Rolling bearing fault diagnosis method, device and storage medium

Abstract

The application provides a rolling bearing fault diagnosis method, device and storage medium, and relates to the technical field of intelligent fault diagnosis. The method comprises the following steps: generating a candidate sample set corresponding to a target fault category based on a generation quota corresponding to the target fault category in a target round of iteration and a trained classification generative adversarial network; screening each candidate target fault category sample in the candidate sample set, and incorporating the qualified sample obtained after screening into a small sample training set to form an expanded training set; training an initial fault diagnosis model based on the expanded training set to obtain a trained fault diagnosis model; evaluating the trained fault diagnosis model based on a verification set, and updating the generation quota of the target fault category in the next round of iteration and a preset screening rule according to the evaluation result to determine whether the next round of iteration is to be performed; and if the next round of iteration is to be stopped, performing fault diagnosis based on the trained fault diagnosis model obtained in the last round of training. The application can improve the diagnosis accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology, and in particular to a method, device and storage medium for diagnosing rolling bearing faults. Background Technology

[0002] Rolling bearings are core supporting components of rotating machinery, and their operating condition directly affects the reliability and safety of port machinery, industrial transmission equipment, and various rotating machinery systems. In actual industrial scenarios, rolling bearings often operate under complex conditions such as high speed, high temperature, heavy load, or variable load. Once a failure occurs, it may cause equipment downtime, which will not only increase maintenance costs but may even lead to safety accidents.

[0003] With the development of deep learning technology, many types of deep models have been widely used in rolling bearing fault diagnosis. However, deep models typically require a large number of labeled fault samples for stable training. In real industrial settings, on the one hand, equipment cannot be artificially operated to a faulty state for extended periods; on the other hand, the data distribution varies significantly depending on the type of fault, the degree of damage, and the operating conditions, making it difficult to collect a large number of fault samples.

[0004] Currently, with the continuous development of generative adversarial networks and their variants for small-sample fault diagnosis, although only a small number of labeled fault samples can be used, the inventors have found that due to the inconsistent quality of the generated samples, the diagnostic accuracy of the trained fault diagnosis model still cannot be guaranteed. Summary of the Invention This invention provides a method, device, and storage medium for diagnosing rolling bearing faults, in order to solve the problem that current fault diagnosis models cannot accurately diagnose rolling bearing faults.

[0005] In a first aspect, embodiments of the present invention provide a method for diagnosing rolling bearing faults, including: Construct a small-sample training set and a validation set for the rolling bearing to be diagnosed; both the small-sample training set and the validation set include real samples of multiple fault categories. The pre-built classification generative adversarial network and the pre-built fault diagnosis model are trained based on the small sample training set to obtain the trained classification generative adversarial network and the initial fault diagnosis model. In the target round iteration, a candidate sample set corresponding to the target fault category is generated based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network. According to a preset screening rule, the candidate target fault category samples in the candidate sample set are screened, and the qualified samples obtained after screening are incorporated into the small sample training set to form an expanded training set. The initial fault diagnosis model is trained based on the expanded training set to obtain a trained fault diagnosis model. The trained fault diagnosis model is evaluated based on the validation set, and the generation quota of the target fault category in the next round iteration is updated according to the evaluation results, and it is determined whether to proceed to the next round iteration to generate samples. If the next iteration is stopped, the rolling bearing is then diagnosed based on the training fault diagnosis model obtained from the previous iteration.

[0006] In one possible implementation, the classification generative adversarial network includes a generator, a discriminator, and an independent classifier. The generator is used to generate candidate fault samples, the discriminator is used to output a realism score, and the independent classifier is used to output a class confidence score and an intermediate layer feature vector. In the target round iteration, generating a candidate sample set corresponding to the target fault category based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network, further includes: In the target round iteration, based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network, the authenticity score of each candidate target fault category sample, the category confidence score of each candidate target fault category sample, and the intermediate layer feature vector of each candidate target fault category sample are generated.

[0007] In one possible implementation, the step of filtering candidate target fault category samples in the candidate sample set according to a preset filtering rule, and merging the qualified samples obtained after filtering into the small sample training set to form an expanded training set, includes: The comprehensive credibility score of the candidate target fault category samples is determined based on the authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity. If the comprehensive credibility score, authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity of the candidate target fault category sample all meet the preset screening rules, then the candidate target fault category sample is determined to be a qualified sample, and the qualified sample is incorporated into the small sample training set to form an expanded training set; otherwise, the candidate target fault category sample is removed.

[0008] In one possible implementation, a comprehensive credibility score for candidate target fault category samples is determined based on the authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity. This score includes: Based on the intermediate layer feature vectors of the candidate target fault category samples and the feature centers of the real samples of the target fault category, the intra-class feature distance is determined; Based on the intermediate layer feature vectors of the candidate target fault category samples and the corresponding easily confused target fault category feature centers, the inter-class boundary distance is determined; wherein, the easily confused target fault category is determined from the confusion matrix obtained by testing the training fault diagnosis model in the previous round of the target round iteration based on the validation set; Spectral or envelope spectrum analysis is performed on the candidate target fault category samples and the real samples of the target fault category to obtain a frequency domain consistency score; Based on the candidate target fault category samples and the qualified samples of the target fault category obtained after screening, the maximum similarity is determined; The comprehensive credibility score of the candidate target fault category samples is determined by a weighted sum of the authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity.

[0009] In one possible implementation, evaluating the trained fault diagnosis model based on the validation set, updating the generation quota of the target fault category in the next iteration based on the evaluation results, and determining whether to proceed with the next iteration to generate samples include: The trained fault diagnosis model is evaluated based on the validation set to obtain the recall rate, confusion matrix, and comprehensive performance index of the trained fault diagnosis model for the target fault category. Based on the recall rate, confusion matrix, and generation quota of the target fault category in the current iteration, as well as the preset minimum generation quantity and preset maximum generation quantity of the target fault category, the generation quota of the target fault category in the next iteration is determined. Based on the difference between the comprehensive performance index of the trained fault diagnosis model and the comprehensive performance index of the trained fault diagnosis model in the previous round, the marginal performance gain is determined. Based on the marginal performance gain, the number of qualified samples of the target fault category, the screening pass rate of the candidate sample set, or the recall rate of the target fault category, determine whether to proceed to the next iteration to generate samples.

[0010] In one possible implementation, determining the generation quota of the target fault category in the next iteration based on the recall rate, confusion matrix, and generation quota of the target fault category in the current iteration, as well as the preset minimum generation quantity and preset maximum generation quantity of the target fault category, includes: Based on the confusion matrix, the confusion intensity of the target fault category is determined; Based on the confusion intensity of the target fault category, the recall rate of the target fault category, and the generation quota of the target fault category in this iteration, the initial generation quota of the target fault category in the next iteration is obtained. Based on the preset minimum and maximum generation quantities of the target fault category, the initial generation quota of the target fault category in the next iteration is adjusted to obtain the generation quota of the target fault category in the next iteration. If the recall rate of the target fault category is less than a preset recall rate threshold and the confusion intensity of the target fault category is greater than a preset intensity threshold, the filtering threshold in the preset filtering rule is updated.

[0011] In one possible implementation, determining whether to proceed to the next iteration to generate samples based on the marginal performance gain, the number of qualified samples of the target fault category, the screening pass rate of the candidate sample set, or the recall rate of the target fault category includes: If the marginal performance gain is less than a preset gain threshold, or the marginal performance gain is negative for multiple consecutive rounds, or the number of qualified samples of the target fault category is greater than or equal to a preset quantity threshold, or the screening pass rate of the candidate sample set is less than a pass threshold, or the recall rate of the target fault category is equal to or greater than a recall threshold, then the next iteration stops; otherwise, the next iteration continues.

[0012] In one possible implementation, the overall performance metric is one or more combinations of prediction accuracy, F1 score, or macro average recall.

[0013] Secondly, embodiments of the present invention provide a rolling bearing fault diagnosis device, comprising: A construction module is used to build a small sample training set and a validation set for the rolling bearing to be diagnosed; both the small sample training set and the validation set include real samples of multiple fault categories. The training module is used to train the pre-built classification generative adversarial network and the pre-built fault diagnosis model based on the small sample training set, so as to obtain the trained classification generative adversarial network and the initial fault diagnosis model. An iterative module is used to generate a candidate sample set corresponding to the target fault category in the target round iteration based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network; to filter the candidate target fault category samples in the candidate sample set according to a preset screening rule, and to incorporate the qualified samples obtained after screening into the small sample training set to form an expanded training set; to train the initial fault diagnosis model based on the expanded training set to obtain a trained fault diagnosis model; to evaluate the trained fault diagnosis model based on the validation set, and to update the generation quota and preset screening rule of the target fault category in the next round iteration according to the evaluation result, and to determine whether to proceed to the next round iteration to generate samples; The diagnostic module is used to diagnose rolling bearing faults based on the training fault diagnosis model obtained in the previous round of training if the next round of iteration is stopped.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0015] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.

[0016] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect or any possible implementation thereof.

[0017] To reduce the impact of low-quality, class-confusing, or redundant generated samples on the fault diagnosis model, this embodiment of the invention filters candidate target fault category samples in the candidate sample set based on preset screening rules. Samples that do not meet the preset screening rules are removed, and only the qualified samples obtained after screening are incorporated into the small sample training set to form an expanded training set. This reduces low-quality, class-confusing, or redundant generated samples and improves the quality of the final expanded training set. Furthermore, by evaluating the trained fault diagnosis model based on the validation set, the generation quota and preset screening rules for the target fault category in the next iteration can be updated according to the evaluation results, and it can be determined whether to proceed to the next iteration to generate samples. Thus, sample generation is no longer a one-time, fixed-quantity expansion, but forms a complete closed-loop control process, thereby ensuring the diagnostic accuracy of the obtained trained fault diagnosis model. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the implementation of the rolling bearing fault diagnosis method provided in this embodiment of the invention. Figure 2 This is a comparison chart of the spectra of real samples and generated samples of moderately correlated faults provided in the embodiments of the present invention; Figure 3 This is a comparison chart of the spectra of real samples and generated samples of severe compound faults provided in the embodiments of the present invention; Figure 4 This is a comparison chart of the spectra of normal real samples and generated samples provided in the embodiments of the present invention; Figure 5 This is a visual comparison chart of the output features of 1DCNN after training with different generative models, provided in the embodiments of the present invention. Figure 6 This is a schematic diagram of the structure of the rolling bearing fault diagnosis device provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0020] As introduced in the background section, Generative Adversarial Networks (GANs) and their variants provide data augmentation strategies for small-sample fault diagnosis. Conditional GANs can generate labeled samples using class labels, and auxiliary classification GANs can further improve the class constraint capability of generated samples. However, in some GAN-based small-sample bearing fault diagnosis, generated samples are usually directly incorporated into the training set or screened only based on the true / false judgment results. The inventors have found that for bearing data with similar categories, similar damage degrees, or complex faults, if the generated samples deviate from the true fault pattern, are close to easily confused categories, or highly overlap with existing samples, the effectiveness of expanding the training set may be reduced.

[0021] To address the problems in the above-mentioned rolling bearing fault diagnosis, such as difficulty in ensuring the quality of generated samples, insufficient physical consistency between generated samples and real vibration signals, the tendency of fixed generation quantity to cause sample redundancy or contamination, and the low diagnostic accuracy of the final trained fault diagnosis model, this invention provides a rolling bearing fault diagnosis method, device, and storage medium.

[0022] See Figure 1 The flowchart illustrating the implementation of the rolling bearing fault diagnosis method provided in this embodiment of the invention is shown below: S110. Construct a small sample training set and validation set for the rolling bearing to be diagnosed.

[0023] The small sample training set and validation set both include real samples of various fault categories.

[0024] In some embodiments, the vibration signal of the rolling bearing to be diagnosed can be acquired, the vibration signal can be preprocessed and sampled using a sliding window to construct an original sample set with fault category labels, and samples can be selected from each fault category in the original sample set to construct a small sample training set, with the remaining samples used as a validation set or test set.

[0025] In this embodiment, the vibration signals of the rolling bearing to be diagnosed under one or more operating conditions are first acquired, and all acquired vibration signals undergo one or more preprocessing steps, including detrending, normalization, filtering, or envelope demodulation. Then, the vibration signals are sampled using a sliding window according to a preset window length to obtain time-series samples of uniform length. Category labels are assigned to normal states, single fault states, complex fault states, or states with different fault degrees. To simulate situations where there are insufficient fault samples in the equipment's operating environment, a small number of real samples can be selected from each fault category to construct a small training set, with the remaining samples used for validation or test sets.

[0026] S120. Train the pre-built classification generative adversarial network and the pre-built fault diagnosis model based on a small sample training set to obtain the trained classification generative adversarial network and the initial fault diagnosis model.

[0027] In some embodiments, the classification generative adversarial network includes a generator, a discriminator, and an independent classifier. The generator generates candidate fault samples, the discriminator outputs a ground truth score, and the independent classifier outputs a class confidence score and intermediate layer feature vectors.

[0028] In this embodiment, the classification generative adversarial network can use MDACGAN, CGAN, WGAN-GP, diffusion generative model, variational autoencoder generative model, or other models that can generate candidate fault samples based on category labels as candidate sample sources.

[0029] For example, the classification generative adversarial network in this application is a Multi-Discriminator Auxiliary Classifier Generative Adversarial Network (MDACGAN). MDACGAN includes a generator, a discriminator, and an independent classifier. The discriminator focuses on true / false classification and outputs a truth score. The independent classifier focuses on fault category identification and outputs a category confidence score and intermediate layer feature vectors, thereby avoiding error propagation caused by the coupling of true / false classification and category identification tasks in traditional ACGANs.

[0030] In this embodiment, the generator's input consists of 100-dimensional random noise and 9-dimensional one-hot encoded class labels. The input is mapped to a high-dimensional feature space via a fully connected layer, reshaped into a feature map, and then input into a multi-branch parallel sampling structure. This structure includes three convolutional branches with kernel sizes of 3, 5, and 7. Each branch includes a convolutional layer, a batch normalization layer, and a LeakyReLU activation function. The feature maps output from the three branches are concatenated along the channel dimension, then subjected to 1×1 convolution for dimensionality reduction, followed by 3×1 convolution and Tanh activation to output a candidate sample set of length 1024. The discriminator outputs a ground truth score (SD), the independent classifier outputs a class confidence score (SC), and provides an intermediate layer feature vector F(x).

[0031] In some embodiments, the receptive field of the multi-branch parallel sampling structure can be determined based on the sampling frequency, the switching frequency, and the theoretical fault characteristic frequency.

[0032] In this embodiment, the point interval of adjacent fault impacts in the sampling sequence can be estimated based on the sampling frequency and the characteristic frequency of the rolling bearing fault to be diagnosed. The kernel size, dilation rate, or multi-layer convolution combination is then selected to ensure that the effective receptive field of at least one branch covers this point interval, and that at least one branch covers the neighborhood range of this point interval. Thus, the generator can simultaneously express local impact morphology and cross-cycle repetitive features.

[0033] In some embodiments, the generator's loss function includes adversarial loss, class consistency loss, feature matching loss, first-order smoothing loss, and second-order smoothing loss, satisfying: ; in, Indicating resistance to loss, This represents the loss of class consistency. Represents feature matching loss, This represents the first-order smoothing loss. This represents the second-order smoothing loss. , , , These are the loss weights, which can be set to 1.0, 10.0, 0.1, and 0.05 respectively.

[0034] In this embodiment, feature matching loss for: ; First-order smoothing loss and second-order smoothing loss They are respectively: ; ; in, This represents the feature map of the l-th layer of the discriminator. Indicates random noise The candidate target fault category samples generated from the target fault category y denoted as the amplitude of the candidate target fault category sample at time t, and N represents the sample length.

[0035] Combat losses and category consistency loss These are all standard expressions, and will not be elaborated further here.

[0036] During the training of the classification generative adversarial network, the Adam optimizer can be used to alternately update the discriminator, independent classifier and generator. When the preset number of training rounds is reached, the loss change is less than the threshold or the confidence pass rate of the generated samples on the validation set is stable, the training of the classification generative adversarial network is stopped.

[0037] S130. In the target round iteration, based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network, a candidate sample set corresponding to the target fault category is generated; according to the preset screening rules, the candidate target fault category samples in the candidate sample set are screened, and the qualified samples obtained after screening are merged into the small sample training set to form an expanded training set; the initial fault diagnosis model is trained based on the expanded training set to obtain the trained fault diagnosis model; the trained fault diagnosis model is evaluated based on the validation set, and the generation quota and preset screening rules of the target fault category in the next round iteration are updated according to the evaluation results, and it is determined whether to proceed to the next round iteration to generate samples.

[0038] In some embodiments, during the target round iteration, the authenticity score, class confidence score, and intermediate layer feature vector of each candidate target fault category sample can be generated based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network.

[0039] In this embodiment, the target round iteration refers to the process of generating candidate sample sets in any round. A candidate sample set is generated for each round of iteration. In each round, the candidate sample sets generated for which target fault categories are determined based on the generation quota for that round. The target fault category is any one of the various fault categories in the small sample training set.

[0040] For example, if the generation quota for fault samples of type y is 10 in this iteration, then 10 candidate fault samples of type y will be generated. If the generation quota for fault samples of type y is 0, then no candidate fault samples of type y will be generated in this iteration.

[0041] In this embodiment, after obtaining the authenticity score, class confidence score, and intermediate layer feature vector of each candidate target fault category sample, the comprehensive credibility score of the candidate target fault category sample can be determined based on the authenticity score, class confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity. If the comprehensive credibility score, authenticity score, class confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity of the candidate target fault category sample all meet the preset screening rules, then the candidate target fault category sample is determined to be a qualified sample, and this qualified sample is incorporated into the small sample training set to form an expanded training set. Otherwise, the candidate target fault category sample is removed.

[0042] For example, for the y-th type of candidate fault sample In this application, instead of directly adding them to the small sample training set, a reliable screening process is required. This involves generating candidate fault samples of class y. At the same time, the discriminator will output candidate fault samples of type y. Authenticity rating The independent classifier will output candidate fault samples of class y. Category confidence score and intermediate feature vectors .

[0043] In this embodiment, the intra-class feature distance can be determined based on the intermediate layer feature vector of the candidate target fault category sample and the feature center of the real sample of the target fault category.

[0044] For example, let's take the candidate fault sample of type y above. For example, intra-class feature distance It can be represented as: ; ; in, Let be the number of true samples in class y. Let y be the set of real samples of class y, where y is the label of the target fault category.

[0045] In this embodiment, the inter-class boundary distance can be determined based on the intermediate layer feature vectors of candidate target fault category samples and the corresponding easily confused target fault category feature centers. The easily confused target fault categories are determined from the confusion matrix obtained by testing the trained fault diagnosis model in the previous iteration of the target round using the validation set.

[0046] For example, let's take the candidate fault sample of type y above. For example, inter-class boundary distance It can be represented as: ; Specifically, the set of easily confused target fault categories can be determined based on the confusion matrix obtained by testing the trained fault diagnosis model in the previous round of the target round using the validation set. , Let be the feature center of the k-th easily confused target fault category. The inter-class boundary distance is used to ensure that candidate target fault category samples are not only close to the target fault category, but also far away from the easily confused category.

[0047] In this embodiment, spectral or envelope spectrum analysis can be performed on candidate target fault category samples and real samples of target fault categories to obtain a frequency domain consistency score.

[0048] When performing frequency domain consistency scoring, short-time Fourier transform, wavelet transform, envelope spectrum, order analysis, or frequency band energy statistics can also be used.

[0049] For example, let's take the candidate fault sample of type y above. For example, the candidate target fault category samples are... Perform spectral or envelope spectrum analysis on real samples of the target fault category to calculate the position deviation of the main frequency peak. Frequency band energy distribution distance Spectral correlation coefficient Frequency domain consistency score is obtained. .

[0050] ; in, For the y-th type of candidate fault sample The spectrum or envelope spectrum, For the first Similar to the frequency domain of real samples, , , This represents the frequency domain scoring weight.

[0051] It should be noted that the position deviation of the main frequency peak Frequency band energy distribution distance Spectral correlation coefficient After normalizing the obtained frequency peak position deviation, main frequency peak position deviation and frequency band energy distribution distance, the resulting values ​​are all dimensionless values ​​in the range of 0-1.

[0052] For example, the range normalization method can be used to pre-set the maximum and minimum values ​​of the main frequency peak position deviation, the maximum and minimum values ​​of the band energy distribution distance, and the maximum and minimum values ​​of the spectral correlation coefficient. This allows for the normalization of the actually calculated frequency peak position deviation, main frequency peak position deviation, and band energy distribution distance, resulting in a dimensionless main frequency peak position deviation. Frequency band energy distribution distance Spectral correlation coefficient .

[0053] For example, , , They can be set to 1.0, 0.5, and 0.5 respectively.

[0054] In this embodiment, the maximum similarity can be determined based on the candidate target fault category samples and the qualified samples of the target fault category obtained after screening.

[0055] For example, let's take the candidate fault sample of type y above. For example, to calculate the sample of candidate target fault categories Qualified samples of the target fault category obtained after screening Maximum similarity The calculation method is as follows: ; like If empty, then Record it as 0.

[0056] After identifying the y-th type of candidate fault samples Authenticity rating Category confidence score Intra-class feature distance Inter-class boundary distance Frequency domain consistency score and maximum similarity Then, a weighted summation can be performed to determine the candidate fault samples of type y. Overall credibility score .

[0057] Among them, the comprehensive credibility score It can be defined as: .

[0058] in, , , , , and These are the weights for authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity, respectively.

[0059] For example, , , , , and The values ​​can be set to 0.20, 0.20, 0.15, 0.15, 0.20, and 0.10 respectively.

[0060] After identifying the y-th type of candidate fault samples Authenticity rating Category confidence score Intra-class feature distance Inter-class boundary distance Frequency domain consistency score Maximum similarity and comprehensive credibility score Then, based on the preset screening rules, the candidate fault samples of type y can be determined. Is it a qualified sample?

[0061] It should be noted that intra-class feature distance Inter-class boundary distance After normalizing the obtained intra-class feature distances and inter-class boundary distances, the resulting values ​​are dimensionless values ​​in the range of 0-1.

[0062] For example, the range normalization method can be used. By pre-setting the maximum and minimum values ​​of the intra-class feature distance, and the maximum and minimum values ​​of the inter-class boundary distance, the calculated intra-class feature distance and inter-class boundary distance can be normalized to obtain dimensionless intra-class feature distances. and inter-class boundary distance .

[0063] In this embodiment, the preset screening rules can be: the comprehensive credibility score is greater than the comprehensive credibility threshold, the authenticity score is greater than or equal to the authenticity score threshold, the category confidence score is greater than or equal to the category confidence threshold, the intra-class feature distance is less than or equal to the intra-class distance threshold, the inter-class boundary distance is greater than or equal to the inter-class distance threshold, the frequency domain consistency score is greater than or equal to the frequency domain score threshold, and the maximum similarity is less than or equal to the similarity threshold. Only when all the above conditions are met can the candidate target fault category sample be determined as a qualified sample; otherwise, the sample is removed.

[0064] Using the candidate fault sample of type y above as an example For example, when Greater than the comprehensive confidence threshold, and , , , , , At that time, the candidate fault samples of type y will be... The sample is deemed acceptable; otherwise, it is discarded.

[0065] In some embodiments, the thresholds in the preset filtering rules are all adaptively determined based on their categories.

[0066] In this embodiment, the true scoring threshold The category confidence threshold can be determined based on the distribution of real-world scores for the target fault category in the discriminator. The intra-class distance threshold can be determined based on the confidence distribution of the independent classifier on the validation set for true samples of the target fault category. The inter-class distance threshold can be determined based on the mean and standard deviation of the distances from real samples of the target fault category to their feature centers. The frequency domain scoring threshold can be determined based on the inter-class feature distance distribution between real samples of the target fault category and real samples of easily confused target fault categories. The similarity threshold can be dynamically adjusted based on the number of qualified samples and feature coverage of the target fault category obtained after screening. The thresholds can be determined based on the spectral distance distribution of real samples in the frequency domain of the target fault category. By using these thresholds, it is possible to simultaneously control authenticity, class consistency, intra-class compactness, inter-class separation, frequency domain physical consistency, and sample diversity, thereby improving the quality of the qualified samples obtained through screening.

[0067] It should be noted that: the true rating threshold Category confidence threshold Intra-class distance threshold Inter-class distance threshold Frequency domain scoring threshold Similarity threshold The overall credibility threshold is not a fixed value set in advance, but is adaptively determined based on the empirical distribution of evaluation indicators corresponding to real samples of each fault category. Among them, the authenticity scoring threshold... Category confidence threshold Inter-class boundary distance threshold and frequency domain scoring threshold Take the 10th percentile of the corresponding real sample indicator, and the intra-class feature distance threshold. Take the 90th percentile; similarity threshold It can be initially set to 0.98.

[0068] By extending the quality control of generated samples from a single true / false judgment to a joint reliable screening of multiple indicators, the generated samples can simultaneously meet the requirements in terms of authenticity, class consistency, intra-class distribution consistency, inter-class boundary clarity, frequency domain physical consistency, and diversity, thereby reducing the risk of false and redundant samples contaminating the training set from the source.

[0069] After selecting qualified samples, all qualified samples can be incorporated into the small sample training set to form an expanded training set.

[0070] In some embodiments, after obtaining the expanded training set, the initial fault diagnosis model can be trained using the expanded training set. The initial fault diagnosis model can employ a deep convolutional neural network (DCNN), a convolutional neural network (CNN), a temporal convolutional network (TCN), a Transformer, or a combination of these networks.

[0071] In this embodiment, the initial fault diagnosis model can be a 1DCNN, including an input layer, convolutional layer, pooling layer, dropout layer, fully connected layer, and output layer. The input is a vibration signal time series of length 1024, and the output is the probability distribution of 9 fault categories.

[0072] In some embodiments, after the initial fault diagnosis model is trained using an expanded training set in each round and the trained fault diagnosis model is obtained, it is also necessary to evaluate the trained fault diagnosis model using a validation set. This can not only update the generation quota and preset screening rules of the target fault category in the next iteration based on the evaluation results, but also determine whether to proceed to the next iteration to generate samples.

[0073] In this embodiment, firstly, based on the evaluation results of the trained fault diagnosis model using the validation set, the recall rate, confusion matrix, and comprehensive performance index of the trained fault diagnosis model for the target fault category are obtained. Then, based on the recall rate, confusion matrix, and generation quota of the target fault category in this iteration, as well as the preset minimum and maximum generation quantities of the target fault category, the generation quota and preset screening rules for the target fault category in the next iteration are determined. Next, based on the difference between the comprehensive performance index of the trained fault diagnosis model and the comprehensive performance index of the trained fault diagnosis model in the previous round, the marginal performance gain is determined. Finally, based on the marginal performance gain, the number of qualified samples of the target fault category, the screening pass rate of the candidate sample set, or the recall rate of the target fault category, it is determined whether to proceed to the next iteration to generate samples.

[0074] The overall performance metric can be one or more combinations of prediction accuracy, F1 score, or macro average recall.

[0075] In this embodiment, firstly, the confusion intensity of the target fault category is determined based on the obtained confusion matrix. Next, based on the confusion intensity, recall rate, and generation quota of the target fault category in the current iteration, the initial generation quota for the target fault category in the next iteration is obtained. Then, based on the preset minimum and maximum generation quantities of the target fault category, the initial generation quota for the target fault category in the next iteration is adjusted to obtain the generation quota for the target fault category in the next iteration. Finally, if the recall rate of the target fault category is less than a preset recall rate threshold and the confusion intensity is greater than a preset intensity threshold, the filtering threshold in the preset filtering rules is updated.

[0076] The confusion intensity of the target fault category can be the sum of the proportions in which the target fault category is misclassified as other fault categories, or it can be the bidirectional confusion ratio between the target fault category and one or more easily confused categories.

[0077] For example, the y-th type of fault in Quota generation in round iteration It can be represented as: ; in, For the first Round The generation quota for fault types, , For adjustment coefficients, Let be the recall rate for the y-th type of fault. Let represent the confusion intensity of the y-th type of fault.

[0078] For example, and They can be 30 and 20 respectively.

[0079] In obtaining Then, it needs to be rounded down and limited to a preset minimum number of N. min With the preset maximum number of generators N max between.

[0080] When the recall rate of type y faults is lower than a preset recall threshold and the confusion intensity is greater than a preset intensity threshold, the generation quota for this type of fault in the next iteration can be increased, along with the class confidence threshold, inter-class distance threshold, or frequency domain scoring threshold, to reduce the number of boundary error samples entering the expanded training set. When the recall rate of type y faults reaches a preset requirement and the candidate sample pass rate is high, the generation quota for this type can be reduced or the diversity constraint can be increased to prevent the continued accumulation of duplicate samples. The threshold adjustment range can be determined by the recall gap, confusion intensity, and marginal performance gain of the previous round, and is limited to a preset upper and lower bound. Through this adjustment method, targeted enhancement can be performed on easily confused categories such as moderate and severe faults, single faults and compound faults, instead of simply generating the same number of samples for all fault categories.

[0081] In this embodiment, the marginal performance gain , can be represented as: ; in, For accuracy, Score, macro average recall, or a combination thereof.

[0082] If the marginal performance benefit is less than a preset benefit threshold, or the marginal performance benefit is negative for multiple consecutive rounds, or the number of qualified samples for the target fault category is greater than or equal to a preset quantity threshold, or the screening pass rate of the candidate sample set is less than a pass threshold, or the recall rate of the target fault category is equal to or greater than a recall threshold, the next iteration will stop if any of the above conditions are met. Otherwise, the process will continue to the next iteration, generating a candidate sample set and repeating the closed-loop process of screening, expansion, diagnosis, and feedback.

[0083] The process of generating candidate sample sets through each iteration establishes a mutually constraining relationship between reliable screening, diagnostic training, and feedback control. Generating candidate sample sets provides expandable samples. Reliable screening eliminates samples with insufficient realism, inconsistent categories, deviations from the feature center of their class, proximity to easily confused categories, inconsistent frequency domain features, or excessive repetition before expanding the sample set, thus ensuring the reliability of samples entering the expanded set. The evaluation results of diagnostic training determine the fault categories that need to be prioritized for enhancement in the next iteration based on recall and confusion intensity. Furthermore, the expansion scale can be controlled based on performance marginal benefits and other indicators, addressing the difficulty in determining the number of generated samples. Therefore, this approach ensures the generation of high-quality samples while increasing the effective training sample size for small-sample fault diagnosis models and reducing the risk of distribution shift caused by excessive or erroneous generation.

[0084] In this embodiment of the invention, by filtering candidate target fault category samples in the candidate sample set based on preset screening rules, samples that do not meet the preset screening rules are eliminated. Only the qualified samples obtained after screening are incorporated into the small sample training set to form an expanded training set. This reduces the generation of low-quality, category-confusing, or redundant samples, thereby improving the quality of the final expanded training set. Furthermore, by evaluating the trained fault diagnosis model based on the validation set, the generation quota and preset screening rules for the target fault category in the next iteration can be updated according to the evaluation results. It also determines whether to proceed with the next iteration to generate samples. Thus, sample generation is no longer a one-time, fixed-quantity expansion, but forms a complete closed-loop control process, ensuring the diagnostic accuracy of the obtained trained fault diagnosis model. By updating the generation quota and preset screening rules for the target fault category in the next iteration based on the evaluation results, the output scale of candidate samples and the sample entry quality threshold can be adjusted in a coordinated manner, solving the technical problem that a single adjustment cannot simultaneously consider both sample quantity and sample quality.

[0085] To facilitate understanding of this invention, specific experimental examples are given below using the HUST rolling bearing dataset.

[0086] In this embodiment, the rolling bearing vibration data used was obtained under a motor speed of 40Hz. The sampling frequency was 25.6kHz, the acquisition time for each fault at each speed was 10.2 seconds, and the number of sampling points for each original record was 262,144. After normalizing the vibration signal in the x-direction, samples were constructed using a sliding window of length 1024. The nine bearing categories and sample composition are shown in Table 1.

[0087] Table 1 HUST Rolling Bearing Failure Types and Sample Composition The specific fault diagnosis process is as follows: S210. Construct an original sample set based on the fault categories in Table 1, and randomly select 15 real samples from each fault category to construct a small sample training set, and randomly select 100 real samples from each fault category to construct a diagnostic set.

[0088] S220. MDACGAN is trained using 100-dimensional random noise and 9-dimensional class-specific one-hot encoding as inputs to a classification generative adversarial network. The generator employs a three-branch parallel sampling structure with convolutional kernel sizes of 3, 5, and 7; the discriminator outputs a realism score; and the independent classifier outputs class confidence and extracts intermediate layer feature vectors.

[0089] S230. Using the trained MDACGAN, generate a candidate sample set for each fault category, and determine the authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity for each candidate sample. Candidate samples that do not meet the comprehensive confidence score and individual thresholds are discarded. Candidate samples that meet the preset screening rules are determined as qualified samples.

[0090] S240. Merge the qualified samples with 15 real small samples to construct expanded training sets with 0, 30, 45, 60, 75, 90, and 105 samples per class, respectively, and train a 1DCNN fault diagnosis model under the same training parameters. Each experiment is repeated 10 times, and the average value is taken as the diagnosis result.

[0091] S250. After each round of augmentation training, evaluate the augmentation effect based on accuracy, precision, recall, and F1 score on the validation set. When the number of generated samples continues to increase while the marginal performance gain decreases or becomes negative, a stop decision is triggered, and generation is stopped.

[0092] To verify the consistency between the generated samples and the real vibration signals, a Fast Fourier Transform was performed on both samples, and their spectral characteristics were compared. Figure 2-4 As shown, the main frequency distribution and spectral morphology of the samples generated by MDACGAN are in good agreement with those of the real samples, indicating that it can learn the frequency domain characteristics of real rolling bearing fault signals.

[0093] In addition, KLD and PCC were used to quantitatively evaluate the quality of the generated samples. A smaller KLD indicates a smaller distribution difference, while a larger PCC indicates a stronger linear correlation. Compared with the traditional ACGAN, MDACGAN significantly reduced KLD for all nine types of faults, and the PCC further approached 1, as shown in Table 2 below.

[0094] Table 2 Comparison of similarity indices between generated samples and real samples Furthermore, the diagnostic results under different numbers of augmented samples are shown in Table 3. As can be seen from Table 3, when there are only 15 real samples per class and no generated samples are added, the 1DCNN diagnostic accuracy is 82.78%. When 30 generated samples are added to each class, the accuracy increases to 86.33%. When 60 generated samples are added to each class, the accuracy increases to 90.33%. When 90 generated samples are added to each class, the accuracy, recall, and F1 score reach 92.78%, 92.78%, and 92.64%, respectively. Further increasing to 105 generated samples per class results in a decrease in accuracy to 92.43%. This result indicates that there is an effective augmentation range for generated samples. This application controls the augmentation scale through a performance marginal benefit stopping condition, which can reduce the imbalance in training data distribution caused by excessive generation.

[0095] Table 3 Diagnostic results with different numbers of expanded samples To further illustrate the performance of MDACGAN compared to other generative models, MDACGAN is compared with WGAN-GP, ACGAN, and CGAN. Each generative model generates samples based on 15 real small samples per class, expands the training samples for each class to 105, and then trains a 1DCNN with the same structure. The results are shown in Table 4.

[0096] Table 4. Diagnostic results of 1DCNN trained with different datasets As shown in Table 4, the accuracy of the 1DCNN trained with augmented samples by MDACGAN reached 92.78%, only 1.11 percentage points lower than the 93.89% achieved with sufficient real samples. Compared to WGAN-GP, ACGAN, and CGAN with augmented samples, the accuracy improved by 4.00, 5.67, and 7.67 percentage points, respectively, indicating that MDACGAN can provide more suitable augmented samples for fault diagnosis model learning.

[0097] Furthermore, the 1DCNN output features were visualized using t-SNE dimensionality reduction, and the results are as follows: Figure 5 As shown, compared with CGAN, ACGAN, and WGAN-GP, the test sample features clustered more compactly after MDACGAN-assisted training, and the boundaries between different fault categories are clearer, indicating that high-confidence generated samples can improve the feature extraction and category separation capabilities of diagnostic models.

[0098] To verify the contributions of multi-branch parallel sampling, independent classifiers, and improved loss functions to diagnostic performance, MDACGAN was ablation validated, and the results are shown in Table 5. Table 5 shows that after introducing multi-branch parallel sampling on top of the traditional ACGAN, the accuracy improved from 87.11% to 89.67%; further introducing an independent classifier improved the accuracy to 92.11%; and adding feature matching loss and smoothing loss further improved the accuracy to 92.78%. Therefore, multi-branch parallel sampling, independent classifiers, and improved loss functions collectively improved the quality of generated samples and provided clear support for the accuracy of fault diagnosis in small samples.

[0099] Table 5 Ablation Experiment Results of the Improved Generative Adversarial Network Depend on Figure 2 and Figure 3 As shown in Tables 1 to 5, under the condition of only 15 real samples per class, this invention can generate candidate fault samples with consistent spectra and high statistical similarity to real samples using MDACGAN. Furthermore, by controlling the quality and quantity of generated samples through reliable screening and diagnostic feedback closed-loop control, the diagnostic accuracy is improved from 82.78% to 92.78%, while providing a stopping criterion when performance degrades due to a continued increase in generated samples.

[0100] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0101] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0102] Figure 6 A schematic diagram of the rolling bearing fault diagnosis device provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 6 As shown, the rolling bearing fault diagnosis device includes: Module 610 is used to construct a small-sample training set and a validation set for the rolling bearing to be diagnosed; both the small-sample training set and the validation set include real samples of multiple fault categories. Training module 620 is used to train a pre-built classification generative adversarial network and a pre-built fault diagnosis model based on a small sample training set, so as to obtain a trained classification generative adversarial network and an initial fault diagnosis model. The iteration module 630 is used to generate a candidate sample set corresponding to the target fault category based on the generation quota corresponding to the target fault category in ... The diagnostic module 640 is used to perform fault diagnosis on the rolling bearing based on the training fault diagnosis model obtained in the previous round of training if the next round of iteration is stopped.

[0103] In one possible implementation, the classification generative adversarial network includes a generator, a discriminator, and an independent classifier. The generator is used to generate candidate fault samples, the discriminator is used to output a truth score, and the independent classifier is used to output a class confidence score and an intermediate layer feature vector. The iteration module 630 is used to generate, in the target round iteration, the authenticity score of each candidate target fault category sample, the class confidence score of each candidate target fault category sample, and the intermediate layer feature vector of each candidate target fault category sample based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network.

[0104] In one possible implementation, the iterative module 630 is used to determine the comprehensive credibility score of the candidate target fault category samples based on the authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity. If the comprehensive credibility score, authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity of the candidate target fault category samples all meet the preset screening rules, then the candidate target fault category samples are determined to be qualified samples, and the qualified samples are incorporated into the small sample training set to form an expanded training set; otherwise, the candidate target fault category samples are removed.

[0105] In one possible implementation, the iterative module 630 is used to determine the intra-class feature distance based on the intermediate layer feature vector of the candidate target fault category samples and the feature center of the real sample of the target fault category; Based on the intermediate layer feature vectors of candidate target fault category samples and the corresponding easily confused target fault category feature centers, the inter-class boundary distance is determined; where the easily confused target fault category is determined from the confusion matrix obtained by testing the training fault diagnosis model in the previous round of the target round iteration on the validation set; Spectral or envelope spectrum analysis is performed on candidate target fault category samples and real samples of target fault categories to obtain frequency domain consistency scores; The maximum similarity is determined based on the candidate target fault category samples and the qualified samples of the target fault categories obtained after screening. The comprehensive credibility score of candidate target fault category samples is determined by a weighted sum of the authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity.

[0106] In one possible implementation, the iteration module 630 is used to evaluate the trained fault diagnosis model based on the validation set to obtain the recall rate, confusion matrix and comprehensive performance index of the trained fault diagnosis model for the target fault category. Based on the recall rate, confusion matrix, and generation quota of the target fault category in this iteration, as well as the preset minimum and maximum generation quantities of the target fault category, determine the generation quota and preset filtering rules of the target fault category in the next iteration. The marginal performance gain is determined by the difference between the comprehensive performance index of the trained fault diagnosis model and the comprehensive performance index of the trained fault diagnosis model in the previous round. Based on the marginal performance gain, the number of qualified samples for the target fault category, the screening pass rate of the candidate sample set, or the recall rate of the target fault category, determine whether to proceed to the next iteration to generate samples.

[0107] In one possible implementation, the iterative module 630 is used to determine the confusion intensity of the target fault category based on the confusion matrix; Based on the confusion intensity of the target fault category, the recall rate of the target fault category, and the generation quota of the target fault category in this iteration, the initial generation quota of the target fault category in the next iteration is obtained. Based on the preset minimum and maximum number of generation for the target fault category, the initial generation quota for the target fault category in the next iteration is adjusted to obtain the generation quota for the target fault category in the next iteration. If the recall rate of the target fault category is less than the preset recall rate threshold and the confusion intensity of the target fault category is greater than the preset intensity threshold, update the filtering threshold in the preset filtering rules.

[0108] In one possible implementation, the iteration module 630 is used to stop the next round of iteration if the marginal performance benefit is less than a preset benefit threshold, or the marginal performance benefit is negative for multiple consecutive rounds, or the number of qualified samples of the target fault category is greater than or equal to a preset quantity threshold, or the screening pass rate of the candidate sample set is less than the pass threshold, or the recall rate of the target fault category is equal to or greater than the recall threshold; otherwise, the next round of iteration continues.

[0109] In one possible implementation, the overall performance metric is one or more combinations of prediction accuracy, F1 score, or macro average recall.

[0110] Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Figure 7 As shown, the electronic device 7 of this embodiment includes a processor 70 and a memory 71. The memory 71 stores a computer program 72. When the processor 70 executes the computer program 72, it implements the steps in the various method embodiments described above. Alternatively, when the processor 70 executes the computer program 72, it implements the functions of each module / unit in the various device embodiments described above.

[0111] For example, computer program 72 may be divided into one or more modules / units, which are stored in memory 71 and executed by processor 70 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 72 in electronic device 7.

[0112] Electronic device 7 may include, but is not limited to, processor 70 and memory 71. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 7 and does not constitute a limitation on electronic device 7. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 7 may also include input / output devices, network access devices, buses, etc.

[0113] The processor 70 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0114] The memory 71 can be an internal storage unit of the electronic device 7, such as a hard disk or RAM. The memory 71 can also be an external storage device of the electronic device 7, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 71 can include both internal and external storage units of the electronic device 7. The memory 71 is used to store the computer program 72 and other programs and data required by the electronic device 7. The memory 71 can also be used to temporarily store data that has been output or will be output.

[0115] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0116] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0117] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0118] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0119] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0120] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for diagnosing rolling bearing faults, characterized in that, include: Construct a small-sample training set and a validation set for the rolling bearing to be diagnosed; both the small-sample training set and the validation set include real samples of multiple fault categories. The pre-built classification generative adversarial network and the pre-built fault diagnosis model are trained based on the small sample training set to obtain the trained classification generative adversarial network and the initial fault diagnosis model. In the target round iteration, a candidate sample set corresponding to the target fault category is generated based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network. According to the preset screening rules, the candidate target fault category samples in the candidate sample set are screened, and the qualified samples obtained after screening are incorporated into the small sample training set to form an expanded training set. The initial fault diagnosis model is trained based on the expanded training set to obtain a trained fault diagnosis model; The trained fault diagnosis model is evaluated based on the validation set. The generation quota and preset screening rules of the target fault category in the next iteration are updated according to the evaluation results, and it is determined whether to proceed to the next iteration to generate samples. If the next iteration is stopped, the rolling bearing is then diagnosed based on the training fault diagnosis model obtained from the previous iteration.

2. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The classification generative adversarial network includes a generator, a discriminator, and an independent classifier. The generator is used to generate candidate fault samples, the discriminator is used to output a realism score, and the independent classifier is used to output a class confidence score and an intermediate layer feature vector. In the target round iteration, generating a candidate sample set corresponding to the target fault category based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network, further includes: In the target round iteration, based on the generation quota corresponding to the target fault category in the target round iteration and the trained classification generative adversarial network, the authenticity score of each candidate target fault category sample, the category confidence score of each candidate target fault category sample, and the intermediate layer feature vector of each candidate target fault category sample are generated.

3. The rolling bearing fault diagnosis method according to claim 2, characterized in that, The step of filtering candidate target fault category samples in the candidate sample set according to preset filtering rules, and merging the qualified samples obtained after filtering into the small sample training set to form an expanded training set, includes: The comprehensive credibility score of the candidate target fault category samples is determined based on the authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity. If the comprehensive credibility score, authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity of the candidate target fault category sample all meet the preset screening rules, then the candidate target fault category sample is determined to be a qualified sample, and the qualified sample is incorporated into the small sample training set to form an expanded training set; otherwise, the candidate target fault category sample is removed.

4. The rolling bearing fault diagnosis method according to claim 3, characterized in that, Based on the authenticity score, class confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity of candidate target fault category samples, a comprehensive credibility score for candidate target fault category samples is determined, including: Based on the intermediate layer feature vectors of the candidate target fault category samples and the feature centers of the true samples of the target fault category, the intra-class feature distance is determined; Based on the intermediate layer feature vectors of the candidate target fault category samples and the corresponding easily confused target fault category feature centers, the inter-class boundary distance is determined; wherein, the easily confused target fault category is determined from the confusion matrix obtained by testing the training fault diagnosis model in the previous round of the target round iteration based on the validation set; Spectral or envelope spectrum analysis is performed on the candidate target fault category samples and the real samples of the target fault category to obtain a frequency domain consistency score; Based on the candidate target fault category samples and the qualified samples of the target fault category obtained after screening, the maximum similarity is determined; The comprehensive credibility score of the candidate target fault category samples is determined by a weighted sum of the authenticity score, category confidence score, intra-class feature distance, inter-class boundary distance, frequency domain consistency score, and maximum similarity.

5. The rolling bearing fault diagnosis method according to claim 1, characterized in that, The process of evaluating the trained fault diagnosis model based on the validation set, updating the generation quota and preset screening rules for the target fault category in the next iteration based on the evaluation results, and determining whether to proceed to the next iteration to generate samples includes: The trained fault diagnosis model is evaluated based on the validation set to obtain the recall rate, confusion matrix, and comprehensive performance index of the trained fault diagnosis model for the target fault category. Based on the recall rate, confusion matrix, and generation quota of the target fault category in this iteration, as well as the preset minimum generation quantity and preset maximum generation quantity of the target fault category, the generation quota and preset filtering rules of the target fault category in the next iteration are determined. Based on the difference between the comprehensive performance index of the trained fault diagnosis model and the comprehensive performance index of the trained fault diagnosis model in the previous round, the marginal performance gain is determined. Based on the marginal performance gain, the number of qualified samples of the target fault category, the screening pass rate of the candidate sample set, or the recall rate of the target fault category, determine whether to proceed to the next iteration to generate samples.

6. The rolling bearing fault diagnosis method according to claim 5, characterized in that, The determination of the generation quota and preset filtering rules for the target fault category in the next iteration, based on the recall rate, confusion matrix, and generation quota of the target fault category in the current iteration, as well as the preset minimum and maximum generation quantities of the target fault category, includes: Based on the confusion matrix, the confusion intensity of the target fault category is determined; Based on the confusion intensity of the target fault category, the recall rate of the target fault category, and the generation quota of the target fault category in this iteration, the initial generation quota of the target fault category in the next iteration is obtained. Based on the preset minimum and maximum generation quantities of the target fault category, the initial generation quota of the target fault category in the next iteration is adjusted to obtain the generation quota of the target fault category in the next iteration. If the recall rate of the target fault category is less than a preset recall rate threshold and the confusion intensity of the target fault category is greater than a preset intensity threshold, the filtering threshold in the preset filtering rule is updated.

7. The rolling bearing fault diagnosis method according to claim 5, characterized in that, The step of determining whether to proceed to the next iteration to generate samples based on the marginal performance gain, the number of qualified samples of the target fault category, the screening pass rate of the candidate sample set, or the recall rate of the target fault category includes: If the marginal performance gain is less than a preset gain threshold, or the marginal performance gain is negative for multiple consecutive rounds, or the number of qualified samples of the target fault category is greater than or equal to a preset quantity threshold, or the screening pass rate of the candidate sample set is less than a pass threshold, or the recall rate of the target fault category is equal to or greater than a recall threshold, then the next iteration stops; otherwise, the next iteration continues.

8. The rolling bearing fault diagnosis method according to any one of claims 5-7, characterized in that, The overall performance metric is one or more combinations of prediction accuracy, F1 score, or macro average recall.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images