Rolling bearing fault diagnosis method in data imbalance scene
Through the combination of multi-scale mutual learning fault diagnosis model and chain generation adversarial network, the problem of low accuracy in rolling bearing fault diagnosis in data imbalance scenarios is solved, and higher diagnostic accuracy and generalization capabilities are achieved.
Patent Information
- Application Number
- CN202510530317.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-25
- Publication Date
- 2025-08-08
AI Technical Summary
The existing rolling bearing fault diagnosis methods in the scenario of data imbalance is difficult to solve the problems of scarcity and unbalanced data distribution simultaneously, resulting in low diagnostic accuracy.
A multi-scale mutual learning fault diagnosis model is adopted, combined with a chain generation adversarial network for data expansion, and the feature extraction capability is improved through multi-scale feature extraction and mutual learning modules, and the focus loss and supervision and comparison loss optimization model performance is introduced.
It improves the accuracy and generalization ability of fault diagnosis, enhances the model's ability to distinguish different categories, and solves the problem of degradation of diagnostic performance in data imbalance scenarios.
Smart Images
Figure CN120445646A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field related to fault signal processing, and more specifically, relates to a rolling bearing fault diagnosis method in a data imbalance scenario. Background Art
[0002] With the rapid development of mechanical equipment, rolling bearings, as key transmission components, play a vital role in industrial production. However, due to their long-term operation under complex conditions, their reliability is difficult to guarantee. Consequently, a growing number of researchers have begun to focus on rolling bearing fault diagnosis. In recent years, numerous scholars at home and abroad have applied deep learning algorithms, represented by convolutional neural networks, to rolling bearing fault diagnosis. By conducting comprehensive and in-depth analysis of vibration signals during bearing operation, they have built and trained various network models to achieve relatively accurate predictions of bearing health status. However, these models assume a balanced distribution of training data and require a large amount of labeled data. In real industrial scenarios, early fault signals are weak and easily drowned out by complex operating noise, resulting in data scarcity and uneven distribution. This is primarily due to the sudden nature of faults and the varying probabilities of different types of faults, making them difficult to promote and apply on a large scale.
[0003] Existing fault diagnosis methods for data imbalance scenarios primarily include: traditional data augmentation methods, which augment data through time-domain shifting, noise addition, or dynamic simulation modeling; generative modeling methods, which learn from the real data distribution to generate identically distributed samples; and cross-domain transfer methods, which leverage readily available data to transfer knowledge and improve classification accuracy. However, these methods all have limitations. Signal processing methods easily destroy the physical consistency of fault characteristics and can only expand identically distributed samples. While simulation methods offer physical interpretability, their modeling accuracy relies on parameter completeness, making them difficult to adapt to assembly errors and wear degradation, and they lack realistic background noise. Generative models are prone to overfitting in small sample sizes, while domain adaptation methods rely heavily on the alignment assumption between the source and target domains, requiring high data diversity. In practical industrial applications, these methods struggle to simultaneously address the dual challenges of data imbalance and small sample sizes. Therefore, the development of more effective rolling bearing fault diagnosis techniques is urgently needed. Summary of the Invention
[0004] In response to the above defects or improvement needs of the prior art, the present invention provides a rolling bearing fault diagnosis method in a data imbalance scenario to solve the problem of low accuracy of the bearing fault diagnosis model.
[0005] To achieve the above objectives, according to one aspect of the present invention, a multi-scale mutual learning fault diagnosis model is provided. The fault diagnosis model includes a multi-scale feature extraction module and a mutual learning module, wherein:
[0006] The multi-scale feature extraction module is used to extract feature information of the input vibration signal;
[0007] The mutual learning module is connected to the multi-scale feature extraction module, which includes a first channel attention model, a first spatial attention model, a second channel attention model, a second spatial attention model and an output result calculation model, wherein: the first channel attention model and the first spatial attention model are both connected to the multi-scale feature extraction module, the first channel attention model and the second channel attention model, and the first spatial attention model and the second spatial attention model are both connected through an attention fusion layer, the output ends of the second channel attention model and the second spatial attention model are both connected to a fully connected layer, the fully connected layer takes the output of the second channel attention model or the second spatial attention model as input, and the output is the probability of predicting the fault type of the vibration signal, the output result calculation model is connected to all fully connected layers, and the probabilities of various fault types of vibration signals predicted by all fully connected layers are averaged, and the obtained average probability of each fault type is the probability of the fault type corresponding to the input vibration signal.
[0008] Further preferably, the calculation formula of the loss function of the fault diagnosis model is as follows:
[0009] L total =L focal +λ1L KL_all +λ2L con
[0010] L KL_all =L KL1 +L KL2 +L KL3
[0011] Among them, λ1 and λ2 are weight coefficients, L focal is the focal loss function between the fault diagnosis model output and the true value, L con is the supervised contrast loss between the fault diagnosis model output and the true value, L KL_all is the KL divergence loss, L KL1 is the KL divergence loss between the output of the first channel attention model and the first spatial attention model, L KL2 is the KL divergence loss between the output of the second channel attention model and the second spatial attention model, L KL3 is the KL divergence loss between the outputs of the respective connected fully connected layers of the second channel attention model and the second spatial attention model.
[0012] Further preferably, the calculation formula of the KL divergence loss is as follows:
[0013] L KL=KL(P H ||P L )+KL(P L ||P H )
[0014] Among them, P H is the output distribution of the high-resolution branch, P L is the output distribution of the low-resolution branch, and KL is the KL divergence.
[0015] More preferably, the L focal The calculation formula is as follows:
[0016] L focal =-∑(α t (1-p t ) γ log(p t ))
[0017] Among them, α t is the balance factor; p t is the predicted probability and γ is the focusing parameter.
[0018] Further preferably, the calculation formula of the supervised contrast loss is as follows:
[0019]
[0020] Among them, sim(·,·) is the cosine similarity function, z i and z j are the sample features of the same category, z k are sample features of different categories, and τ is the temperature parameter.
[0021] According to another aspect of the present invention, a rolling bearing fault diagnosis method in a data imbalance scenario is provided, the method comprising the following steps:
[0022] Using the vibration signals of the rolling bearing during operation and the fault type corresponding to each vibration signal as training data to train the fault prediction model described in the above item;
[0023] The trained fault prediction model inputs the vibration signal of the rolling bearing to be diagnosed and outputs the probability of each fault type of the rolling bearing, thereby realizing rolling bearing fault diagnosis.
[0024] Further preferably, the method performs data expansion on the training data before training the above-mentioned fault prediction model.
[0025] Further preferably, the data expansion adopts a chained generative adversarial network module, which includes a data augmentation unit, a constraint generator and a discriminator, wherein:
[0026] The data enhancement unit simulates and obtains an ideal fault simulation vibration signal according to the measured vibration signal, and then performs data enhancement on the simulated vibration signal and the measured vibration signal;
[0027] The input of the constraint generator is the simulated vibration signal after data enhancement, and the output is the simulated measured vibration signal;
[0028] The discriminator is used to identify whether the input signal is a simulated measured vibration signal or a measured signal. By calculating the loss function, the network parameters of the constraint generator and the discriminator are adjusted, so that the obtained simulated measured vibration signal is close to the measured vibration signal after data enhancement.
[0029] Further preferably, the data enhancement adopts noise injection, random scaling or random masking, and random translation.
[0030] Further preferably, the calculation formula of the loss function is as follows:
[0031]
[0032]
[0033] L PG =||FFT(G(z))-FFT(x)||2
[0034]
[0035] in, is the base loss of the adversarial generation network, L PG and λ PG is the physical guidance loss and its weight coefficient, L 0MR and λ 0MR is the zero-mean regularization loss and its weight coefficient, E is the expected distribution, x is the input measured vibration signal, z is the input simulated vibration signal, D(.) is the output of the discriminator, G(z) is the output of the generator, Pdata is the distribution of real data, Pz is the distribution of simulated vibration signals, FFT is the fast Fourier transform, ||·||2 is the L2 norm, λ is a hyperparameter, p∈[0,1] is the adaptive control regularization strength, and μ is the mean of the features obtained by zero-mean regularization of the discriminator.
[0036] In general, the above technical solutions conceived by the present invention have the following beneficial effects compared with the prior art:
[0037] 1. The fault prediction model constructed by the present invention adopts a channel attention model and a spatial attention model to form a two-branch and two-level feature extraction mode, which fully mines the data-rich fault features, effectively overcomes the problem of decreased diagnostic model performance in the scenario of unbalanced fault data, and improves the accuracy of fault diagnosis.
[0038] 2. The present invention introduces a multi-scale mutual learning network into the fault diagnosis model. The high-resolution branch captures local detail features, and the low-resolution branch obtains overall trend features, which greatly improves its generalization ability and prediction performance, and the extracted multi-scale features are more representative. In the feature fusion process, the introduction of the mutual learning module can achieve effective complementarity of high- and low-resolution features. The introduction of supervised contrast loss can shorten the feature distance of similar samples and push away heterogeneous samples, thereby improving the discrimination of the feature space, which not only accelerates the convergence speed of the network but also enhances the model diagnosis effect.
[0039] 3. The present invention provides a rolling bearing fault diagnosis method in a data imbalance scenario. Compared with traditional diagnosis methods, the mechanism-guided chain generative adversarial network expands the training data set for training the fault prediction model, solves the problem of a significant decline in the performance of the diagnosis model caused by the scarcity of fault data, and improves the accuracy of fault diagnosis.
[0040] 4. Adversarial Loss in the Chained Adversarial Generative Network of the Present Invention It is the basic loss of the adversarial generative network, which helps to combat the training process and the physical guidance loss L PG , which can promote the fault signal output by the generator to have spectral characteristics similar to the real signal. Zero-mean regularization loss L 0MR In order to improve the training stability of the model, it can suppress abnormal oscillations of the gradient.
[0041] 5. Focus loss L in the multi-scale mutual learning diagnosis module of the present invention focal By adjusting the weights of difficult and easy samples, the model can strengthen the learning of difficult samples. KL loss L KL all By constraining the similarity between the features extracted by the two branches, the feature extraction ability of the model is improved. con By shortening the distance between samples of the same category, the model's ability to distinguish different categories is improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 It is a flow chart of the fault diagnosis method model algorithm constructed according to the preferred embodiment of the present invention;
[0043] Figure 2Schematic diagram of simulated vibration signals obtained by a dynamic simulation model constructed according to a preferred embodiment of the present invention, wherein (a) is a normal vibration signal, (b) is a vibration signal containing rolling element loss, (c) is a vibration signal containing an inner race fault, and (d) is a vibration signal containing an outer race fault;
[0044] Figure 3 is a schematic diagram of a fault diagnosis method model constructed according to a preferred embodiment of the present invention;
[0045] Figure 4 is a test case confusion matrix constructed according to a preferred embodiment of the present invention;
[0046] Figure 5 is a visualization result of a confusion matrix constructed according to a preferred embodiment of the present invention;
[0047] Figure 6 It is a T-SNE feature visualization structure constructed according to the preferred embodiment of the present invention. DETAILED DESCRIPTION
[0048] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
[0049] A multi-scale mutual learning diagnosis module includes a multi-scale feature extraction module and a mutual learning module between different network branches.
[0050] (1) Multi-scale feature extraction aims to extract deep features. First, a feature extraction module (Multi-Scale Convolution Module, MS Module) consisting of four groups of basic convolution blocks with different convolution kernel sizes and a parallel residual connection structure is used to extract and fuse feature information of different scales.
[0051] Secondly, the dual-branch structure based on channel attention and spatial attention is used to capture features, and the fusion features are connected through residual links. Finally, the prediction results are output through the fully connected layer, such as Figure 3As shown. Specifically, the mutual learning module is connected to the multi-scale feature extraction module, which includes a first channel attention model, a first spatial attention model, a second channel attention model, a second spatial attention model and an output result calculation model, wherein: the first channel attention model and the first spatial attention model are both connected to the multi-scale feature extraction module, the first channel attention model and the second channel attention model, and the first spatial attention model and the second spatial attention model are connected through an attention fusion layer, the output ends of the second channel attention model and the second spatial attention model are both connected to a fully connected layer, the fully connected layer takes the output of the second channel attention model or the second spatial attention model as input, and outputs the probability of predicting the fault type of the vibration signal, the output result calculation model is connected to all fully connected layers, and the probabilities of various fault types of the vibration signal predicted by all fully connected layers are averaged, and the obtained average probability of each fault type is the probability of the fault type corresponding to the input vibration signal.
[0052] The module adaptively extracts data features and uses these features to predict the health status of rolling bearings. Unlike traditional single-resolution feature extraction, multi-scale feature extraction offers enhanced capabilities, making it suitable for processing complex data and delivering superior results.
[0053] (2) Mutual Learning Fusion Network: A mutual learning loss function is designed to flatten features after each convolutional layer and calculate feature distribution. KL divergence constraints are used to promote bidirectional complementary optimization of features between the two branches to achieve information transfer. Furthermore, the traditional cross entropy loss is replaced with focal loss to enhance the model's learning of fault samples.
[0054] The preprocessed expanded data is input into the multi-scale feature extraction network to obtain high- and low-resolution features; efficient information interaction between high- and low-resolution features is achieved through the mutual learning module; finally, a classifier is used to classify the health status of the fused features to obtain the final diagnosis result.
[0055] The fault prediction model includes multiple loss functions to optimize the module network parameters, which mainly includes the following three parts:
[0056] First, the network module should have a better classification effect on the fault data. Here we define the focus loss function:
[0057]
[0058] Where, α t is a balancing factor used to adjust the weights of different categories; p t is the prediction probability, and γ is the focusing parameter used to adjust the weight of difficult and easy samples.
[0059] Secondly, by minimizing the KL divergence, we promote the two-way complementary optimization of high-resolution and low-resolution features and achieve mutual learning. The formula is as follows:
[0060] L KL =KL(P H ‖P L )+KL(P L ‖P H )
[0061] Where, P H represents the output distribution of the high-resolution branch, P L represents the output distribution of the low-resolution branch, and KL represents the KL divergence.
[0062] In order to further improve the model's ability to distinguish different categories, supervised contrast loss is introduced:
[0063]
[0064] Where sim(·,·) represents the cosine similarity function, z i and z j Represents the sample features of the same category, z k Represents the sample characteristics of different categories, and τ is the temperature parameter.
[0065] The total loss function is expressed as follows:
[0066] L total =L focal +λ1L KL_all +λ2L con
[0067] L KL_all =L KL1 +L KL2 +L KL3
[0068] Where λ1 and λ2 are weight coefficients.
[0069] A rolling bearing fault diagnosis method in a data imbalance scenario includes the following steps:
[0070] Step a, collecting vibration signals during the operation of the bearing and preprocessing the signals;
[0071] The primary analysis targets vibration signals from bearings in various health states during operation. Due to varying operating conditions, the length of each sample data is likely to vary, necessitating data truncation to ensure consistent input sample length within the network.
[0072] Step b: Inputting the simulated data samples and the collected measured data samples obtained from the dynamic model of the same working condition into the chained generative adversarial network module to train the model, and using the module to expand the minority class fault data;
[0073] Step c: inputting the expanded training data into the multi-scale mutual learning diagnosis module to train the fault diagnosis model;
[0074] Step d: Use test data to test the optimized model and obtain the diagnosis results.
[0075] In step b, since early fault signals in actual industrial sites are weak and easily overwhelmed by complex working condition noise, resulting in data scarcity and uneven distribution; however, using only traditional data augmentation methods will cause the physical consistency of fault characteristics to be destroyed and can only expand the same distribution samples. Therefore, a chained generative adversarial network module is used to expand the minority class fault data.
[0076] The Mechanism-Guided Chain GAN (MG-CGAN) module combines data-driven and mechanism modeling to introduce frequency domain prior constraints during training to enhance the physical consistency of the generated signals.
[0077] First, the dynamic simulation model is used to generate a noise-free ideal fault signal vibration signal (without noise caused by external factors such as the real environment), and the simulation data is obtained as follows Figure 2 As shown in the figure, the simulated vibration signal A and the measured vibration signal B under the same working condition are uniformly enhanced through data alignment strategies, such as noise injection, random scaling, random masking, random translation and other enhancement operations;
[0078] Secondly, during the training phase, a physics-guided loss constraint generator is introduced to enable it to learn fault impact characteristics that conform to physical laws; specifically, the simulated vibration signal A is input and the simulated vibration signal C is output;
[0079] Finally, a discriminator is used to determine whether the simulated vibration signal C or the measured vibration signal B is simulated or measured. Zero-mean regularization and adaptive root mean square normalization are introduced into the discriminator to suppress abnormal oscillations in the discriminator gradient. The discriminator is used to determine whether the signal C is simulated or measured. It balances the convergence of the discriminator with the spectral consistency constraint by dynamically adjusting the normalization strength of the feature map. The spectral consistency constraint uses FFT amplitude spectrum alignment to enhance the frequency domain realism of the generated signal. The measured vibration B and the simulated vibration signal C are Fourier transformed, and the difference between the two spectral signal amplitudes is calculated. This difference is then backpropagated to update the network parameters of the generator and discriminator.
[0080] The parameters of the adversarial network module are optimized using minority class fault data. The loss function mainly includes the following three parts:
[0081] Since the mechanism-guided chain-based generative adversarial network module improves the generator's generation ability through the confrontation between the discriminator and the generator, the adversarial loss function between the discriminator and the generator is defined as:
[0082]
[0083] Where D(x) is the discriminator output, which represents the probability that the input sample x is a true sample. G(z) is the generator output, which represents the data distribution of the generated sample G(z). Pdata is the distribution of the true data, and Pz is the noise distribution.
[0084] During the generation process, in order to ensure the physical consistency of the generated signal, the physical guidance loss (PG-Loss) is introduced and defined as follows:
[0085] L PG =||FFT(G(z))-FFT(x)||2
[0086] Among them, FFT represents fast Fourier transform, ||·||2 represents L2 norm, and this loss function is used to constrain the generated signal to match the real signal in the frequency domain.
[0087] In order to improve the stability of the model, zero-mean regularization loss and adaptive root mean square normalization operation are introduced, which are defined as follows:
[0088]
[0089] The total loss function is expressed as follows:
[0090]
[0091] Among them, λ1 and λ2 are weight coefficients.
[0092] After the model training is completed, the module is used to generate minority class fault samples, which to a certain extent makes up for the shortage of minority class fault samples and provides a more balanced data basis for subsequent diagnosis tasks.
[0093] The present invention will be further described below with reference to specific embodiments.
[0094] This embodiment provides a rolling bearing fault diagnosis method in a data imbalance scenario, and its algorithm flow chart is as follows: Figure 1 The specific implementation steps are as follows:
[0095] Vibration data was extracted from the public dataset CWRU. The time-domain vibration acceleration signals of 10 different state types under the actual 0HP working condition at a sampling frequency of 12K were collected. The types covered normal states as well as outer ring, inner ring, and rolling element faults of different fault depths. 1024 points were sampled in the sliding window as training samples.
[0096] The 5s simulation data of the corresponding state is simulated from the mechanism model as the training input of the generator of the subsequent MGCGAN model. Some simulation data are as follows Figure 2 shown.
[0097] To meet the needs of the present invention's scenarios, this paper constructed four datasets for experimental evaluation. Specifically, in dataset A, the proportion of each type of fault sample is 50% (i.e., 30 out of 60 samples are selected for training), so it can be considered a balanced dataset. In the unbalanced datasets B, C, and D, the proportion of normal samples is also set to 50%, but the proportions of the three types of fault samples (BF, IF, and OF) are adjusted to 10%, 25%, and 40%, respectively, to construct different imbalanced configurations. The proportion of fault categories in the test set is uniformly set to 50% to ensure that all models evaluate the single fault diagnosis ability under the same data scale, facilitating comparison of the results, as shown in Table 1.
[0098] Table 1 Imbalanced dataset settings
[0099]
[0100] The MGCGAN module is trained, and its generator architecture is shown in Table 2. The imbalanced training data is supplemented by the above-mentioned generative model, and input into the multi-scale mutual learning module for training, and the final training model parameters are saved. The module framework is as follows Figure 3 shown.
[0101] Table 2 Generator architecture table
[0102]
[0103] In order to verify the superiority of this method, this section compares and analyzes this method with the more advanced diagnostic methods currently available. Specifically, the methods compared in this article are as follows:
[0104] 1DCNN: Based on a one-dimensional convolutional neural network (1DCNN), it performs fault classification by extracting local features of time series signals.
[0105] RNN: Recurrent Neural Network (RNN), which can model temporal correlation and is suitable for processing vibration signal data with time dependence.
[0106] Attention-based Method (ABM): A deep learning method that introduces an attention mechanism, which can dynamically focus on important information during fault feature extraction and improve diagnostic performance.
[0107] Cost-sensitive CNN (CS-CNN): A convolutional neural network for imbalanced data that uses a cost-sensitive learning strategy to reduce the risk of misclassification of minority classes.
[0108] Transformer: A deep learning architecture based on the self-attention mechanism that can effectively capture long-range dependencies
[0109] The evaluation method is as follows:
[0110] Overall accuracy: the number of correctly classified samples divided by the total number of samples.
[0111] Precision: The number of correctly predicted samples for each category divided by the total number of samples predicted for that category.
[0112] Recall: The number of correctly predicted samples in each category divided by the total number of true samples in that category.
[0113] F1 score: The harmonic average of precision and recall, calculated as F1 = 2 × (Precision × Recall) / (Precision + Recall).
[0114] G-mean: The geometric mean of the recall rates of each category, which can better reflect the classification performance of the minority class in an imbalanced dataset.
[0115] Based on the dataset set in Table 1, we trained all the comparison methods and compared the diagnosis results, such as Figure 4 The following is a visualization of the fault diagnosis effects of the five methods under five different data balance conditions. The detailed diagnosis results are attached in Table 3. Figure 5 is the confusion matrix visualization result, Figure 6 It is the T-SNE feature visualization structure.
[0116] The results show that this method has a better detection effect in the case of data imbalance. Since this method is applicable to more scenarios, it has a broader application prospect.
[0117] Table 3 Detailed diagnosis results under data imbalance conditions are as follows
[0118]
[0119] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A multi-scale mutual learning fault diagnosis model, characterized in that: The fault diagnosis model includes a multi-scale feature extraction module and a mutual learning module, where: The multi-scale feature extraction module is used to extract feature information of the input vibration signal; The mutual learning module is connected to the multi-scale feature extraction module, which includes a first channel attention model, a first spatial attention model, a second channel attention model, a second spatial attention model and an output result calculation model, wherein: the first channel attention model and the first spatial attention model are both connected to the multi-scale feature extraction module, the first channel attention model and the second channel attention model, and the first spatial attention model and the second spatial attention model are both connected through an attention fusion layer, the output ends of the second channel attention model and the second spatial attention model are both connected to a fully connected layer, the fully connected layer takes the output of the second channel attention model or the second spatial attention model as input, and the output is the probability of predicting the fault type of the vibration signal, the output result calculation model is connected to all fully connected layers, and the probabilities of various fault types of vibration signals predicted by all fully connected layers are averaged, and the obtained average probability of each fault type is the probability of the fault type corresponding to the input vibration signal.
2. The fault diagnosis model according to claim 1, wherein: The calculation formula of the loss function of the fault diagnosis model is as follows: L total =L focal +λ1L KL_all +λ2L con L KL_all =L KL1 +L KL2 +L KL3 Among them, λ1 and λ2 are weight coefficients, L focal is the focal loss function between the fault diagnosis model output and the true value, L con is the supervised contrast loss between the fault diagnosis model output and the true value, L KL_all is the KL divergence loss, L KL1 is the KL divergence loss between the output of the first channel attention model and the first spatial attention model, L KL2 is the KL divergence loss between the output of the second channel attention model and the second spatial attention model, L KL3 is the KL divergence loss between the outputs of the respective connected fully connected layers of the second channel attention model and the second spatial attention model.
3. The fault diagnosis model according to claim 2, wherein: The calculation formula of the KL divergence loss is as follows: L KL =KL(P H ||P L )+KL(P L ||P H ) Among them, P H is the output distribution of the high-resolution branch, P L is the output distribution of the low-resolution branch, and KL is the KL divergence.
4. The fault diagnosis model according to claim 2 or 3, characterized in that: The L focal The calculation formula is as follows: L focal =-(a t (1-p t ) γ log(p t )) Among them, α t is the balance factor; p t is the predicted probability and γ is the focusing parameter.
5. The fault diagnosis model according to claim 4, wherein: The calculation formula of the supervised contrast loss is as follows: Among them, sim(·,·) is the cosine similarity function, z i and z j are the sample features of the same category, z k are sample features of different categories, and τ is the temperature parameter.
6. A rolling bearing fault diagnosis method in a data imbalance scenario, characterized in that: The method comprises the following steps: Using the vibration signals during the operation of the rolling bearing and the fault type corresponding to each vibration signal as training data to train the fault prediction model according to any one of claims 1 to 5; The trained fault prediction model inputs the vibration signal of the rolling bearing to be diagnosed and outputs the probability of each fault type of the rolling bearing, thereby realizing rolling bearing fault diagnosis.
7. The rolling bearing fault diagnosis method in a data imbalance scenario according to claim 6, characterized in that: This method performs data expansion on the training data before training the fault prediction model described in any one of claims 1 to 5.
8. The rolling bearing fault diagnosis method in a data imbalance scenario according to claim 7, characterized in that: The data expansion adopts a chained generative adversarial network module, which includes a data enhancement unit, a constraint generator and a discriminator, wherein: The data enhancement unit simulates and obtains an ideal fault simulation vibration signal according to the measured vibration signal, and then performs data enhancement on the simulated vibration signal and the measured vibration signal; The input of the constraint generator is the simulated vibration signal after data enhancement, and the output is the simulated measured vibration signal; The discriminator is used to identify whether the input signal is a simulated measured vibration signal or a measured signal. By calculating the loss function, the network parameters of the constraint generator and the discriminator are adjusted, so that the obtained simulated measured vibration signal is close to the measured vibration signal after data enhancement.
9. The rolling bearing fault diagnosis method in a data imbalance scenario according to claim 8, characterized in that: The data enhancement adopts noise injection, random scaling or random masking, and random translation.
10. The rolling bearing fault diagnosis method in a data imbalance scenario according to claim 8, characterized in that: The calculation formula of the loss function is as follows: L PG ||FFT(G(z))-FFT(x)||2 in, is the base loss of the adversarial generation network, L PG and λ PG is the physical guidance loss and its weight coefficient, L 0MR and λ 0MR is the zero-mean regularization loss and its weight coefficient, E is the expected distribution, x is the input measured vibration signal, z is the input simulated vibration signal, D(.) is the output of the discriminator, G(z) is the output of the generator, Pdata is the distribution of real data, Pz is the distribution of simulated vibration signals, FFT is the fast Fourier transform, ||·||2 is the L2 norm, λ is a hyperparameter, p∈[0,1] is the adaptive control regularization strength, and μ is the mean of the features obtained by zero-mean regularization of the discriminator.