Rail train bogie gearbox bearing fault diagnosis method

Abstract

The invention discloses a high-speed train bogie gearbox bearing fault diagnosis method, which comprises the following steps of: acquiring vibration acceleration signals of a bearing in a train bogie gearbox in different health states, preprocessing the vibration acceleration signals into a training sample set and a test sample set, and constructing a fault diagnosis classification model of noise label robustness; and training the classifier model by adopting a self-paced learning strategy, and inputting the test sample set into the trained classification model to obtain a train bogie gearbox bearing fault diagnosis result. According to the method, through the provided adaptive course sample selection strategy, the threshold value of confidence sample selection is associated with the development of the network, and the introduction of a resistance loss function can effectively resist the over-fitting problem easily generated by the model on a noise label. According to the method, the diagnosis accuracy of the data set with the noise label can be improved, and the high-speed train bogie gearbox bearing fault diagnosis precision is improved under the condition of high noise label rate.

Description

Technical Field

[0001] The invention relates to the field of bearing fault diagnosis, and in particular to a method for diagnosing bearing faults of a gearbox of a high-speed rail train bogie. Background Art

[0002] The increase in the speed of high-speed trains and the rapid development of the scale of operation have put forward higher requirements for the safety and stability of train operation. As an important part of high-speed trains, the bogie system includes traction motors, gearboxes, gearboxes, etc. Its performance will directly affect the safety and comfort of train operation. The gearbox is an important part of the high-speed EMU drive system. Due to the long-term high-speed and heavy-load working environment, it is susceptible to damage and failure. Once it fails, it will have a significant impact on the safety and stability of train operation. Therefore, early diagnosis of gearbox bearing failure is of great significance to eliminate safety hazards of high-speed trains and reduce operational risks.

[0003] At present, supervised deep neural networks represented by convolutional neural networks have been widely used in the field of fault diagnosis research. Such supervised learning algorithms usually search for the best classifier model parameters in a specific task by minimizing the error between the predicted label and the correct label, thereby obtaining higher diagnostic accuracy. However, in actual engineering, the labels of fault data may contain noisy labels or errors. The fault data used to supervise the training of fault identification models needs to be collected for a long time, and on this basis, they are labeled according to fault analysis reports and expert experience knowledge. This process is affected by certain human subjectivity. At the same time, in actual engineering, some gearbox bearing faults are constantly evolving and may be accompanied by secondary faults and compound faults. The fuzzy symptom boundaries of its fault categories increase the difficulty of fault data set labeling. The presence of noise labels will directly lead to errors in the calculation of the loss function, seriously mislead the supervised training process, and reduce the diagnostic accuracy and generalization performance of the model. How to overcome the shortcomings in the background technology so that high-speed train gearbox bearings can achieve higher diagnostic accuracy in the noise label scenario is an urgent problem to be solved.

[0004] The above contents are only used to assist in understanding the technical solution of the present invention and do not constitute an admission that the above contents are prior art. Summary of the invention

[0005] The object of the present invention is to provide a high-speed train bogie gearbox bearing fault diagnosis method, the technical solution of the present invention is as follows:

[0006] S1: Collect the vibration acceleration signals of the bearings in the train bogie gearbox under different health conditions, and intercept the samples x through the sliding window sw, and label the samples according to the corresponding operating status of the bearings. Preprocess the training sample set D with uniform noise labels y added O_train , and the test sample set DO_test ;

[0007] Optionally, in step S1, the sliding window sw includes two parameters: a window length and a sliding step length; the window length is greater than the number of sampling points corresponding to a complete cycle of rotation of the bearing with the shaft.

[0008] S2: For D O_train , D O_test The time domain signal is converted into a time-frequency signal using discrete wavelet packets, and a training sample set D consisting of two-dimensional time-frequency images is obtained. train , test sample set D test ;

[0009] S3: Construct a fault diagnosis classification model C that is robust to noise labels, including setting the hyperparameters of each network layer and defining the loss function L total ;

[0010] Optionally, the network layer hyperparameters in step S3 include convolution kernel size, convolution kernel moving step size, and the number of neurons in the fully connected layer.

[0011] S4: Randomly initialize network parameters, input the samples x and their corresponding labels y in step S2 into the diagnostic model C according to the batch size, and output the diagnostic results obtained after the model softmax layer And adopt the self-paced learning strategy. Before the model training T1 round, the training set D train Participate in calculating sample labels y and model diagnostic result labels The cross entropy loss L between CE , back propagation calculates the gradient of each network layer, and the optimizer updates the network C parameters.

[0012] The cross entropy loss formula in self-paced learning is as follows:

[0013]

[0014] In the formula, λ j is the threshold for selecting the confidence sample of the jth category, V represents the course sample, and multiple confidence samples constitute the course sample, W represents the model parameter, B represents the number of samples in each batch during the model training process, and x i is the i-th sample, v i represents the weight of the i-th sample, K represents the number of categories in the classification task, p(x i ) j Represents sample x i The predicted probability value of the label belonging to the jth class; is the predicted label of model C for the i-th sample of the j-th category.

[0015] S5: During the model training process, calculate the probability value given by each category sample in the model softmax layer in each round of training, calculate the average probability value μ and standard deviation σ given by each category in the current stage model, and set the threshold to the average value μ plus k times the standard deviation σ. When the classification diagnosis probability value given by the current model of a sample is greater than the above threshold, it is defined as a confident sample and included in the course information.

[0016] S6: After the model is trained for T1 rounds, the loss cross entropy loss L of the course information composed of the confidence samples in the current model is calculated CE . Calculate the resistance loss L of the training sample in the current model Resist .

[0017] L Resist The calculation formula is as follows:

[0018]

[0019] In the formula, B represents the number of samples in each batch during model training, and p t (x i ) j Indicates that sample x in the tth round of training i The predicted probability value of the label belonging to the jth class, p t-1 (x i ) j Indicates that sample x in the t-1th round of training i The predicted probability value of the label belonging to the jth class, x i is the ith sample.

[0020] S7: L CE With L Resist The weighted summation is calculated to obtain L total , back propagation calculates the gradient of each network layer, and uses the optimizer to update the network C parameters;

[0021] L total The calculation formula is as follows:

[0022] L total =L CE +γL Resist

[0023] In the formula, γ is the weight parameter, L CE is the cross entropy loss, L Resist To resist loss

[0024] S8: Repeat steps S5 to S7, and when the training reaches a preset number of rounds N, save the classification model C;

[0025] S9: Load the classification model C and input the test sample set D test, and obtain the train bogie gearbox bearing fault diagnosis results.

[0026] The present invention has at least the following beneficial effects: the present invention proposes a method for diagnosing the fault of a gearbox bearing of a high-speed train bogie, which selects confidence samples as course information based on the adaptive round-by-round statistical update to provide meaningful supervision information for other training samples, uses the selected confidence samples and the proposed resistance loss function to update the model parameters, and optimizes the model in combination with the original loss function. Through adaptive course sample selection, the threshold of confidence sample selection is linked to the development of the network, and the course is dynamically adjusted to achieve a simple adaptive effect. At the same time, the introduction of the resistance loss function effectively resists the overfitting problem that the model is prone to produce under noise labels. The method of the present invention can improve the diagnostic accuracy of data sets with noise labels, solve the problem that the high-speed train bogie gearbox bearing fault diagnosis model based on deep learning under noise labels is easily damaged by the wrongly labeled samples, resulting in reduced diagnostic accuracy, and improve the high-speed train bogie gearbox bearing fault diagnosis accuracy under high noise label rates. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the main steps of the method of the present invention;

[0028] Figure 2 is a diagnostic schematic diagram of different coefficients k applied to Trainbogie and Paderborn data sets in an embodiment of the method of the present invention; DETAILED DESCRIPTION

[0029] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with embodiments and drawings. It should be understood that the exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in each embodiment of the present invention described below can be combined in any way as long as they do not conflict with each other.

[0030] See also Figure 1 The present invention provides a rolling bearing fault diagnosis method for a rail vehicle bogie gearbox, which is characterized by comprising:

[0031] Step S1: Collect the vibration acceleration signals of the rolling bearings in the train bogie gearbox under different health conditions, and intercept the samples x through the sliding window sw, and label the samples according to the corresponding operating state of the rolling bearings. Preprocess the initial training sample set D with uniform noise labels y added. O_train , and the initial test sample set D O_test .

[0032] Optionally, the interception size of the sliding window sw is set according to the data preprocessing requirements, the size of the sample x is [1×D], and the noise label y can be added in a certain proportion. The sliding window sw includes two parameters: the window length Length and the sliding step length Stride; the Length is greater than the number of sampling points corresponding to a complete cycle of the rolling bearing rotating with the shaft, so that more complete sampling data is obtained; the Stride is the number of sampling points that the window slides after each interception on the signal.

[0033] Step S2: D O_train , D O_test The sample set uses discrete wavelet packets to convert time domain signals into time-frequency signals, and uses wavelet packet transform to perform localized analysis on one-dimensional signals in the time domain and frequency domain, and obtains a training sample set D consisting of two-dimensional time-frequency images. train , test sample set D test ;

[0034] Step S3: Construct a fault diagnosis classification model C that is robust to noise labels, including setting the hyperparameters of each network layer and defining the loss function L total ;

[0035] Optionally, the network layer hyperparameters in step S3 include convolution kernel size, convolution kernel moving step size, and the number of neurons in the fully connected layer.

[0036] Step S4: Randomly initialize the network parameters and convert the training sample set D in step S2 into train Input the batch size into the diagnostic model C to obtain the predicted label of model C And adopt the self-paced learning strategy, when the model training round t<T1 or t≤T1, the training set D train Participate in calculating the noise label y and the predicted label The cross entropy loss L between CE , back propagation calculates the gradient of each network layer, and the optimizer updates the network C parameters.

[0037] In one embodiment, in the self-paced learning strategy in step S4, the model training round T1 can be set according to the training requirements. In each model iterative optimization process, the model parameter W is fixed, and the global weight parameter (i.e., course sample) V is updated according to the size of the sample loss value. The weight parameter V is fixed and the model parameter W is updated.

[0038] The cross entropy loss formula is as follows:

[0039]

[0040] In the formula, λ jis the threshold for selecting the confidence sample of the jth category, V represents the course sample, and multiple confidence samples constitute the course sample, W represents the model parameter, B represents the number of samples in each batch during the model training process, and x i is the i-th sample, v i represents the weight of the i-th sample, K represents the number of categories in the classification task, p(x i ) j Represents sample x i The predicted probability value of the label belonging to the jth class; is the predicted label of model C for the i-th sample of the j-th category.

[0041] is a one-hot encoding form. If the sample x i The true category of is equal to j, then Take it as 1, otherwise take it as 0.

[0042] Step S5: During the model training process, calculate the training sample set D in each round of training train The classification diagnosis probability value P given by the softmax layer of the model for samples of each category, as well as the average probability value μ and standard deviation σ given by each category sample in the current stage model, and the threshold λ are set. When the classification diagnosis probability value P given by the current model for samples of each category is greater than the above threshold λ, it is defined as a confident sample and included in the course information.

[0043] The threshold λ=μ+kσ, where k is the standard deviation coefficient, is adaptively selected according to the model accuracy requirements;

[0044] In one embodiment, as shown in the attached Figure 2 As shown in the figure, when k=1, the model achieves the best diagnostic performance at both 30% and 50% noise label rates.

[0045] The confidence sample weight screening formula is as follows:

[0046]

[0047] In the formula, λ j =μ j +k*σ j The threshold selected for the confidence sample of the jth category; maxf j (x i ; w) is the sample x i The maximum posterior probability of belonging to category j, W represents the model parameters, and f represents the classifier.

[0048] Step S6: When the model training round t>T1 or t≥T1, calculate the cross entropy loss L of the course sample composed of the confidence sample in the current model CE, and the resistance loss L of all samples in the current model Resist .

[0049] L Resist The calculation formula is as follows:

[0050]

[0051] In the formula, B represents the number of samples in each batch during model training, and p t (x i ) j Indicates that sample x in the tth round of training i The predicted probability value of the label belonging to the jth class, p t-1 (x i ) j Indicates that sample x in the t-1th round of training i The predicted probability value of the label belonging to the jth class, x i is the ith sample.

[0052] Step S7: L CE With L Resist The weighted summation is calculated to obtain L total , back propagation calculates the gradient of each network layer, and uses the optimizer to update the network C parameters;

[0053] L total The calculation formula is as follows:

[0054] L total =L CE +γL Resist

[0055] In the formula, γ is a weight parameter, which can be selected according to the model training requirements. CE is the cross entropy loss, L Resist To resist losses.

[0056] Step S8: Repeat steps S5 to S7, and when the training reaches a preset number of rounds N, save the classification model C;

[0057] Optionally, the preset number of training rounds N is selected based on the convergence of model C.

[0058] Step S9: Load the classification model C and input the test sample set D test , and obtain the rolling bearing fault diagnosis results of the train bogie gearbox.

[0059] The working process of an embodiment of the present invention is as follows:

[0060] 1. Data acquisition: The acceleration sensor arranged on the train bogie gearbox collects the vibration signal of the gearbox during operation. The faulty bearings are distributed in the gearbox, and there are nine operating states including rolling element fault, inner ring fault, outer ring fault and cage fault and their corresponding different damage degrees. Bogie working condition: motor speed 50Hz, lateral load 150KG, longitudinal load 1700KG.

[0061] 2. Data preprocessing: The bearing vibration data of nine operating conditions are sampled through a sliding window, the window length is set to 4096, the sliding step is 4096, the ratio of the number of samples in the training set to the test set is 2:1, specifically, the number of training sets for each type of data is 400, and the number of sample sets is 200. The bearing operating conditions are shown in Table 1.

[0062] Table 1 Gearbox bearing operating status

[0063]

[0064] Each sample is subjected to discrete wavelet packet transform, and the one-dimensional vibration data is converted into a 64×64 two-dimensional time-frequency feature map using the db3 wavelet basis and 6-level decomposition, and then input into the model for training. The wavelet transform formula is as follows:

[0065] C(n,m)=m -1 / 2 [x(mt),Ψ * (tn / m)]

[0066] Where: m is the scale factor, n is the time shift factor, m -1 / 2 Ensure that the energy at different m values ​​remains equal during wavelet function scaling.

[0067] 3. Model construction: The model is built based on the Pytorch deep learning framework. The deep learning model parameters generally include weights and biases in the model, which are adjusted during the training process to minimize or optimize the loss function. The model framework parameters of this embodiment are shown in Table 2.

[0068] Table 2 Specific parameter settings of the network structure of the method of the embodiment of the present invention

[0069]

[0070] In data D trainWhen inputting into the network, it is first input into the network layer in batches. The network layer is mainly composed of three major blocks. The first two modules are connected to a maximum pooling layer and a Dropout layer respectively after three convolutional layers for convolution operation, batch regularization, and activation function (LeakyRelu) activation operation to prevent the model from overfitting and enhance its generalization ability. Then the features are input into the third module for convolution, batch regularization, and activation function (LeakyRelu). Finally, the extracted features are input into the global average pooling layer for feature information fusion, and feature extraction is performed after passing through the fully connected layer. Softmax can be used in the last layer of the neural network to convert the original scores output by the neural network into probability distribution to obtain the final classification result.

[0071] 4. Model training: Training is performed on a 64-bit operating system based on Windows 10, the CPU is Intel Xeon Silver4210R@2.4GHz, and the GPU is NVIDIA GeForce RTX 3090. The hyperparameter settings of the training process are shown in Table 3.

[0072] Table 3 Hyperparameter settings

[0073]

[0074] Five methods, namely CCE, Coteaching, JoCor, MentorNet and SPRL, were used to compare with the present invention, and different label noise rates were compared under the high-speed train bogie gearbox bearing data set. The experiment was repeated ten times for each model, and the performance of each model was measured using the average diagnostic accuracy and standard deviation of the test sample set. In order to ensure that the experiment is as close to the actual working conditions as possible, this embodiment adds Gaussian noise with a signal-to-noise ratio of -6dB to the signal in each training set to verify the robustness of the model to signal noise.

[0075] 5. Model testing and analysis:

[0076] Table 4

[0077]

[0078] The test sample set is input into the trained model, and the average diagnostic accuracy and standard deviation of each diagnostic task after 10 repeated experiments are shown in Table 4. When the LNR increases from 0 to 0.5, the diagnostic accuracy of each comparison method decreases significantly. However, the diagnostic accuracy of the method proposed in the present invention only decreases by 6.81% under this data set. Under the LNR of 0.5, the average diagnostic accuracy of the model of the proposed method also reaches 91.51%, which has good label noise robustness. The method of the present invention selects confidence samples as course information based on the adaptive round statistical update to provide meaningful supervision information for other training samples, uses the selected confidence samples and the proposed resistance loss function to update the model parameters, and optimizes the model in combination with the original loss function. Through the adaptive course sample selection strategy, the threshold of confidence sample selection is linked to the update of the network. As the model matures, the course is dynamically adjusted to achieve a simple adaptive effect. At the same time, the introduction of the resistance loss function effectively resists the overfitting problem that the model is prone to produce under noise labels. The method of the present invention can better extract the fault diagnosis classification characteristics under noise labels and improve the fault diagnosis accuracy.

[0079] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A rail train bogie gearbox bearing fault diagnosis method, characterized in that: The following steps are involved: S1: Collect the vibration acceleration signals of the bearings in the train bogie gearbox under different health conditions, and intercept the samples x through the sliding window sw, and label the samples according to the corresponding operating status of the bearings. Preprocess the training sample set D with uniform noise labels y added O_train , and the test sample set D O_test ; S2: For D O_train , D O_test The time domain signal is converted into a time-frequency signal using discrete wavelet packets, and a training sample set D consisting of two-dimensional time-frequency images is obtained. train , test sample set D test ; S3: Construct a fault diagnosis classification model C that is robust to noise labels, including setting the hyperparameters of each network layer and defining the loss function L total ; S4: Randomly initialize the network parameters and convert the training sample set D in step S2 into train Input the batch size into the diagnostic model C to obtain the predicted label of model C And adopt the self-paced learning strategy, when the model training round t≤T1, the training set D train Participate in calculating the noise label y and the predicted label The cross entropy loss L between CE , back propagation calculates the gradient of each network layer, and the optimizer updates the network C parameters; S5: Calculate the training sample set D in each round of training train The classification diagnosis probability value P given by the softmax layer of the model for each category of samples, as well as the average probability value μ and standard deviation σ given by each category of samples in the current stage model, and the threshold λ is set. When the classification diagnosis probability value P given by the current model for each category of samples is greater than the above threshold λ, it is defined as a confidence sample and included in the course information; S6: When the model training round t>T1, calculate the cross entropy loss L of the course sample composed of the confidence sample in the current model CE , and the resistance loss function L of all samples in the current model Resist ; S7: L CE With L Resist The weighted summation is calculated to obtain L total , back propagation calculates the gradient of each network layer, and uses the optimizer to update the network C parameters; S8: Repeat steps S5 to S7, and when the training reaches a preset number of rounds N, save the classification model C; S9: Load the classification model C and input the test sample set D test , and obtain the train bogie gearbox bearing fault diagnosis results.

2. In the rail train bogie gearbox bearing fault diagnosis method as described in claim 1, the sliding window sw in step S1 includes two parameters: window length and sliding step length; the window length is greater than the number of sampling points corresponding to a complete cycle of rotation of the bearing with the shaft.

3. In the rail train bogie gearbox bearing fault diagnosis method as described in claim 1, the network layer hyperparameters in step S3 include the convolution kernel size, the convolution kernel moving step size and the number of neurons in the fully connected layer.

4. The rail vehicle bogie gearbox bearing fault diagnosis method according to claim 1, wherein the cross entropy loss formula is as follows: In the formula, λ j is the threshold for selecting the confidence sample of the jth category, V represents the course sample, and multiple confidence samples constitute the course sample, W represents the model parameter, B represents the number of samples in each batch during the model training process, and x i is the i-th sample, v i represents the weight of the i-th sample, K represents the number of categories in the classification task, p(x i ) j Represents sample x i The predicted probability value of the label belonging to the jth class; is the predicted label of model C for the i-th sample of the j-th category.

5. The rail vehicle bogie gearbox bearing fault diagnosis method according to claim 4, wherein the threshold λ=μ+kσ, k is the standard deviation coefficient; and the confidence sample weight screening formula is as follows: In the formula, λ j =μ j +k*σ j The threshold selected for the confidence sample of the jth category; maxf j (x i ; w) is the sample x i The maximum posterior probability of belonging to category j, W represents the model parameters, and f represents the classifier.

6. The rail vehicle bogie gearbox bearing fault diagnosis method according to claim 5, wherein: Resist The calculation formula is as follows: In the formula, B represents the number of samples in each batch during model training, and p t (x i ) j Indicates that sample x in the tth round of training i The predicted probability value of the label belonging to the jth class, p t-1 (x i ) j Indicates that sample x in the t-1th round of training i The predicted probability value of the label belonging to the jth class, x i is the ith sample.

7. The rail vehicle bogie gearbox bearing fault diagnosis method according to claim 6, wherein: total The calculation formula is as follows: THE total =L CE +γL Resist In the formula, γ is the weight parameter, L CE is the cross entropy loss, L Resist To resist losses.

Images