Rolling bearing fault diagnosis method for single-domain unbalanced data

Abstract

The invention belongs to the field of fault diagnosis, and discloses a rolling bearing fault diagnosis method for single-domain unbalanced data. A single-domain balance data generation module based on decoupling mixing and a multi-domain balance data generation module based on loss driving are designed, single-domain unbalanced data are converted into single-domain balance data, and then the single-domain balance data are converted into multi-domain balance data; and performing feature extraction on the multi-domain balance data after semantic alignment to obtain a health state category. The method provided by the invention not only can solve the training problem caused by multi-domain data distribution difference to the model, but also can process the problems caused by a single source domain and unbalanced data, is superior to other existing methods, has relatively important practical application value, and is suitable for popularization and application. And wide application of the fault diagnosis model in an actual production environment can be further promoted.

Description

Technical Field

[0001] The invention relates to the field of fault diagnosis, and in particular to a rolling bearing fault diagnosis method oriented to single-domain imbalance data. Background Art

[0002] Rotating machinery is widely used in many industries such as aerospace, transportation, petrochemicals, etc., and occupies a key position in various production systems. Among them, rolling bearings are important components of rotating machinery, and their operating status is directly related to the stability and reliability of the entire equipment. Due to the complex and changeable operating environment of rotating equipment, rolling bearings often face relatively harsh working conditions such as high temperature, high pressure, and high load, which significantly increases the risk of failure. Once a rolling bearing fails, it is very likely to cause a shutdown accident of the entire rotating machinery system, which will cause serious property losses and even threaten the safety of personnel life. Therefore, it is of vital practical significance to carry out accurate and timely fault diagnosis of rolling bearings.

[0003] In actual industrial production scenarios, rotating machinery often presents significant diversity and instability in its operating conditions due to the combined influence of multiple factors such as dynamic changes in operating conditions, long-term wear and tear of the equipment itself, and aging degradation. Such changes in operating conditions pose a severe challenge to the effective application of fault diagnosis technology based on deep learning. Especially when facing single-domain data, changes in operating conditions can easily cause a shift in data distribution, resulting in a significant decrease in the performance of deep diagnostic models. What is even more tricky is that the actual collected single-domain fault data often has serious imbalance problems, that is, there are large differences in the number of samples for different fault categories. This imbalance may be caused by a variety of practical limiting factors, such as sensor failure, inconsistent data collection time windows, and low frequency of certain faults.

[0004] Traditional fault diagnosis methods based on deep learning usually rely on the assumption of relatively balanced data distribution. However, in the reality of single-domain unbalanced data, the diagnostic accuracy of these methods will be greatly suppressed and the generalization ability will become very limited. Although in recent years, methods based on domain generalization have made certain progress in the field of cross-domain fault diagnosis. They match the distribution of samples in different domains in high-level subspaces and extract domain-invariant features to deal with cross-domain distribution differences, thus showing good performance in cross-domain fault diagnosis. However, when dealing with single-domain unbalanced data problems, these methods cannot be directly and effectively applied to solve the challenges brought by single-domain unbalanced data because their inherent design mechanisms do not fully consider the actual situation such as label imbalance within a single domain and cross-domain label distribution differences. Therefore, how to design and develop fault diagnosis methods with high diagnostic accuracy and strong generalization ability for the problem of single-domain unbalanced data has become a key problem that needs to be solved in the current field of rotating machinery fault diagnosis. This has important theoretical value and practical significance for ensuring the safe and stable operation of rotating machinery and improving the efficiency and reliability of industrial production.

[0005] At present, most of the generalized research methods in the field of fault diagnosis are methods under single-domain balanced data and methods under multi-domain unbalanced data. Considering that the rolling bearing data collected in real scenarios may show single-domain unbalanced characteristics and the number of samples of different fault categories varies significantly, the present invention combines the above two types of research and proposes a more extreme single-domain unbalanced data condition. At the same time, it proposes a solution to solve the diagnosis problem under this data condition, namely, a rolling bearing fault diagnosis method for single-domain unbalanced data. Summary of the invention

[0006] In view of the above situation, the present invention specifically proposes a rolling bearing fault diagnosis method for single-domain imbalance data. The main idea is to convert single-domain imbalance data into multi-domain balance data. This problem can be considered in two stages: first, converting single-domain imbalance data into single-domain balance data, and then converting single-domain balance data into multi-domain balance data.

[0007] The technical solution of the present invention is as follows: a rolling bearing fault diagnosis method for single-domain imbalance data, first converting the single-domain imbalance data into single-domain balance data, and then converting the single-domain balance data into multi-domain balance data; the multi-domain balance data is semantically aligned and feature extracted to obtain a health status category.

[0008] Furthermore, the conversion of the single-domain unbalanced data into the single-domain balanced data is realized by a single-domain balanced data generation module based on decoupling mixing; the single-domain balanced data generation module based on decoupling mixing includes: a fault extraction branch, a working condition extraction branch, a binary classifier module, a decoder and a feature mixing module;

[0009] The single-domain balanced data generation module based on decoupled mixing is an encoder-decoder structure, and a deep neural network is used as an encoder. The encoder includes a fault extraction branch and a working condition extraction branch. The single-domain unbalanced data is simultaneously input into the fault extraction branch and the working condition extraction branch to extract high-dimensional fault features f respectively. fr and high-dimensional working condition characteristics f wr ; The high-dimensional fault feature f fr Through the feature mixing module, a new high-dimensional fault feature is generated and recorded as f fr_new ; The decoder will fuse the generated high-dimensional fault features f fr_new Compared with the original high-dimensional working condition feature f wr As input, after feature concatenation, the original form of data is reconstructed as single-domain balanced data.

[0010] Further, the fault extraction branch includes a plurality of sub-branches, and the number of sub-branches is consistent with the number of fault categories in the single-domain unbalanced data;

[0011] The binary classifier module includes several submodules, the number of which is consistent with the number of fault categories in the single-domain unbalanced data. Each submodule only inputs the output of its corresponding fault extraction subbranch to promote each fault extraction subbranch to extract relevant fault information.

[0012] Furthermore, the fault extraction branch includes five convolutional layers and five maximum pooling layers, the first convolutional layer uses a 4×1 convolutional kernel, and the remaining convolutional layers all use a 3×1 convolutional kernel; batch normalization and activation operations are used after each convolution operation, and each sub-branch in the fault extraction branch maintains the same structure;

[0013] The working condition extraction branch includes five convolutional layers and five maximum pooling layers; the first convolutional layer uses an 8×1 convolution kernel, and the remaining convolutional layers use a 6×1 convolution kernel. Batch normalization and activation operations are also used after each convolution operation.

[0014] The binary classifier module performs a Flatten operation on the input data to flatten the features into one-dimensional data, then transforms them through several linear layers and finally maps them into two-dimensional outputs;

[0015] The decoder transforms the high-dimensional fault feature f fr and high-dimensional working condition characteristics f wr As the module input, perform element-by-element addition to obtain the fusion feature f fusion , and after several deconvolution operations, the original form of the data is obtained.

[0016] Furthermore, the conversion of the single-domain balanced data into multi-domain balanced data is achieved through a loss-driven multi-domain balanced data generation module; the loss-driven multi-domain balanced data generation module includes a VAE reconstruction network and a semantic consistency module; by building and training a VAE reconstruction network, the deep information of the single-domain balanced data is learned; the VAE reconstruction network parameters are fixed, the single-domain balanced data is set to a computable gradient form, and after the single-domain balanced data is input into the VAE reconstruction network and the semantic consistency module, the loss is calculated and back-propagated to the single-domain balanced data to obtain data containing the new domain, thereby generating multi-domain balanced data.

[0017] Furthermore, the VAE reconstruction network uses the single-domain balanced data as training data to learn the deep semantic information of the original single-domain balanced data; the VAE reconstruction network uses a convolution kernel of size 4×1, and uses maximum pooling, batch normalization and activation operations after each convolution operation;

[0018] The semantic consistency module is a binary classifier module trained in the single-domain balanced data generation module based on decoupled mixing; the multi-domain balanced data reconstructed by the VAE reconstruction network is taken as input and input into the binary classifier of the corresponding category, its classification loss is calculated, and the classification loss is minimized to ensure that the reconstructed multi-domain balanced data can be normally classified into the corresponding category.

[0019] Furthermore, the training and generation process of the loss-driven multi-domain balanced data generation module is as follows:

[0020] Use the single-domain balanced data as the input of the VAE reconstruction network for model training;

[0021] The parameters of the VAE reconstruction network are fixed, and the binary classifier module trained in the single-domain balanced data generation module based on decoupled mixing is used as the semantic consistency module; the single-domain balanced data is set to a computable gradient form, and the total loss is obtained by calculating and maximizing the VAE loss and minimizing the semantic consistency loss; the total loss is back-propagated and optimized to the single-domain balanced data as a data iteration, and a batch of new domain data is obtained after several rounds of iterations; the above iterative process is repeated several times to obtain several batches of new domain data.

[0022] Furthermore, semantic alignment is performed on the multi-domain balanced data using maximum mean difference.

[0023] Beneficial effects of the present invention: The present invention better solves the impact of extreme data conditions on model training in cross-domain fault diagnosis research, and has a certain promoting effect on fault diagnosis research under single-domain unbalanced data. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is the overall framework diagram of the present invention.

[0025] Figure 2 This is a structural diagram of the single-domain balanced data generation module based on decoupled mixing.

[0026] Figure 3 This is a structural diagram of the loss-driven multi-domain balanced data generation module.

[0027] Figure 4 This is the probability density plot of Beta(0.5,0.5).

[0028] Figure 5 This is a comparison chart of the generated data and the original data. (a) is the original data, and (b) is the generated data.

[0029] Figure 6 It is the T-SNE diagram of the original data and the generated multi-domain data.

[0030] Figure 7 It is the result of comparative experiment. DETAILED DESCRIPTION

[0031] The specific implementation modes of the present invention are described in detail below.

[0032] The present invention is mainly based on the following four modules: First, a single-domain balanced data generation module based on decoupled mixing is designed. This module obtains high-dimensional fault information and high-dimensional working condition information respectively through decoupling, and uses a hybrid method to balance categories to obtain single-domain balanced data; secondly, a multi-domain balanced data generation module based on loss-driven is designed. This module first uses the aforementioned single-domain balanced data for training, and then generates data for new working conditions by fixing model parameters and optimizing data; then, a semantic alignment module and a health status classification module are designed to reduce cross-domain distribution differences by learning domain-invariant features, thereby improving the classifier's ability to identify similar faults under different working conditions. In general, the method proposed in the present invention solves the impact of single-domain unbalanced data conditions on the model and enhances the generalization ability of the model.

[0033] The present invention proposes a rolling bearing fault diagnosis method for single-domain unbalanced data. The overall framework includes a single-domain balance data generation module based on decoupled mixing, a multi-domain balance data generation module based on loss drive, a semantic alignment module and a health status classification module. The input of the framework is a single source domain, a one-dimensional original vibration signal with unbalanced categories, and a data health status, and the output is the current health status predicted by the data.

[0034] First, the present invention analyzes the characteristics of the fault diagnosis methods based on domain generalization that are mainly used in existing research. Most of these methods assume that the data set used to train the model usually contains multiple operating conditions, and the ratio between each fault category is the same, that is, multi-domain balanced data, but this data condition is difficult to meet in the actual production environment. As a result, methods for model training based on single-domain balanced data or multi-domain unbalanced data have emerged. Although these methods have considered the problems of the actual production environment to a certain extent, they still leave room, that is, more extreme scenarios may appear, single-domain unbalanced data. In addition, most of the current methods used for single-domain balanced data or multi-domain unbalanced data use data enhancement to generate some samples containing new conditions or new samples containing known class faults to overcome the impact of data problems. Among them, common methods include mixing (Mixup) and generative models. The mixing operation usually selects two samples, performs interpolation mixing operations on them according to a certain ratio, and then generates new samples, but this scheme can only generate data between two samples, and cannot jump out of the two points to generate real new samples, so this method cannot be directly used to generate samples. The generative model method is usually to train a model similar to a generative adversarial network and then use it to generate new data. However, this method uses coupled information when extracting data and cannot generate a specific type of data. Both of these solutions have certain limitations, which limit the diversity of generated samples to a certain extent. The present invention combines the characteristics of the above method to generate data and achieve sample expansion while retaining the fault information in the original data.

[0035] (1) In order to realize the conversion of single-domain unbalanced data into single-domain balanced data and to generate samples with higher purity, given that neural networks have the ability to automatically learn features and can adaptively capture and extract relevant information from original signals, the present invention considers using neural networks to build models, design fault extraction branches and operating condition extraction branches, perform data decoupling on the original data, and decompose the data into fault-related information and operating condition-related information. Subsequently, in view of the characteristics of the Mixup method, that is, it is more suitable for generating samples between two points, a new hybrid scheme is designed to achieve sample category balance, that is, to process high-dimensional fault information and to generate minority categories of faults in a direction by controlling the size of the hybrid parameter λ. The new high-dimensional fault information and the original high-dimensional fault information are randomly combined with the original high-dimensional operating condition information to generate new samples to obtain single-domain balanced data. This part of the module can be found in Figure 2 .

[0036] (2) In order to realize the conversion of single-domain balanced data into multi-domain balanced data, the present invention designs a VAE reconstruction network to learn the characteristics of the above-mentioned single-domain balanced data and grasp the sample information. Subsequently, the data is trained by fixing the model parameters, calculating the model loss, and optimizing the data, thereby realizing the generation of new working condition data. In addition, in order to avoid the loss of semantic information caused by gradient random optimization during the training data process, this part uses a semantic consistency module to calculate the degree of information loss that occurs during the data optimization process, and reduces the semantic information loss by minimizing the semantic consistency loss. This part of the module can be found in Figure 3 .

[0037] (3) In order to complete the domain-invariant feature extraction and health status classification of the data, the method proposed in the present invention can correctly extract cross-domain features and classify the features. The maximum mean difference (MMD) was first proposed for the detection problem of two samples, mainly used to determine whether two distributions p and q are the same. Its basic assumption is: if for all functions f with the sample space generated by the distribution as input, if the means of the corresponding images of enough samples generated by the two distributions on f are equal, then the two distributions can be considered to be the same distribution. Now this method is generally used to measure the similarity between two distributions, which is based on the reproducing kernel Hilbert space theory. By calculating the mean difference between the source domain features and the target domain features in the reproducing kernel Hilbert space, the similarity between their distributions is evaluated. The present invention extracts domain-invariant features through this method; in addition, on this basis, a health status classification module is added as a task module, and the output of the semantic alignment module is used as input, and the health status classification module outputs the health status of the current data according to the input.

[0038] (4) In order to evaluate the performance of the method proposed in this invention, experiments were carried out on two public bearing fault datasets (CWRU dataset and PU dataset). First, comparative experiments were carried out with several single-domain balanced data fault diagnosis methods and several multi-domain unbalanced data fault diagnosis methods to verify the effectiveness and superiority of the proposed method; secondly, ablation experiments and feature visualization were carried out to verify the effectiveness and necessity of each module in the proposed method; finally, comparative experiments with different parameters were carried out to explore the performance of the proposed method under different parameter settings. The experimental results are shown in the figure. Figure 5 ,6,7; among them Figure 5 A comparison chart of the original data and the newly generated data is shown; Figure 6 The T-SNE plots of the original data and the generated multi-domain data are shown; Figure 7The experimental results show that the proposed method effectively solves the single-domain imbalanced data problem in cross-domain fault diagnosis tasks, and outperforms other existing mainstream methods in terms of diagnosis accuracy and generalization performance. It has important practical application value and further promotes the application of deep fault diagnosis models in actual production environments.

[0039] The environment of this implementation is: Windows 10, PyTorch.

[0040] Step 1: Implement each part.

[0041] Step 1: Implement a single-domain balanced data generation module based on decoupled mixing.

[0042] Since the fault-related data and the working condition-related data in the vibration signal data react differently to convolution kernels of different sizes, the fault-related information is more sensitive to small convolution kernels, and the working condition-related information is more sensitive to large convolution kernels. Therefore, the fault extraction branch and the working condition extraction branch use small convolution kernels and large convolution kernels to extract features respectively. The fault extraction branch contains five convolution layers and five maximum pooling layers. Among them, the first convolution layer uses a 4×1 convolution kernel, and then the remaining convolution layers use a 3×1 convolution kernel. This is conducive to deepening the network model, improving the feature extraction capability, and suppressing overfitting. Batch normalization and activation operations are used after each convolution operation, and each sub-branch in the fault extraction branch maintains the same structure. The working condition extraction branch also contains five convolution layers and five maximum pooling layers. Among them, the first convolution layer uses an 8×1 convolution kernel, and the remaining convolution layers use a 6×1 convolution kernel. Batch normalization and activation operations are also used after each convolution operation. The network structure parameters of the fault extraction branch are shown in Table 1. The structural parameters of the branch network extracted from the working condition are shown in Table 2.

[0043] The binary classifier module receives the output of the aforementioned fault sub-branch as input, and promotes each fault sub-branch to extract information corresponding to the fault category. After receiving the output of the aforementioned fault sub-branch, this part performs a Flatten operation on the data to flatten the features into one-dimensional data, and then transforms them through several linear layers, and finally maps them into two-dimensional outputs. The binary classifier network structure parameters are shown in Table 3.

[0044] The decoder is used to fuse high-dimensional fault information and high-dimensional operating condition information and reconstruct them into original data. After receiving the high-dimensional fault information and operating condition information obtained by the above operations, the two are stacked to form fusion features, and after several inverse convolution operations, the original data can be obtained. The decoder network structure parameters are shown in Table 4.

[0045] Table 1. Fault extraction branch network structure parameters

[0046]

[0047] Table 2 Branch network structure parameters extracted under working conditions

[0048]

[0049]

[0050] Table 3. Network structure parameters of binary classifier

[0051]

[0052] The feature mixing module is used to mix high-dimensional fault features. This part controls the data sampled from the Beta (0.5, 0.5) distribution, and then controls the mixing operation to generate a specific category of high-dimensional fault features. Among them, Beta (0.5, 0.5) is symmetric about x = 0.5, such as Figure 4 As shown, in order to enable the sampled λ to control the generation of specific class samples, this scheme calculates the upper α quantile x up and the lower α quantile x low To filter the sampled λ, as shown in formula (1).

[0053]

[0054] Table 4 Decoder network structure parameters

[0055]

[0056] Table 5 VAE reconstruction network parameters

[0057]

[0058]

[0059] Step 2: Implement the loss-driven multi-domain balanced data generation module.

[0060] The purpose of the VAE reconstruction network is to extract relevant information as much as possible, realize data reconstruction, and generate new working condition data. Therefore, this part needs to pay attention to both fault information and working condition information. Therefore, in this part of the module, a convolution kernel of size 4×1 is used, and each layer of convolution operation is followed by a maximum pooling operation, a batch normalization operation, and an activation operation. The VAE network structure parameters are shown in Table 5.

[0061] The purpose of the semantic consistency module is to calculate the semantic loss and reduce the semantic loss by reducing the loss. This part uses the binary classifier module trained in the first step, inputs the sample into the corresponding binary classifier, calculates the classification loss, and minimizes the classification loss, thereby maintaining the semantic content of the data.

[0062] Step 3: Design and implement the loss functions of each part of the present invention. First, for the single-domain balanced data generation module based on decoupled mixing, in this part of the training process, in order to make the fault information and the working condition information as uncorrelated as possible, the orthogonal loss can be used to reduce the correlation between the two. The loss is shown in formula (2). fr is the fault-related feature, x wr Characteristics related to working conditions.

[0063]

[0064] In order to enable different fault branches to extract specific categories of fault information, a binary cross entropy loss is designed for each branch, and its loss function is shown in formula (3). i is the input sample, y m is the label corresponding to the input sample.

[0065]

[0066] In order to make the information between different fault branches as unrelated as possible, orthogonal loss can be used to reduce the correlation between each branch. The loss function is shown in formula (4). fr is the fault-related feature, x wr Characteristics related to working conditions.

[0067]

[0068] In order to make the reconstructed sample as similar as possible to the original sample and reduce data loss, reconstruction loss is used, and its loss function is shown in formula (5). new is the decoded sample, and MSE(·) is the mean square error.

[0069] L recon_1 =MSE(x,x new ) (5)

[0070] Secondly, for the loss-driven multi-domain balanced data generation module, in order to enable the model to reconstruct the data normally, the reconstruction loss is used, and its loss function is shown in formula (6). Where x is the original sample, is the decoded sample, and MSE(·) is the mean square error.

[0071]

[0072] To ensure that the latent variables in the VAE reconstruction model are closer to the normal distribution, KL divergence is used for calculation, and its loss function is shown in formula (7). i is the sample mean, is the sample variance.

[0073]

[0074] In order to optimize the sample in a specific direction to generate multi-domain data, the original sample The features obtained after encoding are recorded as Each time a sample is input into the model The features obtained after encoding are recorded as By maximizing and The Wasserstein distance between them realizes data reverse optimization, and its loss function is shown in formula (8).

[0075]

[0076] In order to reduce the loss of semantic content, the semantic consistency module loss is shown in formula (9). To enhance the sample, y m is the label of the enhanced sample.

[0077]

[0078] For the semantic alignment module, it mainly uses MMD to extract domain invariant features, and its loss function is shown in formula (10).

[0079]

[0080] For the health status classification module, cross entropy loss is used, and the loss function is shown in formula (11).

[0081]

[0082] Step 2: Experimental dataset.

[0083] The present invention uses CWRU dataset and PU dataset as datasets for model training and testing. In order to ensure the smooth progress of the performance, the present invention preprocesses the dataset, and the specific operation is to generate a dataset for training and testing from the collected original signal. It mainly includes three steps: sample segmentation, normalization, and construction of a single-domain unbalanced dataset. During the sample segmentation process, the number of overlapping samples is set to half the sample length. During the normalization process, each sample is normalized using its own mean and variance to limit the amplitude of the sample, which helps to reduce the difficulty of model training. The normalized samples are then used to construct a single-domain unbalanced dataset.

[0084] Step 3: Train the model.

[0085] The first step is to use the deep learning framework PyTorch to build a model and perform initialization and configuration.

[0086] The second step is to load the training data and preprocess the data.

[0087] The third step is to train a single-domain balanced data generation module based on disentangled mixing and use this module to generate single-domain balanced data.

[0088] In the fourth step, the single-domain balanced data obtained in the third step is used to train a loss-driven multi-domain balanced data generation module, and the module is used to generate multi-domain balanced data.

[0089] In the fifth step, use the multi-domain balanced data obtained in the fourth step to train the semantic alignment module and health status classification module and save the model.

[0090] Step 4: Test the model.

[0091] In the first step, we use the PyTorch framework to initialize the classifier network, initialize the parameters, and preprocess the test data.

[0092] The second step is to load the trained model parameters.

[0093] The third step is to input the test data into the model to obtain the test results.

Claims

1. A rolling bearing fault diagnosis method for single-domain imbalance data, characterized in that: First, the single-domain unbalanced data is converted into single-domain balanced data, and then the single-domain balanced data is converted into multi-domain balanced data; the multi-domain balanced data is semantically aligned and feature extraction is performed to obtain health status categories.

2. The rolling bearing fault diagnosis method for single-domain imbalance data according to claim 1 is characterized in that: The conversion of the single-domain unbalanced data into the single-domain balanced data is realized by a single-domain balanced data generation module based on decoupling mixing; the single-domain balanced data generation module based on decoupling mixing includes: a fault extraction branch, a working condition extraction branch, a binary classifier module, a decoder and a feature mixing module; The single-domain balanced data generation module based on decoupled mixing is an encoder-decoder structure, and a deep neural network is used as an encoder. The encoder includes a fault extraction branch and a working condition extraction branch. The single-domain unbalanced data is simultaneously input into the fault extraction branch and the working condition extraction branch to extract high-dimensional fault features f respectively. fr and high-dimensional working condition characteristics f wr ; The high-dimensional fault feature f fr Through the feature mixing module, a new high-dimensional fault feature is generated and recorded as f fr_new ; The decoder will fuse the generated high-dimensional fault features f fr_new Compared with the original high-dimensional working condition feature f wr As input, after feature concatenation, the original data is reconstructed as single-domain balanced data.

3. The rolling bearing fault diagnosis method for single-domain imbalance data according to claim 2 is characterized in that: The fault extraction branch includes a plurality of sub-branches, and the number of sub-branches is consistent with the number of fault categories in the single-domain unbalanced data; The binary classifier module includes several submodules, the number of which is consistent with the number of fault categories in the single-domain unbalanced data. Each submodule only inputs the output of its corresponding fault extraction subbranch to promote each fault extraction subbranch to extract relevant fault information. The feature mixing module is used to target different categories of high-dimensional fault features f fr Perform a hybrid operation and generate a few types of faults in a directional manner by controlling the size of the hybrid parameter λ to obtain a new high-dimensional fault feature f fr_new .

4. The rolling bearing fault diagnosis method for single-domain imbalance data according to claim 3 is characterized in that: The fault extraction branch includes five convolutional layers and five maximum pooling layers. The first convolutional layer uses a 4×1 convolution kernel, and the remaining convolutional layers use a 3×1 convolution kernel. Batch normalization and activation operations are used after each convolution operation, and each sub-branch in the fault extraction branch maintains the same structure. The working condition extraction branch includes five convolutional layers and five maximum pooling layers; the first convolutional layer uses an 8×1 convolution kernel, and the remaining convolutional layers use a 6×1 convolution kernel. Batch normalization and activation operations are also used after each convolution operation. The binary classifier module performs a Flatten operation on the input data to flatten the features into one-dimensional data, then transforms them through several linear layers and finally maps them into two-dimensional outputs; The decoder transforms the high-dimensional fault feature f fr and high-dimensional working condition characteristics f wr As the module input, perform element-by-element addition to obtain the fusion feature f fusion , and after several deconvolution operations, the original form of the data is obtained.

5. The rolling bearing fault diagnosis method for single-domain imbalance data according to claim 4 is characterized in that: The conversion of single-domain balanced data into multi-domain balanced data is realized by a loss-driven multi-domain balanced data generation module; the loss-driven multi-domain balanced data generation module includes a VAE reconstruction network and a semantic consistency module; by building and training a VAE reconstruction network, the deep information of the single-domain balanced data is learned; the VAE reconstruction network parameters are fixed, the single-domain balanced data is set to a computable gradient form, and after the single-domain balanced data is input into the VAE reconstruction network and the semantic consistency module, the loss is calculated and back-propagated to the single-domain balanced data to obtain data containing the new domain, thereby generating multi-domain balanced data.

6. The rolling bearing fault diagnosis method for single-domain imbalance data according to claim 5, characterized in that: The VAE reconstruction network uses the single-domain balanced data as training data to learn the deep semantic information of the original single-domain balanced data; the VAE reconstruction network uses a convolution kernel of size 4×1, and uses maximum pooling, batch normalization and activation operations after each convolution operation; The semantic consistency module is a binary classifier module trained in the single-domain balanced data generation module based on decoupled mixing; the multi-domain balanced data reconstructed by the VAE reconstruction network is taken as input and input into the binary classifier of the corresponding category, its classification loss is calculated, and the classification loss is minimized to ensure that the reconstructed multi-domain balanced data can be normally classified into the corresponding category.

7. The rolling bearing fault diagnosis method for single-domain imbalance data according to claim 5, characterized in that: The training and generation process of the loss-driven multi-domain balanced data generation module is as follows: Use the single-domain balanced data as the input of the VAE reconstruction network for model training; The parameters of the VAE reconstruction network are fixed, and the binary classifier module trained in the single-domain balanced data generation module based on decoupled mixing is used as the semantic consistency module; the single-domain balanced data is set to a computable gradient form, and the total loss is obtained by calculating and maximizing the VAE loss and minimizing the semantic consistency loss; the total loss is back-propagated and optimized to the single-domain balanced data as a data iteration, and a batch of new domain data is obtained after several rounds of iterations; the above iterative process is repeated several times to obtain several batches of new domain data.

8. The rolling bearing fault diagnosis method for single-domain imbalance data according to claim 1, characterized in that: The multi-domain balanced data is semantically aligned using the maximum mean difference.

Images