Subway train transmission system multi-fault diagnosis method based on FFMC model

Abstract

The invention discloses a subway train transmission system multi-fault diagnosis method based on an FFMC model, and belongs to the technical field of subway train fault diagnosis, and the method comprises the steps: obtaining a fault data set of a plurality of parts of a subway train transmission system, and carrying out the preprocessing; dividing the preprocessed fault data sets of the multiple parts, and carrying out normalization processing; an FFMC model is constructed, and parameters of the FFMC model are initialized; training the FFMC model by using the training set; and testing the trained FFMC model by using the test set to realize classification and diagnosis of multiple faults of the subway train transmission system. According to the method, a traditional convolutional neural network is improved, GRU, multi-scale convolution, cavity convolution and residual connection are fused, and a global average pooling layer is used to replace a full connection layer, so that the feature extraction capability and classification accuracy of the model are improved.

Description

Technical Field

[0001] The invention relates to the technical field of subway train fault diagnosis, and in particular to a subway train transmission system multi-fault diagnosis method based on an FFMC model. Background Art

[0002] As an important part of urban rail transit, the reliability and safety of subway train transmission system are directly related to the stability of subway operation and the safety of passengers. However, the transmission system is composed of multiple complex components, which are prone to various types of failures during long-term operation. Therefore, it is particularly important to diagnose faults in the transmission system in a timely and accurate manner to ensure its normal operation. Previous studies on mechanical component failures of subway train transmission systems mainly used traditional methods, such as applying variational mode decomposition, empirical mode decomposition and other algorithms to decompose vibration signals, and then selecting components for further analysis. However, traditional methods rely heavily on expert experience, and excessive manual intervention inevitably has a certain impact on the diagnosis results.

[0003] Nowadays, with the improvement of intelligence level, the speed and magnitude of data collection of various faults have been significantly improved, laying a good foundation for the wide application of deep learning methods in the field of subway train transmission system fault diagnosis. As a widely used intelligent research method, deep learning has a strong adaptive feature extraction capability, which effectively reduces manual intervention and experience errors in the data analysis process. Therefore, it is gradually being applied by more and more researchers in the field of mechanical fault diagnosis. In "Research on Composite Fault Diagnosis of Urban Rail Train Bearings Based on Deep Learning", Yao Dechen et al. converted the bearing vibration signal into a grayscale image, and then used a convolutional neural network to analyze the obtained grayscale image samples, which can effectively identify the bearing fault of the urban rail train. In "Train Bearing Fault Diagnosis Based on Multi-scale Convolution Intra-class Transfer Learning", Shen Changqing et al. first used the ResNet-50 network to extract the intermediate level features of the data, and then extracted the high-level features through the constructed multi-scale feature extractor on this basis, and finally input the classifier for fault diagnosis, which is used to deal with the fault problem of train bearings under variable working conditions. In "Feature Transfer Learning Method for Limited Variable Operating Conditions and Its Application in Fault Diagnosis of Axle Box Bearings of High-Speed ​​Trains", Luo Honglin et al. transferred the signal feature vector set extracted under different operating conditions to the signal features under standard operating conditions through a supervised autoencoder, and then input the migrated signal features into a convolutional neural network pre-trained with the features of the reference operating condition training set for fault identification, thus realizing the fault diagnosis of train bearings under variable operating conditions.

[0004] Although current research has been conducted in the field of mechanical fault diagnosis, it is all aimed at the fault data of a single train component and has few classifications. However, under some complex conditions, multiple components of the train transmission system may have multiple faults. It is difficult to achieve the expected results using ordinary models to identify and diagnose this, and it is easy to produce problems such as overfitting. Some existing complex models may be able to achieve certain diagnostic effects on multiple component faults, but deep learning models have high hardware requirements, and their training process is relatively lengthy and cannot quickly produce diagnostic results, so they are not applicable in practical applications. Summary of the invention

[0005] The purpose of the present invention is to provide a subway train transmission system multi-fault diagnosis method based on the FFMC model. By constructing the FFMC model, the classification and diagnosis of the subway train transmission system multi-faults are realized.

[0006] To achieve the above object, the present invention provides a multi-fault diagnosis method for a subway train transmission system based on the FFMC model, the steps comprising:

[0007] S1. Acquire various component failure data sets of subway train transmission systems, and pre-process the various component failure data sets;

[0008] S2, dividing the preprocessed various component failure data sets into a training set and a test set, and performing normalization processing;

[0009] S3, improve the convolutional neural network, build the FFMC model and initialize the FFMC model parameters;

[0010] S4, use the training set to train the FFMC model;

[0011] S5. Use the test set to test the trained FFMC model to achieve classification and diagnosis of multiple faults in the subway train transmission system.

[0012] Preferably, the preprocessing of the various component failure data sets in step S1 includes: segmenting sample data in the various component failure data sets, creating labels for the segmented data in the data sets and converting the labels into One-hot encodings.

[0013] Preferably, the FFMC model in step S3 includes an input layer, a spatiotemporal feature fusion layer, a multi-scale convolution layer, a hole convolution layer and a classification layer, and the classification layer includes a global average pooling layer and an output layer.

[0014] Preferably, the spatiotemporal feature fusion layer includes a convolutional layer, a batch normalization layer, a GRU layer, a permutation layer, a reshaping layer, a splicing fusion layer and a maximum pooling layer. The spatiotemporal feature fusion layer uses a convolutional layer and a GRU layer to extract the spatial features and temporal features of the input data respectively, and is processed by a batch normalization layer, and a permutation layer and a reshaping layer are used to perform dimensional conversion and reshaping so that its output shape is the same as the convolutional layer output, and finally a splicing fusion layer is used to fuse the temporal features and the spatial features.

[0015] Preferably, the multi-scale convolution layer includes convolution layers of different scales, residual connection layers, batch normalization layers and maximum pooling layers. The multi-scale convolution layer uses convolution layers with convolution kernels of different sizes to extract features of different scales, then performs feature fusion through the residual connection layer, and uses the batch normalization layer for normalization, and finally uses the maximum pooling layer to reduce the spatial dimension.

[0016] Preferably, the atrous convolution layer includes convolution layers with different atrous rates, residual connection layers, batch normalization layers and maximum pooling layers. The atrous convolution layer extracts features of data with different receptive fields through convolution layers with different atrous rates, performs feature fusion through the residual connection layer, and uses the batch normalization layer for normalization, and finally uses the maximum pooling layer to reduce the spatial dimension.

[0017] Preferably, in step S4, the cross entropy function is used to train the FFMC model, and the minimum loss model parameters on the training set are continuously saved during the training process until the training is completed, and the final minimum loss model is the optimal FFMC model.

[0018] Preferably, in step S5, the generated optimal FFMC model is tested using a test set to achieve classification and diagnosis of multiple faults in the subway train transmission system.

[0019] Therefore, the present invention adopts the above-mentioned subway train transmission system multi-fault diagnosis method based on the FFMC model, which has the following beneficial effects:

[0020] (1) The traditional convolutional neural network was improved by integrating GRU, multi-scale convolution, dilated convolution and residual connection, and using the global average pooling layer instead of the fully connected layer, which improved the feature extraction ability and classification accuracy of the FFMC model;

[0021] (2) A spatiotemporal feature fusion layer is introduced, GRU is used as the temporal feature extractor, and the two-dimensional convolution kernel is used as the spatial feature extractor to fuse the extracted temporal and spatial features. The multi-scale convolution layer and the dilated convolution layer are fused through residual connection, which effectively improves the accuracy of feature extraction, reduces the risk of overfitting, and ensures the stability of the FFMC model.

[0022] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of the method of the present invention;

[0024] Figure 2 This is a diagram of the structure of a common convolutional neural network;

[0025] Figure 3 It is the overall structure diagram of the FFMC model of the present invention;

[0026] Figure 4 Schematic diagram of multi-scale convolution;

[0027] Figure 5 Schematic diagram of dilated convolution;

[0028] Figure 6 Schematic diagram of residual connection;

[0029] Figure 7 This is the structure diagram of the gated recurrent unit GRU;

[0030] Figure 8 This is the structure diagram of the fault simulation test bench for the PHM dataset;

[0031] Fig. 9 It is a time domain diagram of a single sample of 13 types of data;

[0032] Fig.10 It is the training and test accuracy curve;

[0033] Fig.11 Loss curves for training and testing;

[0034] Fig.12 To test the confusion matrix diagram;

[0035] Fig.13 It is the time domain diagram of 13 samples;

[0036] Fig.14 Visualize the feature extraction process;

[0037] Fig.15 It is a three-dimensional visualization of the weighted normal distribution;

[0038] Fig.16 It is a two-dimensional visualization of the weighted normal distribution;

[0039] Fig.17 This is a KPCA dimensionality reduction visualization diagram of the original data distribution;

[0040] Fig.18 KPCA dimension reduction visualization diagram of data distribution after training classification;

[0041] Fig.19 Test accuracy curves for different diagnostic methods;

[0042] Fig. 20 Testing loss curves for different diagnostic methods;

[0043] Fig.21 Confusion matrix diagram for new test data;

[0044] Fig. 22 The time domain diagram of the signal with different signal-to-noise ratio Gaussian white noise added;

[0045] Fig.23 This is a line chart of the test accuracy of different models in the ablation experiment;

[0046] Fig.24 A bar chart showing the accuracy of the dataset divided into different proportions for different diagnostic methods. DETAILED DESCRIPTION

[0047] Example

[0048] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0049] Reference Figure 1 The present invention provides a multi-fault diagnosis method for a subway train transmission system based on an FFMC model, the steps comprising:

[0050] S1. Obtain a dataset of various component failures of a subway train transmission system, and preprocess the dataset. First, divide the sample data in the dataset into 1024 points per sample data point, and then create labels for the divided data in the dataset and convert the labels into One-hot encoding.

[0051] S2. Divide the preprocessed various component failure data sets into training sets and test sets in proportion, and perform normalization.

[0052] S3. Improve the convolutional neural network by integrating GRU, multi-scale convolution, dilated convolution and residual connection, and use the global average pooling layer to replace the fully connected layer, build the FFMC model and initialize the FFMC model parameters.

[0053] The traditional convolutional neural network consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Figure 2 shown.

[0054] The improved model, namely the FFMC model, includes an input layer, a spatiotemporal feature fusion layer, a multi-scale convolution layer, a hole convolution layer and a classification layer. The classification layer includes a global average pooling layer and an output layer.

[0055] The schematic diagram of the FFMC model structure is as follows: Figure 3 The schematic diagram of FFMC model parameters is shown in Table 1.

[0056] Table 1 FFMC model parameter settings

[0057]

[0058] The spatiotemporal feature fusion layer uses a 2D convolution kernel (Conv2D) with a size of 8×8 to extract the spatial features of the data, and then processes it through a batch normalization (BN) layer to improve the expressiveness of the features. The GRU layer with a unit number of 32 is used to extract the temporal features of the data, and then processes it through a BN layer. Finally, the permutation layer and the reshaping layer are used to permute and reshape the dimensions so that the shape of the GRU layer output is consistent with that of the convolution layer output. The extracted temporal and spatial features are fused through the splicing fusion layer, and then a maximum pooling layer with a kernel size of (2×2) and a stride of 1 is added, which helps to reduce the spatial dimension of the feature map while retaining key feature information. The multi-scale convolution layer includes three parallel 2D convolution kernels (Conv2D) with sizes of 3×3, 6×6, and 9×9, respectively, which are used to extract features of different scales of the data. The output is fused through a residual connection layer, and then processed through a BN layer to improve the expressiveness of the features. Finally, a maximum pooling layer with a kernel size of (3×3) and a stride of 2 is used to further reduce the spatial size of the feature map and reduce the computational complexity. The dilated convolution layer includes three parallel two-dimensional convolution kernels (Conv2D) with a size of 3×3 and dilation rates of DR=1, DR=3, and DR=5, respectively, which are used to extract features of different receptive fields of the data. The output is fused through the residual connection layer, and then processed through the batch BN layer to improve the expressiveness of the features. Finally, the maximum pooling layer with a kernel size of (3×3) and a stride of 2 is used to further reduce the size of the feature map. The classification layer consists of a global average pooling layer and an output layer (Softmax). The global average pooling layer replaces the fully connected layer, integrates the global feature information while reducing the dimension, and Softmax outputs the fault diagnosis results.

[0059] Multi-scale convolution and dilated convolution:

[0060] When a mechanical component fails, the fault types are diverse, the transmission paths are complex, and it is easy to couple with other components, resulting in the vibration signal showing multi-scale characteristics. Vibration data reflects multi-scale fault characteristics. Multi-scale convolution uses convolution kernels of different sizes to operate in parallel and can extract features of different scales, such as Figure 4 shown.

[0061] Dilated convolution is also called dilated convolution. It is a convolutional neural network method that expands the receptive field of the output unit without increasing the number of parameters. It is implemented by inserting holes between adjacent elements of the convolution kernel, which is equivalent to adding zero elements between these adjacent elements. The number of inserted holes is called the dilation rate and is recorded as DR. Taking a 3×3 convolution kernel as an example, dilated convolutions with different dilation rates are as follows: Figure 5 As shown in the figure, the hole convolution can expand the receptive field of the output unit, thereby enhancing the network's ability to extract features of different scales.

[0062] Residual Connection:

[0063] Residual connections solve the problem of gradient vanishing or gradient exploding during training as the network depth increases. Through the skip connection of the shortcut path, the network can directly pass the input data to the subsequent layers of the network. Figure 6 As shown in Figure 1, this connection method forms an identity mapping that allows the network to learn the residual between the input and output. The residual block in the figure mainly includes the main path and the shortcut path. The main path includes several convolutional layers, activation functions, etc. for extracting features; the shortcut path directly passes the input data to the output of the residual block. The output of the residual block is the element-by-element addition of the output of the main path and the output of the shortcut path. The operation process is as follows:

[0064] H(x)=F(x)+x

[0065] Among them, x represents the input data, F(x) represents the features after processing by the main path, and H(x) represents the final output of the residual block.

[0066] Batch Normalization (BN):

[0067] Batch normalization is a method used for data normalization. The process of batch normalization for any output data in training is as follows:

[0068] Calculate the average of a batch of data:

[0069]

[0070] Calculate the variance of batch data:

[0071]

[0072] Normalized data:

[0073]

[0074] Scaling and offsetting:

[0075]

[0076] Where m represents the batch size; ε represents a constant term to ensure numerical stability; γ and β represent the scale factor and offset factor learned by the network, respectively; Represents the output of the BN layer. Through batch normalization, the output data of each layer can maintain a normal distribution, thereby accelerating convergence and reducing the offset of internal covariates during training, improving the generalization ability and training speed of the model.

[0077] Global Average Pooling:

[0078] Traditional convolutional neural networks usually use multiple fully connected layers to reduce feature dimensionality, but the fully connected layer uses a fully connected method, which generates a large number of parameters during the training process, resulting in high model calculation complexity and prone to overfitting problems. The global average pooling layer (GAP) associates categories with feature maps, so that each channel has actual category meaning. Each feature map can be regarded as a confidence map of the category, and there are no parameters to be optimized inside it. While reducing the dimensionality, it greatly reduces the parameters of the network model and integrates the global feature information, which can effectively avoid overfitting.

[0079] Gated Recurrent Unit (GRU):

[0080] GRU has a reset gate and an update gate. The unique gate structures of the two can fully utilize the long-term dependencies of the input data. Figure 7 As shown. The reset gate determines whether to discard information in the candidate hidden state to achieve the forgetting function through the input and the hidden state of the previous moment; the update gate controls whether the current input and the hidden state information of the previous moment are updated to the candidate hidden state. The calculation formula is as follows:

[0081] Update gate expression:

[0082] Z t =σ(W Z *[h t-1 , x t ])

[0083] Reset the gate expression:

[0084] R t =σ(W r *[h t-1 , x t ])

[0085] Candidate hidden state expressions:

[0086]

[0087] Hidden state expression:

[0088]

[0089] Among them, W, W Z , W r represents the learnable weight, and σ represents the Sigmoid activation function.

[0090] S4. Use the cross entropy function to train the FFMC model. During the training process, the minimum loss model parameters on the training set are continuously saved until the end of the training. The final minimum loss model is the optimal FFMC model.

[0091] S5. Use the test set to test the optimal FFMC model generated in step S4 to achieve classification and diagnosis of multiple faults in the subway train transmission system.

[0092] Example 1

[0093] The PHM dataset is used to visualize the classification results, feature extraction, and weight distribution of the model. The PHM dataset is compared with three typical models: AlexNet, GhostNet, and MobileNetV2 to verify the classification effect, robustness, and noise resistance of the FFMC model.

[0094] The PHM dataset is collected from the subway train transmission system fault simulation test bench, such as Figure 8 As shown in the figure, the test bench is based on the real subway bogie and is reduced and adjusted at a ratio of 1:2. A single transmission chain in the test bench includes a motor, a reduction gearbox and an axle box. The transmission chain is driven by a three-phase asynchronous AC motor, and the load is applied by a hydraulic device. The motor bearing model is SKF6205-2RSH; the gear of the reduction gearbox is a helical gear, the number of teeth of the driving gear is 16, and the number of teeth of the driven gear is 107; the driving gear support bearing model is Harbin Bearing Factory 32305; the axle box bearing model is Harbin Bearing Factory 352213. The data set contains 9 working conditions, namely WC1, WC2, WC3, WC4, WC5, WC6, WC7, WC8 and WC9. The train running speed is simulated by different motor speeds, and the train straight driving and cornering are simulated by different lateral loads. The experiment collects two types of signals, three-way vibration and three-phase current, with a total of 21 channels, and the sampling frequency of each channel is 64kHz.

[0095] The selected components are the traction motor, gearbox, and axle box bearing, and the data sets are measured in the x-direction with three-way vibration of WC1, WC2, and WC3. The data collected from the driving end of the traction motor under normal, short circuit, broken bar, and shaft bending conditions are used as experimental data; the data collected from the input shaft of the gearbox under normal, tooth root cracks, tooth surface wear, broken teeth, and inner and outer ring failures of the driving gear support shaft are experimental data; the data collected under the inner and outer ring failures of the left axle box bearing are experimental data, and their label settings are shown in Table 2. Take 256,000 sampling points for each selected data set and divide them into 250 groups. Each group is used as 1 sample, and each sample has 1024 sampling points. The time domain of a single sample of each type of data is as follows: Fig. 9 shown.

[0096] Table 2 Data types and label numbers

[0097]

[0098]

[0099] The deep learning framework Tensorflow was used, and the computer configuration selected Core(TM) i5-13400F CPU processor and NVIDIA GeForce RTX4060 graphics card. It was divided into training and test groups in an 8:2 ratio, with 200 groups put into training and the remaining 50 groups for testing. Batch Size was set to 256, the number of iterations was set to 200, Adam was selected as the optimizer of the model, the learning rate was set to 0.001, and the cross entropy loss function was used as the loss function of the model.

[0100] The training and test accuracy curves are as follows Fig.10 As shown, the loss curve is Fig.11 As shown in the figure, the accuracy curve of the training set has completely converged after 30 iterations, and the accuracy reaches 100%. The loss curve decreases rapidly with the iteration and completely converges after 45 iterations, and the loss is infinitely close to 0. The accuracy curve of the test set has completely converged after 108 iterations, reaching 99.38%. The loss curve decreases rapidly with the iteration and completely converges after 95 iterations, and is infinitely close to 0.

[0101] The confusion matrix of the test set is as follows Fig.12 As shown in the figure, the horizontal axis represents the predicted label and the vertical axis represents the actual label. The main diagonal basically reaches 50, which is equal to the number of test sets for each data, and the recognition accuracy rate reaches 100%. There are slight deviations in the categories with labels 3 and 9, and the number of correct predictions is 49 and 48, achieving an extremely high recognition accuracy rate.

[0102] Select the first sample of the 13 types of data in the test set, convert each sample into a two-dimensional array of 1024 data points and a data structure of (32,32,1) in order from left to right and from top to bottom to a one-dimensional array of (1024,1), and draw the time domain graph in sequence with the horizontal axis being the number of data points and the value corresponding to the point, as shown below: Fig.13 shown. Fig.14 This is a visualized line graph of the feature extraction process of all convolutional layers and pooling layers for the 13 samples selected above during training. It can be seen from the figure that as the training process progresses, the original different types of feature points are mixed together, the feature points of the same type are gradually reduced, and the feature points of different types are gradually dispersed. Finally, a small number of feature points with relatively high discrimination are extracted to basically separate the 13 samples.

[0103] The convolutional layers are initialized using a random normal distribution. The weight values ​​and density distribution of each convolutional layer are visualized in three dimensions as the number of iterations changes. Fig.15 Its two-dimensional visualization is shown in Fig.16 As shown. By comparing the two graphs, we can get the following information:

[0104] The overall distribution range of the convolution kernel weight values ​​of the spatiotemporal feature fusion layer (Layer-1) is (-0.14, 0.16), and the main distribution range is (-0.06, 0.06); the overall density distribution range of the weight values ​​is (0, 15), and the main density distribution range is (12, 15). The overall distribution range of the convolution kernel weight values ​​of the multi-scale convolution layer (Layer-2) is (-0.21, 0.13), and the main distribution range is (-0.07, 0.07); the overall density distribution range of the weight values ​​is (0, 10), and the main density distribution range is (8, 10). The overall distribution range of the convolution kernel weight values ​​of the atrous convolution layer (Layer-3) is (-0.18, 0.18), and the main distribution range is (-0.10, 0.10); the overall density distribution range of the weight values ​​is (0, 6.5), and the main density distribution range is (5, 6.4).

[0105] The distribution range of weight values ​​of each layer gradually increases with the number of iterations and remains basically unchanged around iteration 185. The maximum density distribution of weight values ​​of the spatiotemporal feature fusion layer, multi-scale convolution layer and dilated convolution layer decreases successively, proving that the model learns richer subtle features.

[0106] Fig.17 , Fig.18 The KPCA dimensionality reduction visualizations are the original data distribution and the KPCA dimensionality reduction visualization after training and classification. The comparison shows that the distribution of the original data in the figure is relatively scattered, and various data are mixed together. As the training process progresses, different types of data points gradually disperse, while the same type of data points gradually gather and finally separate completely, achieving excellent classification results.

[0107] Classification effect

[0108] Fig.19 and Fig. 20 The figure is a comparison diagram of the test accuracy and loss curves of the FFMC model and three typical models, AlexNet, GhostNet and MobileNetV2. It can be seen from the figure that the FFMC model has the best effect. The accuracy curve and loss curve converge the fastest among various methods. The accuracy reaches 99.38%, the highest among various methods, and the loss is the lowest among various methods, infinitely close to 0. In contrast, the effect of direct diagnosis using methods such as AlexNet, GhostNet and MobileNetV2 is poor. The test accuracy of the three methods can only be maintained at about 95% at most, and the test loss is also much higher than the FFMC model. In addition, when the batch size and the number of iterations are the same, the test curves of the three typical methods fluctuate greatly, making it difficult to converge effectively, and the effect is poor.

[0109] robustness

[0110] Select new data of the same data type to build a new test data set. Take 102400 sampling points for each type of data, divide each type of data into 100 samples, and each sample has 1024 sampling points. The label setting is the same as Table 2. Save the trained and converged models of FFMC, AlexNet, GhostNet and MobileNetV2 in turn, and use the new test data to conduct 5 experiments respectively.

[0111] like Fig.21 As shown in the figure, it is a schematic diagram of the confusion matrix of the first test using the FFMC model. The main diagonal basically reaches 100, which is equal to the number of test samples for each data type, and the recognition accuracy rate reaches 100%. There is a slight deviation in the label 4 category, and the number of correct predictions is 99, but it also achieves an extremely high recognition accuracy rate.

[0112] The overall average accuracy of each test result of different models is shown in Table 3. As can be seen from the table, the test accuracy of the FFMC model is stable at more than 99%. The test accuracy of the other three models is unstable and the diagnostic effect is poor. The highest test accuracy is 96.62%, which is much lower than the model of the present invention.

[0113] Table 3 Test accuracy of different models

[0114]

[0115] Noise immunity

[0116] Gaussian white noise with signal-to-noise ratio (SNR) of -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, and 6dB is added to the new test data set. The single sample of the gearbox driving gear support shaft inner ring fault signal is disturbed by the noise. Fig. 22 As shown in the figure. As the noise signal-to-noise ratio decreases, the bearing vibration signal is gradually covered by noise and is almost distorted below -4dB. The models saved by training FFMC, AlexNet, GhostNet and MobileNetV2 without adding noise data were tested. The test data with different noises were tested 5 times. The average test accuracy of each noise is the average of the accuracy of the 5 experiments, which is shown in Table 4. As can be seen from the table, the average accuracy of the FFMC model is higher than that of other models in both the noise-free and noise-containing cases, which fully verifies that the model has excellent robustness.

[0117] Table 4 Test accuracy of different noise data

[0118]

[0119] Example 2

[0120] In order to verify the rationality of the components of the model of the present invention, an ablation experiment was conducted and 9 models were set for comparative experiments. The specific settings are as follows:

[0121] Model 1: The FFMC model used.

[0122] Model 2: Average pooling layer replaces the max pooling layer.

[0123] Model 3: Remove the GRU layer in the spatiotemporal feature fusion layer.

[0124] Model 4: Ordinary convolutional layers replace multi-scale convolutional layers.

[0125] Model 5: Ordinary convolutional layers replace dilated convolutional layers.

[0126] Model 6: Ordinary concatenated feature fusion layer replaces the residual connection feature fusion layer.

[0127] Model 7: The convolution kernel sizes of the multi-scale convolutional layers are set to (3×3), (7×7), and (12×12) respectively.

[0128] Model 8: The dilation rates of the atrous convolutional layers are set to DR=1, DR=2, and DR=3 respectively.

[0129] Model 9: Fully connected layers replace global average pooling layers.

[0130] The accuracy results of each model after 5 tests are as follows Fig.23As shown in the figure, the accuracy of FFMC is higher than that of the other eight models, which fully verifies the rationality and superiority of the components of the proposed model.

[0131] Example 3

[0132] To verify the generalization of the FFMC method of the present invention, the data set taken in Example 1 was divided into a training set and a test set according to the ratio of 7:3 and 6:4 respectively. During the training, the batch size was set to 256 and the number of iterations was 200. Fig.24 As shown in the figure, as the data set division ratio changes, the test accuracy of the diagnosis using the FFMC model remains stable at 99%, while the test accuracy of the diagnosis methods of the AlexNet, GhostNet and MobileNetV2 models fluctuates greatly compared with the previous 8:2 division of the data set, proving that the FFMC model has excellent classification effects in fault diagnosis of various components.

[0133] Therefore, the present invention improves the classification effect of fault diagnosis of various components of the subway train transmission system and improves the classification accuracy by constructing the FFMC model through the above-mentioned subway train transmission system multi-fault diagnosis method based on the FFMC model.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the preferred embodiments have described the present invention in detail, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-fault diagnosis method for subway train transmission system based on FFMC model, characterized in that the steps include: S1. Acquire various component failure data sets of subway train transmission systems, and pre-process the various component failure data sets; S2, dividing the preprocessed various component failure data sets into a training set and a test set, and performing normalization processing; S3, improve the convolutional neural network, build the FFMC model and initialize the FFMC model parameters; S4, use the training set to train the FFMC model; S5. Use the test set to test the trained FFMC model to achieve classification and diagnosis of multiple faults in the subway train transmission system.

2. The multi-fault diagnosis method for subway train transmission system based on FFMC model according to claim 1 is characterized in that: The preprocessing of the various component failure data sets in step S1 includes: segmenting the sample data in the various component failure data sets, creating labels for the segmented data in the data sets, and converting the labels into One-hot codes.

3. The multi-fault diagnosis method for subway train transmission system based on FFMC model according to claim 1 is characterized in that: The FFMC model in step S3 includes an input layer, a spatiotemporal feature fusion layer, a multi-scale convolution layer, a hole convolution layer and a classification layer, and the classification layer includes a global average pooling layer and an output layer.

4. The method for multi-fault diagnosis of subway train transmission system based on FFMC model according to claim 3 is characterized in that: The spatiotemporal feature fusion layer includes a convolutional layer, a batch normalization layer, a GRU layer, a permutation layer, a reshaping layer, a splicing fusion layer and a maximum pooling layer. The spatiotemporal feature fusion layer uses a convolutional layer and a GRU layer to extract the spatial features and temporal features of the input data respectively, and is processed by a batch normalization layer. A permutation layer and a reshaping layer are used to perform dimensional conversion and reshaping so that its output shape is the same as the convolutional layer output. Finally, a splicing fusion layer is used to fuse the temporal features and the spatial features.

5. The method for multi-fault diagnosis of subway train transmission system based on FFMC model according to claim 3 is characterized in that: The multi-scale convolution layer includes convolution layers of different scales, residual connection layers, batch normalization layers and maximum pooling layers. The multi-scale convolution layer uses convolution layers with convolution kernels of different sizes to extract features of different scales, then performs feature fusion through the residual connection layer, and uses the batch normalization layer for normalization, and finally uses the maximum pooling layer to reduce the spatial dimension.

6. The method for multi-fault diagnosis of subway train transmission system based on FFMC model according to claim 3 is characterized in that: The atrous convolution layer includes convolution layers with different atrous rates, residual connection layers, batch normalization layers and maximum pooling layers. The atrous convolution layer extracts features of different receptive fields of data through convolution layers with different atrous rates, performs feature fusion through the residual connection layer, and uses the batch normalization layer for normalization. Finally, the maximum pooling layer is used to reduce the spatial dimension.

7. The method for multi-fault diagnosis of subway train transmission system based on FFMC model according to claim 1 is characterized in that: In step S4, the cross entropy function is used to train the FFMC model. During the training process, the minimum loss model parameters on the training set are continuously saved until the training is completed. The final minimum loss model is the optimal FFMC model.

8. The method for multi-fault diagnosis of subway train transmission system based on FFMC model according to claim 7 is characterized in that: In step S5, the generated optimal FFMC model is tested using a test set to achieve classification and diagnosis of multiple faults in the subway train transmission system.

Images