Subway bearing label confusion data fault diagnosis method based on DSS-CNN model

Abstract

The invention relates to the technical field of bearing fault diagnosis, and particularly discloses a subway bearing label confusion data fault diagnosis method based on a DSS-CNN model. Comprising the following steps: S1, carrying out data preprocessing, dividing a data set and carrying out normalization processing; s2, initial training and dynamic sample selection: inputting a whole data set which is constructed by PHM experimental data and contains a small number of label error samples into a DSS-CNN for initial training until the model converges; s3, performing final training and model optimization to obtain an optimal network model; s4, visually displaying a result; and S5, evaluating the diagnosis effect of the model. According to the method, the problem of tag confusion possibly occurring in the current signal acquisition process is solved, the influence of tag confusion on a diagnosis model is reduced, fault diagnosis of the subway bearing containing a small amount of tag error sample data is realized, the data has availability, and resources and cost are saved.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a subway bearing label confusion data fault diagnosis method based on a DSS-CNN model. Background Art

[0002] Railway transportation is the main artery of the national economy and plays an extremely important role in the national social development. In the subway system, the gearbox, as the core component of the transmission system, carries the important task of transmitting the power of the motor to the wheels. The health of the gearbox bearing directly affects the running safety and stability of the train. Therefore, it is extremely important to timely and effectively identify and diagnose the faults of the subway gearbox bearings.

[0003] Previous research on gearbox bearing faults mainly used traditional methods, such as applying variational mode decomposition, empirical mode decomposition and other algorithms to decompose the bearing vibration signal, and then selecting the components for further analysis. However, traditional methods rely heavily on expert experience, and too much manual intervention will inevitably have a certain impact on the diagnosis results. Nowadays, with the improvement of the level of intelligence, the collection speed and order of magnitude of various types of bearing data have been significantly improved, laying a good foundation for the widespread application of deep learning methods in the field of bearing 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 process of bearing data analysis. Therefore, it is gradually being applied by more and more researchers to the field of bearing fault diagnosis. For example:

[0004] In the study "Research on Complex 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 urban rail train bearing faults.

[0005] In "Train Bearing Fault Diagnosis Based on Multi-Scale Convolutional 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 constructed a multi-scale feature extractor to extract high-level features. Finally, the features were input into the classifier for fault diagnosis, which is used to deal with the failure problems of train bearings under variable operating conditions.

[0006] 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.

[0007] In "High-speed train bearing fault diagnosis based on improved VMD and APSO-SVM", Zhang Qingsong et al. performed variational modal decomposition on the original signal, then calculated the parameter-optimized Hurst index eigenvalues ​​of each modal component, and finally input the eigenvector into a support vector machine for analysis, which was used to identify the weak fault characteristic information of high-speed train wheelset bearings.

[0008] The current research mainly focuses on the situation where the various labels of bearing data correspond to the signal type completely correctly. However, in some cases, due to human errors, errors will occur when setting labels for the collected signals, resulting in multiple samples of a type of signal containing some samples with incorrect labels, which will greatly interfere with the feature extraction of the signal. It is also difficult to achieve the expected results for the identification and diagnosis of such data using traditional deep learning models. The data collection of gearbox bearings of subway trains is relatively difficult. Collecting new data again due to a small number of incorrectly labeled samples will waste a lot of time and resources. Therefore, it is necessary to solve the problem of label confusion that may occur during signal collection, weaken the impact of label confusion on the diagnostic model, and make such data available. Summary of the invention

[0009] In view of the above-mentioned problems in the prior art, the present invention provides a subway bearing label confusion data fault diagnosis method based on the DSS-CNN model, solves the problem of label confusion that may occur in the current signal acquisition process, reduces the impact of label confusion on the diagnosis model, and realizes fault diagnosis of subway bearings containing a small amount of label error sample data.

[0010] The technical solution adopted by the present invention is as follows:

[0011] A subway bearing label confusion data fault diagnosis method based on a DSS-CNN model includes the following steps:

[0012] S1. Perform data preprocessing and data set division. Divide the data set into training set and test set according to a certain ratio and perform normalization processing.

[0013] S2, initial training and dynamic sample selection, the entire dataset containing a small number of mislabeled samples constructed by PHM experimental data is input into DSS-CNN for initial training until the model converges;

[0014] S3, final training and model optimization, use the screened high-quality training and test samples to continue training the model, and keep saving the minimum loss model parameters on the test set throughout the training process until the end of training. The final minimum loss model is the optimal network model;

[0015] S4. Visualize the model’s classification results, feature extraction, weight distribution, and the correct and incorrect samples that have been screened;

[0016] S5. After training is completed, save the trained model and use a new labeled correct data set for testing to evaluate the model's diagnostic effect.

[0017] Preferably, in step S1, the constructed data with a small amount of label confusion is read, the data is divided into 1024 points of a single sample data, labels are created for the data set and the labels are converted into One-hot encoding.

[0018] Preferably, the DSS-CNN model in step S2 consists of five parts: a screening layer, a traditional convolution layer, a multi-scale convolution layer, a hole convolution layer and a classification layer.

[0019] Preferably, the classification layer of the DSS-CNN model in step S2 is divided into a global average pooling layer and a Softmax.

[0020] Preferably, in step S2, the data set is input into DSS-CNN for initial training to achieve model convergence, comprising the following steps:

[0021] After the model converges in the initial training phase S21, the screening layer is activated, and the model trained in the initial phase is called to predict each type of data and dynamically freeze samples with inconsistent labels, retaining samples with consistent prediction results and labels;

[0022] S22 uses convolution kernels to extract features in the traditional convolution layer, and combines it with BN processing, adding a maximum pooling layer with KernelSize of (2×2) and Stride of 1 to retain key feature information;

[0023] S23 performs three parallel convolution operations in the multi-scale convolution layer to capture the features of two-dimensional data at different scales. The output is fused through residual connection features and combined with BN processing. Finally, it passes through the maximum pooling layer with KernelSize of (3×3) and Stride of 2.

[0024] S24 performs parallel convolution operations with dilation rates of DR=1, DR=3, and DR=5 in the dilated convolution layer to capture the features of different receptive fields of two-dimensional data. The output is fused through residual connection features, followed by batch BN processing, and finally passes through the maximum pooling layer with KernelSize of (3×3) and Stride of 2.

[0025] S25 integrates global feature information at the classification layer and outputs fault diagnosis results.

[0026] Preferably, the traditional convolution layer in step S22 uses a convolution kernel with a size of 8×8.

[0027] Preferably, the three convolution sizes of the multi-scale convolution layer in step S23 are 3×3, 6×6, and 9×9.

[0028] Preferably, the three convolution sizes of the atrous convolution layer in step S24 are 3×3.

[0029] Preferably, in step S25, the global average pooling layer replaces the fully connected layer, and Softmax outputs the fault diagnosis result.

[0030] Preferably, in step S21, the screening layer is located after the input layer, and the model dynamically screens the samples, retaining only the samples whose prediction results are consistent with the labels for final training, and no sample screening is performed in the final training stage.

[0031] The present invention has the following characteristics and advantages:

[0032] (1) The present invention has achieved excellent classification and diagnosis results for subway bearing fault diagnosis with a small amount of sample data with label errors, and can obtain extremely high diagnostic accuracy in a relatively short time. The PHM experimental data was used to construct a dataset with label errors and input it into the DSS-CNN model for fault identification and diagnosis, achieving excellent performance;

[0033] (2) Based on the traditional convolutional layer, multi-scale convolution, dilated convolution and residual connection are integrated, and the fully connected layer is replaced by the global average pooling layer. This design improves the model's feature extraction ability and classification accuracy;

[0034] (3) The screening layer introduced after the output layer can reduce the interference of label errors on model training through a dynamic sample selection mechanism.

[0035] (4) The distribution, feature extraction, weight distribution, and classification process of the correct and incorrect samples screened are visualized, which improves the interpretability of the model and further verifies the effectiveness and robustness of the model. The extracted feature vectors are output as classification results through the classification layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of the implementation method of the present invention;

[0037] Figure 2 It is the overall structure diagram of the DSS-CNN model of the present invention;

[0038] Figure 3 This is a diagram of the structure of a common convolutional neural network;

[0039] Figure 4 It is a schematic diagram of multi-scale convolution;

[0040] Figure 5 This is a schematic diagram of a dilated convolution;

[0041] Figure 6 It is a schematic diagram of residual connection.

[0042] Figure 7 This is the structure diagram of the fault simulation test bench for the PHM dataset;

[0043] Figure 8 It is the time domain graph of the label-confused raw data;

[0044] Fig. 9 It is the training and test accuracy curve;

[0045] Fig.10 is the training and testing loss curve;

[0046] Fig.11 It is a t-SNE dimension reduction visualization diagram of the screened sample distribution;

[0047] Fig.12 It is the time domain diagram of all correct and incorrect samples screened;

[0048] Fig.13 It is the time domain diagram of 5 samples;

[0049] Fig.14 It is a line chart that visualizes the feature extraction process;

[0050] Fig.15 It is a three-dimensional visualization of the weighted normal distribution;

[0051] Fig.16 It is a two-dimensional visualization of the weighted normal distribution;

[0052] Fig.17 This is the KPCA dimensionality reduction visualization diagram of the original data distribution;

[0053] Fig.18 It is a KPCA dimensionality reduction visualization diagram of data distribution after training classification;

[0054] Fig.19 It is a curve chart of test accuracy of different diagnostic methods;

[0055] Fig. 20 It is a graph of test loss curves of different diagnostic methods;

[0056] Fig.21 is the initial training model test confusion matrix diagram;

[0057] Fig. 22 This is the final training model test confusion matrix diagram;

[0058] Fig.23 It is the time domain diagram of the signal with different signal-to-noise ratio Gaussian white noise added;

[0059] Fig.24 This is a line chart of the test accuracy of different models in the ablation experiment;

[0060] Fig.25 It is a bar chart of the accuracy of the test of equal-proportion label confusion data. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments.

[0062] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0063] The implementation method flow chart of a subway bearing label confusion data fault diagnosis method based on DSS-CNN model is as follows: Figure 1 As shown, the following steps are included:

[0064] S1. Perform data preprocessing: read the constructed data with a small amount of label confusion, split the data according to the number of single sample data points of 1024, label the data set and convert the labels into one-hot encoding, divide the data set into training set and test set according to a certain ratio and perform normalization;

[0065] S2 Initial training and dynamic sample selection: The entire dataset containing a small number of incorrectly labeled samples constructed from PHM experimental data is input into DSS-CNN for initial training until the model converges, and the training samples are dynamically screened using the screening layer. By analyzing the consistency between the predicted results and the actual labels, samples that do not meet the labels are dynamically frozen, and samples whose predicted results are consistent with the labels are retained, entering the final training stage;

[0066] S3, Final training and model optimization: Use the selected high-quality training and test samples to continue training the model. The training samples are used to train the model fitting parameters, and the test samples are used to adjust the hyperparameters of the network model. During the entire training process, the minimum loss model parameters on the test set are continuously saved until the end of the training. The final minimum loss model is the optimal network model.

[0067] S4. Result visualization: Visualize the model’s classification results, feature extraction, weight distribution, and the correct and incorrect samples that have been screened, so as to deeply analyze the model’s performance and improvement effect in processing label confusion samples;

[0068] S5. Model evaluation: After training is completed, save the trained model and use a new labeled correctly dataset for testing to evaluate the model's diagnostic effect.

[0069] The overall structure of the DSS-CNN model of the present invention is as follows Figure 2The model parameters are shown in Table 1:

[0070] Table 1DSS-CNN model parameter settings

[0071]

[0072] The data is processed and analyzed using the CNN model. The typical CNN structure consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Figure 3 shown.

[0073] The function of the input layer is to receive signals from the neural network.

[0074] The function of the convolution layer is to perform convolution operation on the input signal to extract important features. Its expression is as follows:

[0075]

[0076] Where x i is the current input feature matrix, x i+1 is the calculated feature matrix, ω i is the convolution kernel weight parameter, b i is the bias parameter, is the convolution operation.

[0077] The pooling layer is generally used to reduce the dimension of the features extracted by the convolution layer, increase the anti-distortion ability of the model, and effectively reduce the feature space and network parameters after convolution. There are two common pooling operations in convolutional neural networks, namely maximum pooling and average pooling. The present invention adopts the maximum pooling layer to perform the maximum pooling operation on the feature information, reducing the data dimension while reducing the parameters; the maximum statistic is selected to obtain the local features of the area.

[0078] The fully connected layer is responsible for converting the two-dimensional feature matrix output after the previous series of processing into a one-dimensional vector. The output layer is located after the fully connected layer and classifies and outputs the features obtained from the fully connected layer.

[0079] When a bearing 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. Traditional CNN can only extract single-scale features, which affects the model effect. Multi-scale convolution uses convolution kernels of different sizes to operate in parallel, which can extract different scale features such as Figure 4 The present invention selects convolution kernels with sizes of 3×3, 6×6, and 9×9 to construct a multi-scale convolution layer.

[0080] 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. It can be seen that the dilated convolution can expand the receptive field of the output unit, thereby enhancing the network's ability to extract features of different scales. The present invention selects dilated convolutions with expansion rates of 1, 3, and 5 to construct the dilated convolution layer.

[0081] 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 the figure, this connection method forms an identity mapping, allowing 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:

[0082] H(x)=F(x)+x(2)

[0083] Among them, x is the input data, F(x) is the feature after processing by the main path, and H(x) is the final output of the residual block.

[0084] Batch Normalization (BN) is a method used for data normalization. The process of batch normalization for any output data in training is as follows:

[0085] Calculate the average of a batch of data:

[0086]

[0087] Calculate the variance of batch data:

[0088]

[0089] Normalized data:

[0090]

[0091] Scaling and offsetting:

[0092]

[0093] Among them, m represents the size of the batch, ε is a constant term to ensure numerical stability, γ and β are scale factors and offset factors that can be learned by the network, respectively. is 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.

[0094] Traditional convolutional neural networks usually use multiple fully connected layers to reduce feature dimensionality. However, since the fully connected layers use a fully connected approach, a large number of parameters are generated during the training process, resulting in high model computational complexity and prone to overfitting. 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. While reducing the dimensionality, it greatly reduces the parameters of the network model and integrates global feature information, which can effectively avoid overfitting.

[0095] PHM experimental data analysis:

[0096] The PHM dataset is collected from the subway train transmission system fault simulation test bench. Figure 7 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. The 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.

[0097] This experiment selects working conditions WC1, WC2 and WC3, and the driving gear support bearing data measured in the x direction beside the gearbox output shaft is used as the experimental data. The data under normal conditions, inner ring fault, outer ring fault, rolling element fault, and cage fault are selected to construct different proportions of label confusion data sets. 256,000 sampling points are taken for each type of data, and each type of data is divided into 250 groups, each group is used as 1 sample, and each sample has 1024 sampling points. Figure 8 The following is a time domain diagram of the constructed label obfuscated data. The label settings are shown in Table 3, and the detailed label setting process is as follows:

[0098] (1) When the label is 0, it is defined as normal data, but it contains 225 sets of normal data and 25 sets of outer ring fault data;

[0099] (2) When the label is 1, it is defined as inner race fault data, but it contains 200 sets of inner race fault data and 50 sets of rolling element fault data;

[0100] (3) When the label is 2, it is defined as outer race fault data, but it contains 175 sets of outer race data and 75 sets of cage fault data;

[0101] (4) When the label is 3, it is defined as rolling element fault data, but it contains 200 rolling element data and 50 sets of normal data;

[0102] (5) When the label is 4, it is defined as cage failure data, but it contains 175 sets of cage data and 75 sets of inner ring failure data.

[0103] Table 2 Data types and label numbers

[0104] Scale settings Data Types Tag Number 9:1 225 groups of normal samples + 25 groups of outer ring fault samples 0 8:2 200 sets of inner ring fault samples + 50 sets of rolling element fault samples 1 7:3 175 sets of outer ring failure samples + 75 sets of cage failure samples 2 8:2 200 sets of rolling element failure samples + 50 sets of normal samples 3 7:3 175 sets of cage failure samples + 75 sets of inner ring failure samples 4

[0105] The deep learning framework used in the experiment is Tensorflow, and the computer configuration is Core(TM) i5-13400F CPU processor and NVIDIA GeForce RTX4060 graphics card. It is divided into training group and test group in a ratio of 8:2, of which 200 groups are put into training and the remaining 50 groups are tested. BatchSize is set to 256, Epoch is set to 500, Adam is selected as the optimizer of the model, the learning rate is set to 0.001, and the cross entropy loss function is selected as the loss function of the model. The model training goal is to minimize the validation set loss function.

[0106] Accuracy and loss curve:

[0107] The training and test accuracy curves for the initial and final training phases are shown in Figure 2. Fig. 9 As shown, the loss curve is Fig.10 shown.

[0108] In the initial training phase, the model converged after 200 iterations. The training set accuracy curve converged quickly after 30 iterations, reaching 81.90%. The loss curve fully converged after 50 iterations, and the loss value stabilized at around 0.298. The test set accuracy reached convergence after 150 iterations, with an accuracy of 61.16%. Due to the influence of the wrong samples, the test set loss curve converged after 180 iterations, and the loss fluctuated around 1.68.

[0109] In the final training stage, the model training and test accuracy significantly improved at 201 iterations because the filter layer froze the wrong samples of the training set and test set after 200 iterations. The training set accuracy curve fully converged after 220 iterations and reached 100%. The loss curve also decreased rapidly after 220 iterations and approached 0. The test set accuracy fully converged after 320 iterations and reached 96.75%. The loss curve fully converged after 230 iterations, and the loss stabilized at around 0.075.

[0110] Sample screening visualization:

[0111] When the model converges in the initial training phase, the screening layer is activated, and the training samples and test samples of each type of label data are screened in turn. The distribution of correct and incorrect samples screened is visualized using t-SNE dimensionality reduction. Fig.11 Table 3 is a statistical analysis of the screening results of all samples of each type of label data. From the experimental results, we can see that:

[0112] (1) There are 205 correct samples and 45 wrong samples in the sample data with label 0. From the quantity, it can be seen that there are 20 more wrong samples than the actual number.

[0113] (2) For the sample data with label 1, there are 218 correct samples and 32 wrong samples. From the quantity, we can see that the wrong samples are 18 less than the actual number.

[0114] (3) For the sample data with label 2, there are 178 correct samples and 72 wrong samples. From the quantity, we can see that there are 3 fewer wrong samples than the actual number.

[0115] (4) For the sample data with label 3, there are 197 correct samples and 53 wrong samples. From the quantity, we can see that there are 3 more wrong samples than the actual number.

[0116] (5) For the sample data with label 4, there are 177 correct samples and 73 incorrect samples. From the quantity, it can be seen that the number of incorrect samples is 2 less than the actual number.

[0117] Table 3 Number of correct and incorrect samples screened

[0118]

[0119] The above is only a quantitative analysis. It is difficult to see whether the screening layer is screening the correct and incorrect samples correctly, and whether there is any screening confusion. Therefore, all the correct and incorrect samples screened by each type of label data are plotted in the time domain as follows: Fig.12As shown in the time domain diagram, it can be seen that the correct samples of the training group and the test group of each type of label data have extremely high similarity; the wrong samples of the training group and the test group have extremely high similarity, and there are only a few screening confusions, which proves that the screening results are relatively ideal.

[0120] Visualization of feature extraction during training:

[0121] Select the first sample of the five types of data in the test set respectively, convert each sample into a two-dimensional array of 1024 data points and a data structure of (32,32,1) in the 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 in the figure below: Fig.13 shown.

[0122] Fig.14 This is a visualized line chart of the feature extraction process of all convolutional layers and pooling layers of the 5 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 same type of feature points gradually decrease, and different types of feature points gradually disperse. Finally, a small number of feature points with relatively high discrimination are extracted to basically separate the 5 types of samples. Due to the large number of samples, only the feature extraction of a single sample of the 5 data types in the training process is visualized, so the feature points finally obtained have some overlap, which also reflects that the deep learning model needs a large number of samples to learn more features during the training process.

[0123] Visualization of weight distribution during training:

[0124] 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:

[0125] (1) The overall distribution range of the convolution kernel weight value of the ordinary convolution layer is (-0.10, 0.11), and the main distribution range is (-0.05, 0.05); the overall density distribution range of the weight value is (0, 15), and the main density distribution range is (10, 15);

[0126] (2) The overall distribution range of the convolution kernel weight values ​​of the multi-scale convolution layer is (-0.13, 0.13), and the main distribution range is (-0.07, 0.07); the overall density distribution range of the weight value is (0, 10.3), and the main density distribution range is (8, 10.3);

[0127] (3) The overall distribution range of the convolution kernel weight values ​​of the atrous convolution layer is (-0.17, 0.17), and the main distribution range is (-0.10, 0.10); the overall density distribution range of the weight values ​​is (0, 6.4), and the main density distribution range is (4, 6.4).

[0128] The distribution range of weight values ​​of each layer gradually increases with the number of iterations and remains basically unchanged around 400 iterations, which once again verifies that it is reasonable to set the total number of iterations to 500. The maximum density distribution of weight values ​​of ordinary convolutional layers, multi-scale convolutional layers, and dilated convolutional layers decreases in turn, proving that the model learns more subtle features.

[0129] The overall classification process is reduced to visualization:

[0130] Fig.17 , Fig.18 The KPCA dimensionality reduction visualizations of the original data distribution and after training classification are shown in Figure 1. As can be seen from the comparison, the distribution of the original data in the figure is relatively scattered, and various data are mixed together. As the training process proceeds, the screening layer freezes the erroneous samples, and data points of different types gradually disperse, while data points of the same type gradually gather and finally separate completely, achieving excellent classification results. This proves that the method of the present invention has excellent classification performance for bearing label confusion data.

[0131] Comparison of test accuracy and loss of different fault diagnosis methods:

[0132] To verify the superiority of the proposed method, it is compared with three typical diagnostic methods: AlexNet, GhostNet, and MobileNetV2. During the training process, the batch size is uniformly set to 256 and the number of iterations is 500. The final test accuracy and loss curve are visualized, as shown in Figure 2. Fig.19 and Fig. 20 As shown in the results, the method proposed in the present invention has the best effect. Even in the initial training stage, the accuracy curve and the loss curve converge the fastest among various methods. The accuracy reaches 61.16%, the highest among various methods, and the loss is 1.68, the lowest among various methods. In the final training stage, the accuracy curve of the test set fully converges after 320 iterations, reaching 96.75%. The loss curve decreases rapidly with the iteration and fully converges after 230 iterations, and the loss fluctuates around 0.075.

[0133] In contrast, the diagnostic effect of AlexNet, GhostNet and MobileNetV2 is poor. The test accuracy of the three models is below 60%, and the test loss is much higher than that of the DSS-CNN model. In addition, when the batch size and 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.

[0134] Comparison of new data test accuracy of different fault diagnosis methods:

[0135] In order to verify the test accuracy of the model at different training stages, new data under normal state, inner ring fault, outer ring fault, rolling element fault, and cage fault conditions are selected to construct a new test data set. 102,400 sampling points are taken for each type of data, and each type of data is divided into 100 samples, each sample has 1024 sampling points, and its label settings are shown in Table 4.

[0136] Table 4 Test data types and label numbers

[0137] Data Types Total number of samples Tag Number Normal data 100 0 Inner race fault 100 1 Outer ring failure 100 2 Rolling element failure 100 3 Cage failure 100 4

[0138] The model that converged in the initial training phase is saved and tested using the new test data set taken above. The test confusion matrix is: Fig. 20 As shown in the figure, the test accuracy of the five types of data is about 85%, among which the accuracy of label 2 is the highest, 97%, and the overall average accuracy is 87.6%. Save the model that converges in the final training stage, and use the new test data set taken above for testing. The test confusion matrix is Fig.21 As shown in the figure, the test accuracy of the five types of data basically reached about 90%, among which the highest accuracy of label 2 was 100%, and the overall average accuracy was 92%. In the end, the accuracy was improved by 4.4% compared with the average accuracy of the initial training model test, proving that the screening layer played a key role. Freezing the wrong samples of each type of data from participating in the subsequent model training can reduce the interference of wrong samples and learn rich correct features.

[0139] In order to verify the effectiveness and robustness of the DSS-CNN model, it was trained with the AlexNet, GhostNet and MobileNetV2 models and the final converged model was saved. Five tests were performed using the new test data taken above, and the overall average accuracy of each test result is shown in Table 5. As can be seen from the table, the test accuracy of the DSS-CNN model is stable at around 92%. The test accuracy of the other three models is unstable and the diagnostic effect is poor, and the test accuracy is much lower than that of the DSS-CNN model.

[0140] Table 5 Test accuracy of different models

[0141]

[0142] Experiment on noise immunity of different fault diagnosis methods:

[0143] In order to verify the diagnostic effect of the DSS-CNN model under noise conditions, Gaussian white noise with signal-to-noise ratios (SNR) of -6dB, -4dB, -2dB, 0dB, 2dB, 4dB, and 6dB was added to the new test data set. Fig.23 As shown in Table 7. 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 converged in the training phase of DSS-CNN, AlexNet, GhostNet and MobileNetV2 were saved respectively, and the new test data with noise were tested 5 times. The average prediction accuracy of each noise is the average of the accuracy of the 5 experiments, which is shown in Table 7. As can be seen from the table, the average accuracy of the DSS-CNN 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.

[0144] Table 7 Test accuracy of different noise data

[0145]

[0146] DSS-CNN model ablation experiment:

[0147] In order to verify the rationality of the components of the model of the present invention and explore their role in the entire model, the following 9 models are set up for comparative experiments. The specific settings are as follows:

[0148] Model 1: DSS-CNN model of the present invention;

[0149] Model 2: Average pooling layer replaces the maximum pooling layer;

[0150] Model 3: remove the normal convolutional layer of the model;

[0151] Model 4: ordinary convolutional layers instead of multi-scale convolutional layers;

[0152] Model 5: Ordinary convolutional layer instead of dilated convolutional layer;

[0153] Model 6: The ordinary concatenation feature fusion layer replaces the residual connection feature fusion layer;

[0154] Model 7: The convolution kernel sizes of the multi-scale convolutional layers are set to (3×3), (7×7), and (12×12) respectively;

[0155] Model 8: The dilation rates of the atrous convolutional layers are set to DR=1, DR=2, and DR=3 respectively;

[0156] Model 9: Fully connected layers replace global average pooling layers.

[0157] The accuracy results of each model in 5 tests on the new test data set are as follows Fig.24 As shown in the figure, the accuracy of DSS-CNN is higher than that of the other eight models, which fully verifies the rationality and superiority of the model components of the present invention.

[0158] Comparison of the accuracy of the equal-proportion label confusion experiment of different fault diagnosis methods:

[0159] In order to test the generalization of DSS-CNN, the number of samples with incorrect labels was adjusted, and the ratio of samples with correct labels to samples with incorrect labels was set to 8:2 and 7:3 for analysis (see Table 8).

[0160]

[0161] The DSS-CNN, AlexNet, GhostNet and MobileNetV2 models are trained using the two types of label confusion data. The data set division ratio, batch size, and number of iterations are the same as above. The final converged model is saved, and 5 test experiments are performed using a new test data set with correct labels. The test accuracy is the average of the 5 experimental accuracies. Fig.24 As shown in the figure, when the ratio is set from 8:2 to 7:3, the test accuracy of the DSS-CNN model can still reach 89%, while the test accuracy of the AlexNet, GhostNet and MobileNetV2 models drops significantly. This once again verifies that the model of the present invention has an excellent effect in the classification of bearing faults under the condition of a small amount of label confusion data.

[0162] In summary, this method has achieved excellent classification and diagnosis results for subway bearing fault diagnosis with a small amount of sample data with label errors, and can obtain extremely high diagnostic accuracy in a relatively short time. The PHM experimental data was used to construct a dataset with label errors and input it into the DSS-CNN model for fault identification and diagnosis, which solved the problem of label confusion that may occur in the current signal acquisition process, reduced the impact of label confusion on the diagnosis model, and realized the fault diagnosis of subway bearing sample data with a small amount of label errors, making this type of data available and saving resources and costs.

[0163] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A subway bearing label confusion data fault diagnosis method based on DSS-CNN model, characterized in that: The following steps are involved: S1. Perform data preprocessing and data set division. Divide the data set into training set and test set according to a certain ratio and perform normalization processing. S2, initial training and dynamic sample selection, the entire dataset containing a small number of mislabeled samples constructed by PHM experimental data is input into DSS-CNN for initial training until the model converges; S3, final training and model optimization, use the screened high-quality training and test samples to continue training the model, and keep saving the minimum loss model parameters on the test set throughout the training process until the end of training. The final minimum loss model is the optimal network model; S4. Visualize the model’s classification results, feature extraction, weight distribution, and the correct and incorrect samples that have been screened; S5. After training is completed, save the trained model and use a new labeled correct data set for testing to evaluate the model's diagnostic effect.

2. The method according to claim 1, characterized in that In the step S1, the constructed data with a small amount of label confusion is read, the data is divided into 1024 points of a single sample data, labels are created for the data set and the labels are converted into one-hot encoding.

3. The method according to claim 1, characterized in that: The DSS-CNN model in step S2 consists of five parts: a screening layer, a traditional convolutional layer, a multi-scale convolutional layer, a hole convolutional layer and a classification layer.

4. The method according to claim 1, characterized in that: The classification layer of the DSS-CNN model in step S2 is divided into a global average pooling layer and a Softmax.

5. The method according to claim 1, characterized in that In step S2, the data set is input into DSS-CNN for initial training to achieve model convergence, including the following steps: After the model converges in the initial training phase S21, the screening layer is activated, and the model trained in the initial phase is called to predict each type of data and dynamically freeze samples with inconsistent labels, retaining samples with consistent prediction results and labels; S22 uses convolution kernels to extract features in the traditional convolution layer, and combines it with BN processing, adding a maximum pooling layer with KernelSize of (2×2) and Stride of 1 to retain key feature information; S23 performs three parallel convolution operations in the multi-scale convolution layer to capture the features of two-dimensional data at different scales. The output is fused through residual connection features and combined with BN processing. Finally, it passes through the maximum pooling layer with KernelSize of (3×3) and Stride of 2. S24 performs parallel convolution operations with dilation rates of DR=1, DR=3, and DR=5 in the dilated convolution layer to capture the features of different receptive fields of two-dimensional data. The output is fused through residual connection features, followed by batch BN processing, and finally passes through the maximum pooling layer with KernelSize of (3×3) and Stride of 2. S25 integrates global feature information at the classification layer and outputs fault diagnosis results.

6. The method according to claim 5, characterized in that The traditional convolution layer in step S22 uses a convolution kernel with a size of 8×8.

7. The method according to claim 5, characterized in that The three convolution sizes of the multi-scale convolution layer in step S23 are 3×3, 6×6, and 9×9.

8. The method according to claim 5, characterized in that The three convolution sizes of the hole convolution layer in step S24 are 3×3.

9. The method according to claim 5, characterized in that In step S25, the global average pooling layer replaces the fully connected layer, and Softmax outputs the fault diagnosis result.

10. The method according to claim 5, characterized in that In step S21, the screening layer is located after the input layer, and the model dynamically screens the samples, retaining only the samples whose prediction results are consistent with the labels for final training, and no sample screening is performed in the final training stage.

Images