Rotary machinery data enhancement method based on SCF-CVAE-GAN network

Abstract

The invention relates to the field of rotating machine fault diagnosis, in particular to a rotating machine data enhancement method based on an SCF-CVAE-GAN network, and the method comprises the steps: obtaining rotating machine vibration signals in different fault modes; the rotating machinery vibration signals are converted into a two-dimensional time-frequency image, and the SCF-CVAE-GAN network is trained by adopting the two-dimensional time-frequency image; generating a two-dimensional time-frequency image sample by adopting the trained SCF-CVAE-GAN, calculating the structural similarity and the distance score of the generated sample, and evaluating the quality of the generated sample; mixing the generated sample meeting the quality requirement and the two-dimensional time-frequency image to obtain an enhanced data set; according to the method, a double-self-correction network is provided, a frequency domain regularization strategy is integrated, global information and local textures of data are deeply mined, and the diversity and fidelity of generated samples are ensured by constraining distribution of the samples in a frequency domain.

Description

Technical Field

[0001] The present invention relates to the field of rotating machinery fault diagnosis, and in particular to a rotating machinery data enhancement method based on an SCF-CVAE-GAN network. Background Art

[0002] Due to the low occurrence rate of rotating machinery failures, fault samples are scarce, which easily leads to data imbalance problems during fault diagnosis.

[0003] To solve the above problems, the existing technical solutions mainly include: data-level methods: undersampling may lead to the loss of valuable information, while oversampling may introduce noise due to the repeated use of minority class samples; algorithm-level methods: they rely heavily on expert experience and are not suitable for fault diagnosis under imbalanced conditions, and the decision-making process lacks transparency; data enhancement methods: based on generative adversarial networks, by using neural networks to generate more fault data, fault diagnosis under sample imbalance conditions has been widely used.

[0004] Existing data augmentation methods have achieved certain results in generating samples, but they do not fully consider the detailed texture features, global structural information and frequency domain distribution characteristics of the data, resulting in the generated samples being difficult to truly reflect the fault mode, which in turn affects the effectiveness of data augmentation and the overall performance of fault diagnosis. Summary of the Invention

[0005] In view of this, the present invention discloses a rotating machinery data enhancement method based on the SCF-CVAE-GAN network to solve the above problems; the method comprises:

[0006] S1. Obtain vibration signals of rotating machinery with different fault modes;

[0007] S2. Use continuous wavelet transform to convert the rotating machinery vibration signal into a two-dimensional time-frequency image, and use the two-dimensional time-frequency image to train the SCF-CVAE-GAN network; the SCF-CVAE-GAN network is used to generate two-dimensional time-frequency image samples similar to the rotating machinery fault mode, and the training adopts a frequency domain regularization strategy;

[0008] S3. Generate two-dimensional time-frequency image samples using the trained SCF-CVAE-GAN and calculate the structural similarity and distance scores of the generated samples. Set structural similarity and distance score thresholds and select generated samples that meet the quality requirements based on the structural similarity and distance score thresholds.

[0009] S4. Mix the generated samples that meet the quality requirements with the two-dimensional time-frequency image to obtain an enhanced dataset.

[0010] The beneficial effects of the present invention include:

[0011] A dual self-correction network for global and local feature extraction is proposed. This network can self-calibrate to expand the receptive field of local texture features and enhance the perception of global structural information. The generated samples can reflect the actual fault mode.

[0012] A frequency domain regularization strategy is proposed to constrain the distribution difference between generated samples and real samples in the frequency domain space, so as to guide the model to maintain frequency domain consistency when generating samples.

[0013] A novel SCF-CVAE-GAN network is proposed. By integrating dual self-correcting networks and frequency domain regularization strategies, it can deeply mine the global information and local texture of the data. By constraining the distribution of samples in the frequency domain, it ensures the diversity and fidelity of the generated samples, thereby improving the accuracy of fault diagnosis based on the generated samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of the structure of the SCF-CVAE-GAN network in the present invention;

[0015] Figure 2 Schematic diagram of the method for rotating machinery fault diagnosis using the SCF-CVAE-GAN network of the present invention;

[0016] Figure 3 Schematic diagram of the structure of the dual self-correction network in the present invention;

[0017] Figure 4 This is a comparison chart of sample quality assessment based on the bearing dataset generated in the present invention;

[0018] Figure 5 A sample quality assessment comparison curve generated based on a bearing data set in the present invention;

[0019] Figure 6 This is a comparison chart of sample quality assessment based on the gear dataset in the present invention;

[0020] Figure 7 A sample quality assessment comparison curve based on the gear dataset generated in the present invention;

[0021] Figure 8 is the accuracy comparison curve obtained based on the bearing data set in the present invention;

[0022] Figure 9 This is the accuracy comparison curve obtained based on the gear data set in the present invention. DETAILED DESCRIPTION

[0023] In order to make the objectives, technical solutions, features and advantages of the present invention more clearly understood, the present invention is further described below with reference to the accompanying drawings and embodiments.

[0024] A rotating machinery data enhancement method based on an SCF-CVAE-GAN network, comprising:

[0025] S1. Obtain vibration signals of rotating machinery with different fault modes.

[0026] Specifically, the vibration signals of the rotating machinery under different fault conditions are collected by an accelerometer, and the corresponding true category labels are obtained.

[0027] S2. The continuous wavelet transform is used to convert the vibration signal of the rotating machinery into a two-dimensional time-frequency image, and the two-dimensional time-frequency image is used to train the SCF-CVAE-GAN network; the SCF-CVAE-GAN network is used to generate two-dimensional time-frequency image samples similar to the fault mode of the rotating machinery, and the training adopts a frequency domain regularization strategy.

[0028] Specifically, the vibration signal of the rotating machinery is randomly sampled in sections to obtain a plurality of samples with a length of 784, and the samples are normalized.

[0029] To more clearly characterize the signal characteristics, the continuous wavelet transform is used to convert the samples into a two-dimensional time-frequency image. The continuous wavelet transform uses a finite-length decaying wavelet as the basis function. By convolving the signal with the wavelet basis, it can effectively extract rich time-frequency feature information.

[0030] Furthermore, if Figure 1 As shown, the SCF-CVAE-GAN network includes:

[0031] The generator based on the dual self-correcting network, Figure 1 The generator is used to process the two-dimensional time-frequency image as a real sample to obtain a generated sample.

[0032] The generator based on the dual self-correction network includes: an encoder, a decoder, and a dual self-correction network; the data processing of the generator based on the dual self-correction network includes:

[0033] Step 1: Perform two-dimensional convolution on the real sample to obtain a shallow feature map with 16 channels.

[0034] Step 2: Use an encoder to encode the shallow feature map into a deep feature map with 32 channels, map the deep feature map to the latent space, and obtain the latent vector.

[0035] Step 3: Use the true category label as conditional information, concatenate the true category label with the latent vector, and use a fully connected layer to map it into a conformal feature map that is consistent with the shape of the deep feature map.

[0036] Step 4: Use a decoder to decode the conformal feature map into a low-dimensional feature map with 16 channels.

[0037] Step 5: Use transposed convolution to reconstruct the low-dimensional feature map into a generated sample with the same size as the real sample.

[0038] The dual self-correction network processes the shallow feature map, deep feature map and low-dimensional feature map respectively to obtain the corrected features, which are used to train the SCF-CVAE-GAN network.

[0039] Furthermore, if Figure 3 As shown in the figure, the dual self-correction network includes a local self-correction module and a global self-correction module, which effectively utilizes the local details and global structure of the data through multi-scale information fusion and global context modeling.

[0040] Furthermore, the local self-correction module adopts a dual-branch structure based on the self-calibration mechanism, including a local enhancement branch and a self-correction convolution branch;

[0041] The local self-correction module processes data by dividing the number of channels of the input feature map X of the dual self-correction network into two, keeping the height and width unchanged, to obtain two feature maps X1 and X2 of the same size.

[0042] Furthermore, convolution is performed on X1 to obtain local detail information. The formula is:

[0043] Y1=F1(X1)=X1*K1

[0044] Among them, F1() represents the convolution operation, the convolution kernel is K1, and Y1 represents local detail information.

[0045] The self-correcting convolution branch uses a convolution layer with a convolution kernel of K. By decomposing the convolution kernel K into four convolution kernels K1, K2, K3, and K4, data features are extracted and fused at different scales to enhance the local feature expression capability. The formula for convolution kernel decomposition is:

[0046] K=K1+K2+K3+K4

[0047] Furthermore, the self-correction convolution branch processes data including:

[0048] Step 1: Use average pooling downsampling to reduce the resolution of X2 and extract the global low-frequency information F2(Z1) through convolution operation.

[0049]

[0050] Among them, F2() represents the convolution operation, the convolution kernel is K2, AP() represents average pooling, and Z1 represents the low-resolution feature obtained after downsampling.

[0051] Step 2: Restore the global low-frequency information to the same scale as X2 through upsampling to obtain the upsampled feature X′2.

[0052] X′2=UP(F2(Z1))

[0053] Among them, UP() represents upsampling, which is used to reconstruct the low-resolution feature map to the same spatial resolution as the original feature, and X′2 represents the feature map restored to the original resolution by upsampling.

[0054] Step 3: Convolve X2 to obtain detail features.

[0055] F3(X2)=X2*K3

[0056] Among them, F3() represents the convolution operation, and the convolution kernel is K3.

[0057] Step 4: Fuse the upsampled features, X2, and detail features, and obtain more receptive fields through the convolution operation F4, so as to ensure that the output feature map has richer local details and higher resolution, and obtain the optimized feature map Y2. The formula is:

[0058] Y′2=F3(X2)·τ(X2+X′2)

[0059] Y2=F4(Y′2)=Y′2*K4

[0060] Among them, τ represents the activation function, τ(X2+X′2) represents the self-calibration weight, F4() represents the convolution operation, the convolution kernel is K4, Y′2 represents the feature map after fusion under the guidance of the self-calibration weight, and Y2 represents the optimized feature map after self-calibration.

[0061] The local detail information is spliced ​​with the optimized feature map to obtain the local self-correction feature Y, the formula is:

[0062] Y=Y1+Y2

[0063] Furthermore, in order to enhance the model's ability to express long-range dependencies, the present invention further combines global context information on the basis of local self-correction, associates local self-correction features with global context, and introduces a global self-correction module. The global self-correction module processes data in the following ways:

[0064] Step 1: Convolve the local self-correction feature Y to obtain the original feature, and perform global average pooling on the original feature to obtain the global information T1.

[0065] F5(Y)=Y*K5

[0066] T1=GAP(F5(Y))

[0067] Among them, F5() represents the convolution operation, K5 represents the convolution kernel of the convolution operation F5, and GAP() represents global average pooling.

[0068] Step 2: Perform compression convolution and extended channel dimension convolution on the global information in sequence to generate an adaptive global perception weight G.

[0069] F6(T1)=T1*K6

[0070] G=F7(F6(T1))=F6(T1)*K7

[0071] Among them, F6() represents compressed convolution, F7() represents expanded convolution, K6 and K7 represent the convolution kernel of compressed convolution and the convolution kernel of expanded convolution, respectively.

[0072] Step 3: Multiply the global perception weight G and the original feature element by element to obtain the corrected feature Y′, which is the output of the dual self-correction network.

[0073] Y′=F5(Y)·G

[0074] The discriminator is used to discriminate the generated samples and output the discrimination results.

[0075] Specifically, the discriminator uses two two-dimensional convolution and maximum pooling layers to process the samples generated by the generator, extract key features, and use a fully connected layer to map the key features into probability values ​​between 0 and 1. If the probability value is close to 1, it means that the generated sample and the real sample are highly similar. If the probability value is close to 0, it means they are not similar.

[0076] The classifier uses a fully connected layer to classify the generated samples and outputs the predicted category label of the generated samples.

[0077] The discrimination results and predicted category labels are used to train the SCF-CVAE-GAN network.

[0078] In the present invention, the output of each convolutional layer is processed by the LeakyReLU activation function to enhance the nonlinear expression ability and alleviate the gradient vanishing problem.

[0079] Furthermore, during the training process, the present invention introduces a frequency-domain regularization strategy into the loss function. By constraining the distribution of generated and real samples in the frequency domain, this ensures the consistency and fidelity of the generated samples in the frequency domain. This addresses the problem that traditional CVAE-GAN generative models primarily rely on time-domain reconstruction loss and KL divergence when optimizing generated samples, lacking effective constraints on samples in other domains.

[0080] Specifically, frequency domain regularization uses Fourier transform to map real and generated samples to the frequency domain, calculating and reducing the error between them, thereby ensuring that the generator maintains consistency in frequency domain information during training. Frequency domain regularization consists of two key components: a basic frequency domain loss term and a multi-scale frequency domain consistency term.

[0081] Among them, the basic frequency domain loss term calculates the mean square error between the generated samples and the real samples at the original scale, and the multi-scale frequency domain consistency loss term calculates the difference between the two after downsampling at multiple scales, thereby enhancing the model's ability to model frequency domain features at different scales. The formula used is:

[0082]

[0083]

[0084] Among them, L freq Represents the basic frequency domain loss, L multi represents the multi-scale frequency domain consistency loss, x and Represent the real sample and the generated sample respectively, x s represents the downsampled samples at scale s, F() represents Fourier transform, and S={2,4} represents the multi-scaling factor.

[0085] Furthermore, the design of the total loss function of the generator includes the weighted sum of frequency domain regularization, reconstruction loss, KL divergence, adversarial loss and classification loss.

[0086] The reconstruction loss measures the difference between the generated samples and the real samples in the spatial domain to ensure the spatial consistency of the generated samples. The KL divergence loss constrains the distribution of the latent space to make it close to the standard normal distribution to improve the quality of the generated samples.

[0087]

[0088] L=L recon +L KLD +αL freq +βL multi +L D +L C

[0089] Among them, L recon represents the reconstruction loss, L KLD represents the divergence loss, L D Denotes the discriminator loss, L C represents the classifier loss, L represents the final loss function, x and represent the real samples and the reconstructed samples respectively; P r and P zrepresent the real data distribution and the reconstructed data distribution, μ and σ represent the mean and standard deviation of the encoder output, z represents random noise, G(Z) represents the generated sample, y1 and Represent the true label and the classifier predicted label respectively, E[] represents the mathematical expectation; α and β represent the weights of the basic frequency domain loss and the multi-scale frequency domain consistency loss, which are trained as hyperparameters during the training process. The weight is used to adaptively adjust according to the training process to optimize the frequency domain characteristics of the generated samples.

[0090] The loss function of the classifier measures the classification accuracy of the model for the input sample by calculating the difference between the predicted category and the true label. The loss function of the classifier L C for:

[0091]

[0092] Among them, y2 represents the true category label, Represents the predicted output of the classifier.

[0093] The loss function of the discriminator consists of the discriminant loss of the real sample and the discriminant loss adversarial loss of the generated sample, which aims to measure the discriminator's ability to distinguish between real samples and generated samples. The loss function of the discriminator is:

[0094]

[0095] Among them, D(x) and Represent the output of the discriminator for real samples and reconstructed samples, P r and P g represent the real data distribution and the reconstructed data distribution respectively.

[0096] S3. Use the trained SCF-CVAE-GAN to generate two-dimensional time-frequency image samples and calculate the structural similarity and distance scores of the generated samples; set the structural similarity and distance score thresholds, and filter out the generated samples that meet the quality requirements based on the structural similarity and distance score thresholds.

[0097] Specifically, the present invention uses Structural Similarity Index Measure (SSIM) and Frechet Inception Distance (FID) as indicators for quantitatively evaluating the quality of generated samples and sample distribution.

[0098] SSIM comprehensively evaluates the structural similarity between generated samples and real samples from three perspectives: brightness, contrast, and structure. The formula is:

[0099]

[0100] Among them, x and y represent the real image and the generated image respectively, μ x and μ y are the means of x and y, σ x and σ y represent the variance of x and y respectively, σ xy represents the covariance of x and y. Constants A1, A2, and A3 are used to avoid the denominator being zero. In this embodiment, the values ​​of A1, A2, and A3 are preferably (0.01×P) 2 、(0.03×P) 2 and 0.5A2, where P represents the dynamic range of pixel values ​​(0 to 255). The SSIM value range is 0 to 1, and a larger value indicates a higher structural similarity between the two images.

[0101] FID measures the distribution of generated samples by calculating the statistical difference between generated images and real images in the feature space. The formula is:

[0102]

[0103] Among them, μ r and μ g Represent the mean of the real image features and the mean of the generated image features, Tr() represents the trace of the matrix, which is used to measure the overall difference between the covariance matrices, ∑ r and ∑ g They represent the covariance matrix of the real image features and the covariance matrix of the generated image features respectively. The smaller the FID value, the closer the distribution of the generated image in the feature space is to the real image.

[0104] The structural similarity threshold is 0.6, the distance score threshold is 120, and the generated samples with a value greater than the structural similarity threshold and less than the distance score threshold are selected as generated samples that meet the quality requirements.

[0105] S4. Mix the generated samples that meet the quality requirements with the two-dimensional time-frequency image to obtain an enhanced dataset.

[0106] Furthermore, if Figure 2 As shown in the figure, the test is conducted based on the bearing dataset and the gear dataset respectively. In this embodiment, the standard CNN neural network is used to perform fault diagnosis on the enhanced dataset. The rotating machinery fault data enhancement model based on the generative adversarial network in the prior art is selected as a comparison. Specifically, the comparison models include: MACGAN, WGAN-GP, MCVAE-GAN, AAE and DCGAN. The comparison results of the quality and distribution of the generated samples of some selected bearing and gear data are shown in the figure. Figure 4 、 Figure 5 、 Figure 6 、 Figure 7As shown, it can be seen intuitively:

[0107] In terms of sample quality, the SSIM values ​​of the samples obtained by data augmentation based on the SCF-CVAE-GAN of the present invention are significantly higher than those of other models, indicating that the generated samples are closest to the real samples in terms of brightness, contrast, and structure. This is mainly due to the use of a dual self-correction network in SCF-CVAE-GAN, which can capture detailed texture features and global structural features during model training and optimize the feature extraction method through a self-calibration mechanism, thereby effectively improving the quality of the generated samples.

[0108] Furthermore, the SSIM values ​​of the enhanced data obtained using MACGAN, WGAN-GP, MCVAE-GAN, AAE, and DCGAN are all above 0.5. This indicates that these models have achieved certain results in generating sample structure and details by improving feature extraction capabilities and loss functions. However, compared with SCF-CVAE-GAN, these models still have certain shortcomings in the richness of texture details and global consistency. This is because SCF-CVAE-GAN introduces a more advanced feature extraction module and adopts a frequency domain regularization strategy, which can more finely constrain and optimize the detail level and global structure of the generated samples, thereby further improving the authenticity and quality of the samples.

[0109] In terms of sample distribution, SCF-CVAEGAN achieves the smallest FID value, indicating that its generated samples are most similar to real samples. This is mainly due to the fact that SCF-CVAE-GAN imposes dual constraints on generated samples in both the spatial and frequency domains, thereby improving the distribution consistency between generated and real samples.

[0110] Figure 8 and Figure 9 The fault diagnosis accuracy curves for the two datasets are shown. As can be seen, fault diagnosis accuracy gradually increases with the addition of more generated samples, indicating that the number of samples in a dataset is a key factor influencing the performance of deep learning-based fault diagnosis. Furthermore, as the sample size increases, the DSFRCVAE-GAN performance curve remains consistently the highest, further demonstrating that the feature distribution of its generated samples is highly similar to that of real samples, effectively improving fault diagnosis accuracy.

[0111] Comparative analysis shows that SCF-CVAE-GAN demonstrates significant performance advantages on both datasets. This superior performance is primarily attributed to its innovative model architecture, which employs a frequency-domain regularization strategy to effectively measure the distribution differences between generated and real samples in the frequency domain. Through a dual self-correcting network, it adaptively extracts global and local features from real samples, generating higher-quality fake samples.

[0112] Finally, it should be noted that the above only describes some embodiments of the present invention. For those skilled in the art, it is conceivable that various changes, modifications, substitutions and deformations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents, and the above-mentioned actions should be covered within the scope of protection of the present invention.

Claims

1. A rotating machinery data enhancement method based on SCF-CVAE-GAN network, characterized in that: include: S1. Obtain vibration signals of rotating machinery with different fault modes; S2. Use continuous wavelet transform to convert the rotating machinery vibration signal into a two-dimensional time-frequency image, and use the two-dimensional time-frequency image to train the SCF-CVAE-GAN network; the SCF-CVAE-GAN network is used to generate two-dimensional time-frequency image samples similar to the rotating machinery fault mode, and the training adopts a frequency domain regularization strategy; S3. Generate two-dimensional time-frequency image samples using the trained SCF-CVAE-GAN and calculate the structural similarity and distance scores of the generated samples. Set the structural similarity and distance score thresholds, and filter out the generated samples that meet the quality requirements based on the structural similarity and distance score thresholds; S4. Mix the generated samples that meet the quality requirements with the two-dimensional time-frequency image to obtain an enhanced dataset.

2. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 1 is characterized in that: The SCF-CVAE-GAN network includes: A generator based on a dual self-correcting network is used to process two-dimensional time-frequency images as real samples to obtain generated samples; The discriminator is used to discriminate the generated samples and output the discrimination results; Classifier, used to classify the generated samples and output the predicted category label of the generated samples; The discrimination results and predicted category labels are used to train the SCF-CVAE-GAN network.

3. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 2 is characterized in that: The generator based on the dual self-correction network includes: an encoder, a decoder, and a dual self-correction network; the data processing of the generator based on the dual self-correction network includes: Step 1: Perform two-dimensional convolution on the real sample to obtain a shallow feature map; Step 2: Use an encoder to encode the shallow feature map into a deep feature map, map the deep feature map to the latent space, and obtain a latent vector; Step 3: Get the true category label corresponding to the real sample, concatenate the true category label with the latent vector, and use the fully connected layer to map it into a conformal feature map with the same shape as the deep feature map; Step 4: Use a decoder to decode the conformal feature map into a low-dimensional feature map; Step 5: Use transposed convolution to reconstruct the low-dimensional feature map into a generated sample with the same size as the real sample; The dual self-correction network is used to process the shallow feature map, deep feature map and low-dimensional feature map separately to obtain the corrected features, which are used to train the SCF-CVAE-GAN network.

4. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 3 is characterized in that: The dual self-correction network includes a local self-correction module and a global self-correction module; the local self-correction module is used to process features to obtain local self-correction features; the global self-correction module is used to perform global context association on the local self-correction features to obtain correction features.

5. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 4 is characterized in that: The local self-correction module includes a local enhancement branch and a self-correction convolution branch. The local self-correction module processes data by: dividing the number of channels of the input feature map X of the dual self-correction network into two, with the height and width unchanged, to obtain two feature maps X1 and X2 of the same size. The local enhancement branch performs convolution processing on X1 to obtain local detail information; the self-correction convolution branch processes X2 to obtain an optimized feature map; and the local detail information is spliced ​​with the optimized feature map to obtain local self-correction features.

6. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 5 is characterized in that: The self-correction convolution branch processes X2 including: Step 1: Use average pooling downsampling to reduce the resolution of X2 and extract global low-frequency information through convolution operation; Step 2: Restore the global low-frequency information to the same scale as X2 by upsampling to obtain the upsampled features; Step 3: Convolve X2 to obtain detail features; Step 4: Fuse the upsampled features, X2 and detail features, and perform convolution to obtain the optimized feature map.

7. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 4 is characterized in that: The global self-correction module processes data including: Step 1: Convolve the local self-correction features to obtain the original features, and perform global average pooling on the original features to obtain global information; Step 2: Perform compression convolution and expansion channel dimension convolution on the global information in sequence to generate adaptive global perception weights; Step 3: Multiply the global perception weight and the original feature element by element to obtain the corrected feature.

8. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 2 is characterized in that: The output of each convolutional layer of the SCF-CVAE-GAN network is processed by the LeakyReLU activation function.

9. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 1 is characterized in that: The structural similarity threshold is 0.6, the distance score threshold is 120, and the generated samples with a value greater than the structural similarity threshold and less than the distance score threshold are selected as generated samples that meet the quality requirements.

10. The rotating machinery data enhancement method based on the SCF-CVAE-GAN network according to claim 1, characterized in that: The frequency domain regularization strategy includes basic frequency domain loss and multi-scale frequency domain consistency loss. The loss function L used in training is: L=L recon +L KLD +αL freq +βL multi +L D +L C Among them, L recon represents the reconstruction loss, L KLD represents the divergence loss, L freq Represents the basic frequency domain loss, L multi represents the multi-scale frequency domain consistency loss, L D Denotes the discriminator loss, L C Represents the classifier loss, α and β represent the basic frequency domain loss weight and the multi-scale frequency domain consistency loss weight, respectively, which are trained as hyperparameters during the training process.

Images