Aiming at the problem that there are many types of faults in the current track circuit and it is difficult to obtain high-quality fault label data, this paper proposes a fault diagnosis method for ZPW-2000R track circuit based on OC-SVM and DNN. In this method, OC-SVM is first used to classify and identify 48 monitoring data collected by ZPW-2000R track circuit to collect signal data of different fault types; then the DNN method is used to train the collected fault data to automatically classify the known fault data. A large number of experiments were carried out on this method using the data of ZPW-2000R track circuit, the diagnostic experimental results show that the proposed fault diagnosis method can effectively automatically collect the novel fault data and accurately detect the fault type, and the accuracy of fault diagnosis can reach more than 98%.
Research Article
Research on Fault Diagnosis of ZPW-2000R Track Circuit Based on OC-SVM and DNN
https://doi.org/10.21203/rs.3.rs-1854870/v2
This work is licensed under a CC BY 4.0 License
Version 2
posted
You are reading this latest preprint version
Fault diagnosis
Deep learning
Single classification support vector machine
Track circuit
In recent years, the rapid development of railway construction in China has brought great convenience to people ' s life and travel, and also promoted the development of national economy to a certain extent, it is of great value to study how to ensure the safe and efficient operation of railway. ZPW-2000 series non-insulating frequency shift track circuit is the basic equipment of train operation control system, the normal operation of track circuit is the key factor to ensure the safe and efficient operation of trains(Yu Dong and Xing Chen, 2018). However, the composition of track circuit system is complex, the working environment is bad, the fault rate is high and the fault types are more. Therefore, in order to improve the safe and efficient operation of trains, it is necessary to conduct in-depth research on the fault diagnosis of track circuits.
The technical problems of processing track circuit faults have gone through three stages, namely artificial diagnosis, signal processing diagnosis and artificial intelligence diagnosis(Fei Yu et al., 2021). Manual diagnosis stage: the staff mainly judge and repair the faults of track circuit by virtue of maintenance experience and maintenance technology, which has high blindness, low efficiency and high labor intensity; signal processing diagnosis stage: rely on signal centralized monitoring system for data acquisition and analysis, its anti-interference ability is poor, low precision; artificial intelligence diagnosis stage: on the basis of traditional diagnosis, combined with many algorithms of machine learning to diagnose track circuit faults, it not only improves the diagnostic efficiency, but also improves the diagnostic accuracy. Therefore, it is very valuable to strengthen the research of artificial intelligence algorithm in this field.
At present, some scholars have achieved certain results in the diagnosis stage of artificial intelligence. SUN Y et al.(Sun et al., 2007) proposed a new method for analog circuit fault diagnosis based on SVM. SUN S et al.(Sun and Zhao, 2013) adopts the one-to-one strategy of multi-classification support vector machine, and puts forward a fault diagnosis system for the electrical insulation joint in the key parts of railway track circuit. MA Q et al.(Ma et al., 2016) proposes a new decision tree method for analog circuit fault diagnosis using binary support vector machine trained by different data sets. LI J et al.(Li et al., 2019) proposed a fault detection and diagnosis system of HVAC based on statistical machine learning technology. In order to further improve the accuracy of fault diagnosis, Guangxin Han (Guangxin Han, 2021) used differential evolution algorithm to further optimize the support vector machine, and proposed a fault diagnosis system of high-speed rail choke adapter transformer based on optimized support vector machine. YUAN X et al.(Yuan et al., 2019) proposed an improved hybrid particle swarm optimization algorithm to optimize SVM, and applied it to analog circuit fault diagnosis. ZHANG Y et al.(Zhang et al., 2017) proposed a new fault diagnosis method for oil-immersed transformer based on SVM and improved imperialist competition algorithm. When the support vector machine method is used to train large-scale data, the efficiency is low, and the support vector machine method is sensitive to the selection of parameters and kernel functions and missing data, the optimized support vector machine method is helpful to improve the accuracy of fault diagnosis. Jiao Lu et al.(Jiao Lu et al., 2021) proposed a fault diagnosis model of ZPW-2000R track circuit based on deep convolution neural network to solve the intelligent self-diagnosis problem of ZPW-2000R track circuit fault. For the fault diagnosis of chemical process, WU H et al.(Wu and Zhao, 2018) proposed a fault diagnosis method based on DCNN model composed of convolution layer, pooling layer, dropout layer and fully connected layer. WAZIRALILAH et al.(Waziralilah et al., 2019) mainly studied the application of convolutional neural network in the fault diagnosis of rolling bearing. Fenxia Tian et al.(Fenxia Tian et al., 2020), aiming at the problem of ignoring the influence of tuning zone fault on train safety in existing fault diagnosis, a fault diagnosis system for tuning zone of non-insulating track circuit based on improved convolutional neural network is established. LU J et al.(Lu et al., 2021) proposed an intelligent diagnosis method based on neural network for the situation that the fault monitoring was not timely and the accuracy of fault diagnosis was low due to the lack of outdoor monitoring equipment and only indoor monitoring equipment. ZHAO H et al.(Zhao et al., 2018) proposed a fault diagnosis method based on long short-term memory neural network. When using neural network method to diagnose faults, it is necessary to collect a large number of experimental data for model training and learning, due to the complex composition of track circuit system equipment, it has great limitations to rely solely on a single network model for fault diagnosis of the system(Wenbo Zhu and Xiaomin Wang, 2018). Guangwu Chen et al.(Guangwu Chen et al., 2021), aiming at the problems of low accuracy and unstable diagnosis results of traditional insulationless track circuit fault diagnosis, a method combining simulated annealing algorithm and particle swarm least squares support vector machine is proposed to diagnose the fault of track circuit. LI D et al.(Li et al., 2020), a novel battery fault diagnosis method is proposed by combining long-short memory recurrent neural network with equivalent circuit model. Considering fault feature extraction and occurrence time delay, a new fault diagnosis method is proposed, which consists of sliding window processing and CNN-LSTM model based on convolutional neural network and long-short-term memory network(Huang et al., 2021). Qiushi Wang et al.(Qiushi Wang and Xiaomin Wang, 2017) took the red-band fault of ZPW-2000 track circuit without insulation frequency shift as the research object, and proposed an intelligent fault diagnosis method for track circuit based on fault tree analysis and improved BP neural network in view of the diversity and complexity of its fault. Xiaoying Yu et al.(Xiaoying Yu et al., 2021) proposed a track circuit fault diagnosis system based on grey correlation analysis, simulation comprehensive evaluation and back propagation BP neural network.
However, most of the current research on track circuit fault diagnosis uses neural network method or other machine learning methods to conduct supervised learning on some fault type data samples. This method requires a large number of labeled fault data for algorithm training. Once unknown or new fault data are encountered, the accuracy of fault diagnosis will be reduced, affecting the application of the model in the actual scene. This paper provides a fault diagnosis method for ZPW-2000R track circuit based on OC-SVM and DNN, which can not only effectively solve the automatic identification problem of unknown or new fault data, but also improve the accuracy of fault diagnosis. A large number of experiments were carried out on this method using the data of ZPW-2000R track circuit. The diagnostic experimental results show that the proposed fault diagnosis method can effectively automatically collect the novel fault data and accurately detect the fault type.
2.1 Support vector machine
Support vector machine (SVM) is a kind of generalized linear classifier for binary classification of data according to supervised learning method, its decision boundary is the maximum margin hyperplane for solving learning samples. Single classification support vector machine is a single classification algorithm, which is mainly used for outlier detection(Steinwart and Christmann, 2008), compared with the traditional support vector machine, the algorithm belongs to unsupervised learning method, and does not need to manually label the dataset. The algorithm forms a minimum hypersphere by training the observed high-dimensional sample data. The hypersphere contains a large number of background data, while the data falling outside the hypersphere are abnormal data, thus realizing the identification of specific values. Support vector machine algorithm in practical application, will encounter the problem of positive and negative data imbalance, that is, more normal data, less abnormal data(Osuna et al., 1997). For example, in the fault diagnosis of track circuit studied in this paper, it is difficult to obtain high-quality fault label data, and the traditional support vector machine algorithm is no longer applicable. However, OC-SVM can effectively solve the problem of imbalance between abnormal data and normal data, and can effectively collect signal data of different fault categories.
Set sample data \(\left\{{\chi }_{1},{\chi }_{2},\dots ,{\chi }_{l}\right\}\in {X}^{n}\), \(l\) number of samples, \(\varphi \left(\chi \right)\) is the function of mapping samples to feature space, ω and ρ represent the normal vector and offset of separating hyperplane in feature space, The expression of separation hyperplane is \({\omega }^{\text{{\rm T}}}\varphi \left(\chi \right)-\rho =0\). The objective is to maximize the distance between the separation hyperplane and the origin, then OC-SVM needs to solve the following optimization problems:
$$\underset{\omega ,\rho ,\xi }{\text{min}}\frac{1}{2}{‖\omega ‖}^{2}+\frac{1}{\upsilon \iota }\sum _{i=1}^{\iota }{\xi }_{i}-\rho$$
$$s.t. {\omega }^{\text{{\rm T}}}\varphi \left({\chi }_{i}\right)\ge \rho -{\xi }_{i}$$
$$\begin{array}{c}{\xi }_{i} \ge 0,i=\text{1,2},\dots \dots ,l\#（1）\end{array}$$
In the expression, \({\xi }_{i}\) is a slack variable, which means that outliers are allowed to exist, and ν is a parameter that controls the upper limit of the number of outliers and the lower limit of the number of all support vectors.
By using Lagrange multiplier method, the dual problem of the above optimization problem can be obtained, namely:
$$\underset{\alpha }{\text{min}}\frac{1}{2}\sum _{i}^{\iota }\sum _{j}^{\iota }{\alpha }_{i}{\alpha }_{j}\kappa \left({\chi }_{i},{\chi }_{j}\right)$$
$$s.t. \sum _{i=1}^{\iota }{\alpha }_{i}=\text{1,0}\le {\alpha }_{i}\le \frac{1}{\upsilon \iota }$$
$$\begin{array}{c}i=\text{1,2},\dots \dots ,l\#\left(2\right)\end{array}$$
Where \({\alpha }_{i}\) is the Lagrange coefficient corresponding to the sample \({\chi }_{i}\). kernel function \(k\left({\chi }_{i},{\chi }_{j}\right)\) = \(⟨\varphi \left({\chi }_{i}\right),\varphi \left({\chi }_{j}\right)⟩\), replaces the inner product in the feature space.
After solving optimization problem (2), the sample \({\chi }_{i}\) corresponding to Lagrange coefficient \({\alpha }_{i}>0\) becomes the support vector, Normal vectors \(\omega =\sum _{i=1}^{\iota }{\alpha }_{i}\varphi \left({\chi }_{i}\right)\) and hyperplane offsets 
can be determined by these support vectors, \({\chi }_{SV}\) represents a support vector, classification decision functions are then obtained :
$$\begin{array}{c}f\left(x\right)=sgn\left[{\omega }^{\text{{\rm T}}}\varphi \left(\chi \right)-\rho \right]\\ =sgn\left[\sum _{i}^{\iota }{\alpha }_{i}\kappa \left({\chi }_{i},\chi \right)-\rho \right]\#\left(3\right)\end{array}$$
The track circuit data are tested and substituted into Eq. (3). When the result is + 1, this group of data can be considered as a normal point; if the result is -1, the group of data is outliers (Chapelle et al., 2002, Awad and Khanna, 2015, Maldonado et al., 2021).
2.2 Deep learning
Deep learning is a feature learning method with multi-level representation, which converts raw data into higher level and more abstract representation through some simple nonlinear models. Deep learning is actually developed from deep neural networks (DNN). The number of hidden layers (the depth of the network) of deep neural networks is large. Increasing the depth of the network can reduce the number of features to be fitted in each layer, and represent complex functions with fewer parameters, which can extract high-level feature information. Therefore, deep neural networks have been widely used(Mukherjee et al., 2021, Attique Khan et al., 2021, Zhang et al., 2021).
By learning a deep nonlinear network structure, DNN realizes the approximation of complex functions and obtains the distributed expression of input data. As shown in Fig. 1, the neurons are connected in the form of acyclic graph, and the upper output is the representation of the input of the lower neurons, the input value propagates forward from the input layer neurons through the weighted connection layer by layer, passes through the hidden layer, and finally reaches the output layer to obtain the output; the output layer calculates the loss function to measure the difference between the actual output and the expected output of the network; the loss function is propagated forward layer by layer from the output end, and the gradient of the loss function with respect to intermediate variables is calculated, the gradient values of all parameters are obtained by chain method, the network adjusts the parameters according to the obtained gradient until the loss function reaches the minimum(Sharma and Singh, 2017).
Although the DNN model structure looks complex, studying its local model will find that it is a linear relationship \(\text{{\rm Z}}=\sum {\omega }_{i}{\chi }_{i}+b\) and an activation function \(\sigma \left(z\right)\). DNN includes two processes: forward propagation and backward propagation: the forward propagation algorithm uses the output of the previous layer to calculate the output of the next layer, and the backward propagation is the feedback of the original data.
The so-called forward propagation algorithm of DNN is to use several weight coefficient matrix w and bias vector b to carry out a series of linear operations and activation operations with the input value vector x, and calculate layer by layer from the input layer and get the output result. The commonly used activation functions include sigmoid function, softmax function, tanh function and ReLU function(Ertuğrul, 2018). Sigmoid activation function is commonly used in binary classification problems, when neurons are in a saturated state, the gradient will disappear, that is, when the output is 0 or 1, the gradient is almost 0, and the convergence rate is slow; Softmax activation function is commonly used in multi-classification problems, the input is mapped to a probability value, and the node corresponding to the maximum value is taken as the prediction target; The tanh activation function converges faster than sigmoid activation function, but when the input of neurons is large positive or small negative, the neurons will be saturated and the gradient disappears; ReLU activation function is the most commonly used activation function, and its calculation speed is faster, and its convergence speed is about six times faster than tanh function and sigmoid function. According to the data characteristics of this paper, the activation function is selected as follows: The hidden layer uses ReLU activation function;The output layer uses softmax activation function.
The so-called back propagation algorithm of DNN is the process of iteratively optimizing the loss function of DNN to obtain the minimum value(Hecht-Nielsen, 1992). Before reverse propagation, it is necessary to determine a loss function to measure the loss between the output data and the real data. In this paper, we choose cross entropy loss as a function to measure the loss, that is, for each group of data, we expect to minimize the following:
$$\begin{array}{c}L\left(\widehat{y},y\right)=-{\sum }_{j=1}^{C}{y}_{i}\text{log}{\widehat{y}}_{i}\#\left(4\right)\end{array}$$
Among them, \({\text{y}}_{\text{i}}\) is the target value, \({\widehat{\text{y}}}_{\text{i}}\) is the predicted value.
The expression of total loss function is:
$$\begin{array}{c}J=\frac{1}{m}{\sum }_{i=1}^{m}L\left(\widehat{y},y\right)\#\left(5\right)\end{array}$$
In this paper, Adam optimizer is used for gradient optimization of w, b, and the calculation process is as follows(Misra, 2019):
$$\begin{array}{c} sdw=\beta sdw+\left(1-\beta \right){\left(dw\right)}^{2}\#\left(6\right)\end{array}$$
$$\begin{array}{c} sdb=\beta sdb+\left(1-\beta \right){\left(db\right)}^{2}\#\left(7\right)\end{array}$$
$$\begin{array}{c} w=w-\alpha \frac{dw}{\sqrt{sdw+ϵ}}\#\left(8\right)\end{array}$$
$$\begin{array}{c}b=b-\alpha \frac{db}{\sqrt{sdb+ϵ}}\#\left(9\right)\end{array}$$
Among them, \(\beta\) is the weight value, \(\alpha\) is the learning rate, \(ϵ\) is the minimum constant.
The resulting w, b is fed back to the forward propagation process, so looped until the maximum number of iterations or minimum error requirements are met.
3.1 Data collection and processing
The data used in this paper are 48 monitoring data collected by ZPW-2000R track circuit, as shown in Table 1. The collected data are divided into five sectors: 1–3 sector(sector 2 is the fault sector, sector 1 is the rear sector, sector 3 is the front sector), 6–7 sector(sector 7 is the fault sector, and sector 6 is the rear sector), 10–11 sector(sector 11 is the fault sector, and sector 10 is the rear sector), 12–13 sector(sector 12 is the fault sector, and sector 13 is the front sector) and 25 sector. The data in each sector were processed, for example, the back section 1, fault section 2 and front section 3 in section 1–3 are integrated into a set of experimental data; then the data feature extraction; finally, the experimental data are obtained, as shown in Table 2.
|
Serial |
Number Data Name |
Serial |
Number Data Name |
|---|---|---|---|
|
1 |
Main engine output voltage |
25 |
BP long external current at the sending end |
|
2 |
Main Engine Power Current |
26 |
Short internal current of BP transmitter |
|
3 |
Upper side frequency of main transmitting carrier frequency |
27 |
BP Short External Current |
|
4 |
Downside frequency of main transmitting carrier frequency |
28 |
Side current of receiving end BP cable |
|
5 |
Main Sending Low Frequency |
29 |
BP long internal current at receiving end |
|
6 |
standby power output voltage |
30 |
Long external current of receiving end BP |
|
7 |
The standby power output current |
31 |
Short internal current of receiving end BP |
|
8 |
Sending carrier frequency upside frequency |
32 |
BP short external current at receiving end |
|
9 |
Downlink frequency for sending carrier frequency |
33 |
Send BA long main current |
|
10 |
Sending low frequency |
34 |
Long external main current of BA transmitter |
|
11 |
Transmission of Lightning Protection Analog Network Cable Side Voltage |
35 |
Send BA Short Internal Main Current |
|
12 |
Transmission of Lightning Protection Analog Network Cable Side Current |
36 |
Short external main current of sending end BA |
|
13 |
Main Voltage of Receiving Lightning Protection Analog Network Cable Side |
37 |
Send BA long internal current modulation |
|
14 |
Receiving Lightning Protection Analog Network Cable Side Voltage Regulation |
38 |
Send BA long external current modulation |
|
15 |
Receiving Lightning Protection Analog Network Cable Side Current |
39 |
Sending BA short internal current modulation |
|
16 |
input voltage received by attenuator |
40 |
BA short external current at the sending end |
|
17 |
Host Main Access Voltage |
41 |
The long internal main current of receiving end BA |
|
18 |
Host Adjust Access Voltage |
42 |
Long external main current of receiving end BA |
|
19 |
Main receiving low frequency |
43 |
Receiving BA Short Internal Main Current |
|
20 |
Parallel main access voltage |
44 |
short external main current of receiving end BA |
|
21 |
On - line voltage regulation |
45 |
Long internally regulated current of receiving end BA |
|
22 |
And receive low frequency |
46 |
Receiving BA Long External Current |
|
23 |
Side current of BP cable |
47 |
The receiving end BA short internal current regulation |
|
24 |
BP long internal current at the sending end |
48 |
Short external current of receiving end BA |
Table 2 Fault experimental data of each sector
Table 2 shows that there are 15 kinds of experimental data in which the symbol 0 represents the normal data, and the symbol 1–14 represent one fault type respectively(Tong Sun et al., 2019). The specific fault type and part of the track circuit fault setting diagram are shown in Tables 3 and Fig. 2. In sector 25, the fault type experimental data corresponding to fault code 6 and 8 are missing.
|
Serial |
Failure mode |
|---|---|
|
1 |
Power output to the transmission of lightning protection simulation network disc disconnection |
|
2 |
Main Power Amplifier Level Sealing Break |
|
3 |
Level Sealing Break of Power Amplifier |
|
4 |
The receiving lightning protection analog network disk to the attenuator discontinuity line |
|
5 |
Failure of attenuator or failure of main receiver (Main access anomaly) |
|
6 |
Failure of attenuator or failure of main receiver (Call access exception) |
|
7 |
Failure of attenuator or parallel receiver (Main access anomaly) |
|
8 |
Failure of attenuator or parallel receiver (Call access exception) |
|
9 |
Transmitting Lightning Protection Analog Network Disk to Simulate Terminal Break |
|
10 |
Transmitting Lightning Protection Analog Network Disk to Simulate Terminal Break |
|
11 |
Transmitting Lightning Protection Analog Network Disk to Simulate Terminal Break |
|
12 |
Received Lightning Protection Analog Network Disk Analog Terminal Break |
|
13 |
Received Lightning Protection Analog Network Disk Analog Terminal Break |
|
14 |
Received Lightning Protection Analog Network Disk Analog Terminal Break |
3.2 Model framework
The track circuit fault diagnosis model framework proposed in this project is shown in Fig. 3.
The specific process is as follows:
Step 1 Obtains the signal data of track circuit at a certain time and preprocesses the data;
Step 2 Loads the processed data into One-Class SVM model for prediction;
Step 3 If the prediction result is 1, the signal data of the track circuit at that time are loaded into the DNN model for fault classification and prediction. If the prediction result is − 1 and appears for 10 times or more, it is added to the training set and added a new label; otherwise, load to the DNN model for fault classification prediction;
Step 4 Loads the updated training set into One-Class SVM model and DNN model for training to generate a new prediction model;
3.3 Model experiment process
3.3.1 One-Class SVM model experimental steps
Step 1 Loads the experimental data and converts the format of the experimental data into array type;
Step 2 Get a combination of experimental data types using itertools.combination ( ) functions;
Step 3 Experimental data division: the experimental data corresponding to each combination are used as training samples, and the remaining experimental data are used as test samples;
Step 4 Constructs the framework of One-Class SVM, the parameter setting is nu = 0.01, kernel = ' rbf ', gamma = 0.01;
Step 5 Each time a combination of training data is taken out, the model is trained, and the test sample is loaded into the trained model for prediction until all the combination training is completed;
Step 6 Obtains the minimum, maximum and average values of the accuracy of the test sample.
3.3.2 Experimental steps of DNN model
Step 1 Loads the experimental data, converts the format of the experimental data into array type, normalizes the experimental data, and each experimental data type is divided into 70% training samples and 30% test samples;
Step 2 Divides the feature columns and label columns of training samples and test samples, and converts the label column into one-hot encoding;
Step 3 Build a deep neural network model: the number of hidden layer nodes is 22, and the activation function is ReLU function; the output layer node number is 15, the activation function uses softmax function; the cross entropy loss function is used to calculate the loss; Adam algorithm as optimizer; the number of iterations is set to 500; calculation model accuracy; the model after training is preserved;
Step 4 Loads the test samples into the trained model for prediction and records the results.
4.1 OC-SVM model
The parameter setting of svm.OneClassSVM ( ) function affects the accuracy of OC-SVM model. According to reference(Li et al., 2003, Xiao et al., 2014a, Xiao et al., 2014b, Roobaert et al., 2006, Raman et al., 2017), the parameters of svm.OneClassSVM ( ) function are set as follows : nu = 0.01, kernel = ' rbf ', gamma = 0.01. The training samples of ZPW-2000R track circuit signal data are loaded into the OC-SVM model for training, and the test samples are predicted. The prediction results of each segment are obtained, as shown in Fig. 4.
It can be seen from Fig. 4 that the average accuracy of the test samples of sector 1–3, sector 10–11 and sector 12–13 reaches 100%, the average accuracy of the test samples of sector 6–7 remains above 98%, and the average accuracy of the test samples of sector 25 remains above 96%. The experimental results show that the OC-SVM model can accurately and efficiently identify the alien data, and provide technical support for the subsequent addition of labels to unknown data or new data.
4.2 DNN model
The parameter setting of DNN model has an important influence on the fault diagnosis of ZPW-2000R track circuit. In order to obtain the optimal DNN model structure, this paper uses the ' experience method ' and ' trial and error method ' to set the parameters of the training model(Zheng et al., 2020). For the DNN model, the data partition of training samples and test samples for each segment is shown in Table 4.
Table 4 The training samples and prediction samples of experimental data types in each section
4.2.1 Impact of the number of hidden layer nodes
In order to realize the advanced feature extraction of input data, it is necessary to determine the number of hidden layer nodes in the DNN model. The determination method can refer to the empirical formula (10) (Wanas et al., 1998):
$$\begin{array}{c}S=\sqrt{m+n}+c\#\left(10\right)\end{array}$$
In the formula: m is the number of neurons in the input layer; n is the number of neurons in the output layer; c is the positive integer between [1,10]; S is the number of hidden layer nodes. For five different sectors, in empirical formula (1), m is 144, 96, 96, 96 and 48; n in turn 15, 15, 15, 15 and 13. So the range of hidden layer node number S is [13,22], [11,20], [11,20], [11,20] and [8,17].
In the case of one hidden layer and 500 iterations, the average accuracy of fault diagnosis in the training process and the accuracy of prediction model are analyzed for different hidden layer nodes in five sections. The experimental results are shown in Fig. 5.
The experimental results show that when the number of hidden layer nodes is 22, in sector 1–3, the average accuracy of fault diagnosis in the training process of DNN model is 98.57%, and the accuracy of prediction process is 98.86%. When the number of hidden layer nodes is 20, in sector 6–7,10–11,12–13, the average accuracy of fault diagnosis in the training process of DNN model is 97.8%, 98.24% and 96.79%, and the accuracy of prediction process is 99.45%, 98.8% and 98.09%, respectively. In sector 25, when the number of hidden layer nodes is 15, the average accuracy of fault diagnosis in the training process of DNN model is 94.84%, and the accuracy of prediction process is 98.53%. The experimental results show that the number of hidden layer nodes in each segment is 22,20,20 and 15, respectively.
4.2.2 Impact of hidden layers
The hidden layer number of DNN model has certain influence on the training process and prediction process of the model. If the number of hidden layers is too small, the accuracy of the training process model will be low; if the number of hidden layers exceeds a certain number, it is prone to over-fitting, which affects the prediction results of the DNN model(Choldun et al., 2019).So it needs to be studied separately.
In the five sectors with 500 iterations and hidden layer nodes of 22, 20, 20, 20 and 15, respectively, the average accuracy of fault diagnosis in the training process and the accuracy of prediction model are analyzed for the DNN model with 1–5 hidden layers. The experimental results are shown in Fig. 6.
The experimental results show that with the increase of hidden layers, the accuracy of fault diagnosis in the training process of DNN model will increase slightly, but the accuracy of prediction process is negatively correlated with the increase of hidden layers(Zhang and Morris, 1998). In the case of high accuracy of fault diagnosis in the training process, in order to ensure the maximum accuracy of the prediction process, the analysis results show that the number of hidden layers selects one layer.
4.2.3 Impact of iterations
In the training process of DNN model, the selection of iteration number is also an important factor affecting the accuracy of fault diagnosis(Zhang et al., 2008). In five sectors with hidden layers and nodes of 22, 20, 20, 20 and 15, the relationship between the number of iterations and the average accuracy of fault diagnosis in the training process and the accuracy of the prediction model is analyzed. The results are shown in Fig. 7.
The experimental results show that with the increase in the number of iterations, the model results have a tendency to increase. Before the number of iterations is less than 500, the accuracy of fault diagnosis in the training process will reach 100%. Considering the factors of more iterations and longer time consumption, the maximum number of iterations is set to 500, which not only ensures the accuracy of fault diagnosis, but also improves the efficiency of fault diagnosis.
Through the above analysis of the number of hidden layer nodes, the number of hidden layers and the number of iterations, the DNN model structure of the five sectors is determined as follows: the number of hidden layer nodes 22, 20, 20, 20 and 15; the number of hidden layers is 1 layer; the number of iterations is 500. The track circuit signal data are loaded into the above DNN model for training, and the model accuracy and loss function of each segment in the training process are obtained, as shown in Figs. 8–9. Load the prediction sample into the trained DNN model to evaluate the accuracy of the DNN model(Wang, 1994).
In order to better evaluate the effect of DNN model, this paper introduces four indicators of precision, recall, accuracy and F1-score to analyze the performance of DNN model(Shim et al., 2021). Precision refers to the ratio of the number of samples correctly classified by the model to the total number of samples in the test set, which can evaluate the overall classification performance of the model. Recall rate refers to the proportion of the target class samples identified by the model in the total target class, which measures the recall rate of the diagnostic model. Accuracy refers to the percentage of correct prediction results in the total sample. F1-score is a probability value derived from precision and recall, which reflects the classification performance of the model for fault types. The calculation formulas of precision rate, recall rate, accuracy rate and F1 - score are shown in Equations (11)–(14).
$$\begin{array}{c} Precision P=\frac{TP}{TP+FP}\#\left(11\right)\end{array}$$
$$\begin{array}{c}Recall R=\frac{TP}{TP+FN}\#\left(12\right)\end{array}$$
$$\begin{array}{c}Acc A=\frac{TP+TN}{TP+FN+FP+TN}\#\left(13\right)\end{array}$$
$$\begin{array}{c} F1-score=\frac{2*TP}{2*TP+FP+FN}\\ =\frac{2*P*R}{P+R}\#\left(14\right)\end{array}$$
TP is a positive sample predicted by the model ; FP is a negative sample predicted to be positive by the model ; FN is a positive sample predicted to be negative by the model ; TN is a negative sample predicted to be negative by the model(Qingcheng Meng et al., 2021). TP, FP, FN, TN constitute confusion matrix. According to the formula (11)-(14), the prediction accuracy of DNN model can reach more than 98%. The confusion matrix of each segment is shown in Fig. 10.
It can be seen from Fig. 10 that for the fault types in ZPW-2000R track circuit sectors 1–3, 6–7, 10–11, 12–13 and 25, the DNN model constructed in this paper can achieve 100% recognition rate for some fault types. However, there are also some misjudgments: in sector 1–3, one data with fault code 6 is misjudged as normal data; in sector 10–11, there is one normal data, which is misjudged as the fault type with fault code 6; in sector 12–13, a data with fault code 8 was misjudged as normal. In short, the proposed method improves the accuracy of track circuit fault diagnosis, which proves the effectiveness of the method.
Aiming at the complexity and randomness of track circuit fault, this paper proposes a fault diagnosis method of ZPW-2000 R track circuit based on OC-SVM and DNN. According to the structure of ZPW-2000R track circuit, 48 monitoring data are selected, and 14 kinds of fault types and one normal data are obtained as the experimental data of this model. The OC-SVM model is used to identify the abnormal data, and the DNN model is used to diagnose the fault of track circuit. The simulation results show that the fault diagnosis method of ZPW-2000R track circuit based on OC-SVM and DNN has high accuracy and good performance, which provides reliable guidance for fault maintenance of track circuit and further ensures the safe and efficient operation of railway. However, the OC-SVM method has high sensitivity, and misjudgment will occur for the data that are difficult to distinguish features, thus affecting the accuracy of fault diagnosis. Therefore, subsequent studies will consider trying other classification models to improve the generalization of classification methods.
Funding None.
Conflict of interest The authors declare that have no conflict of interest.
Data availability statements The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.
- ATTIQUE KHAN M, SHARIF M, AKRAM T, et al. 2021. A two‐stream deep neural network‐based intelligent system for complex skin cancer types classification. International Journal of Intelligent Systems [J].
- AWAD M, KHANNA R 2015. Support vector machines for classification [M], Efficient learning machines. Springer: 39-66.
- CHAPELLE O, VAPNIK V, BOUSQUET O, et al. 2002. Choosing multiple parameters for support vector machines. Machine learning [J], 46: 131-159.
- CHOLDUN I, SANTOSO J, SURENDRO K 2019. Determining the number of hidden layers in neural network by using principal component analysis [C] //, Springer; City. 490-500.
- ERTUĞRUL Ö F 2018. A novel type of activation function in artificial neural networks: Trained activation function. Neural Networks [J], 99: 148-157.
- HECHT-NIELSEN R 1992. Theory of the backpropagation neural network [M], Neural networks for perception. Elsevier: 65-93.
- HUANG T, ZHANG Q, TANG X, et al. 2021. A novel fault diagnosis method based on CNN and LSTM and its application in fault diagnosis for complex systems. Artificial Intelligence Review [J]: 1-27.
- LI D, ZHANG Z, LIU P, et al. 2020. Battery fault diagnosis for electric vehicles based on voltage abnormality by combining the long short-term memory neural network and the equivalent circuit model. IEEE Transactions on Power Electronics [J], 36: 1303-1315.
- LI J, GUO Y, WALL J, et al. 2019. Support vector machine based fault detection and diagnosis for HVAC systems. International Journal of Intelligent Systems Technologies and Applications [J], 18: 204-222.
- LI K-L, HUANG H-K, TIAN S-F, et al. 2003. Improving one-class SVM for anomaly detection [C] //, IEEE; City. 3077-3081.
- LU J, YU J, HUANG C, et al. 2021. Research on ZPW-2000R Track Circuit Fault Diagnosis Based on Neural Network [C] //, IOP Publishing; City. 012077.
- MA Q, HE Y, ZHOU F 2016. A new decision tree approach of support vector machine for analog circuit fault diagnosis. Analog Integrated Circuits and Signal Processing [J], 88: 455-463.
- MALDONADO S, LóPEZ J, VAIRETTI C 2021. Time-weighted Fuzzy Support Vector Machines for classification in changing environments. Information Sciences [J], 559: 97-110.
- MISRA D 2019. Mish: A self regularized non-monotonic activation function. arXiv preprint arXiv:1908.08681 [J].
- MUKHERJEE H, GHOSH S, DHAR A, et al. 2021. Deep neural network to detect COVID-19: one architecture for both CT Scans and Chest X-rays. Applied Intelligence [J], 51: 2777-2789.
- OSUNA E, FREUND R, GIROSI F 1997. An improved training algorithm for support vector machines [C] //, IEEE; City. 276-285.
- RAMAN M G, SOMU N, KIRTHIVASAN K, et al. 2017. An efficient intrusion detection system based on hypergraph-Genetic algorithm for parameter optimization and feature selection in support vector machine. Knowledge-Based Systems [J], 134: 1-12.
- ROOBAERT D, KARAKOULAS G, CHAWLA N V 2006. Information gain, correlation and support vector machines [M], Feature extraction. Springer: 463-470.
- SHARMA P, SINGH A 2017. Era of deep neural networks: A review [C] //, IEEE; City. 1-5.
- SHIM K, DO T N, NGUYEN T-V, et al. 2021. Enhancing PHY-Security of FD-Enabled NOMA Systems Using Jamming and User Selection: Performance Analysis and DNN Evaluation. IEEE Internet of Things Journal [J], 8: 17476-17494.
- STEINWART I, CHRISTMANN A 2008. Support vector machines [M]. Springer Science & Business Media.
- SUN S, ZHAO H 2013. Fault diagnosis in railway track circuits using support vector machines [C] //, IEEE; City. 345-350.
- SUN Y, CHEN G, LI H 2007. Analog circuits fault diagnosis using support vector machine [C] //, IEEE; City. 1003-1006.
- WANAS N, AUDA G, KAMEL M S, et al. 1998. On the optimal number of hidden nodes in a neural network [C] //, IEEE; City. 918-921.
- WANG C 1994. A theory of generalization in learning machines with neural network applications [M]. University of Pennsylvania.
- WAZIRALILAH N F, ABU A, LIM M, et al. 2019. A review on convolutional neural network in bearing fault diagnosis [C] //, EDP Sciences; City. 06002.
- WU H, ZHAO J 2018. Deep convolutional neural network model based chemical process fault diagnosis. Computers & chemical engineering [J], 115: 185-197.
- XIAO Y, WANG H, XU W 2014a. Parameter selection of Gaussian kernel for one-class SVM. IEEE transactions on cybernetics [J], 45: 941-953.
- XIAO Y, WANG H, ZHANG L, et al. 2014b. Two methods of selecting Gaussian kernel parameters for one-class SVM and their application to fault detection. Knowledge-Based Systems [J], 59: 75-84.
- YUAN X, LIU Z, MIAO Z, et al. 2019. Fault diagnosis of analog circuits based on IH-PSO optimized support vector machine. IEEE Access [J], 7: 137945-137958.
- ZHANG J, MORRIS A J 1998. A sequential learning approach for single hidden layer neural networks. Neural Networks [J], 11: 65-80.
- ZHANG Y, MA W, CAI B 2008. From Zhang neural network to Newton iteration for matrix inversion. IEEE Transactions on Circuits and Systems I: Regular Papers [J], 56: 1405-1415.
- ZHANG Y, WEI H, LIAO R, et al. 2017. A new support vector machine model based on improved imperialist competitive algorithm for fault diagnosis of oil-immersed transformers. Journal of Electrical Engineering and Technology [J], 12: 830-839.
- ZHANG Y, XIE Y, ZHANG Y, et al. 2021. The adoption of deep neural network (DNN) to the prediction of soil liquefaction based on shear wave velocity. Bulletin of Engineering Geology and the Environment [J], 80: 5053-5060.
- ZHAO H, SUN S, JIN B 2018. Sequential fault diagnosis based on LSTM neural network. IEEE Access [J], 6: 12929-12939.
- ZHENG Z, DAI S, XIE X 2020. Research on fault detection for ZPW-2000A jointless track circuit based on deep belief network optimized by improved particle swarm optimization algorithm. IEEE Access [J], 8: 175981-175997.
- Chen Guangwu, Gao Yali, Jiao Xiangmeng 2021. Track circuit fault diagnosis based on adaptive mutation SAPSO-LSSVM. Journal of Beijing Jiaotong University [ J ], 45 : 1-7.
- Dong Yu, Chen Xing 2018. Fault diagnosis of ZPW-2000 track circuit based on rough set and fuzzy cognitive map. Journal of Railways [ J ], 40 : 83-89.
- Han Guangxin 2021. Fault diagnosis of high-speed rail choke adapter transformer based on optimized support vector machine [ M ]. Beijing Jiaotong University.
- Lu Jiao, Yu Jianli, Huang Chunlei, et al. 2021. Intelligent fault diagnosis method for ZPW-2000R track circuit based on convolutional neural network. Industrial engineering [ J ], 24 : 127-133.
- Meng Qingcheng, Wanda, Wu Haojie, et al. 2021. Concrete crack image recognition method based on convolutional neural network. Journal of Shenyang Jianzhu University ( Natural Science Edition ) [ J ], 37 : 832-840.
- Sun Tong, Chu Junying, Liu Yujie, et al. 2019. Research on fault diagnosis method of ZPW-2000 track circuit. New industrialization [ J ], 9 : 31-35.
- Tian Fenxia, Yang Shiwu, Cui Yong, et al. 2020. Fault diagnosis of non-insulating track circuit tuning area based on improved convolutional neural network. Railway computer application [ J ], 29 : 58-63 + 74.
- Wang Qiushi, Wang Xiaomin 2017. Diagnosis method of track circuit red light band based on FTA and improved neural network. Railway standard design [ J ], 61 : 147-153.
- Yu Xiaoying, Dong Yu, Dong Yu 2021. Fault diagnosis of track circuits based on multi-method evidence fusion. Journal of Railways [ J ], 43 : 86-94.
- Fault diagnosis model design of track circuit based on big data analysis. Communication power technology [ J ], 38 : 80-82.
- Zhu Wenbo, Wang Xiaomin 2018.Research on fault diagnosis method of insulationless track circuit based on combination decision tree.Railway Journal [ J ], 40 : 74-79.
No competing interests reported.