Railway train bogie rolling bearing fault diagnosis method

Abstract

The application provides a rail train bogie rolling bearing fault diagnosis method, an interactive generation learning strategy is used to train a self-encoder, so that the model fully learns the common features in the bogie bearing fault samples; meanwhile, in the self-encoder generation stage, the synthetic minority over-sampling technique is introduced into the hidden layer space with a larger Euclidean distance, the diversity of the self-encoder generated samples is improved, rich incremental information is brought to the fault diagnosis model, and the diagnosis accuracy of the deep learning fault diagnosis model in the sample imbalance scene is improved. The application provides a new idea for the rail train bogie rolling bearing fault diagnosis based on deep learning under the sample imbalance, and has important significance for the application of the deep learning model in the rolling bearing fault diagnosis field.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rolling bearing fault diagnosis, and in particular to a rolling bearing fault diagnosis method for a bogie of a rail train.

BACKGROUND

[0002] With the continuous increase of train operation speed, the continuous expansion of allocation scale and vehicle types, and the influence of complex service environments such as large load of rail trains and multi-physical field coupling, the requirements for train safety guarantee and economic operation and maintenance continue to increase. As an important component connecting the wheels and the car body, the bogie usually plays a role in bearing load and guiding the train. Under long-time operation, it often causes wear and tear of the transmission components (such as rolling bearings) and even causes failure. Once the rolling bearing of the bogie of the train fails, the safe operation of the train cannot be effectively guaranteed, and in severe cases, it may cause economic losses and even casualties. Therefore, reliable diagnosis of the rolling bearing of the bogie of the train at an early stage of failure and timely repair or replacement can effectively reduce the occurrence of serious accidents.

[0003] Although the deep learning-based fault diagnosis method has achieved satisfactory results in the field of rolling bearing fault diagnosis, it is mostly based on the assumption that each type of data is balanced. The actual operation of the train is different from the laboratory data collection process. The train cannot run in a fault state for a long time. The data collected in actual application usually presents a long-tail distribution, that is, the normal state data is much more than the fault type data. This data imbalance problem often makes the diagnosis model pay more attention to the majority class samples, forming a skewed decision boundary, which affects the final diagnosis accuracy. Therefore, bearing fault diagnosis under unbalanced data is a problem that needs to be solved in practical application.

[0004] From the perspective of data enhancement, some oversampling techniques, such as SMOTE and K-mean-SMOTE algorithms, have a certain promoting effect on bearing fault diagnosis under unbalanced data, but such methods usually oversample in the original space and are blind, providing limited incremental information for the diagnosis model. The current sample generation method based on generative adversarial network (GAN) provides a solution for bearing fault diagnosis under unbalanced data, such as the train bearing fault diagnosis method based on improved generative adversarial network disclosed in Chinese patent CN202210308771.1, which improves the sample generation quality by introducing the Pearson correlation coefficient to train the generative adversarial network. However, this method still faces the problems of mode collapse, difficulty in training, and limited application in the field of fault diagnosis.

[0005] In summary, existing bearing fault diagnosis methods for imbalanced samples can alleviate the impact of imbalanced samples on deep learning models to some extent. However, since trains operate under normal conditions for a long time and the probability of failure is low, existing generation methods still suffer from insufficient versatility and low diversity of generated samples when data imbalance occurs. [Summary of the Invention]

[0006] To address the aforementioned shortcomings, this disclosure proposes a fault diagnosis method for rolling bearings of railway bogies. This method extracts the amplitude-frequency characteristics of vibration signals during bogie operation, trains an autoencoder using an interactive generative learning strategy, and finally augments the imbalanced data to mitigate the impact of sample imbalance on the fault diagnosis model.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for diagnosing rolling bearing faults in a railway bogie, characterized by comprising:

[0009] Step 1: Collect vibration acceleration signals of rolling bearings of train bogies under different health conditions, and extract samples x of size [1×D] through a sliding window w. Label the samples according to the corresponding health conditions of the rolling bearings to construct an imbalanced dataset D between health condition categories. I ;

[0010] Step 2: Use Fast Fourier Transform to convert the time-domain signal into an amplitude-frequency signal, and then normalize the amplitude-frequency signal;

[0011] Step 3: Construct the autoencoder network model, including setting the hyperparameters of each network layer in encoder E and decoder D, and defining the loss function L. total ;

[0012] Step 4: Randomly initialize network parameters. Input the sample x from Step 1 into the encoder E according to the batch size. Output the hidden layer feature h of size [1×d]. Input h into the decoder D. Output the reconstructed sample with the same size as x. x and All dimensions are [1×D]. Calculate x and... Mean square error reconstruction loss L MSE ;

[0013] Step 5: From D I A sample set x is constructed by randomly selecting a sample of a certain health status. c And train using an interactive generative learning strategy; the interactive generative learning strategy includes using x c The feature set h is obtained by feature extraction from the encoder E. c , for hc Perform a reordering operation to obtain the feature set. Input the decoder D to reconstruct the sample, and obtain the result with x. c Reconstructed sample sets with different orders Calculate x c and Reordering penalty and reconstruction loss L penalty The reordering operation includes moving the first feature sample in the feature space to the end and shifting the index of the remaining feature samples one position to the front.

[0014] Step 6: Place L MSE With L penalty L is obtained by weighted summation. total Backpropagation is used to calculate the gradients of each network layer, and the E and D network parameters are updated using the optimizer.

[0015] Step 7: Repeat steps 4 to 6 until the preset number of rounds M is reached, then save the trained encoder. With decoder Network parameters;

[0016] Step 8: Transfer the imbalanced dataset D I Input the completed training Feature extraction is performed to obtain the feature set h of each health status sample. i Then, the synthetic minority class oversampling technique is used to analyze h. i The minority class feature set in the model is expanded to make the number of features in each health state feature set consistent, resulting in a balanced feature set h. B Then h B Input to Decoding is performed to obtain reconstructed samples for expanding D. I This forms a new balanced training set D. B ;

[0017] Step 9: Construct fault diagnosis classification model C, and combine it with the model from Step 8. B The model is trained using the input, and the classification model C is saved when the training reaches a preset number of rounds N.

[0018] Step 10: Load the classification model C, input the bogie operation data, and output the train bogie rolling bearing fault diagnosis results.

[0019] Furthermore, the hyperparameters of the network layer in step 3 include the kernel size, kernel stride, and number of neurons in the fully connected layer.

[0020] Furthermore, the weighted summation calculation formula in step 6 is as follows:

[0021] L total=(1-λ)L MSE +λL penalty

[0022] In the formula, L MSE It is the mean square error reconstruction loss, L penalty The reordering penalty reconstruction loss is λ, where λ is the weight of the penalty term.

[0023] The formula for the mean square error loss is as follows:

[0024]

[0025] In the formula, m is the number of samples in each batch. It is a reconstructed sample;

[0026] The reordering penalty reconstruction loss L penalty The formula is shown below:

[0027]

[0028] In the formula, n is the number of samples drawn from a certain type of sample set. It is the reconstructed sample after the reordering operation.

[0029] Furthermore, in step 8, the synthetic minority class oversampling technique is used to calculate the Euclidean distance between a certain feature sample and its K nearest neighbor feature samples, and then randomly interpolates between the two feature samples to obtain a synthetic feature sample; the random interpolation formula is as follows:

[0030] h new =h i +rand(0,1)×(h j -h i )

[0031] In the formula, h new It is a synthetic feature sample, h i It is a randomly selected feature sample, h j It is the j-th nearest neighbor feature sample (j = 1, 2, 3, ... K), and rand(·) is the random number generator.

[0032] Furthermore, in step 1, the sliding window w includes two parameters: window length and sliding step size. The length is greater than the number of sampling points corresponding to one complete cycle of the rolling bearing rotating with the shaft. The stride is the number of sampling points that the window slides after each segmentation of the signal.

[0033] Furthermore, the normalization formula in step 2 is as follows:

[0034]

[0035] In the formula, y max It is the expected maximum value of the normalized value, y min It is the expected normalized minimum value, where x is the frequency amplitude obtained after the time-domain signal undergoes a Fast Fourier Transform. max It is the maximum value of the frequency amplitude, x min It is the minimum value of the frequency amplitude.

[0036] Compared with existing technologies, the advantages of this invention are as follows: This invention proposes a method for diagnosing rolling bearing faults in railway bogies. This method utilizes an interactive generative learning strategy to train an autoencoder, enabling the generative model to learn the common intra-class features of samples. Furthermore, it expands the number of minority class features in the hidden layer during the generation stage, thereby improving the diversity of generated samples and providing rich incremental information for the fault diagnosis model. This invention effectively improves the quality of samples generated by the generator in multiple imbalanced sample scenarios, enhancing the accuracy of deep learning models in identifying minority class faults. Therefore, this invention has certain application prospects and engineering value. [Attached Image Description]

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a comparison between real samples and generated samples of bearings in various health states in the embodiments of the present invention;

[0039] Figure 3 This is a comparison of the similarity between samples generated by each method in the embodiments of the present invention and real samples;

[0040] Figure 4 This is a bar chart showing the diagnostic accuracy of each method under different imbalance ratios in the embodiments of the present invention.

Detailed Implementation Methods

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. It should be understood that the illustrative embodiments and descriptions of this invention are for explaining the invention only and are not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined in any way as long as they do not conflict with each other.

[0042] See Figure 1 This invention provides a method for diagnosing faults in rolling bearings of railway bogies, characterized by comprising:

[0043] Step 1: Collect vibration acceleration signals of rolling bearings of train bogies under different health conditions, and extract samples x of size [1×D] through a sliding window w. Label the samples according to the corresponding health conditions of the rolling bearings (normal condition, rolling element failure, outer ring failure, inner ring failure, and cage failure) to construct an imbalanced dataset D between health condition categories. I ;

[0044] Step 2: Use Fast Fourier Transform to convert the time-domain signal into an amplitude-frequency signal (amplitude-frequency signals are more stable than time-domain vibration signals, which can reduce the difficulty of generation), and normalize the amplitude-frequency signal.

[0045] Step 3: Construct the autoencoder network model, including setting the hyperparameters of each network layer in encoder E and decoder D, and defining the loss function L. total ;

[0046] Step 4: Randomly initialize network parameters. Input the sample x from Step 1 into the encoder E according to the batch size. Output the hidden layer feature h of size [1×d]. Input h into the decoder D. Output the reconstructed sample with the same size as x. x and All dimensions are [1×D]. Calculate x and... Mean square error reconstruction loss L MSE ;

[0047] Step 5: From D I A sample set x is constructed by randomly selecting a sample of a certain health status. c And train using an interactive generative learning strategy; the interactive generative learning strategy includes using x c The feature set h is obtained by feature extraction from the encoder E. c , for h c Perform a reordering operation to obtain the feature set. Input the decoder D to reconstruct the sample, and obtain the result with x. c Reconstructed sample sets with different orders Calculate x c and Reordering penalty and reconstruction loss L penalty The reordering operation includes moving the first feature sample in the feature space to the end and shifting the index of the remaining feature samples forward by one position (e.g., if the index is 1, 2, 3, 4, the reordered index is 2, 3, 4, 1).

[0048] Step 6: Place L MSE With L penalty L is obtained by weighted summation. total Backpropagation is used to calculate the gradients of each network layer, and the E and D network parameters are updated using the optimizer.

[0049] Step 7: Repeat steps 4 to 6 until the preset number of epochs M is reached (ensuring model convergence), then save the trained encoder. With decoder Network parameters;

[0050] Step 8: Transfer the imbalanced dataset D I Input the completed training Feature extraction is performed to obtain the feature set h of each health status sample. i Then, the synthetic minority class oversampling technique is used to analyze h. i The minority class feature set in the model is expanded to make the number of features in each health state feature set consistent, resulting in a balanced feature set h. B Then h B Input to Decoding is performed to obtain reconstructed samples for expanding D. I This forms a new balanced training set D. B ;

[0051] Step 9: Construct fault diagnosis classification model C, and combine it with the model from Step 8. B The model is trained using the input model C. When the training reaches a preset number of rounds N (ensuring model convergence), the classification model C is saved (for subsequent loading).

[0052] Step 10: Load the classification model C, input the bogie operation data, and output the train bogie rolling bearing fault diagnosis results.

[0053] Specifically, the hyperparameters of the network layer in step 3 include the kernel size, kernel stride, and number of neurons in the fully connected layer.

[0054] Specifically, the weighted summation calculation formula in step 6 is as follows:

[0055] L total =(1-λ)L MSE +λL penalty

[0056] In the formula, L MSE It is the mean square error reconstruction loss, L penalty It is the reordering penalty reconstruction loss, where λ is the weight of the penalty term (λ is preferably 0.2);

[0057] The formula for the mean square error loss is as follows:

[0058]

[0059] In the formula, m is the number of samples in each batch. It is a reconstructed sample;

[0060] The reordering penalty reconstruction loss L penalty The formula is shown below:

[0061]

[0062] In the formula, n is the number of samples drawn from a certain type of sample set. It is the reconstructed sample after the reordering operation.

[0063] Specifically, in step 8, the synthetic minority class oversampling technique calculates the Euclidean distance between a feature sample and its K nearest neighbor feature samples, and then randomly interpolates between the two feature samples to obtain a synthetic feature sample; the random interpolation formula is as follows:

[0064] h new =h i +rand(0,1)×(h j -h i )

[0065] In the formula, h new It is a synthetic feature sample, h i It is a randomly selected feature sample, h j It is the j-th nearest neighbor feature sample (j = 1, 2, 3…K) (K is 5), and rand(·) is the random number generator.

[0066] Specifically, in step 1, the sliding window w includes two parameters: window length and sliding step size. The length is greater than the number of sampling points corresponding to one complete cycle of the rolling bearing rotating with the shaft (ensuring that each signal sample contains sufficient health status information). The stride is the number of sampling points that the window slides after each segmentation of the signal.

[0067] Specifically, the normalization formula in step 2 is as follows:

[0068]

[0069] In the formula, y max It is the expected maximum value of the normalized value, y min It is the expected normalized minimum value, where x is the frequency amplitude obtained after the time-domain signal undergoes a Fast Fourier Transform. max It is the maximum value of the frequency amplitude, x min It is the minimum value of the frequency amplitude.

[0070] Working process of this invention:

[0071] The present invention provides a method for diagnosing rolling bearing faults in railway bogies, which is illustrated in conjunction with specific embodiments.

[0072] I. Data Acquisition:

[0073] Vibration signals during train operation were collected by acceleration sensors mounted on the axle boxes of the train bogies. Faulty bearings were located in both the axle boxes and gearboxes, and each contained four types of faults: rolling element fault, inner ring fault, outer ring fault, and cage fault, totaling nine health states. Bogie operating conditions: motor speed 3000 r / h, vertical load 1200 N.

[0074] II. Data Preprocessing and Slicing:

[0075] Nine types of bearing vibration data under healthy conditions were sampled through a sliding window. The window length and sliding step size were set to 4096, and four types of sample imbalance ratios were set. Detailed settings are shown in Table 1.

[0076] Table 1 Detailed settings for imbalanced sample data

[0077]

[0078] For each sample, perform a Fast Fourier Transform to extract amplitude-frequency features, and normalize the frequency amplitude s. The normalization method is as follows:

[0079]

[0080] III. Model Construction:

[0081] The model is built based on the PyTorch deep learning framework. The experimental equipment configuration is as follows: Windows 10 64-bit operating system, Intel Xeon Silver 4210R CPU @ 2.4GHz, and NVIDIA GeForce RTX 3090 GPU. The specific parameters of the network structure of this embodiment are shown in Table 2.

[0082] Table 2. Specific parameter settings for the network structure of the method in this embodiment of the invention.

[0083]

[0084] The decoder's last deconvolutional layer is sequentially connected to the Sigmoid activation function (no parameters need to be set, limiting the generated amplitude to [0,1]); the encoder's convolutional layer is sequentially connected to the LeakyRelu activation function (the negative activation slope is set to 0.2).

[0085] IV. Model Training:

[0086] The network model uses the Adam optimizer for parameter optimization, with the encoder and decoder learning rate set to 0.0002, batch size set to 64, and trained for 1000 epochs. The final convolutional neural network classifier has a learning rate set to 0.005, batch size set to 128, and trained for 100 epochs. This embodiment of the invention selects the direct diagnostic method (WDCNN), oversampling methods (SMOTE, K-mean-SMOTE), and deep generative methods (WAE-MMD, CVAE-GAN) as comparative methods for comparative experiments. Except for WDCNN, the other four methods address the sample imbalance problem from the data augmentation level by mixing generated samples with original samples during training.

[0087] V. Model Testing:

[0088] Load the parameters of the convolutional neural network classifier model, input the test set from the dataset into the trained model, and output the fault identification results.

[0089] VI. Analysis of Beneficial Effects:

[0090] See Figure 2 The paper presents the amplitude-frequency signals of eight types of faulty bearings and compares them with real samples and generated samples. It shows that the method of this invention, through interactive generative learning strategy training, can effectively learn the distribution characteristics and intra-class common features of real samples, and generate samples with high similarity to the original samples. At the same time, it exhibits certain differences in frequency amplitude, thus enabling the generated samples to possess diversity. (See also...) Figure 3 It can be seen that the samples generated by the method of this invention outperform the comparative methods in both Pearson correlation coefficient and cross-correlation coefficient. For the accuracy of each method on datasets with different imbalance ratios, please refer to [link to relevant documentation]. Figure 4 It can be seen that the method of the present invention has the best augmentation effect on imbalanced data and the most significant improvement in diagnostic accuracy. In the S4 dataset, the method of the present invention improves the accuracy by 6.5%, 5.26%, 3.58%, and 4.06% respectively compared with the other four sample augmentation methods.

[0091] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for diagnosing faults in rolling bearings of railway bogies, characterized in that, include: Step 1: Collect vibration acceleration signals of rolling bearings of train bogies under different health conditions, and extract samples x of size [1×D] through a sliding window w. Label the samples according to the corresponding health conditions of the rolling bearings to construct an imbalanced dataset D between health condition categories. I ; Step 2: Use Fast Fourier Transform to convert the time-domain signal into an amplitude-frequency signal, and then normalize the amplitude-frequency signal; Step 3: Construct the autoencoder network model, including setting the hyperparameters of each network layer in encoder E and decoder D, and defining the loss function L. total ; Step 4: Randomly initialize network parameters. Input the sample x from Step 1 into the encoder E according to the batch size. Output the hidden layer feature h of size [1×d]. Input h into the decoder D. Output the reconstructed sample with the same size as x. x and All dimensions are [1×D]. Calculate x and... Mean square error reconstruction loss L MSE ; Step 5: From D I A sample set x is constructed by randomly selecting a sample of a certain health status. c And train using an interactive generative learning strategy; the interactive generative learning strategy includes using x c The feature set h is obtained by feature extraction from the encoder E. c , for h c Perform a reordering operation to obtain the feature set. Input the decoder D to reconstruct the sample, and obtain the result with x. c Reconstructed sample sets with different orders Calculate x c and Reordering penalty and reconstruction loss L penalty The reordering operation includes moving the first feature sample in the feature space to the end and shifting the index of the remaining feature samples one position to the front. Step 6: Place L MSE With L penalty L is obtained by weighted summation. total Backpropagation is used to calculate the gradients of each network layer, and the E and D network parameters are updated using the optimizer. Step 7: Repeat steps 4 to 6 until the preset number of rounds M is reached, then save the trained encoder. With decoder Network parameters; Step 8: Transfer the imbalanced dataset D I Input the completed training Feature extraction is performed to obtain the feature set h of each health status sample. i Then, the synthetic minority class oversampling technique is used to analyze h. i The minority class feature set in the model is expanded to make the number of features in each health state feature set consistent, resulting in a balanced feature set h. B Then h B Input to Decoding is performed to obtain reconstructed samples for expanding D. I This forms a new balanced training set D. B ; Step 9: Construct fault diagnosis classification model C, and combine it with the model from Step 8. B The model is trained using the input, and the classification model C is saved when the training reaches a preset number of rounds N. Step 10: Load the classification model C, input the bogie operation data, and output the train bogie rolling bearing fault diagnosis results.

2. The method for diagnosing rolling bearing faults in a railway bogie as described in claim 1, characterized in that, In step 3, the network layer hyperparameters include the kernel size, kernel stride, and the number of neurons in the fully connected layer.

3. The method for diagnosing rolling bearing faults in a railway bogie as described in claim 1, characterized in that, The weighted summation formula in step 6 is as follows: THE total =(1-λ)L MSE +λL penalty In the formula, L MSE It is the mean square error reconstruction loss, L penalty The reordering penalty reconstruction loss is λ, where λ is the weight of the penalty term. The formula for the mean square error loss is as follows: In the formula, m is the number of samples in each batch. It is a reconstructed sample; The reordering penalty reconstruction loss L penalty The formula is shown below: In the formula, n is the number of samples drawn from a certain type of sample set. It is the reconstructed sample after the reordering operation.

4. The method for diagnosing rolling bearing faults in a railway bogie as described in claim 1, characterized in that, In step 8, the synthetic minority class oversampling technique is used to calculate the Euclidean distance between a certain feature sample and its K nearest neighbor feature samples, and then randomly interpolate between the two feature samples to obtain a synthetic feature sample. The random interpolation formula is as follows: h new =h i +rand(0,1)×(h j -h i ) In the formula, h new It is a synthetic feature sample, h i It is a randomly selected feature sample, h j It is the j-th nearest neighbor feature sample (j = 1, 2, 3, ... K), and rand(·) is the random number generator.

5. The method for diagnosing rolling bearing faults in a railway bogie as described in claim 1, characterized in that, In step 1, the sliding window w includes two parameters: window length and sliding step size. The length is greater than the number of sampling points corresponding to one complete cycle of the rolling bearing rotating with the shaft. The stride is the number of sampling points that the window slides after each segmentation of the signal.

6. The method for diagnosing rolling bearing faults in a railway bogie as described in claim 1, characterized in that, The normalization formula in step 2 is as follows: In the formula, y max It is the expected maximum value of the normalized value, y min It is the expected normalized minimum value, where x is the frequency amplitude obtained after the time-domain signal undergoes a Fast Fourier Transform. max It is the maximum value of the frequency amplitude, x min It is the minimum value of the frequency amplitude.

Images