Fault diagnosis sample generation method based on physical constraint adaptive migration network and application and equipment

Abstract

This invention belongs to the field of fault diagnosis technology. It discloses a fault diagnosis sample generation method and its application and device based on a physical constraint adaptive transfer network. The steps are as follows: (1) converting the original vibration signals of the reference device and the monitoring device into a time-series spectrum map through continuous wavelet transform; (2) processing the time-series spectrum map using the spectrum feature encoder of the physical constraint adaptive transfer network to extract multi-scale spectrum features; wherein, the loss function of the physical constraint adaptive transfer network includes content loss, style loss and frequency band energy loss; (3) processing the multi-scale spectrum features using the adaptive style normalization module to fuse the fault content features of the reference device with the machine style features of the monitoring device; (4) processing the fused features using a feature reconstruction decoder to reconstruct and generate synthetic samples. This invention solves the problem of insufficient data authenticity and physical consistency.

Description

Technical Field

[0001] This invention belongs to the technical field of fault diagnosis, and more specifically, relates to a fault diagnosis sample generation method, application, and device based on a physical constraint adaptive migration network. Background Technology

[0002] With the increasing complexity and automation of modern industrial systems, mechanical equipment is being used on a large scale in fields such as power, transportation, aerospace, and manufacturing. The operational safety and reliability of mechanical equipment are directly related to production efficiency and economic benefits; therefore, timely and accurate diagnosis of equipment failures is of great significance.

[0003] Existing mechanical fault diagnosis methods mainly fall into two categories: (1) Physical model-based method: Fault signals are simulated by establishing a dynamic model or finite element model of the mechanical system, and then the working state of the equipment is analyzed. Although this method can reflect the dynamic characteristics of the equipment to a certain extent, it requires accurate structural parameters and boundary conditions, the modeling process is complex, and the simulated signals often differ greatly from the actual operating data, which limits the diagnostic effect.

[0004] (2) Data-driven approach: With the development of deep learning, more and more studies are using models such as convolutional neural networks and generative adversarial networks to learn from historical data, thereby achieving automatic feature extraction and fault identification. However, in real-world scenarios, due to differences in structure, sensor placement, and operating conditions between different devices, cross-device fault migration is often difficult. At the same time, many devices operate stably and have a low probability of failure, resulting in extremely limited fault samples for monitoring devices, making model training prone to overfitting and significantly reducing diagnostic performance.

[0005] Some studies have attempted to generate fault samples and perform data augmentation using methods such as Generative Adversarial Networks (GANs), but existing methods still have the following shortcomings: (1) Lack of physical consistency: Traditional generative adversarial networks (GANs) often operate as a "black box" when generating samples, lacking clear physical prior knowledge constraints. This results in the generated signals being visually similar, but their inherent energy distribution principles may contradict the physical signals in the real world. For example, key fault characteristic frequencies and their harmonic components may be distorted or lost, which will directly affect the reliability of downstream diagnostic tasks.

[0006] (2) Poor diversity and quality of generated samples: Existing generative models are prone to generating samples that lack diversity or fail to capture the true and specific statistical style of the target machine under small sample conditions, resulting in low quality of generated data and limited improvement in diagnostic performance.

[0007] (3) The challenge of "style" alignment in cross-device migration: Due to differences in structural tolerances, sensor layout and operating load, mechanical equipment exhibits a unique "machine style" in its vibration signals. Existing methods struggle to accurately convert the "style" of one device to the "style" of another while preserving the core fault information, thus limiting the accuracy of cross-device diagnosis.

[0008] Therefore, how to generate high-fidelity synthetic samples with physical constraints under small sample conditions and achieve stable fault diagnosis across devices has become a technical problem that urgently needs to be solved. Summary of the Invention

[0009] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a fault diagnosis sample generation method, application and device based on physical constraint adaptive migration network, which aims to solve the problem of insufficient data authenticity and physical consistency in the existing methods.

[0010] To achieve the above objectives, according to one aspect of the present invention, a method for generating fault diagnosis samples based on a physically constrained adaptive migration network is provided, the method comprising the following steps: (1) Convert the original vibration signals of the reference device and the monitoring device into a time-series spectrum using continuous wavelet transform; (2) The spectral feature encoder of the physical constraint adaptive transfer network is used to process the time series spectrogram to extract multi-scale spectral features; wherein, the physical constraint adaptive transfer network also includes an adaptive style normalization module and a feature reconstruction decoder, and its loss function is a multi-objective loss function, which includes content loss, style loss and frequency band energy loss. (3) The adaptive style normalization module is used to process the multi-scale spectral features to fuse the fault content features of the reference device with the machine style features of the monitoring device; (4) The fused features are processed by a feature reconstruction decoder to reconstruct and generate synthetic samples.

[0011] Furthermore, the original vibration signals from the reference and monitoring devices are transformed into time-series spectrograms using complex Morlet wavelets. The corresponding transformation formula is as follows:

[0012] In the formula, and These are the scale and translation parameters, respectively. The complex conjugate of the mother wavelet, is the original vibration signal; k is the time variable, representing the continuous time axis of the original vibration signal.

[0013] Furthermore, the transformation formula of the adaptive style normalization module is:

[0014] In the formula, and These are content features and style features, and These are the results of instance normalization and layer normalization, respectively. and Affine parameters are dynamically generated from style features. It is a learnable interpolation factor, and .

[0015] Furthermore, the formula for the content loss is:

[0016] in For fixed encoders, For decoder, The target features are normalized using AdaSN; The formula for the style loss is:

[0017] in For the activation function of the i-th layer of the pre-trained VGG-19, and These are the mean and standard deviation, respectively. This represents the time-frequency representation of the style samples; The formula for energy loss in the aforementioned frequency band is:

[0018] in This is a time-series spectrum diagram; The frequency components are represented in time-frequency representation; The multi-objective loss function is:

[0019] in, , and These are hyperparameters used to balance content loss, style loss, and frequency band energy loss, respectively.

[0020] Furthermore, among them , , .

[0021] Furthermore, the spectral feature encoder extracts multi-scale features through multiple convolutional and pooling layers; the adaptive style normalization module uses learnable... Dynamic balancing instance normalization and layer normalization.

[0022] Furthermore, the feature reconstruction decoder adopts a symmetrical structure, replacing the pooling layer with nearest neighbor upsampling.

[0023] The present invention also provides a method for diagnosing equipment faults. The method uses the fault diagnosis sample generation method based on physical constraint adaptive transfer network as described above to generate synthetic samples. The synthetic samples and real samples are then combined to form an enhanced training set. The fault diagnosis network is trained based on the enhanced training set, and the trained fault diagnosis network is then used for fault diagnosis.

[0024] The present invention also provides a fault diagnosis sample generation system based on a physical constraint adaptive migration network. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the fault diagnosis sample generation method based on a physical constraint adaptive migration network as described above.

[0025] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the fault diagnosis sample generation method based on physical constraint adaptive migration network as described above, or the device fault diagnosis method as described above.

[0026] In summary, compared with the prior art, the fault diagnosis sample generation method, application, and device based on physical constraint adaptive migration network provided by this invention have the following beneficial effects: 1. The physically constrained adaptive transfer network further includes an adaptive style normalization module and a feature reconstruction decoder, whose loss function is a multi-objective loss function, which includes content loss, style loss, and frequency band energy loss. , and These are hyperparameters used to balance content loss, style loss, and frequency band energy loss, respectively. Content loss constrains content fidelity, style loss ensures style consistency, and frequency band energy loss ensures physical authenticity, making the energy distribution of the generated samples more consistent with physical laws, thus guaranteeing the physical consistency of the generated samples and improving the authenticity of the samples and the reliability of downstream diagnostic tasks.

[0027] 2. The AdaSN module can adaptively balance the normalization of global and local features, achieving accurate imitation of the target machine signal style while perfectly preserving fault content, thus achieving high-precision style transfer, which is superior to traditional GANs methods.

[0028] 3. Compared with traditional GANs methods (such as CycleGAN and DualGAN), the network structure of this invention converges faster, has smaller performance variance, and exhibits stronger training stability and generalization ability, making the training process more stable and efficient.

[0029] 4. By generating a large number of high-quality synthetic samples, the training dataset is effectively expanded, overcoming the bottleneck of data scarcity under small sample conditions. Experiments show that this method can improve the accuracy of cross-machine fault diagnosis by an average of 2.6%-5.8%.

[0030] 5. The framework proposed in this invention does not depend on specific equipment physical models and can be widely applied to cross-machine and small-sample fault diagnosis scenarios of various rotating machinery (such as bearings and gearboxes), demonstrating strong versatility.

[0031] 6. The AdaSN module uses learnable... Dynamically balancing IN and LN achieves style-content fusion. During actual operation, the network continuously adjusts based on the characteristics of the input data. The value of . When the machine style characteristics of the monitoring equipment are relatively prominent, it is necessary to rely more on LN to normalize the features to maintain style consistency. The value will decrease accordingly, making In fusion computing, it plays a larger role; conversely, when the fault content features of the reference device have a more critical impact on the generated samples, IN is needed to better highlight the content features. The value increases, The effect is enhanced. Through this dynamic balancing mechanism, key features of different devices can be effectively integrated to generate samples that conform to the characteristics of the monitoring devices.

[0032] 7. In multi-objective loss functions, , , Sample quality is improved through joint optimization. During training, content loss... The main constraint is the consistency of generated samples with real samples in terms of content features, ensuring that the generated samples accurately reflect the key features of the fault. Style loss. Ensure that the generated samples match the style of the monitoring equipment in terms of style, maintaining the equipment's stylistic characteristics. Frequency band energy loss. Ensuring the spectral energy distribution of the generated samples matches the real signal enhances the samples' realism from a physical perspective. Adjusting the values ​​of these three hyperparameters balances the importance of different loss terms in the optimization process. For example, when the generated samples deviate significantly in terms of content features, the hyperparameters can be appropriately increased. This strengthens the impact of content loss on model training; if the style of the generated samples differs significantly from that of the monitoring equipment, then the impact should be increased. In practical applications, through multiple studies and verifications, it was determined that... , , This set of hyperparameters can achieve good sample generation results in different datasets and scenarios.

[0033] Specifically, the content loss does not directly compare the generated image with the original input image, but rather compares the features extracted from the generated image again by a fixed encoder with the target features t normalized by AdaSN. This approach aligns the optimization target with the stylized latent representation, which can effectively accelerate model convergence.

[0034] Style loss is calculated using the channel-level mean and standard deviation of the matching feature maps. Compared to traditional methods based on Gram matrices, this approach is more computationally efficient, more interpretable, and can preserve global and local appearance features in a controllable manner.

[0035] Frequency band energy loss is key to achieving physical constraints, ensuring that the energy of the generated samples matches the real samples of the target machine not only in visual texture but also in frequency domain distribution. This is crucial for correctly transferring characteristic frequencies caused by different bearing models and rotation speeds, ensuring that key diagnostic information such as fault harmonics and sidebands are physically plausible, thus solving the fundamental problem of traditional GAN-generated samples being 'similar in form but not in essence'.

[0036] 8. The feature reconstruction decoder adopts a symmetrical structure and replaces the pooling layer with nearest neighbor upsampling to avoid checkerboard artifacts.

[0037] 9. The adaptive style normalization module consists of multiple convolutional layers and pooling layers, which can effectively capture transient and steady-state fault components. Attached Figure Description

[0038] Figure 1 This is an overall architecture diagram of PCASNet in an embodiment of the present invention; Figure 2 This is a flowchart of data preprocessing in an embodiment of the present invention, where (a), (b), and (c) correspond to specific steps respectively; Figure 3 This is the training and inference pipeline diagram in this embodiment of the invention; Figure 4 This is a comparison diagram of confusion matrices under different combinations of loss functions in embodiments of the present invention; Figure 5 The present invention provides a flowchart of a fault diagnosis sample generation method based on a physical constraint adaptive migration network. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0040] Please see Figure 1 and Figure 5 This invention provides a method for generating fault diagnosis samples based on a physically constrained adaptive migration network. The method mainly includes the following steps: Step 1: Convert the original vibration signals from the reference device and the monitoring device into a time-series spectrum using continuous wavelet transform.

[0041] This implementation uses complex Morlet wavelets to transform the original vibration signals of the reference device and the monitoring device into a time-series spectrum diagram. The corresponding transformation formula is as follows:

[0042] In the formula, and These are the scale and translation parameters, respectively. The complex conjugate of the mother wavelet, is the original vibration signal; k is the time variable, representing the continuous time axis of the original vibration signal.

[0043] In one implementation, during the data preprocessing stage, the original vibration signal is first divided into fixed-length windows of 1024 points. This operation helps to transform the continuous vibration signal into discrete data blocks that are easy to process later, with each window reflecting the operating status of the equipment within a specific short period. Subsequently, the segmented data is standardized to eliminate the influence of differences in signal amplitude, ensuring all data are at the same dimension, facilitating subsequent model learning. After standardization, the data is converted into a 224×224 time-series spectrum using continuous wavelet transform (CWT). When using complex Morlet wavelets for CWT, the center frequency setting is crucial and needs to be adjusted according to the actual signal characteristics. For example, for equipment signals dominated by high-frequency vibration, the center frequency can be appropriately increased to better capture high-frequency fault characteristics; while for cases with abundant low-frequency signals, the center frequency should be correspondingly decreased.

[0044] Step 2: The time-series spectrogram is processed using a spectral feature encoder of a physically constrained adaptive transfer network to extract multi-scale spectral features. The physically constrained adaptive transfer network also includes an adaptive style normalization module (AdaSN) and a feature reconstruction decoder. Its loss function is a multi-objective loss function, which includes content loss, style loss and frequency band energy loss.

[0045] The adaptive style normalization module consists of multiple convolutional and pooling layers, effectively capturing transient and steady-state fault components. The multi-scale spectral features include both transient and steady-state fault components.

[0046] The transformation formula for the adaptive style normalization module is:

[0047] In the formula, and These are content features and style features, and These are the results of instance normalization and layer normalization, respectively. and Affine parameters are dynamically generated from style features. It is a learnable interpolation factor, and .

[0048] As a learnable interpolation factor, dynamically balance instance normalization and layer normalization, and through The constraint is within the range [0,1]. For learning rate, The gradient update vector is calculated. The formula for the content loss is:

[0049] in For fixed encoders, For decoder, The target features are normalized using AdaSN; The formula for the style loss is:

[0050] in For the activation function of the i-th layer of the pre-trained VGG-19, and These are the mean and standard deviation, respectively. This is the time-frequency representation of the style sample.

[0051] The formula for energy loss in the aforementioned frequency band is:

[0052] in This is a time-series spectrum diagram; The frequency component is represented by the time-frequency representation.

[0053] The multi-objective loss function is:

[0054] in, , and These are hyperparameters used to balance content loss, style loss, and frequency band energy loss, respectively; content loss constrains content fidelity, style loss ensures style consistency, and frequency band energy loss ensures physical authenticity.

[0055] In one implementation, please refer to Figure 2 , Figure 3 and Figure 4The spectral feature encoder extracts multi-scale features through multiple convolutional and pooling layers. Taking Conv1 (3, 64, 3×3) + ReLU as an example, this layer uses a 3×3 convolutional kernel, with 3 input channels (corresponding to RGB image channels or similar feature channels), and outputs 64 feature maps. Convolutional operations can extract signal features within local regions, while the ReLU activation function introduces non-linearity to the model, enhancing its ability to express complex features. Subsequent layers such as Conv2 and Conv3 follow the same pattern, gradually extracting higher-level and more abstract features through different convolutional kernel sizes, channel counts, and the application of the ReLU activation function. Pooling layers, such as MaxPool2d 2×2, reduce the resolution of the feature maps, decrease computational cost, and retain key feature information, helping the model learn translation-invariant features and improving its ability to identify fault features at different locations.

[0056] The AdaSN module is learnable. Dynamically balancing IN and LN achieves style-content fusion. During actual operation, the network continuously adjusts based on the characteristics of the input data. The value of . When the machine style characteristics of the monitoring equipment are relatively prominent, it is necessary to rely more on LN to normalize the features to maintain style consistency. The value will decrease accordingly, making In fusion computing, it plays a larger role; conversely, when the fault content features of the reference device have a more critical impact on the generated samples, IN is needed to better highlight the content features. The value increases, The effect is enhanced. Through this dynamic balancing mechanism, key features of different devices can be effectively integrated to generate samples that conform to the characteristics of the monitoring devices.

[0057] The feature reconstruction decoder employs a symmetrical structure, restoring feature dimensions through upsampling to generate synthetic samples. In this process, taking Upsample1 (nearest neighbor upsampling × 2) as an example, it enlarges the feature map output from the previous layer in spatial dimension, doubling the resolution. Subsequent convolutional layers such as Conv10 and Conv11 further process the enlarged feature map, integrating and refining the features through convolution operations to gradually reconstruct and generate complete synthetic samples. Using nearest neighbor upsampling instead of pooling layers effectively avoids checkerboard artifacts, ensuring the image quality and feature accuracy of the generated samples.

[0058] In multi-objective loss functions, , , Sample quality is improved through joint optimization. During training, content loss... The main constraint is the consistency of generated samples with real samples in terms of content features, ensuring that the generated samples accurately reflect the key features of the fault. Style loss. Ensure that the generated samples match the style of the monitoring equipment in terms of style, maintaining the equipment's stylistic characteristics. Frequency band energy loss. Ensuring the spectral energy distribution of the generated samples matches the real signal enhances the samples' realism from a physical perspective. Adjusting the values ​​of these three hyperparameters balances the importance of different loss terms in the optimization process. For example, when the generated samples deviate significantly in terms of content features, the hyperparameters can be appropriately increased. This strengthens the impact of content loss on model training; if the style of the generated samples differs significantly from that of the monitoring equipment, then the impact should be increased. In practical applications, through multiple studies and verifications, it was determined that... , , This set of hyperparameters can achieve good sample generation results in different datasets and scenarios.

[0059] Step 3: The adaptive style normalization module is used to process the multi-scale spectral features to fuse the fault content features of the reference device with the machine style features of the monitoring device.

[0060] The AdaSN module uses learnable channel coefficients to dynamically balance instance normalization (IN) and layer normalization (LN).

[0061] Step four: The fused features are processed using a feature reconstruction decoder to reconstruct and generate synthetic samples.

[0062] The feature reconstruction decoder adopts a symmetrical structure and replaces the pooling layer with nearest neighbor upsampling to avoid checkerboard artifacts.

[0063] This invention employs the CWRU, BJTU-RAO, and HUSTbearing datasets, setting up two cross-device scenarios. In the first scenario, device A from the CWRU dataset is used as the reference device, and device B from the BJTU-RAO dataset is used as the monitoring device. Fault samples for device B are generated using PCASNet and combined with a small number of real samples from device B to form an augmented dataset, which is then used to train a classifier for fault diagnosis. In the second scenario, device C from the HUSTbearing dataset is used as the reference device, and cross-device sample generation and fault diagnosis are performed again for device B. Experimental results show that, compared with existing GAN-based methods, the samples generated by this invention significantly improve diagnostic accuracy, with an average improvement of 2.6%–5.8% in diagnostic accuracy across different scenarios, fully validating the effectiveness and superiority of the proposed method.

[0064] The present invention also provides a method for diagnosing equipment faults. The method uses the fault diagnosis sample generation method based on physical constraint adaptive transfer network as described above to generate synthetic samples. The synthetic samples and real samples are then combined to form an enhanced training set. The fault diagnosis network is trained based on the enhanced training set, and the trained fault diagnosis network is then used for fault diagnosis.

[0065] This invention also provides a fault diagnosis sample generation system based on a physically constrained adaptive migration network, which includes: The data preprocessing module is used to convert the raw vibration signals from the reference device and the monitoring device into a time-series spectrogram through continuous wavelet transform; A spectrum feature encoder is used to process time-series spectrograms to extract multi-scale spectrum features. The spectrum feature encoder is a component of the physical constraint adaptive transfer network, which also includes an adaptive style normalization module (AdaSN) and a feature reconstruction decoder. Its loss function is a multi-objective loss function, which includes content loss, style loss and frequency band energy loss. An adaptive style normalization module is used to process multi-scale spectral features to fuse the fault content features of the reference device with the machine style features of the monitoring device. The feature reconstruction decoder is used to process the fused features to reconstruct and generate synthetic samples.

[0066] The present invention also provides a fault diagnosis sample generation system based on a physical constraint adaptive migration network. The system includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the fault diagnosis sample generation method based on a physical constraint adaptive migration network as described above.

[0067] The present invention also provides a computer-readable storage medium storing machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the fault diagnosis sample generation method or equipment fault diagnosis method based on physical constraint adaptive migration network as described above.

[0068] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating fault diagnosis samples based on a physically constrained adaptive migration network, characterized in that, The method includes the following steps: (1) Convert the original vibration signals of the reference device and the monitoring device into a time-series spectrum using continuous wavelet transform; (2) The spectral feature encoder of the physical constraint adaptive transfer network is used to process the time series spectrogram to extract multi-scale spectral features; wherein, the physical constraint adaptive transfer network also includes an adaptive style normalization module and a feature reconstruction decoder, and its loss function is a multi-objective loss function, which includes content loss, style loss and frequency band energy loss. (3) The adaptive style normalization module is used to process the multi-scale spectral features to fuse the fault content features of the reference device with the machine style features of the monitoring device; (4) The fused features are processed using a feature reconstruction decoder to reconstruct and generate synthetic samples; The transformation formula for the adaptive style normalization module is: In the formula, and These are content features and style features, and These are the results of instance normalization and layer normalization, respectively. and For affine parameters dynamically generated from style features, It is a learnable interpolation factor, and .

2. The fault diagnosis sample generation method based on physical constraint adaptive migration network as described in claim 1, characterized in that: The original vibration signals from the reference and monitoring devices are transformed into time-series spectrograms using complex Morlet wavelets. The corresponding transformation formula is as follows: In the formula, and These are the scale and translation parameters, respectively. The complex conjugate of the mother wavelet, The original vibration signal is represented by k, which is a time variable representing the continuous time axis of the original vibration signal. These are the target features after AdaSN normalization.

3. The fault diagnosis sample generation method based on physically constrained adaptive migration networks as described in claim 1, characterized in that: The formula for the content loss is: in For fixed encoders, For decoder, The target features are normalized using AdaSN; The formula for the style loss is: in For the activation function of the i-th layer of the pre-trained VGG-19, and These are the mean and standard deviation, respectively. This represents the time-frequency representation of the style samples; The formula for energy loss in the aforementioned frequency band is: in This is a time-series spectrum diagram; The frequency components are represented in time-frequency representation; The multi-objective loss function is: in, , and These are hyperparameters used to balance content loss, style loss, and frequency band energy loss, respectively.

4. The fault diagnosis sample generation method based on physical constraint adaptive migration network as described in claim 3, characterized in that: in , , .

5. The fault diagnosis sample generation method based on physically constrained adaptive migration networks as described in any one of claims 1-4, characterized in that: The spectral feature encoder extracts multi-scale features through multiple convolutional and pooling layers; the adaptive style normalization module uses learnable... Dynamic balancing instance normalization and layer normalization.

6. The fault diagnosis sample generation method based on physically constrained adaptive migration networks as described in any one of claims 1-4, characterized in that: The feature reconstruction decoder adopts a symmetrical structure and replaces the pooling layer with nearest neighbor upsampling.

7. A method for diagnosing equipment faults, characterized in that: The diagnostic method uses the fault diagnosis sample generation method based on physical constraint adaptive transfer network as described in any one of claims 1-6 to generate synthetic samples, then combines the synthetic samples with real samples to form an enhanced training set, and trains the fault diagnosis network based on the enhanced training set, and then uses the trained fault diagnosis network to perform fault diagnosis.

8. A fault diagnosis sample generation system based on a physically constrained adaptive migration network, characterized in that: The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the fault diagnosis sample generation method based on physical constraint adaptive migration network as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the fault diagnosis sample generation method based on a physical constraint adaptive migration network as described in any one of claims 1-6 or the equipment fault diagnosis method as described in claim 8.

Images