Switch Fault Diagnosis Method Based on Hybrid Deep Learning Model

Abstract

The present invention discloses a turnout fault diagnosis method based on a hybrid deep learning model, specifically: preprocess and normalize the turnout power data, analyze the fault types corresponding to each type of data, and label the fault types reflected by each sample data in the sample dataset according to set rules; use the improved Smote algorithm to expand the fault data and divide the dataset; establish a deep neural network classification model; use the cross-entropy loss function and the Adam optimizer to train the model. According to the idea of using dilated convolution for feature extraction and the LeNN-attention mechanism fusion model as a classifier, a train turnout fault diagnosis method based on the dilated convolution and LeNN-Attention hybrid model is designed. This method has higher diagnostic accuracy than traditional single neural networks and machine learning algorithms, indicating that the model obtained by the hybrid deep learning algorithm has very good effects in practical applications.

Description

Technical Field

[0001] The present invention belongs to the technical field of train turnout fault diagnosis and artificial intelligence, and particularly relates to a turnout fault diagnosis method based on a hybrid deep learning model. Background Art

[0002] A turnout is one of the basic devices of a railway signal system, and is used to switch the advancing direction of a train to realize the line change and cross-line operation of the train. With the rapid development of China's railway, the scale of the railway network is constantly expanding, and the application demand for turnouts is also increasing. However, due to the constraints of its own mechanical structure and actuating mechanism, the mechanical strength of each part of the turnout is generally lower than that of other devices on the line, and it is very easy to have fatigue changes, thus triggering accidents such as train turnout jamming and even derailment. In addition, turnouts are generally installed outdoors, and their working environment is greatly affected by the weather. Any change in temperature or humidity may affect the safe operation of the turnout. Therefore, the timely discovery, accurate diagnosis and early warning of turnout faults play a crucial role in preventing major accidents and improving the equipment utilization rate.

[0003] At present, the turnout is still maintained by combining manual regular inspections with the monitoring of the turnout action status by a microcomputer monitoring system. Maintenance personnel regularly browse the turnout action current and power curves collected by the microcomputer monitoring system, and identify turnout faults by comparing with the monitoring data of normally operating turnouts. However, manual detection has high requirements for the experience of maintenance personnel, and situations often occur where new employees misdiagnose or miss diagnoses due to lack of experience. Moreover, the thresholds in the detection system are set by maintenance experts, and the ambiguity of expert experience also makes most threshold rules unable to work effectively for a long time and be effectively promoted.

[0004] Therefore, in order to improve the turnout maintenance efficiency and ensure the efficient and safe operation of trains, it is urgent to change the traditional turnout fault diagnosis and maintenance methods, adopt highly intelligent diagnosis methods, and realize the transformation from the maintenance mode based on fault repair and periodic repair to the state repair mode. The working mechanism of turnouts is complex, and how to extract fault features based on the turnout action power curve and realize fault diagnosis is a very challenging problem. Summary of the Invention

[0005] The purpose of the present invention is to provide a turnout fault diagnosis method based on a hybrid deep learning model, which solves the problems of high complexity and uncertainty brought by manual feature extraction and selection in traditional methods, and improves the accuracy of turnout fault diagnosis.

[0006] The technical solution adopted by the present invention is that the turnout fault diagnosis method based on a hybrid deep learning model is specifically implemented according to the following steps:

[0007] Step 1: Preprocess and normalize the switch power data, analyze the fault types corresponding to each type of data, and label the fault types reflected by each sample data in the sample dataset according to the set rules;

[0008] Step 2: Use the improved Smote algorithm to augment the fault data and divide the dataset;

[0009] Step 3: Establish a deep neural network classification model;

[0010] Step 4: Use the cross-entropy loss function and the Adam optimizer to train the model.

[0011] The features of the present invention also lie in:

[0012] In Step 1, specifically:

[0013] Step 1.1: The switch action power curve is one-dimensional time series data. Each data takes 210 sampling points. For the data with less than 210 sampling points, zero-padding is performed at the end, and for the data with more than 210 sampling points, it is truncated to obtain one-dimensional time series data with the same length; the data is normalized;

[0014] Step 1.2: Obtain the normalized switch action power data. Taking the switch action time as the abscissa and the discrete power data as the ordinate, generate the switch action power curve, and corresponding power curves are formed according to the different actions in each time period;

[0015] Step 1.3: Through Step 1.2, all the power time series data is converted into power curve graphs. According to the different trends of the curves in each stage, six different types of power curves are manually selected from the original sample set, corresponding to six different faults in the switch conversion process, namely abnormal power fluctuation in the starting stage, power rising and remaining in the conversion stage, power directly dropping to 0 in the stage, large power fluctuation in the stage, power normally dropping and remaining unchanged but not 0, and abnormal power fluctuation in the conversion stage.

[0016] In Step 2, specifically:

[0017] Step 2.1: For each type of fault data, randomly select one as the minority class sample X i of this type of data, and randomly select another 5 pieces of data of this type as its nearest neighbor samples Y i , use the method of random linear interpolation to generate new samples, and finally add the newly generated data to the original dataset;

[0018] Step 2.2: Use the Smote algorithm to increase each of the 6 types of fault data to 500 pieces, add classification labels to the 500 pieces of normal data and the 500 pieces of fault data in the order of 0 - 6, representing 7 different fault categories. The overall 3500 pieces of data are used as the original data set. After shuffling the order, it is divided into a training set and a test set according to a ratio of 7:3.

[0019] In step 3, specifically:

[0020] Step 3.1: The network model is divided into two parts according to its functions, namely a feature extractor and a classifier. Among them, the feature extractor is served by a dilated convolutional layer, and the classifier is composed of a first convolutional layer, a max-pooling layer, a second convolutional layer, a ReLU activation function layer, a Convolutional Block Attention Module (CBAM) layer for fusion attention mechanism, a fully connected layer, and a DropOut layer;

[0021] Step 3.2: Input the training set data into the feature extractor. The convolution kernel size of this layer is 2*2, the stride is 2, and the dilation rate is set to 2, which means there is an empty position between every two data points, filled with 0. The data after feature extraction is used as the input of the classifier;

[0022] Step 3.3: After extracting features through the dilated convolutional layer, the dimension of the data becomes (1, 8, 8, 19). Input this data into the classifier for classification;

[0023] Step 3.4: Input the output result of the entire classifier into the Convolutional Block Attention Module (CBAM) layer for a second-stage processing; CBAM includes a channel attention module and a spatial attention module;

[0024] Step 3.5: Input the output result into the fully connected layer for the last-stage processing; The first fully connected layer inputs a vector of 16 * 3 * 8 = 768 dimensions, calculates the dot product between the input vector and the weight vector and then adds a bias term. The result L1 is output after being mapped by the ReLU function, with a dimension of 120; The second fully connected layer receives L1 and continues to output a 78-dimensional vector L2 through the ReLU function; Before inputting L2 into the last fully connected layer, first use the Dropout function for processing. The final output result of this layer is D(1, 68); The last fully connected layer has a total of 7 neurons, representing the numbers 0 - 7 respectively, which are the classification labels added to the fault data and normal data. After inputting D(1, 68) into this layer, a vector L3 with a dimension of (1, 7) is obtained; Finally, use the Softmax function to map the output results of the seven neurons to the interval (0, 1).

[0025] In step 3.3, specifically: First, perform the first convolution operation on the input data. Use 6 convolution kernels of size 2*2. During the convolution process, no edge padding is performed, and the stride = 1. A single convolution kernel calculates and outputs a feature map of size 7*18. This layer uses a total of 6 convolution kernels, and after overall calculation, 6 feature maps of size 7*18 are obtained. Each convolution kernel contains a bias term, and a total of 6*(2*2 + 1) = 30 parameters need to be learned, mainly achieved through weight sharing; after all the feature maps are processed by the ReLU activation function, the final output result C1(1, 6, 7, 18) of the first layer of convolution operation is obtained;

[0026] Then perform the second convolution operation. After the second convolution layer receives the output result C1 of the first layer, use 16 convolution kernels of size 2*2 with a stride of 1 to calculate and obtain 16 feature maps of size 6*17; after weight sharing, the total number of parameters is 6*(2*2*3 + 1) + 9*(2*2*4 + 1) + 1*(2*2*6 + 1) = 256. After passing through the ReLU function in the same way, the final result C2(1, 16, 6, 17) of the second layer of convolution operation is obtained; C2 is received by the pooling layer. This layer uses a pooling unit of size 2*2 with a stride of 2. For each data within the 2*2 area in C2, sum them up, multiply by a weight coefficient, and then add a bias. Then the total number of parameters to be learned in this layer is (1 + 1)*16 = 32; the final output result is P1(1, 16, 3, 8).

[0027] In step 3.4, the channel attention module compresses the spatial dimension of the input feature map, and at the same time uses the average pooling feature and the max pooling feature. The feature map M c (F) of the channel attention module is calculated by the formula:

[0028]

[0029] where F is the input feature map, AvgPool and MaxPool represent the average pooling operation and the max pooling operation respectively, F c avg and F c max represent the average pooling feature and the max pooling feature respectively;

[0030] The spatial attention module is a supplement to the channel attention. Apply the average pooling and max pooling operations along the channel direction, splice the results, and use a convolution layer to generate the spatial attention map Ms(F). The formula is:

[0031]

[0032] The channel attention module and the spatial attention module calculate a one-dimensional channel attention map Mc and a two-dimensional spatial attention map Ms in sequence. The final calculation formula of CBAM is as follows:

[0033]

[0034] The beneficial effects of the present invention are as follows: For the turnout fault diagnosis method of the present invention, after analyzing in detail the structures and principles of dilated convolution, LeNN network, and fusion attention mechanism, according to the idea of using dilated convolution for feature extraction and using the LeNN-attention fusion model as a classifier, a train turnout fault diagnosis method based on a hybrid model of dilated convolution and LeNN-Attention is designed. The method proposed by the present invention has higher diagnosis accuracy than traditional single neural network and machine learning algorithms, which indicates that the model obtained by the hybrid deep learning algorithm has very good effects in practical applications. Description of the Drawings

[0035] Figure 1 is a schematic flow chart of the fault diagnosis process in the turnout fault diagnosis method based on the hybrid deep learning model of the present invention;

[0036] Figure 2 is the network structure diagram in the turnout fault diagnosis method based on the hybrid deep learning model of the present invention;

[0037] Figure 3 is the normal turnout action power curve diagram in the turnout fault diagnosis method based on the hybrid deep learning model of the present invention;

[0038] Figure 4 is the action power curve diagrams of 6 types of faulty turnouts in the turnout fault diagnosis method based on the hybrid deep learning model of the present invention. Detailed Embodiments

[0039] The present invention will be described in detail below in conjunction with the drawings and specific embodiments.

[0040] The turnout fault diagnosis method based on the hybrid deep learning model of the present invention has a flow chart as Figure 1 shown, and is specifically implemented according to the following steps:

[0041] Step 1: Preprocess and normalize the turnout power data, analyze the fault types corresponding to each type of data, and label the fault types reflected by each sample data in the sample dataset according to the set rules. Specifically:

[0042] Step 1.1: The switch action power curve is one-dimensional time series data. The normal action time is about 7 seconds, and each data takes 210 sampling points. To unify the input tensor of the model, zero-padding is performed at the end of the data with less than 210 sampling points, and the data with more than 210 sampling points is truncated to obtain one-dimensional time series data of the same length. Before inputting the model for training, the data is normalized, and the formula is:

[0043]

[0044] where x is the original data, and x * is the normalized data, x max and x min are the maximum and minimum values of each data in the original sample;

[0045] Step 1.2: Obtain the normalized switch action power data. Taking the switch action time as the abscissa and the discrete power data as the ordinate, generate the switch action power curve. As Figure 3 shown, the normal action process is divided into five stages: start, unlocking, conversion, locking, and indication. According to the different actions in each period, the corresponding power curves are formed;

[0046] Step 1.3: Through Step 1.2, all the power time series data is converted into power curve graphs. According to the different trends of the curves in each stage, six different types of power curves are manually selected from the original sample set, which respectively correspond to six different faults during the switch conversion process. As Figure 4 shown, the characteristics and selection criteria of the power curve graphs are as follows: (1) The power fluctuates abnormally in the start stage; (2) The power rises and remains in the conversion stage; (3) The power directly drops to 0 in the indication stage; (4) The power fluctuates greatly in the indication stage; (5) The power remains unchanged but is not 0 after normal decrease in the indication stage; (6) The power fluctuates abnormally in the conversion stage.

[0047] Step 2: Use the improved Smote algorithm to expand the fault data and divide the data set. Specifically:

[0048] Step 2.1: Based on the six types of fault data obtained in Step 1, use the improved Smote algorithm to correct the problem of sample imbalance. For each type of fault data, randomly select one of them as the minority class sample X i of this class of data, and randomly select another 5 pieces of data of this class as its nearest neighbor samples Y i . Use the method of random linear interpolation to generate new samples, and finally add the newly generated data to the original data set. The random linear interpolation formula is:

[0049] S new = X i+rand(0,1)*(Y i -X i )

[0050] where S new represents the newly generated sample, and rand(0,1) represents a random number between (0,1).

[0051] Step 2.2: Use the Smote algorithm to increase each of the 6 types of fault data to 500 pieces, add classification labels to the 500 pieces of normal data and the 500 pieces of fault data in the order of 0-6, representing 7 different fault categories. The overall 3500 pieces of data are used as the original dataset, and after shuffling the order, they are divided into a training set and a test set according to a ratio of 7:3.

[0052] Step 3: Establish a deep neural network classification model. The overall structure is as Figure 2 shown, specifically as follows:

[0053] Step 3.1: Determine the overall deep neural network classification model. This network model is divided into two parts, a feature extractor and a classifier, according to its functions. Among them, the feature extractor is served by a dilated convolutional layer, and the classifier consists of a first convolutional layer, a max pooling layer, a second convolutional layer, a ReLU activation function layer, a convolutional block attention module (CBAM) layer, a fully connected layer, and a DropOut layer.

[0054] Step 3.2: Input the training set data into the feature extractor. Compared with the ordinary convolutional layer, the dilated convolution used in this feature extractor introduces a new parameter called "dilation rate", which defines the distance between values when the convolutional kernel processes data. The size of the convolutional kernel in this layer is 2*2, the stride is 2, and the dilation rate is set to 2, representing that there is an empty position between every two data, filled with 0. The data after feature extraction is used as the input of the classifier.

[0055] Step 3.3: After extracting features through the dilated convolutional layer, the dimension of the data becomes (1, 8, 8, 19). Input this data into the classifier to achieve classification; this part includes the most basic convolutional module, pooling module, and fully connected module in deep learning, with a total of 7 layers of network structure, and each layer contains trainable parameters.

[0056] First, perform the first convolution operation on the input data. Use 6 convolution kernels of size 2*2. During the convolution process, no padding is done, and the stride stride = 1. A single convolution kernel calculates and outputs a feature map of size 7*18. A total of 6 convolution kernels are used in this layer. After overall calculation, 6 feature maps of size 7*18 are obtained. Each convolution kernel contains a bias term. A total of 6*(2*2 + 1) = 30 parameters need to be learned, mainly achieved through weight sharing. After all feature maps are processed by the ReLU activation function, the final output result C1(1, 6, 7, 18) of the first-layer convolution operation is obtained. The calculation formula for the convolution output size is:

[0057] C = (W - F + 2P) / S + 1

[0058] Among them, W is the input size, F is the convolution kernel size, S is the convolution kernel stride, and P is the padding width.

[0059] Then perform the second convolution operation. After the second convolution layer receives the output result C1 of the first layer, use 16 convolution kernels of size 2*2 with a stride of 1, and calculate to obtain 16 feature maps of size 6*17. It should be noted that 6 feature maps are input in this layer and 16 feature maps are output, indicating that they are not fully connected, but a local connection method is adopted. Specifically, six of the output feature maps are only connected to three of the input feature maps, nine feature maps are connected to four feature maps, and one is connected to six. After weight sharing, the total number of parameters is 6*(2*2*3 + 1) + 9*(2*2*4 + 1) + 1*(2*2*6 + 1) = 256. After passing through the ReLU function again, the final result C2(1, 16, 6, 17) of the second-layer convolution operation is obtained.

[0060] So far, the convolution operation is all completed. C2 is received by the pooling layer. The pooling layer is used for downsampling, and the purpose is to reduce the amount of data processed by the model, and then fully extract the information features of the data. This layer uses a pooling unit of size 2*2 with a stride of 2. For each data within the 2*2 area in C2, sum them up, multiply by a weight coefficient, and then add a bias. The pooling process is similar to convolution and uses the same weight parameters. Then the total number of parameters to be learned in this layer is (1 + 1)*16 = 32. The final output result is P1(1, 16, 3, 8). The calculation formula for the output size P' of the pooling layer is:

[0061] P' = (L - M) / N + 1

[0062] L is the input size of the pooling layer, M is the pooling size, and N is the pooling stride.

[0063] Step 3.4: Input the output result P1 of the first phase of the entire classification model into the Convolutional Block Attention Module (CBAM) layer for the second-phase processing. CBAM includes a channel attention module and a spatial attention module, and these two modules are arranged in series;

[0064] Among them, the channel attention module compresses the spatial dimension of the input feature map, and at the same time uses the average pooling feature and the maximum pooling feature. The feature map M c (F) of the channel attention module is calculated by the formula:

[0065]

[0066] Among them, F is the input feature map, AvgPool and MaxPool respectively represent the average pooling operation and the maximum pooling operation, F c avg and F c max respectively represent the average pooling feature and the maximum pooling feature. These two descriptors are input into the same shared network to generate the channel attention map Mc. The shared network consists of a multi-layer perceptron (MLP), which includes a hidden layer. After applying the shared network to each descriptor, the output feature vectors are summed and combined. σ represents the sigmoid function. This Mc(F) is multiplied element-wise with F to obtain F’.

[0067] The spatial attention module is a supplement to the channel attention, and the focus is on which part of the channel is the most informative. Apply the average pooling and maximum pooling operations along the channel direction, splice the results, and use a convolutional layer to generate the spatial attention map Ms(F), and the formula is:

[0068]

[0069] Aggregate the results of the average pooling and the maximum pooling into the channel information of a feature map, σ represents the sigmoid function, and f 7×7 represents a convolutional operation with a convolutional kernel size of 7×7.

[0070] The channel attention module and the spatial attention module calculate a one-dimensional channel attention map Mc and a two-dimensional spatial attention map Ms in sequence. The final CBAM calculation formula is:

[0071]

[0072] Among them, F′ is the output result of the first phase of CBAM, represents the matrix dot multiplication operation, and F″ is the final output result of CBAM.

[0073] Step 3.5: Input the output result into the fully connected layer for the last stage of processing. The first fully connected layer takes a vector of 16 * 3 * 8 = 768 dimensions as input. After calculating the dot product between the input vector and the weight vector and adding a bias term, the result L1 is output after being mapped by the ReLU function, with a dimension of 120. The second fully connected layer receives L1 and continues to output a 78-dimensional vector L2 through the ReLU function.

[0074] To prevent overfitting, the Dropout function is used to process the data before L2 is input into the last fully connected layer. During the forward propagation of the data, this function can make the value of a certain neuron stop working with a certain probability p and stretch the remaining data, so that the model does not overly rely on some local features, playing a role in enhancing the generalization ability of the model. The final output result of this layer is D(1, 68).

[0075] The last fully connected layer is also called the output layer. It has 7 neurons, which respectively represent the numbers 0 - 7, that is, the classification labels added to the faulty data and normal data. After inputting D(1, 68) into this layer, a vector L3 with a dimension of (1, 7) is obtained.

[0076] Finally, the Softmax function is used to map the output results of the seven neurons to the interval (0, 1). The sum of these values is 1, satisfying the properties of probability. The label corresponding to the maximum probability is our final classification result. The formula is:

[0077]

[0078] where Z represents each element in the vector, K represents the dimension of the vector, and S j represents the Softmax value of this element.

[0079] Step 4: Use the cross-entropy loss function and the Adam optimizer to train the model. Cross-entropy can measure the distribution difference between the true probability and the predicted probability. The smaller the value, the better the prediction effect of the model. It is often used in combination with the Softmax function to calculate the final loss value.

[0080] Under experimental conditions, the training data was input into the model. After 25 iterations of training, the loss value approached 0.01 and remained stable. At this time, the model state was optimal and the training was stopped. Then, 1000 test samples (including six types of fault samples and one normal sample) were input into the model to test its diagnostic effect. The test classification accuracy reached 94%. Under the same experimental conditions, the classification accuracy achieved by the ordinary convolutional neural network in the model feature extraction layer was 80%, and the classification accuracy achieved by removing the spatial attention layer in the fusion attention mechanism was 89%. The comparison results are shown in Table 1. It can be seen that the proposed method finally achieved accurate classification of train turnout faults.

[0081] Table 1 Classification accuracies of various methods

[0082] Method Diagnostic accuracy Replace dilated convolution with ordinary convolution 80％ Remove the spatial attention layer 89％ LeNN-Attention 94％

[0083] The method of the present invention overcomes the disadvantages of manual feature extraction and heavy dependence on expert knowledge and prior knowledge in traditional data-driven diagnostic methods by automatically extracting features. A new deep learning model is constructed by connecting multiple neural networks in series, so as to intelligently identify the turnout conversion power data and achieve the purpose of automatic diagnosis of turnout faults.

Claims

1. A turnout fault diagnosis method based on a hybrid deep learning model, characterized in that The implementation is specifically carried out according to the following steps: Step 1: Preprocess and normalize the turnout power data, analyze the fault types corresponding to each type of data, and label the fault types reflected by each sample data in the sample dataset according to the set rules; specifically: Step 1.1, the turnout action power curve is one-dimensional time series data. Each data takes 210 sampling points. For the data with less than 210 sampling points, zero-padding is performed at the end, and for the data with more than 210 sampling points, it is intercepted to obtain one-dimensional time series data with the same length; the data is normalized; Step 1.2: Obtain the normalized turnout action power data. Use the turnout action time as the abscissa and the discrete power data as the ordinate to generate the turnout action power curve, and corresponding power curves are formed according to the different actions in each period; Step 1.3: Through Step 1.2, all the power time series data is converted into power curve graphs. According to the different trends of the curves in each stage, six different types of power curves are manually selected from the original sample set, corresponding to six different faults in the turnout conversion process, namely abnormal power fluctuation in the starting stage, power rising and remaining in the conversion stage, power directly dropping to 0 in the stage, large power fluctuation in the stage, power remaining unchanged after normal decline in the stage but not being 0, and abnormal power fluctuation in the conversion stage; Step 2: Use the improved Smote algorithm to expand the fault data and divide the dataset; specifically: Step 2.1: For each type of fault data, randomly select one as the minority class sample X of this type of fault data i , and randomly select another 5 as its nearest neighbor samples Y i , generate new samples using the method of random linear interpolation, and finally add the newly generated data to the original dataset; Step 2.2, use the Smote algorithm to increase each of the 6 types of fault data to 500 pieces, add classification labels to the 500 pieces of normal data and the 500 pieces of fault data in the order of 0 - 6, representing 7 different fault categories. The overall 3500 pieces of data are used as the original dataset, and after shuffling, they are divided into a training set and a test set according to a ratio of 7:3; Step 3: Establish a deep neural network classification model; specifically: Step 3.1: The network model is divided into two parts, a feature extractor and a classifier, according to its functions, where the feature extractor is served by a dilated convolutional layer, and the classifier consists of a first convolutional layer, a max pooling layer, a second convolutional layer, a ReLU activation function layer, a Convolutional Block Attention Module (CBAM) layer for fusion attention mechanism, a fully connected layer, and a DropOut layer; Step 3.2: Input the training set data into the feature extractor. The convolution kernel size of this layer is 2*2, the stride is 2, and the dilation rate is set to 2, which means there is an empty position between every two data points and is filled with 0. The data after feature extraction is used as the input of the classifier; Step 3.3: After feature extraction by the dilated convolutional layer, the dimension of the data becomes (1, 6, 8, 19). Input this data into the classifier for classification; Step 3.4: Input the output result of Step 3.3 into the Convolutional Block Attention Module (CBAM) layer for second-stage processing; CBAM includes a channel attention module and a spatial attention module; Step 3.5: Input the output result into the fully connected layer for the last stage of processing; the first fully connected layer inputs a vector of 16 * 3 * 8 = 768 dimensions, calculates the dot product between the input vector and the weight vector and then adds a bias term, and the result L1 is output after being mapped by the ReLU function, with a dimension of 120; the second fully connected layer receives L1 and continues to output a 78-dimensional vector L2 through the ReLU function; before L2 is input into the last fully connected layer, the Dropout function is used for processing, and the final output result of this layer is D(1, 68); the last fully connected layer, with a total of 7 neurons representing the numbers 0 - 7 respectively, which are the classification labels added for the fault data and the normal data. After inputting D(1, 68) into this layer, a vector L3 with a dimension of (1, 7) is obtained; finally, the Softmax function is used to map the output results of the seven neurons to the interval (0, 1). Step 4: Use the cross-entropy loss function and the Adam optimizer to train the model.

2. The turnout fault diagnosis method based on a hybrid deep learning model according to claim 1, wherein In the said Step 3.3, specifically: First, perform the first convolution operation on the input data. Use 6 convolution kernels of size 2 * 2. During the convolution process, no edge padding is performed, and the stride = 1. A single convolution kernel calculates and outputs a feature map of size 7 * 18. This layer uses a total of 6 convolution kernels, and after overall calculation, 6 feature maps of size 7 * 18 are obtained. Each convolution kernel contains a bias term, and a total of 6 * (2 * 2 + 1) = 30 parameters need to be learned, mainly achieved through weight sharing; after all the feature maps are processed by the ReLU activation function, the final output result C1(1, 6, 7, 18) of the first convolution operation is obtained. Then perform the second convolution operation. After the second convolutional layer receives the output result C1 of the first layer, it uses 16 convolution kernels of size 2 * 2 with a stride of 1, and calculates to obtain 16 feature maps of size 6 * 17; after weight sharing, the total number of parameters is 6 * (2 * 2 * 3 + 1) + 9 * (2 * 2 * 4 + 1) + 1 * (2 * 2 * 6 + 1) = 256. After passing through the ReLU function in the same way, the final result C2(1, 16, 6, 17) of the second convolution operation is obtained; C2 is received by the pooling layer. This layer uses a pooling unit of size 2 * 2 with a stride of 2. For each data in the 2 * 2 area of C2, the sum is calculated and then multiplied by a weight coefficient and then added with a bias. Then the total number of parameters to be learned in this layer is (1 + 1) * 16 = 32; the final output result is P1(1, 16, 3, 8).

3. The turnout fault diagnosis method based on the hybrid deep learning model according to claim 1, wherein, In step 3.4, the channel attention module compresses the spatial dimension of the input feature map, and at the same time uses the average pooling feature and the maximum pooling feature. The calculation formula of the feature map M c c (F) is as follows: Among them, F is the input feature map, AvgPool and MaxPool represent average pooling operation and max pooling operation respectively, F c avg and F c max represent average pooling feature and max pooling feature respectively; σ represents the sigmoid function; MLP is a multi-layer perceptron; The spatial attention module is a supplement to the channel attention. Apply average pooling and max pooling operations along the channel direction, splice the results, and use a convolutional layer to generate the spatial attention map Ms(F). The formula is: The channel attention module and the spatial attention module calculate a one-dimensional channel attention map Mc and a two-dimensional spatial attention map Ms in sequence. The final formula of CBAM is: f 7×7 represents a convolution operation with a convolution kernel size of 7×7; F′ is the output result of the first stage of CBAM, represents a matrix dot product operation, and F″ is the output result of the final CBAM.

Images