Rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination

Abstract

The application discloses a rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination, and comprises the following steps: acquiring an original vibration signal of a rolling bearing and converting the original vibration signal into a time-frequency domain signal; a BD-ACGAN network model based on boundary auxiliary discrimination is constructed; a Shuffle Attention attention mechanism module and a self-adaptive weight loss module are introduced into the BD-ACGAN network model; samples are randomly extracted from an unbalanced data set constructed as inputs of the BD-ACGAN network model; the generator and the discriminator are alternately trained by a gradient descent method until Nash equilibrium is reached; then, the trained generator is used to augment the training data set to obtain a fusion data set; and the obtained fusion data set is used to train an auxiliary classifier of the BD-ACGAN network model to detect fault diagnosis performance. The application applies a double-time scale update rule of a stochastic gradient descent to the improved BD-ACGAN network model, so that the stability of the network can be improved, the convergence time of the network can be further reduced, and high-quality samples can be generated more quickly.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bearing fault diagnosis, and particularly relates to a rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination. BACKGROUND

[0002] Rolling bearings, as important components of mechanical equipment, play a crucial role in ensuring the safe and stable operation of equipment. Rolling bearings are widely used in various mechanical equipment, such as automobiles, aircraft, industrial machinery, etc., to support and reduce rotational friction, ensuring the smooth operation of equipment. Rolling bearings can withstand heavy loads and high-speed rotation, improving the efficiency and reliability of equipment. However, long-term high-speed rotation and heavy load can cause problems such as wear, fatigue, deformation, and protrusion of rolling bearings, which will reduce their service life, making them lose normal operation accuracy and stability, and have a serious impact on the safety and reliability of equipment operation. Therefore, ensuring the accuracy of rolling bearing fault diagnosis is crucial to improving the performance and extending the service life of equipment

[0003] However, in actual bearing fault diagnosis, since bearings are usually in normal state during long-term operation, it is relatively difficult and costly to obtain high-quality fault samples. This leads to a serious imbalance between normal samples and fault samples in the bearing fault data set in real scenarios, making fault diagnosis face severe challenges. With the continuous development of artificial intelligence technology, researchers have proposed an unsupervised data enhancement method based on generative adversarial networks (GAN). Through the "dynamic game" process of the generator and the discriminator, GAN can generate samples with the same distribution as the real samples, providing a new solution to the problem of insufficient original data. However, GAN has some shortcomings in the training process, such as unstable training and mode collapse.

[0004] To overcome these problems, researchers improve GAN and propose a series of improved variant models, such as WGAN (Wasserstein GAN) uses the Wasserstein distance instead of the JS distance to measure the similarity of the probability distribution between the generated samples and the real samples, which solves the problems of unstable training and gradient disappearance to a certain extent; ACGAN (Auxiliary Classifier GAN) introduces an auxiliary classifier to distinguish between true and false samples and classify the generated samples by category, improving the quality and diversity of the generated samples; DCGAN (Deep Convolutional GAN) replaces the generator and discriminator network in the traditional GAN with a deep network structure based on the convolutional neural network CNN (Convolutional Neural Network), which can better process image data and generate higher quality image samples, and a series of methods provide new methods for the field of fault diagnosis.

[0005] Although these data enhancement methods proposed for data imbalance have made some progress, they still face problems such as high training complexity, unstable model training, overfitting, class imbalance, and difficulty in controlling generated samples, so the method for solving the data imbalance problem still needs to be further improved. SUMMARY

[0006] In view of the problems of fault data imbalance commonly faced by rolling bearings and the current problems of unstable model training and limited authenticity of generated samples, the present application provides a rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination. This method extracts the boundary features of fault samples based on a boundary auxiliary discriminator, making the generated samples more similar to real samples and effectively solving the problem of data imbalance. At the same time, through a multi-task adaptive weight module, the network can better adapt to different task requirements during training, adjust the weight of the loss function, improve the training effect of the classifier, and make the fault diagnosis network have good classification accuracy in different fault data sets.

[0007] The technical scheme adopted by the present application is:

[0008] The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination comprises the following steps:

[0009] Step 1: Obtain the original vibration signal of the rolling bearing and convert it to a time-frequency domain signal through continuous wavelet transform to extract the frequency spectrum features of different fault categories;

[0010] Step 2: Divide the time-frequency domain signal obtained in step 1 into a training set and a test set according to a ratio of 8:2, and construct four different imbalance rate data sets to better train and evaluate the fault diagnosis model.

[0011] Step 3: A BD-ACGAN network model based on boundary auxiliary discrimination is constructed, a Shuffle Attention attention mechanism module and an adaptive weight loss module are introduced into the BD-ACGAN network model, and the learning ability and generation ability of the BD-ACGAN network model are improved;

[0012] Step 4: After the construction of the BD-ACGAN network model is completed, samples are randomly extracted from the unbalanced data set constructed in step 2 as the input of the BD-ACGAN network model, the generator and the discriminator are alternately trained by the gradient descent method until the Nash equilibrium, then the trained generator is used to augment the training data set, and four fusion data sets are obtained;

[0013] Step 5: The fusion data set obtained in step 4 is used to train the auxiliary classifier of the BD-ACGAN network model, and unknown samples are used for prediction, and real rolling bearing data are used for fault diagnosis experiments to detect the fault diagnosis performance.

[0014] The step 1 comprises the following steps:

[0015] S1.1: Obtain the original vibration signal of the rolling bearing in different fault states through the vibration acceleration sensor on the rolling bearing, and form an original vibration signal data set;

[0016] S1.2: Since continuous wavelet transform can capture short-time characteristics of signals with high time resolution and analyze frequency characteristics of signals with high frequency resolution, the original vibration signal is preprocessed by continuous wavelet transform, and the time-frequency domain signal after continuous wavelet transform is used as the input of the network; the formula of wavelet transform is as follows:

[0017]

[0018] In the formula: X W (a,b) is the transform value when the frequency is a and the time is b after wavelet transform of the signal x(t); t represents the time variable; x(t) is the original signal; is the wavelet basis function; a and b are the scale parameter and the translation parameter of the wavelet basis function respectively; the Cmor wavelet is selected as the basis function of the continuous wavelet transform to process the original vibration signal data.

[0019] In the step 2, in order to verify the generation ability and fault diagnosis ability of the BD-ACGAN network model and simulate the actual bearing fault sample, the real sample data set obtained in step 1 is set to have an imbalance rate of 1:1, 2:1, 5:1 and 10:1 respectively in the experimental simulation.

[0020] The BD-ACGAN network model in step 3 is composed of a generator G, a main discriminator D1 and an auxiliary discriminator D2. The generator G is composed of four deconvolution layers, and the main discriminator D1 and the auxiliary discriminator D2 are composed of six convolution layers;

[0021] Step 3 includes the following steps:

[0022] Step S3.1: In order to make the BD-ACGAN network model more accurately extract the boundary features of the fault samples, a Sobel operator is added in the boundary auxiliary discriminator; since the image boundary retains the underlying texture information, the extraction of boundary feature information can guide the network to generate samples with more rich texture features. The Sobel operator can detect the boundaries in the image by calculating the gradient of the gray value of the pixel points in the image.

[0023] The Sobel operator contains two groups of 3x3 convolution factors, which can perform horizontal and vertical plane convolution on the samples to detect the vertical and horizontal edges in the image, respectively, to obtain the horizontal and vertical brightness difference approximation values. Among them, the vertical edge detection convolution kernel is which can detect the edges in the vertical direction of the image; the horizontal edge detection convolution kernel is which can detect the edges in the horizontal direction of the image.

[0024] In addition, the process of the Sobel operator is derivable, which can be combined with the backpropagation process of the network, so that the BD-ACGAN network model can effectively learn and optimize the boundary features.

[0025] Step S3.2: In order to better capture the pixel-level pair relationship and channel dependency relationship, improve the performance of the BD-ACGAN network model, and as much as possible reduce the additional calculation overhead generated, a Shuffle Attention attention mechanism module is added before the fourth deconvolution layer of the generator, the fourth convolution layer of the main discriminator and the auxiliary discriminator, respectively;

[0026] In the Shuffle Attention attention mechanism module, the input fault sample feature mapping X ∈ R C×H×W The feature map is grouped into a plurality of non-overlapping three-dimensional tensors along the channel dimension, wherein: X represents the input signal, R represents the number of non-overlapping regions divided, C, H, and W represent the number of channels, spatial height, and width of the feature map, respectively;

[0027] The local attention mechanism is constructed in each region, the attention weights between different positions in the region are calculated by convolution operation, which is used to capture the sample feature information of different positions in the region; the sample features in the region are reorganized or mixed using the calculated regional attention weights; finally, channel shuffling and shuffle are used to realize the information communication between different sub-features.

[0028] The hyperparameters of the Shuffle Attention attention mechanism module only include the scaling channel statistical information parameter W1, the moving channel statistical information parameter b1, the scaling space statistical information parameter W2, the moving space statistical information parameter b2, and GroupNorm, which are much smaller than the number of parameters of the BD-ACGAN network model, so the additional calculation caused by them can be ignored, and parallel processing is performed on multiple sub-features, which greatly improves the calculation performance.

[0029] Step S3.3: In order to make the training process of the BD-ACGAN network model converge quickly, an adaptive weight loss module is introduced. The adaptive weight loss module can self-learn and adjust the loss weight through gradient backpropagation update according to the characteristics of each task, so as to balance the importance and features of each task and further improve the performance of the BD-ACGAN network model. The adaptive weight loss of the BD-ACGAN network model is:

[0030]

[0031] Wherein: W is the weight parameter; σ is the noise parameter; is the loss function of the main discriminator D1, is the loss function of the boundary auxiliary discriminator D2, Loss C is the classifier loss function, represents the adaptive weight of the main discriminator D1 loss, represents the adaptive weight of the boundary auxiliary discriminator D2 loss, σ C represents the adaptive weight of the classifier loss. Each loss parameter is initialized to 1, and the parameter is updated during the training process, so as to adapt to the needs of each task, so that the network converges stably and quickly;

[0032] S3.4: In order to improve the learning efficiency of the BD-ACGAN network model, and at the same time make the less number of discriminators iterations in each generator iteration update, the two time scale update rule (TTUR) of random gradient descent is added in the BD-ACGAN network model. That is, different learning rates are set for discriminators and generators, and the convergence speed of discriminators is faster, that is, the learning rate of discriminators is generally set to be larger than that of generators. Using the TTUR method, the generation network and the discrimination network can be updated at a speed of 1:1, and better results can be produced in the same time. The learning rate of the generator in the BD-ACGAN network model is set to 0.0001, and the learning rate of the discriminator is set to 0.0004, which can solve the instability problem and the long training time problem in the training process of the BD-ACGAN network model.

[0033] The step 4 includes the following steps:

[0034] S4.1: Joint training of generation and classification:

[0035] At the beginning of training, the parameters of the randomly initialized generator and discriminator are randomly initialized, and thereafter, in each iteration process, a batch of samples are randomly selected from the different imbalance rate samples constructed in step 2 for the training of the generator, and the generated samples under the corresponding imbalance rate are output; then, the discriminator is trained by discriminating the true and false of the selected real samples and the corresponding generated samples and classifying the samples. In the training process of the discriminator, the loss function of Gaussian maximum likelihood is used to calculate the loss and parameter update to improve the quality of the generated samples; the generator and the discriminator are alternately trained by the gradient descent method to constantly optimize the parameters. In this way, the generator can generate more realistic samples, and the discriminator can more accurately discriminate and classify the samples. The loss function of the discriminator in the training of the BD-ACGAN network model is based on Gaussian maximum likelihood:

[0036] wherein f W (x) is the output of the BD-ACGAN network model, W is the weight parameter, and σ is the noise parameter.

[0037] S4.2: Data set augmentation:

[0038] In this stage, the trained BD-ACGAN network model is used to augment the training data set;

[0039] First, fix the parameters of the generator, and use the generator to generate a large number of samples. The generated samples can be processed by data enhancement techniques, including random flipping, random rotation, random size adjustment, etc. to produce diversified augmented samples.

[0040] Subsequently, the enhanced samples are combined with the real samples to form a new fusion sample dataset. Finally, the fusion sample dataset is used to train the classifier to improve the fault diagnosis capability and generalization capability of the BD-ACGAN network model.

[0041] In step 5, the fusion dataset obtained in step 4 provides more comprehensive fault samples, which can enable the auxiliary classifier to learn more features and patterns of fault categories, and at the same time enhance the generalization capability of the auxiliary classifier and improve the classification capability of similar feature samples. After training with the fusion dataset, the auxiliary classifier can better adapt to the distribution of fault data. In addition, inference prediction using unknown samples can effectively evaluate the generalization capability of the BD-ACGAN network model, verify the robustness, and improve the BD-ACGAN network model in time, which helps to improve the reliability, adaptability and accuracy of the BD-ACGAN network model.

[0042] Step 5 includes the following steps:

[0043] S5.1: Fault classifier training:

[0044] In this stage, only the classifier branch of the main discriminator D1 is adjusted in weight, and the weight of the auxiliary classifier parameter of the main discriminator D1 is initialized using the weight of the generation and classification combined training stage at the beginning of training; subsequently, in each iteration process, samples are randomly selected from the fusion sample dataset obtained in step 4 to train the auxiliary classifier. In the training process of the auxiliary classifier, the cross-entropy loss function is used to calculate the classification loss, and the parameters of the auxiliary classifier are updated to improve the accuracy of the BD-ACGAN network model diagnosis. Wherein, the cross-entropy loss function of the auxiliary classifier is:

[0045]

[0046] Wherein, p(x) represents the real distribution probability, q(x) represents the approximate distribution probability, and x represents a random variable.

[0047] S5.2: Inference test:

[0048] In the actual inference test stage, only the encoder and classification branch of the main discriminator D1 need to be retained, and the weight parameters of the fault classifier training stage are loaded. The one-dimensional fault data of unknown categories are preprocessed by CWT to convert them into two-dimensional fault images. Then the two-dimensional images are encoded and input into the classifier, and the classification branch obtains the fault category prediction result according to the image features.

[0049] The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination has the following technical effects:

[0050] 1) The improved BD-ACGAN network model is combined with the Shuffle Attention attention mechanism module and the adaptive weight loss module, so that the generator can generate more realistic samples and obtain a balanced fusion dataset through the generated samples, which can be used for subsequent training of the auxiliary classifier.

[0051] 2) In the construction of the improved BD-ACGAN network model, the Shuffle Attention attention mechanism module and the adaptive weight loss module are introduced into the generator and the discriminator. The Shuffle Attention attention mechanism module enhances the influence of important sub-features and promotes information transmission between features within the group, thereby improving the expression ability of the features; the adaptive weight loss module can self-learn and adjust the loss weight through gradient backpropagation update according to the characteristics of each task, so as to balance the importance and features of each task.

[0052] 3) With the deepening of the improved BD-ACGAN network model, the double time scale update rule of stochastic gradient descent is applied to the improved BD-ACGAN network model, which not only improves the stability of the network, but also further reduces the convergence time of the network compared to the BD-ACGAN network model without adding the double time scale update rule, and generates high-quality samples faster. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is the structure diagram of the BD-ACGAN network model of the present application.

[0054] Figure 2 is the structure diagram of the generator of the present application.

[0055] Figure 3(a) is a structure diagram of the main discriminator of the present application.

[0056] Figure 3(b) is a structure diagram of the auxiliary discriminator of the present application.

[0057] Figure 4 is the network training strategy diagram of the present application.

[0058] Figure 5 is the comparison diagram of the generated samples and the real samples under different unbalanced rate datasets of the present application.

[0059] Figure 6 is the diagnostic accuracy of the balanced dataset after sample fusion of the present application.

[0060] Figure 7 is the t-SNE visualization dimension reduction result of the distribution of the generated samples and the real samples of the present application.

[0061] Figure 8 is the confusion matrix diagram after the generated sample is classified and diagnosed in the application.

[0062] Figure 9 is a structural schematic diagram of the Shuffle Attention attention mechanism module in the application.

[0063] Figure 10 is a data set augmentation schematic diagram of the application.

[0064] Figure 11 is a reasoning test process schematic diagram of the application. DETAILED DESCRIPTION

[0065] The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination first acquires vibration signals in different states through a sensor, converts the signals into time-frequency signals by using wavelet transform, and generates data samples with different balance rates according to actual conditions; then, the data samples are input into a BD-ACGAN network model for training to realize Nash balance; then, data with specific labels are generated from the generator to supplement the unbalanced data set; finally, the balanced data are input into an auxiliary classifier to realize feature learning of different fault categories.

[0066] As shown in Figure 1 , it is a BD-ACGAN network model structure schematic diagram of the application, and the BD-ACGAN network model is composed of a generator G, a main discriminator D1 and an auxiliary discriminator D2. The first four layers of the generator G adopt deconvolution layers, a Shuffle Attention attention mechanism module is introduced after the fourth layer to enhance the expression ability of the features, and then three deconvolution layers are used to further extract the features. The first six layers of the main discriminator D1 and the auxiliary discriminator D2 adopt convolution layers, a Shuffle Attention attention mechanism module is introduced after the fourth layer, and then two convolution layers are used to further extract the features.

[0067] In the BD-ACGAN network model, the main discriminator D1 is used for true or false judgment and sample classification of the generated sample, and the auxiliary discriminator D2 judges the authenticity of the generated sample by detecting the boundary information of the generated sample. The BD-ACGAN network model takes a random noise vector Z and label information as input, and generates corresponding samples in the learning iteration process of the generator G. The main discriminator D1 compares the main features of the generated sample with the actual sample, and outputs the discrimination and classification results; the auxiliary discriminator D2 extracts the boundary features of the generated sample and compares them with the actual sample, and outputs the discrimination results.

[0068] In the BD-ACGAN network model, the boundary information of the sample is extracted using the Sobel operator. The Sobel operator is a commonly used image edge detection algorithm that detects the boundaries in the image by calculating the gradient of the gray value of the pixel points in the image. Since the image boundary retains the underlying texture information, extracting the boundary feature information can guide the network to generate samples with more rich texture features. In addition, the process of Sobel operator is derivable, which can be combined with the back propagation process of the network, so that the BD-ACGAN network model can effectively learn and optimize the boundary features.

[0069] As shown in Figure 2 The generator model structure diagram of the application is shown in the figure, which mainly consists of four deconvolution layers. The deconvolution layer can better extract the features in the image. A Shuffle Attention attention mechanism module is added before the fourth deconvolution to improve the BD-ACGAN network model's ability to capture the global dependence of the features when generating images, so as to improve the quality, clarity and aesthetics of the images. This makes the texture details of the image more prominent, and improves the perceptual quality of the image from multiple aspects such as texture, brightness and clarity. In addition, it improves the performance of the generator, reduces the artifacts and noise that may appear in the generated image, and increases the controllability of the generation process, so that the generated image has higher quality and realism.

[0070] The generator execution flow is as follows: set the noise vector dimension of each sample to 100 and the label dimension to 10, map each label to a 100-dimensional embedding vector through label embedding, and multiply it with the noise vector to fuse the condition information. Then, a linear layer is used to map the fused vector to 768 dimensions, and then four deconvolution layers are used to reconstruct the generated features. Finally, the Shuffle Attention attention mechanism module is used to enhance the representation of the generated features, and finally converted to a 64x64x3 feature map. If the input image size is 256x256, it will also be trained through three additional deconvolution layers.

[0071] As shown in Figures 3(a) and 3(b), the network model structure diagrams of the main discriminator and the auxiliary discriminator of the application are shown. The main discriminator is used to generate the true and false judgment and sample classification, while the auxiliary discriminator is used to detect the boundary information of the generated sample to determine its authenticity. Both of them mainly consist of six convolutional layers, and the activation function is LeakyRelu. A Shuffle Attention attention mechanism module is added before the fourth deconvolution to improve the discrimination efficiency and quality of the discriminator.

[0072] The discriminator execution flow: the real sample data and the generator generated sample data are input into the main discriminator and the auxiliary discriminator, after dimension reduction through four convolutional layers, the long distance dependence relationship is captured through the Shuffle Attention attention mechanism module, and finally the channel number is increased to 512 through two convolutional layers, batch normalization, LeakyReLU activation function and Dropout regularization are applied in this process. If the size of the input image is 256*256, additionally pass through two convolutional layers, keep the channel number at 512, and also apply batch normalization, LeakyReLU activation function and Dropout regularization operation.

[0073] As shown in Figure 4 The training strategy of the network of the present application is shown in the schematic diagram, which includes four stages: joint training of generation and classification, data set augmentation, fault classifier training and inference prediction. In the joint training stage of generation and classification, the generator and the discriminator are alternately trained, so that the generator generates more real samples, and the discriminator can accurately distinguish and classify the samples; in the data set augmentation stage, a large number of samples are generated by using the generator, and are processed by data enhancement technology to form a fusion sample data set, which is used to train the classifier to improve the fault diagnosis ability and generalization ability; in the fault classifier training stage, the weight of the classifier branch of the main discriminator is adjusted, and the auxiliary classifier is trained through the cross entropy loss function, so as to improve the accuracy of the BD-ACGAN network model. In the inference test stage, the encoder and the classification branch of the main discriminator are used to preprocess and classify the fault data of unknown categories.

[0074] The experimental platform of the CWRU data set used in the present application. The CWRU data set comes from the bearing data center of Case Western Reserve University in the United States, the bearing rotating speed is 1797r / min, and the sampling frequency is 12kHz. Data acquisition is carried out when the motor load is 1.5kW, including inner ring fault, rolling body fault, outer ring fault and normal working condition, each fault type contains 3 kinds of fault sizes: 0.007mm, 0.014mm, 0.021mm. Therefore, the samples can be sequentially labeled as 1 to 10 according to the fault type and the fault size.

[0075] In the present application, the dataset of the drive end bearing housing under 0 hp load scenario is used. The vibration signal is collected at a sampling rate of 12000 samples per second, with a sampling interval of 300 points. In the experiment, 400 samples are collected for each state for study. Considering that the number of fault samples is often less than the number of normal data samples in actual engineering, in order to simulate the actual bearing fault sample situation, the present application designs four kinds of unbalanced rate datasets, which are 1:1, 2:1, 5:1 and 10:1. The unbalanced ratio refers to the ratio of the number of normal samples to the number of samples of each fault type. Taking the unbalanced ratio of 5:1 as an example, each normal state contains 400 samples, and the number of samples of each fault type is 80, and finally the training set and test set are constructed according to the ratio of 8:2. Table 1 provides more detailed information about the dataset.

[0076] Table 1 Number of samples of rolling bearing dataset under unbalanced rate

[0077]

[0078] According to Table 1, the number of samples of each fault category in the training set and the number of test set samples of each dataset can be obtained. The improved network is trained by the minority class samples, and new samples are generated. In this way, the balanced minority class samples can be combined with the majority class samples to form a balanced dataset. Then, the balanced dataset is input into the fault diagnosis classifier for training. After training, the classifier can classify faults by inputting test set samples and outputting fault diagnosis results.

[0079] As Figure 5 shown, the present application generates a comparison chart of samples and real samples under different unbalanced rate datasets. In order to compare the differences between the images generated by the generator in the present application under different balancing rates and the original images, after training the improved BD-ACGAN network model on the training set of each dataset, 5 different generated fault samples are randomly selected for comparison with the original images. From Figure 5 It can be seen that when the unbalanced rate is 1:1 and 2:1, the overall features show high similarity. However, compared with the unbalanced rate of 1:1, the generated sample features of the unbalanced rate of 2:1 are relatively weakened. When the unbalanced rate is 5:1 and 10:1, the generated samples still retain part of the obvious features.

[0080] In summary, the BD-ACGAN network model can effectively learn the boundary features of the samples and has the ability to expand the original dataset. Therefore, the data enhancement method of the present application can solve the problem of data imbalance.

[0081] To objectively evaluate the generation ability of the generator, three indicators are used to analyze the characteristics of the generated samples, including FID (Frechet Inception Distance), SSIM (Structural Similarity Index), and PSNR (Peak Signal-to-Noise Ratio). To verify the influence of the imbalance rate on the BD-ACGAN network model, 40 pairs of bearing fault samples are randomly selected from each imbalance rate to test the diagnostic accuracy of the BD-ACGAN network model and calculate the mean of the three indicators. The specific results are shown in Table 2.

[0082] Table 2 Classification accuracy and generated sample quality under different imbalance rates

[0083]

[0084] As can be seen from Table 2, when the imbalance rate is 1:1, the accuracy of the BD-ACGAN network model is 98.83%, indicating that it has high diagnostic accuracy and the similarity between the generated samples and the real samples is high. When the imbalance rate is 2:1, the accuracy of the model remains at 97.11%, indicating that the model still maintains good diagnostic performance on the imbalanced data set. When the imbalance rate is 5:1 and 10:1, although the accuracy decreases, the similarity between the generated samples and the real samples is still good, indicating that the BD-ACGAN network model still has certain sample generation ability and diagnostic ability when the data set imbalance rate is high.

[0085] To verify the sample generation ability of the BD-ACGAN network model and the effect of the added module, a control test and an ablation experiment are conducted, and the results are shown in Table 3. Among them, the network model only adding the Shuffle Attention attention mechanism module is denoted as AC-SAGAN, and the network model only adding the boundary auxiliary discriminator is denoted as AC-BDGAN.

[0086] Table 3 Ablation experiment generated sample quality

[0087]

[0088] As can be seen from Table 3, compared with the ACGAN network model, the AC-SAGAN network model improves the distribution difference between the generated samples and the real samples, improves the similarity degree, but there is still room for further optimization in the distortion degree. Compared with the ACGAN network model, the AC-BDGAN network model effectively improves the distribution difference between the generated samples and the real samples, significantly improves the similarity and structural similarity, and also improves the distortion degree. Compared with the other three network models, the BD-ACGAN network model has made significant improvements in the quality, similarity and distortion degree of the generated samples, and provides an effective solution for generating high-quality samples.

[0089] As Figure 6 shown, the present application is compared with other network models in the diagnosis accuracy of the balanced data set after sample fusion. To verify the diagnosis ability of the BD-ACGAN network model, the BD-ACGAN network model is compared with other network models. To ensure the fairness of the experiment, the same data set is used for each network model, and the data set is balanced after fusion. As can be seen from the figure, in the fault diagnosis task, the BD-ACGAN network model shows stronger adaptability than other network models. Whether in the imbalance rate of 1:1, 2:1, 5:1 or 10:1, the BD-ACGAN network model has higher accuracy, which shows that the BD-ACGAN network model can effectively cope with the unbalanced data problem. In contrast, the VIT and AlexNet network models have relatively low accuracy when the imbalance rate is high, and the GaussianNB network model performs well at low imbalance rate, but the accuracy decreases significantly when the imbalance rate increases.

[0090] The experimental platform of the XJUT-SY data set used in the present application is composed of an alternating current motor, a motor speed controller, a rotating shaft, a supporting bearing, a hydraulic loading system and a test bearing. It can perform accelerated degradation test on the bearing under different operating conditions and obtain complete data from operation to failure. The XJTU-SY data set includes 15 groups of data of known failure positions, including inner ring wear, retainer fracture, outer ring wear and other types of faults. To ensure the integrity of the experiment, five types of data, outer ring, inner ring, inner ring and outer ring, retainer and full fault (inner ring, outer ring, retainer, rolling element), are selected for experiment, and the fault samples are marked as 0 to 4 in turn, as shown in Table 4.

[0091] Table 4 Fault type and label

[0092]

[0093] To verify the advantages of the BD-ACGAN network model in processing actual fault data, the generated samples using the XJTU-SY data set are analyzed, and the results are shown in Table 5.

[0094] Table 5 Quality of generated samples

[0095]

[0096] As can be seen from Table 5, the generated samples of the ACGAN network model perform poorly in three indicators, and the generated image quality is low, and there is a significant distribution difference with the real samples. The generated samples of the AC-SAGAN and AC-BDGAN network models have improved to some extent in the FID indicator, and have improved to some extent in the SSIM and PSNR indicators. Compared with the other three network models, the BD-ACGAN network model performs best among all network models, and the quality of the generated samples is obviously improved through the indicators.

[0097] As shown in Figure 7 , the results of the t-SNE visualized dimension reduction of the generated samples and the real sample distribution of the application. As can be seen from Figure 7 , in the same fault type samples, the features of most generated samples are highly consistent with the original data features, and only the features of individual class samples are slightly discrete. This clearly shows that the features have good clustering, and further verifies the effectiveness of the proposed data enhancement method. This means that through adversarial training, synthetic data with similar probability distribution as the original data can be generated, which can make up for the unbalanced data set. In addition, the generated data features of different fault types in the chart can also be clearly distinguished, which shows that the data generated by the data augmentation method can be successfully used for bearing fault classification tests. Therefore, it can be concluded that the proposed data enhancement method is feasible and effective in improving the fault classification performance.

[0098] As shown in Figure 8 , the confusion matrix after classification and diagnosis of the generated samples of the application. The abscissa of the confusion matrix represents the predicted fault type, and the ordinate represents the actual fault type. Each matrix element (i, j) represents that the i-th state sample is incorrectly classified as the j-th state. As can be seen from the confusion matrix, when using the XJTU-SY data set, the overall diagnostic accuracy meets the actual engineering requirements, but there is still a certain diagnostic error, which is mainly concentrated in the inner ring fault and the inner and outer ring faults. This is because in the bearing test test, the outer ring of the bearing is fixed and the inner ring rotates, and when the inner ring of the rolling bearing fails, the inner ring fault signal may be affected by the bearing rotation frequency modulation phenomenon. In addition, the amplitude of the inner ring fault feature at the theoretical fault characteristic frequency is not obvious, and there is a series of rotation frequency harmonics and sidebands.

[0099] like Figure 9 The diagram shows the structure of the Shuffle Attention mechanism module used in this invention. The Shuffle Attention module divides the input feature map into multiple groups and uses a Shuffle unit to integrate channel attention and spatial attention into a block within each group. Then, all sub-features are aggregated, and channel shuffling and mixing are used to achieve communication between different sub-features. The Shuffle Attention module uses channel segmentation to process the sub-features of each group in parallel. For the channel attention branch, GAP is used to generate channel statistics, and then a pair of parameters is used to scale and shift the channel vectors. For the spatial attention branch, group norm is used to generate spatial statistics, and then a compact feature similar to the channel branch is created. These two branches are then concatenated. Afterwards, all sub-features are aggregated, and finally, channel shuffling and mixing are used to achieve communication between different sub-features.

[0100] The channel attention mechanism in the Shuffle Attention module maps the features of the input fault samples to X∈R. C×H×W The feature map is divided into a three-dimensional tensor of g non-overlapping regions along the channel dimension, i.e., X = [X1, X2, ..., X]. g ], X g ∈R C / G×H×W Where X represents the input signal, R represents the number of non-overlapping regions, and C, H, and W represent the number of channels, spatial height, and width of the feature map, respectively. Then, the attention module generates corresponding importance coefficients for each sub-feature. Spatial attention complements channel attention, using Group Norm (GN) to obtain spatial statistics, and then connecting these two branches to make the number of channels the same as the number of inputs.

[0101] like Figure 10 The diagram illustrates the dataset augmentation process of this invention. In this stage, firstly, based on the trained generator, the generator parameters are fixed to generate a large number of synthetic samples. These synthetic samples include samples from normal operating conditions and samples from various fault modes to ensure the diversity and complexity of the generated samples. Secondly, to improve the generalization ability and robustness of the BD-ACGAN network model, methods such as random flipping, random rotation, and random resizing are used to augment the generated samples, adding more diverse sample data to the original generated samples. Then, the augmented generated samples are fused with real samples to form a new fused sample dataset.

[0102] like Figure 11As shown, it is a reasoning test schematic diagram of the application. In the actual reasoning test phase of fault diagnosis, first, the weight parameters of the encoder of the main discriminator and the classification branch of the pre-trained BD-ACGAN network model need to be loaded. The vibration data collected in the actual engineering all exist in the form of one-dimensional time series, so in order to convert them into a form suitable for network model processing, CWT is used to convert the vibration data into two-dimensional images. Secondly, the two-dimensional images processed by CWT are input into the classification branch after image feature extraction, and the classification branch obtains the fault category prediction result according to the image feature.

Claims

1. A method for fault feature enhancement and diagnosis of rolling bearings based on boundary auxiliary discrimination, characterized in that The method comprises the following steps: Step 1: obtaining the original vibration signal of the rolling bearing, and converting the original vibration signal into a time-frequency domain signal through continuous wavelet transform; Step 2: dividing the time-frequency domain signal obtained in step 1 into a training set and a test set, and constructing an unbalanced rate dataset, Step 3: constructing a BD-ACGAN network model based on boundary auxiliary discrimination, introducing a Shuffle Attention attention mechanism module and an adaptive weight loss module in the BD-ACGAN network model, and improving the learning ability and generation ability of the BD-ACGAN network model; Step 4: after the construction of the BD-ACGAN network model is completed, randomly extracting samples from the unbalanced dataset constructed in step 2 as the input of the BD-ACGAN network model, and alternately training the generator and the discriminator through the gradient descent method until Nash equilibrium, then using the trained generator to augment the training dataset to obtain a fusion dataset; Step 5: training the auxiliary classifier of the BD-ACGAN network model using the fusion dataset obtained in step 4 to detect the fault diagnosis performance; In step 3, in order to make the training process of the BD-ACGAN network model quickly converge, an adaptive weight loss module is introduced, which can learn and adjust the loss weight through gradient back propagation update according to the characteristics of each task, so as to balance the importance and characteristics of each task and further improve the performance of the BD-ACGAN network model; wherein the adaptive weight loss of the BD-ACGAN network model is: ； wherein: W is a weight parameter; is a noise parameter; is a loss function of the main discriminator , is a loss function of the boundary auxiliary discriminator , is a classifier loss function, represents an adaptive weight of the loss of the main discriminator , represents an adaptive weight of the loss of the boundary auxiliary discriminator , represents an adaptive weight of the classifier loss; each loss parameter is initialized to 1, and the parameter is updated during the training process, so as to adapt to the needs of each task, and make the network stable and fast convergence.

2. The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination according to claim 1, characterized in that: The step 1 comprises the following steps: S1.1: obtaining the original vibration signal of the rolling bearing under different fault states through the vibration acceleration sensor on the rolling bearing to form an original vibration signal dataset; S1.2: pre-processing the original vibration signal by using continuous wavelet transform, and taking the time-frequency domain signal after the continuous wavelet transform as the input of the network; the formula of the wavelet transform is as follows: ； In the formula: is the signal is the wavelet transform value of the signal at the frequency and the time ; t represents the time variable; is the original signal; is the wavelet base function; are the scale parameter and the translation parameter of the wavelet base function respectively; the Cmor wavelet is selected as the base function of the continuous wavelet transform to process the original vibration signal data.

3. The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination of claim 1, characterized in that: In step 2, in order to verify the generation ability and fault diagnosis ability of the BD-ACGAN network model and simulate the actual bearing fault sample situation, the real sample dataset obtained in step 1 is set to have unbalanced rates of 1:1, 2:1, 5:1 and 10:1 respectively in the experimental simulation.

4. The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination of claim 1, characterized in that: In step 3, the BD-ACGAN network model is composed of a generator G , a main discriminator and an auxiliary discriminator ; the generator G is composed of four deconvolutional layers, the main discriminator and the auxiliary discriminator are composed of six convolutional layers.

5. The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination of claim 4, characterized in that: Step 3 comprises the following steps: Step S3.1: adding a Sobel operator in the boundary auxiliary discriminator, which can detect the boundary of the image by calculating the gradient of the gray value of the pixel points in the image; The Sobel operator includes two sets of 3*3 convolution factors, which can perform horizontal and vertical plane convolution on the sample, and are used for detecting the vertical and horizontal edges in the image respectively, so as to obtain the horizontal and vertical brightness difference approximation values respectively; wherein the vertical edge detection convolution kernel is , which can detect the edges in the vertical direction of the image; and the horizontal edge detection convolution kernel is , which can detect the edges in the horizontal direction of the image. Step S3.2: adding a Shuffle Attention attention mechanism module before the fourth deconvolution layer of the generator, the main discriminator and the fourth convolution layer of the auxiliary discriminator respectively; In the Shuffle Attention attention mechanism module, the input fault sample features are mapped grouping the feature map into a plurality of non-overlapping regions along the channel dimension, wherein: denotes an input signal, denotes the number of non-overlapping regions into which the input signal is divided, denote the number of channels, spatial height and width of the feature map, respectively; Step S3.3: calculating the adaptive weight loss of the BD-ACGAN network model; Step S3.4: adding a double time scale update rule of stochastic gradient descent in the BD-ACGAN network model, and setting different learning rates for the discriminator and the generator.

6. The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination of claim 1, characterized in that: The step 4 comprises the following steps: S4.1: generating and training in combination: At the beginning of training, the parameters of the initial generator and discriminator are randomly generated, and thereafter in each iteration process, a batch of samples are randomly selected from the different imbalance rate samples constructed in step 2 for the training of the generator, and the generated samples under the corresponding imbalance rate are output; subsequently, the discriminator is trained through the true-false discrimination and sample classification of the selected real samples and the corresponding generated samples, and in the training process of the discriminator, the loss and parameter update are calculated by using the loss function of Gaussian maximum likelihood to improve the quality of the generated samples; the generator and the discriminator are alternately trained by the method of gradient descent to continuously optimize the parameters; during the training of the BD-ACGAN network model, the loss function of the discriminator is based on Gaussian maximum likelihood and is: ; wherein: is the output of the BD-ACGAN network model, is a weight parameter, is a noise parameter; S4.2: Data set augmentation: In this stage, the trained BD-ACGAN network model is used to augment the training data set; First, fix the parameters of the generator, use the generator to generate a large number of samples, and the generated samples are processed through data enhancement techniques, including random flipping, random rotation, and random size adjustment methods, to produce diversified augmented samples; Subsequently, the augmented samples are combined with the real samples to form a new fusion sample data set; finally, the fusion sample data set is used to train the classifier.

7. The rolling bearing fault feature enhancement and diagnosis method based on boundary auxiliary discrimination of claim 1, characterized in that: The step 5 includes the following steps: S5.1: Fault classifier training: In this stage, only the weight of the classifier branch of the main discriminator is adjusted At the beginning of the training, the parameters of the auxiliary classifier of the main discriminator are initialized using the weights generated in the joint training stage of generation and classification; then, in each iteration process, samples are randomly selected from the fusion sample dataset obtained in step 4 to train the auxiliary classifier; in the training process of the auxiliary classifier, the classification loss is calculated using the cross-entropy loss function, and the parameters of the auxiliary classifier are updated to improve the accuracy of the BD-ACGAN network model diagnosis; wherein the cross-entropy loss function of the auxiliary classifier is: ； wherein: represents a true distribution probability, represents an approximate distribution probability, represents a random variable; S5.2: Inference test: In the actual inference test stage, only the encoder and classification branch of the main discriminator D1 are retained, and the weight parameters of the fault classifier training stage are loaded; the one-dimensional fault data of unknown categories is preprocessed through CWT to convert it into a two-dimensional fault image; then the two-dimensional image is encoded and input into the classifier, and the classification branch obtains the fault category prediction result according to the image features.

Images