Railway train bogie gear box bearing fault diagnosis method

Abstract

The application discloses a high-speed train bogie gear box bearing fault diagnosis method, which comprises the following steps: collecting vibration acceleration signals of bearings in a train bogie gear box under different health states, preprocessing the vibration acceleration signals into a training sample set and a test sample set, constructing a noise label robustness fault diagnosis classification model, training a classifier model by using a self-paced learning strategy, inputting the test sample set into the trained classification model, and obtaining a train bogie gear box bearing fault diagnosis result. The adaptive course sample selection strategy can effectively resist the overfitting problem of the model caused by the noise label by linking the threshold value of the confidence sample selection with the development of the network and resisting the introduction of the loss function. The method can improve the diagnosis accuracy of the data set with noise labels, and improve the fault diagnosis precision of the high-speed train bogie gear box bearing under the condition of high noise label rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bearing fault diagnosis, and particularly relates to a bearing fault diagnosis method for a gear box of a bogie of a high-speed rail train. BACKGROUND

[0002] The increase of the speed of high-speed trains and the rapid development of the operation scale put forward higher requirements for the safety and stability of train operation. The bogie system, as an important part of the high-speed train, includes a traction motor, a gear box, a gear box, etc. The performance thereof will directly affect the safety and comfort of train operation. The gear box is an important part of the drive system of the high-speed train, and is prone to damage and failure due to long-term operation in a high-speed and heavy-load environment. Once the gear box fails, it will have a significant impact on the safety and stability of train operation. Therefore, early diagnosis of gear box bearing failure is of great significance to eliminate the safety hazards of high-speed trains and reduce the operation risk.

[0003] At present, supervised deep neural networks represented by convolutional neural networks are widely used in the field of fault diagnosis. 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, so as to obtain higher diagnostic accuracy. However, in actual engineering, the labels of fault data may have noise labels or errors. The fault data used for training the supervised fault recognition model need to be collected for a long time, and on this basis, the fault data are labeled according to the fault analysis report and expert experience knowledge. This process is affected by certain human subjectivity. Meanwhile, in actual engineering, some gear box bearing faults are in continuous development and change, which may be accompanied by secondary faults and compound faults. The fuzzy boundary of the signs of the fault categories aggravates the difficulty of labeling the fault data set. The existence of noise labels will directly lead to errors in loss function calculation, 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 art and make the gear box bearing of the high-speed train achieve high diagnostic accuracy in the noise label scene is a problem to be solved.

[0004] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The purpose of the present application is to provide a bearing fault diagnosis method for a gear box of a bogie of a high-speed train. The technical solutions of the present application are as follows:

[0006] S1: Collect the vibration acceleration signals of the bearings in the gear box of the bogie of the train in different health states, and intercept the samples x through a sliding window sw, and label the samples according to the corresponding running state of the bearings. The pre-processing is a training sample set D with uniform noise labels y O_train , and a test sample set DO_test ;

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

[0008] S2: obtaining a training sample set D O_train , D O_test by converting the time-domain signal into a time-frequency signal using a discrete wavelet packet, and obtaining a training sample set D train , a test sample set D test ;

[0009] S3: constructing a fault diagnosis classification model C with noise label robustness, including setting network layer hyperparameters and defining a loss function L total ;

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

[0011] S4: randomly initializing network parameters, inputting the samples x and their corresponding labels y in step S2 into the diagnosis model C in batches, and obtaining a diagnosis result after outputting the softmax layer of the model , and adopting a self-paced learning strategy. Before T1 rounds of model training, the training set D train participates in the calculation of the cross-entropy loss L CE between the sample label y and the model diagnosis result label , the gradients of each network layer are calculated by back propagation, and the network C parameters are updated by the optimizer.

[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 confident samples of the jth class, V represents a curriculum sample, a plurality of confident samples form a curriculum sample, W represents model parameters, B represents the number of samples in each batch during the model training process, x i is the ith sample, v i represents the weight of the ith sample, K represents the number of classes in the classification task, p(x i ) j represents the label prediction probability value of the sample x i belonging to the jth class; is the predicted label of the ith sample of the jth class by the model C.

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

[0016] S6: After T1 rounds of model training, the loss cross-entropy loss L CE of the course information composed of confident samples in the current model is calculated. Resist .

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

[0018]

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

[0020] S7: L CE and L Resist are weighted and summed to obtain L total , the gradients of each network layer are calculated by back propagation, and the network C parameters are updated by the optimizer;

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

[0022] L total = L CE + γL Resist

[0023] In the formula, γ is a weight parameter, L CE is the cross-entropy loss, and L Resist is the resistance loss

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

[0025] S9: Load the classification model C, input the test sample set D test, to obtain the train bogie gear box bearing fault diagnosis result.

[0026] The present application has at least the following beneficial effects: the present application provides a high-speed train bogie gear box bearing fault diagnosis method, based on the adaptive update of the round statistics to select the confidence sample as the course information to provide meaningful supervision information for other training samples, and the selected confidence sample and the proposed resistance loss function are used to update the model parameters, and the original loss function is combined to optimize the model. Through adaptive course sample selection, the threshold of confidence sample selection is related 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 of the model in the noise label. The method of the present application can improve the diagnostic accuracy of the data set with noise labels, solve the problem that the high-speed train bogie gear box bearing fault diagnosis model based on deep learning is easily damaged by error labeled samples, and reduce the diagnostic accuracy. In the case of high noise label rate, the diagnostic accuracy of the high-speed train bogie gear box bearing is improved. BRIEF DESCRIPTION OF DRAWINGS

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

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

[0029] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with embodiments and drawings. It should be understood that the illustrative embodiments of the present application and their descriptions are only used to explain the present application and are not as limitations on the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined in any way as long as they do not conflict with each other.

[0030] Referring to Figure 1 , the present application provides a rail train bogie gear box rolling bearing fault diagnosis method, characterized in that it comprises:

[0031] Step S1: collect vibration acceleration signals of rolling bearings in different health states of the train bogie gear box, and intercept samples x through a sliding window sw, and label the samples according to the corresponding running state of the rolling bearings. The pre-processing is to add an initial training sample set D O_train with uniform noise label y O_test .

[0032] Optionally, the size of the sliding window sw is set according to the data preprocessing requirements. The size of sample x is [1×D], and noise labels y can be added in a certain proportion. The sliding window sw 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, thus obtaining more complete sampling data. The step size is the number of sampling points that the window slides after each segmentation of the signal.

[0033] Step S2: For D O_train D O_test The sample set utilizes discrete wavelet packets to convert the time-domain signal into a time-frequency signal. Wavelet packet transform is then used to perform localized analysis of the one-dimensional signal in both the time and frequency domains, resulting in a training sample set D composed of two-dimensional time-frequency images. train Test sample set D test ;

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

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

[0036] Step S4: Randomly initialize network parameters, using the training sample set D from step S2. train Input the batch size into the diagnostic model C to obtain the predicted labels of model C. It employs a self-paced learning strategy, where the training set D is updated when the model training rounds t < T1 or t ≤ T1. train Involved in calculating noise label y and predicted label Cross-entropy loss L CE Backpropagation is used to calculate the gradients of each network layer, and the optimizer updates the network C parameters.

[0037] In one embodiment, the self-learning strategy in step S4 allows the model training round T1 to be set according to training requirements. During each model iteration optimization process, the model parameters W are fixed, and the global weight parameters (i.e., course samples) V are updated according to the magnitude of the sample loss value. The weight parameters V are fixed, and the model parameters W are updated.

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

[0039]

[0040] In the formula, λ jis the threshold value of the selection of the confident sample of the jth category, V represents the curriculum sample, the plurality of confident samples constitute the curriculum sample, W represents the model parameter, B represents the number of samples per batch in the model training process, x i is the ith sample, v i is the weight of the ith sample, K represents the number of categories in the classification task, p(x i ) j is the label prediction probability value of the sample x i belonging to the jth category. is the predicted label of the model C of the ith sample of the jth category.

[0041] is in one-hot encoding form, if the real category of the sample x i is equal to j, then is 1, otherwise, it is 0.

[0042] Step S5: In the process of model training, the classification diagnosis probability value P of the sample set D train of each category given by the softmax layer of the model, and the average probability value μ and the standard deviation σ of each category sample given by the current stage model are calculated, and a threshold value λ is set, when the classification diagnosis probability value P of each category sample given by the current model is greater than the threshold value λ, it is defined as a confident sample, and is included in the curriculum information.

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

[0044] In an embodiment, as shown in the accompanying Figure 2 , when k = 1, the model achieves optimal diagnostic performance under the noise label rate of 30% and 50%.

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

[0046]

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

[0048] Step S6: When the model training round t > T1 or t ≥ T1, the cross-entropy loss L CEAnd 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 In the t-th round of training, sample x i The predicted probability value of the label belonging to class j, p t-1 (x i ) j In the (t-1)th round of training, sample x i The predicted probability value of the label belonging to class j, x i Let i be the i-th sample.

[0052] Step S7: Place L CE With L Resist L is obtained by weighted summation. total Backpropagation is used to calculate the gradients of each network layer, and the network C parameters are updated using the optimizer.

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

[0054] L total =L CE +γL Resist

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

[0056] Step S8: Repeat steps S5 to S7. When the training reaches the 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 The fault diagnosis results of the rolling bearings in the train bogie gearbox were obtained.

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

[0060] I. Data acquisition: The vibration signal of the gear box during operation is collected by the acceleration sensor arranged on the train bogie gear box. The fault bearing is distributed in the gear box, and includes nine running states of rolling body fault, inner ring fault, outer ring fault, cage fault and their corresponding different damage degrees. Bogie working condition: motor speed 50Hz, lateral load 150KG, longitudinal load 1700KG.

[0061] II. Data preprocessing: The bearing vibration data of nine running states is intercepted by sliding window samples, the window length is set to 4096, the sliding step is 4096, the sample number ratio of training set and test set is 2:1, and the specific number of training set of each kind of data is 400, and the sample set is 200. The bearing running state is shown in Table 1.

[0062] Table 1 Gear box bearing running state

[0063]

[0064] Discrete wavelet packet transform is performed on each sample, db3 wavelet basis and 6-level decomposition are adopted, one-dimensional vibration data is converted into 64*64 two-dimensional time-frequency feature map input model for training. The wavelet transform formula is as follows:

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

[0066] In the formula, m is a proportional factor, n is a time shift factor, m -1 / 2 Ensure that the energy under different m values remains equal during the wavelet function scaling process.

[0067] III. Model construction: The model is constructed based on the Pytorch deep learning framework. The deep learning model parameters include the weight and bias in the model, which are adjusted to minimize or optimize the loss function during the training process. The model framework parameters of the method of the embodiment are shown in Table 2.

[0068] Table 2 Specific parameter settings of network structure of the embodiment of the application

[0069]

[0070] In the data D trainWhen inputting the network, first input it into the network layer in batches. The network layer is mainly composed of three parts. The first two modules are sequentially subjected to convolution operation, batch normalization, and activation function (Leaky Relu) activation operation, and then connected with a max pooling layer and a Dropout layer to prevent model overfitting and enhance its generalization ability. Then the features are input into the third module for convolution, batch normalization, and activation function (Leaky Relu). Finally, the extracted features are input into the global average pooling layer for feature information fusion, and then the features are extracted through the fully connected layer. Softmax can be used in the last layer of the neural network to convert the original score output by the neural network into a probability distribution, and the final classification result is obtained.

[0071] Four, model training: based on the 64-bit operating system of 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] The five methods of CCE, Coteaching, JoCor, MentorNet, and SPRL are used for comparison with the present application, and different label noise rates are compared under the bearing data set of the gear box of the high-speed train bogie. Each model is repeated ten times, and the average diagnostic accuracy and standard deviation of the test sample set are used to measure the performance of each model. In order to ensure that the experiment is as close to the real working condition as possible, the signal in each training set is added with Gaussian noise with a signal-to-noise ratio of-6dB to verify the robustness of the model to signal noise.

[0075] Five, model testing and analysis:

[0076] Table 4

[0077]

[0078] The average diagnostic precision and standard deviation of each diagnostic task repeated 10 times are shown in Table 4. When the LNR rises from 0 to 0.5, the diagnostic precision of each comparative method decreases significantly. The diagnostic precision of the method disclosed in the application only decreases by 6.81% under the data set, and the average diagnostic precision of the model reaches 91.51% under the LNR of 0.5, which has good label noise robustness. The method disclosed in the application selects the confidence sample as the course information based on the adaptive round statistical update to provide meaningful supervision information for other training samples, updates the model parameters by using the selected confidence sample and the proposed resistance loss function, and optimizes the model by combining the original loss function. By the adaptive course sample selection strategy, the threshold of the confidence sample selection is related to the network update, and with the maturity of the model, 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 of the model under the noise label. The method disclosed in the application can better extract the fault diagnosis classification characteristics under the noise label and improve the fault diagnosis precision.

[0079] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the application. It should be understood that the above description is only for specific embodiments of the application and is not used to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application should be included in the protection scope of the application.

Claims

1. A method for diagnosing bearing faults in a bogie gearbox of a railway train, characterized in that, The method comprises the following steps: S1: Collect the vibration acceleration signals of bearings in different health states of the train bogie gear box, and intercept the samples x through a sliding window sw, label the samples according to the corresponding running state of the bearing, and preprocess into a training sample set D with uniform noise label y O_train , and a test sample set D O_test ; S2: to D O_train , D O_test The time-domain signal is converted into a time-frequency signal by using a discrete wavelet packet, and a training sample set D composed of two-dimensional time-frequency images is obtained train , a test sample set D test ; S3: Constructing a fault diagnosis classification model C with noise label robustness, including setting network layer hyperparameters and defining loss function L total ; S4: randomly initialize network parameters, input the training sample set D in step S2 into the diagnostic model C to obtain the predicted label of the model C train Input the batch size into the diagnostic model C to obtain the predicted label of the model C And adopt a self-paced learning strategy, when the model training round t≤T1, the training set D train Participate in the calculation of the cross-entropy loss L between the noise label y and the predicted label CE , back-propagation to calculate the gradient of each network layer, and update the network C parameters by the optimizer;​ S5: calculating the training sample set D in each round of training train The classification diagnosis probability value P given by the model softmax layer for each category of sample, and the average probability value μ and standard deviation σ given by the model in the current stage for each category of sample, and a threshold value λ is set, when the classification diagnosis probability value P given by the current model for each category of sample is greater than the above threshold value λ, it is defined as a confident sample, which is included in the course information. S6: When model training round t > T1, calculate the cross-entropy loss L of the course sample consisting of the confidence sample in the current model CE , and the resistance loss function L of all samples in the current model Resist , L Resist The calculation formula is shown as follows: In the formula, B represents the number of samples in each batch during model training, p t (x i ) j represents the label prediction probability value of the sample x i belonging to the jth class in the tth training round, p t-1 (x i ) j represents the label prediction probability value of the sample x i belonging to the jth class in the t-1th training round, and x i is the ith sample. S7: L CE with L Resist L total , backpropagation to calculate the gradient of each network layer, and update the network C parameters with the optimizer; S8: repeating steps S5 to S7, when the training reaches a preset number of rounds N, saving the classification model C; S9: load the classification model C, input the test sample set D test , get the train bogie gearbox bearing fault diagnosis result.

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

3. The rail train bogie gearbox bearing fault diagnosis method according to claim 1, wherein the network layer hyperparameters in step S3 comprise a convolution kernel size, a convolution kernel moving step length, and a number of neurons in a fully connected layer.

4. The rail train 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 value of the selection of the jth category of confidence sample, V represents the course sample, a plurality of confidence samples constitute the course sample, W represents the model parameter, B represents the number of samples per batch in the model training process, x i is the ith sample, v i represents the weight of the ith sample, K represents the number of categories in the classification task, p(x i ) j represents the label prediction probability value of the sample x i belonging to the jth category; is the predicted label of the model C of the ith sample of the jth category.

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

6. The rail car truck gear case bearing fault diagnostic method of claim 5, L total The calculation formula is shown as follows: L total = L CE + γL Resist where γ is a weight parameter, L CE is the cross-entropy loss, L Resist is the adversarial loss.

Images