layer of the upper-level auto-encoder is used as the input layer of the next-level auto-encoder to form a deeper network structure. The network training of SAE firstly initializes the parameters of the network, and then trains the first layer of AE network to realize the process from encoding to decoding, finally, uses the feature vector in the hidden layer as the input of the next layer of AE network. By keeping the progress in this way until outcomes are output, the effect of layer-by-layer training on the network is achieved[15]. The SAE network structure is shown in Fig.3.

Fig.3 SAE network structure

SAE is an unsupervised network. In order to make its powerful feature extraction ability be used in data sample classification, we combine SAE network with softmax classifier. Using the activation valuea(l)of the deepest hidden unit as the input of the softmax classifier, thus the supervised training has the ability to mapa(l)to digital tags. After comparing the classification output value with the actual value of the sample label, the wrong classified gradient value is backpropagated to the encoding layer. Then by iteratively optimizing the parameters of the entire network, we obtain a SAE model with classification ability[16]. The model is used to extract fault features to achieve effective classification of track circuit faults.

2.2 KPCA

The main idea of KPCA is to nonlinearly transform the sample, and perform the principal component analysis in the original space by principal component analysis in the low-dimensional space to express the original dataset information with the least number of features for the purpose of data dimensionality reduction. The method maps the input data matrixXn×mto the high-dimensional feature spaceH={G(X)} by a nonlinear kernel functionG, wherextis thetth sample of the input data matrix[17]. The covariance matrix of the high dimensional feature space is expressed as

(6)

The key to implementing KPCA is to find the mapping directionRthat can best characterize the variance characteristics of each feature of the original data matrix. The equation is expressed as

ζR=UR,

(7)

whereζis the eigenvalue.

Thus, in the original data sample (x1,x2,…,xn), the mapping equation of the data in the directionRis expressed as

(8)

The commonly used kernel functions of KPCA include sigmoid kernel function and radial basis function(RBF). Because RBF has the characteristics of simple process and good classification performance, it is selected in our work[18]. The equation is expressed as

(9)

The KPCA method realizes the nonlinear projection from the input space to the high-dimensional feature space by the kernel function inner product operation.σis the width parameter of the function, and the size of its value has a large impact on the performance of KPCA. Therefore, when KPCA is used for feature dimension reduction, the kernel width parameterσneeds to be optimized to improve the separability of the feature data. The optimization selection process ofσis as follows：

2) The internal distance and the interclass distance of thekkinds of kernel principal component are defined as Eqs.(10) and (11) respectively as

(10)

(11)

3) The smaller the internal distances of different types of feature data of kernel principal component, the bigger the interclass distance is and the better the separability of feature data. The established optimization function is expressed as

(12)

The maximum ofHis the optimal parameter.

3 Fault diagnosis model

3.1 Fault diagnosis model based on KPCA-SAE

The historical fault data of the site are taken as the sample data. The fault samples used for the establishment of the KPCA-SAE fault diagnosis model are the actual fault data collected under complex interference factors such as multiple weather conditions and different aging phenomena of equipment. Test data also use actual operational data from the site. The data have certain applicability to the fault diagnosis of the track circuit in the actual station. Therefore, the track circuit fault diagnosis model established can be applied in the actual scene. The transmission power value, the work output voltage, the sending distribution board voltage, the rail surface voltage, the rail surface current, the receiving end cable side voltage, the receiving end device voltage, XGJ voltage, rail relay voltage, main rail output voltage are regarded as the original characteristic parameters.

The matrix form of the above original data is mapped to the high-dimensional feature space, and a set of bases of the original data in the high-dimensional space is obtained. The kernel width parameter is optimized, and the linear representation of the original data in the high-dimensional subspace is obtained to realize the dimensionality reduction of the original fault data. The establishment process of the track circuit fault diagnosis model is shown in Fig.4.

Fig.4 Establishment process of track circuit fault diagnosis model

Three characteristic parameters of transmission power value, XGJ voltage, and track relay voltage are deleted by reduction. The reduced characteristic parameters are taken as inputs and then to train the SAE network. Finally, the SAE network is constructed.

The KPCA-SAE fault diagnosis model designed uses a 4-layer network structure[19], which includes a input layer, two hidden layers and an output layer. At present, there is no fixed standard for the setting of the number of neurons in the SAE network, and the selection of the number of nodes has strong subjectivity. In this work, several sets of neuron node numbers are designed according to “ascension type”, “descension type” and “concave type” involved in Ref.[20]. The random values obeying the Gaussian distribution are used as the initialization settings of the SAE-KPCA fault diagnosis model parameterswandb. Let the learning rate be 0.1, we randomly select 70% of the sample as the training set of the network, the remnant as the testing set of the network, and repeat the experiment 10 times. The average accuracy of each combination is shown in Table 1.

Table 1 Comparison of node number combination of hidden layer

According to comparison results, we select the combination node numbers 7-8-14-11. It can be concluded that the 10-time average diagnostic accuracy of the method is 93.04%.To see the effect of convergence and the number of iterations on the experimental results during the training process, the average accuracy of the classifier is analyzed in the cases of iterations 0, 50, 100, 150, 200, 250, 300 and 350, respectively. Fig.5 shows the average accuracy for theiterations.

Fig.5 Average accuracy under different numbers of iterations

Fig.5 shows that when the model is iterated to about 250 times, the accuracy of the model gradually becomes stable, so the number of iterations selected in our work is 250.

3.2 Direction of the data

The data processed by KPCA are used as characteristic parameters, including the work output voltage, the sending distribution board voltage, the rail surface voltage, the rail surface current, the receiving end cable side voltage, the receiving device voltage, and the main rail output voltage. A total of 1 100 sets of fault data are sorted out. According to the fault data provided by the site and related literatures consulted, 10 kinds of faults that are more common in the field are used as the output fault type. In addition to the normal situation, a total of 11 types are used as the output of the network model. The fault types are shown in Table 2.

Table 2 ZPW-2000 track circuit fault types

Some actual sample parameters of sites after processing reduction by KPCA are shown in Table 3.

Table 3 Partial actual sample parameters after KPCA reduction

4 Experimental results and analysis

In the existing literature, the fault classification of ZPW-2000 track circuit is mostly carried out by BP neural network and support vector machine (SVM). Based on the background of the same fault data,the classification accuracy values are compared among KPCA-SAE network model, BP neural network, SVM and SAE network. The BP neural network adopts a network structure of 7-14-11, which has an input layer node, a hidden layer node and an output layer node. The SVM uses the RBF. After 10-fold cross-validation, the network optimization results in a penalty factorC=4.35 and a kernel function radiusσ=0.37. The classification accuracy results of four algorithms are shown in Fig.6.

Fig.6 Comparison of 10 experimental diagnostic accuracy values of different methods

Because the established track circuit fault diagnosis model adopts the layer-by-layer training and the BP algorithm method is fine-tuned, the efficiency and accuracy of fault diagnosis are improved. The experimental results show that the proposed method has higher ability to mine data features, higher accuracy and relatively stable network performance than those of other three methods. Besides, the fluctuation of accuracy is small. The reduction of fault parameters by KPCA helps the SAE track circuit fault diagnosis network improve efficiency and accuracy. As for the diagnosis time, although the proposed method is longer than those of the two traditional diagnostic methods, the difference between them is small which can be negligible. What’s more, the diagnosis results can be obtained quickly, which has a good application prospect. The average accuracy and average diagnostic time for the four methods are shown in Table 4.

It can be seen from Table 4 that the average accuracy based on the KPCA-SAE fault diagnosis model is 93.04%, and the average diagnosis time is 0.028 s. It shows that the classification performance of the proposed model is good. Compared with the diagnosis by manual analysis data, the diagnostic efficiency is improved and the labor intensity of maintenance personnel is reduced.

Table 4 Comparison of experimental results of four algorithms

5 Conclusions

The fault diagnosis model of track circuit based on KPCA-SAE network is proposed and established.The data parameters are used as an input of SAE after being reduced and preprocessed by KPCA. The data are extracted and classified to realize fault location. Compared with the original SAE fault diagnosis model, the average accuracy is improved, and the average diagnostic time used is relatively short. The designed fault diagnosis model overcomes the dependence on the staff's experience, and makes the fault diagnosis more intelligent and efficient. The simulation results show that the diagnostic accuracy of the KPCA-SAE model is higher than those of the original SAE fault diagnosis algorithm, the traditional fault diagnosis algorithm and the manual analysis method. Besides, the stability is good. When the on-site duty personnel conduct daily inspection and maintenance on the track circuit signal equipment, staff members can use the proposed fault diagnosis model to quickly locate the fault after collecting the parameter data of the track circuit, and cooperate with the experienced field staff to reduce the fault diagnosis rate, which increases the efficiency of the staff. Therefore, the proposed model can provide effective help for railway signal fault diagnosis, and has a high theoretical value and practical application prospects.