A motor fault data enhancement method based on spectral normalization conditional generative adversarial network and high-order statistics screening

Abstract

This invention discloses a method for enhancing motor fault data based on a spectral normalization-conditional generative adversarial network (GAN) and high-order statistics for screening, belonging to the field of motor fault diagnosis and data augmentation. The method includes: first, constructing a filter incorporating three aspects: energy efficiency, physical stability, and fault saliency, and a generative adversarial network (GAN) incorporating spectral normalization; second, generating initial fault signal samples using the generator and filtering them through the filter; inputting the filtered fault signal samples and collected real fault signal data into a discriminator for discrimination; and updating the GAN parameters based on the discriminator's discrimination results to obtain a trained generator; finally, using the filter and the trained generator, specifying the fault type, yields new fault signal samples corresponding to that fault type. This invention effectively enhances the diversity of data samples, providing a reference and approach for the practical deployment and optimization of deep learning models in industry.

Description

Technical Field

[0001] This invention belongs to the field of motor fault diagnosis and data augmentation, specifically involving a motor fault data augmentation method based on spectral normalization conditional generative adversarial networks and high-order statistical screening. Technical Background With the advancement of intelligent manufacturing, the health status management of motors, which serve as the core of industrial power in fields such as semiconductor production and precision instrument processing, has become crucial for ensuring production safety and efficiency. Currently, data-driven fault diagnosis methods based on deep learning, due to their powerful feature extraction capabilities, have gradually replaced traditional signal processing methods and become the mainstream technology. However, in practical situations, the performance of deep learning models is highly dependent on the scale and quality of the training data. Ideally, the training set should contain sufficient and class-balanced data. However, during the entire life cycle of a motor, faults such as early bearing pitting or inner ring cracks are extremely rare, and if they do occur, the motor will immediately shut down to protect the equipment due to its protection mechanism. Deliberately damaging the equipment to obtain fault signals presents high costs, making it difficult to collect large-scale, high-quality, labeled fault signals. This results in an excessively high false negative rate in the fault detection system, failing to meet industrial standards.

[0002] Currently, in the field of data augmentation for motor fault diagnosis, before the widespread adoption of deep learning, the most commonly used methods in industry were SMOTE (Synthetic Minority Over-sampling Technique) and physical augmentation methods based on signal processing. The principle of SMOTE is to select any point on the line connecting two fault signal samples in the feature space as a new sample, which is equivalent to linear interpolation. Common physical augmentation methods include adding Gaussian white noise, time shifting, amplitude scaling, and resampling. After the widespread adoption of deep learning technology, methods such as Deep Convolutional Adversarial Networks (DCGANs) are used. These networks extract fault signal features and then generate new samples through a game between the generator and the discriminator.

[0003] However, because motor vibration signals are nonlinear and non-stationary time series, the samples generated by SMOTE often do not conform to physical laws and may even destroy the periodic impact characteristics in the fault signal. The signals in the early stages of a motor fault are often weak and easily masked by minor disturbances. Physical enhancement methods can easily mask fault features (such as weak periodic pulses), leading to a decrease in the signal-to-noise ratio. Furthermore, neither of these methods essentially creates new sample patterns for the network to learn; they simply add "filters" or "transitions" to the existing signals, easily causing overfitting. Deep learning methods, such as the DCGAN network, are prone to generating samples with class imbalance. The network may focus on generating fault signal types that are easy to fool the discriminator, while losing other fault signal types, resulting in class imbalance and causing the deep learning model to lose its ability to distinguish the missing fault signal types.

[0004] Therefore, there is an urgent need to propose a new data augmentation method for motor fault diagnosis. Summary of the Invention

[0005] To address the existing technical problems, this invention provides a method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening. Under the premise of scarce motor fault signal samples, the fault signals are enhanced through conditional adversarial networks.

[0006] The specific technical solution is as follows: S1 introduces spectral normalization to construct an adversarial generative network containing a generator and a discriminator; and constructs a filter from three aspects: energy efficiency, physical stability and fault saliency. S2, collects and labels fault signals during motor operation to obtain real fault signal data containing fault type labels; S3, the generator randomly generates multiple pseudo-fault signals according to the specified fault type, and inputs all pseudo-fault signal data into the filter for filtering to obtain high-quality pseudo-fault signal data; S4, input the fault signal dataset consisting of real fault signal data and high-quality pseudo fault signal data into the discriminator to obtain the final score used to distinguish between real fault signal data and high-quality pseudo fault signals; the discriminator further determines, based on the threshold and the final score, whether each fault signal data in the fault signal dataset comes from real fault signal data or high-quality pseudo fault signal data. S5. Based on the discriminator's discrimination result, the parameters of the adversarial generative network are updated in reverse and S3-S4 are repeated until the network converges. At this point, the trained generator and the finally generated high-quality pseudo-fault signal data are obtained. The finally generated high-quality pseudo-fault signal data is the new fault signal sample after data augmentation.

[0007] Furthermore, in S1, the generator is used to receive a specified fault type and the number of pseudo fault signals generated, and output a specified number of pseudo fault signals of the corresponding fault type; the generator contains several cascaded AdaINResBlocks, which are used to control the fault type of the generated fault signals.

[0008] Furthermore, the specific process of AdaINResBlock is as follows: After receiving the input data, AdaINResBlock first encodes the fault type label to obtain the corresponding variance and mean, and then uses the variance, mean and input data to perform the first modulation and upsampling to obtain the upsampled output features. Secondly, the input data and the upsampled output features are concatenated through residuals and then normalized to obtain the input data for secondary modulation. Based on the secondary modulation input data, a second modulation is performed to obtain the final output result.

[0009] Furthermore, the first modulation specifically includes: ; in, For input data, For the modulation output result, These are the variance and mean after the fault type labels are encoded, respectively. The method for the second modulation is the same as that for the first modulation, wherein The input data uses secondary modulation, and the variance and mean are obtained by re-encoding the fault type label.

[0010] Furthermore, the discriminator specifically comprises: After receiving the fault signal dataset, the discriminator first maps each fault signal data in the dataset to a feature vector. Then, the feature vector is sequentially processed through several cascaded ResDownBlocks for perception extraction. The results of the perception extraction are then sequentially processed through an activation layer and a global average pooling layer. The feature vector output by the global average pooling layer is used to calculate the final score of each fault signal data.

[0011] Furthermore, in step S3, for each pseudo-fault signal data generated by the generator, the filtering process is as follows: First, the filter receives pseudo-fault signal data and calculates the mean of each pseudo-fault signal data. Secondly, the corrected pseudo-fault signal is obtained by subtracting the corresponding mean from the pseudo-fault signal data. Finally, the filter sequentially screens the corrected pseudo-fault signals based on three aspects: energy efficiency, physical stability, and fault significance, ultimately obtaining high-quality pseudo-fault signal data.

[0012] Furthermore, the specific process by which the filter selects the corrected pseudo-fault signals based on three aspects—energy efficiency, physical stability, and fault significance—is as follows: If the root mean square of the corrected pseudo-fault signal is less than the energy efficiency threshold, the corresponding pseudo-fault signal is deleted; otherwise, it enters the physical stability screening. If the absolute value of the peak value of the corrected pseudo-fault signal is greater than the physical stability threshold, the corresponding pseudo-fault signal is deleted; otherwise, it enters the fault significance screening. If the excess kurtosis of the corrected pseudo-fault signal is less than the fault significance threshold, the corresponding pseudo-fault signal is deleted; otherwise, it is retained.

[0013] Furthermore, the root mean square (RMS) specifically refers to: ; in, Let be the signal value at time t, and L be the total signal duration. It is the root mean square; The excess kurtosis is specifically as follows: ; Where D is the standard kurtosis of the signal. This is an excessive kurtosis.

[0014] Furthermore, in S4, the final score of each fault signal data is obtained by weighted sum of an unconditional score and a conditional score; the unconditional score is obtained by calculating the confidence level of the fault signal data; the conditional score is obtained by calculating the similarity between the fault signal data and the corresponding fault type label.

[0015] Furthermore, in S4, the discrimination process of the discriminator is specifically as follows: The scoring thresholds for real fault signal data and spurious fault signal data are set as the first scoring threshold and the second scoring threshold, respectively. The discriminator uses a first score threshold to distinguish between real and fake fault signal data. If the final score of the real fault signal data is higher than the first score threshold, the fault signal data is determined to be real fault signal data. Otherwise, the fault signal data is determined to be fake fault signal data, indicating that the discriminator's judgment result is incorrect. The discriminator uses a second scoring threshold to distinguish between genuine and fake fault signal data. If the final score of the fake fault signal data is lower than the second scoring threshold, the fault signal data is determined to be fake fault signal data; otherwise, the fault signal data is determined to be genuine fault signal data, indicating that the discriminator's judgment result is incorrect.

[0016] The beneficial effects of this invention are: (1) The neural network in this invention learns the physical structure and laws of the fault signal while learning the fault signal features. Compared with the simple interpolation generation in SMOTE, the new sample generated by this invention is more in line with the real physical laws. Moreover, this invention does not cover up physical characteristics such as periodic physical impacts when performing data augmentation. Compared with the signal-to-noise ratio decrease caused by adding interference in traditional physical augmentation methods, the diagnostic network can more easily extract effective features from the augmented signal generated by this invention.

[0017] (2) The generator network of the present invention generates a new signal with fault signal characteristics from noise, rather than transforming the original signal, thus greatly enhancing the diversity of data samples.

[0018] (3) Compared with classic adversarial generative networks such as DCGAN, this invention relies on the introduction of the AdaINResBlock module and quality filter to achieve effective control over the generation of fault signal types and quality, and solves the problems of signal class imbalance and unstable generation effect in the process of data augmentation by generative networks. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the AdaINResBlock structure constructed according to the present invention.

[0020] Figure 2 This is a schematic diagram of the generator structure constructed according to the present invention.

[0021] Figure 3 This is a schematic diagram of the ResDownBlock structure constructed in this invention.

[0022] Figure 4 This is a schematic diagram of the discriminator structure constructed according to the present invention.

[0023] Figure 5 This is an overall framework diagram of the method of the present invention. Detailed Implementation

[0024] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0025] If a function For any input , The following inequalities are satisfied: Where K is the Lipschitz constant.

[0026] Then we call It satisfies K-Lipschitz continuity. Based on this, when the function... When using a neural network structure, since a neural network is a composite of multiple networks, for an input x, the final output of the neural network can be expressed as: in, This is the output of the Lth layer of the neural network.

[0027] Therefore, the Lipschitz constant of the neural network can be obtained. Less than or equal to the product of the Lipschitz constants of each layer: in, Let be the Lipschitz constant of the Lth layer of the neural network.

[0028] Therefore, controlling the Lipschitz constant of a neural network is equivalent to controlling the linear transformation of the weight matrix in each layer (convolutional or fully connected layer) of the neural network. The Lipschitz constant.

[0029] weight matrix spectral norm (i.e., matrix) The largest singular value measures the maximum factor by which the layer can "stretch" the input value. Spectral norm The calculation formula is: in, Let be the vector input from the previous layer in each iteration. It is an L2 norm.

[0030] It can be observed that if the spectral norm of a certain layer The magnitude is very large; even if the input changes very little, the output may change drastically, which can lead to training instability (such as gradient explosion and mode collapse).

[0031] To achieve spectral normalization, the weight matrix is... Each element in the matrix is ​​divided by the spectral norm corresponding to that layer. This yields a new weight matrix. Used to replace the current weight matrix : This is the core formula for spectral normalization. After this operation, the new matrix... The spectral norm is always equal to 1.

[0032] After performing the above operations on each layer of the neural network, the Lipschitz constant of the entire neural network is constrained to a computable upper bound (calculated by the product of the spectral norms of each layer). Since the spectral norm of each layer is normalized to 1, the upper bound of this product is effectively controlled, thereby stabilizing training, preventing gradient explosion or gradient vanishing, and ensuring network convergence.

[0033] Based on the above principles, this invention addresses the problem of poor fault identification capabilities of diagnostic networks in industrial production scenarios where neural networks are used for motor fault diagnosis due to the scarcity of signal samples and high annotation costs. It proposes a motor fault data augmentation method based on spectral normalization conditional generative adversarial networks and high-order statistical screening. By extracting features from limited fault signals, a large number of fault signals with similar features are generated, thereby expanding the training set and enhancing the diagnostic network's identification capabilities.

[0034] like Figure 5 As shown, the enhancement method of the present invention specifically includes the following four parts: I. Prepare the training dataset; In this section, the same acquisition equipment and acquisition parameters are used to collect vibration data of multiple motors. Since all the fault signals actually collected are manually labeled according to the fault type, fault signals containing type information are obtained.

[0035] II. Constructing an adversarial generative network; In this section, the first step is to build a data loading module, which is used to convert the acquired fault signal data into a data type that can be read by the generative adversarial network through code.

[0036] Secondly, a spectral normalized conditional generative adversarial network is constructed, consisting of a generator, a discriminator, and an optimizer. The discriminator and optimizer are only used during the network training phase, and all convolutional layers, fully connected layers, and coding layers in the network are subject to spectral normalization.

[0037] The generator is constructed using AdaINResBlock and a self-attention mechanism. It receives the specified fault data type and the number of samples to be generated, and outputs the specified number of fault samples. The overall structure is as follows: Figure 2As shown, it includes a fully connected layer, several AdaINResBlocks (4 AdaINResBlocks in this specific embodiment), an instance normalization layer, an activation layer and a convolutional layer connected in sequence. The inter-layer activation function is LeakyReLU, and a self-attention layer is added before the last AdaINResBlock to enhance the network's ability to model long sequences.

[0038] AdaINResBlock is used to control the type of generated signal to solve the problem of imbalance between signal classes. The overall structure is as follows: Figure 1 As shown, it mainly consists of two coding layers composed of multilayer perceptrons, two instance normalization layers, two type information modulation modules, two activation layers, one upsampling layer and one convolutional layer.

[0039] The specific operation process of AdaINResBlock is as follows: First, the type information of the fault signal is extracted by passing it through an encoding layer composed of a multiple perceptron (MLP1), and the resulting long vector includes two parts: variance and mean. Used for primary modulation; Secondly, after eliminating the statistical features of the input data using an instance normalization layer, add... The feature completes one modulation, and the modulation method is as follows: in, As input features, The output characteristics are those obtained after one modulation. Upsampling is performed to obtain the upsampled output features.

[0040] Finally, using the same method, the type information of the fault signal is re-passed through the coding layer composed of a multi-sensor MLP2 to obtain... Then, the input statistical feature data and the upsampled output features are connected through residuals (i.e., the two data are added by channel and then divided by 2) to ensure network stability, and finally the input data of secondary modulation is obtained.

[0041] After performing instance normalization on the second-modulated input data, combined with The signal is modulated twice and then passed through an activation layer and a convolutional layer to obtain the output of AdaINResBlock.

[0042] During the training phase, the discriminator receives the training set and new samples from the generator that have been filtered by the filter, and outputs the final score for each new sample. The overall structure is as follows: Figure 4As shown, it consists of a convolutional layer, several ResDownBlocks (4 ResDownBlocks in this specific embodiment), an activation layer, an average pooling layer, and a decision module. The inter-layer activation function uses LeakyReLU, and a self-attention layer is added in the middle of all ResDownBlocks to perceive and judge the global consistency features of the signal, thereby enhancing the decision-maker's ability to judge long sequences.

[0043] After receiving input data, the discriminator first maps the signal features to a high-dimensional space using a one-dimensional convolutional layer. Then, it sequentially uses several ResDownBlock modules, each containing a self-attention layer, to extract and perceive the data features. The extracted results are then passed through an activation layer and input into a global average pooling layer. The feature vector output by the global average pooling layer is used as the basis for decision-making and fed into the decision module. The decision module evaluates and scores the data using both unconditional and conditional scores, and the final score is determined by both.

[0044] Among them, the unconditional scoring first passes the feature vector through a fully connected layer, and then calculates the confidence of the new sample. It is responsible for judging whether the signal features conform to the physical characteristics of the motor fault signal, regardless of the category to which the fault signal belongs, thereby preventing the generator from generating signals that have no physical meaning. Conditional scoring evaluates whether a signal conforms to its class characteristics by judging the similarity between the signal's feature vector and the feature vector of the signal's class label in the feature space. The pooled feature vector is encoded by an encoding layer, and then the similarity with the signal's class label in the training set is calculated to obtain the conditional score.

[0045] The final score of the new sample is obtained by weighting the unconditional and conditional scores proportionally. The discriminator aims to distinguish between real and generated fault signals, outputting the highest possible score for real fault signals and the lowest possible score for generated signals. Therefore, this invention sets a threshold for both real and generated fault signals. For real signals, the score should be higher than the real-signal threshold, and for generated signals, the score should be lower than the low-signal threshold. This provides guidance for the generator's training. If the scores do not meet expectations, the discriminator should continue training iterations.

[0046] The overall structure of ResDownBlock is as follows: Figure 3As shown, it consists of two activation layers, two convolutional layers, and one average pooling layer. The input data for the second activation layer is obtained by concatenating the output data of the first convolutional layer with the original input through residuals. Unlike commonly used residual downsampling modules, the ResDownBlock residual downsampling module in this invention does not use instance normalization or batch normalization methods for constraints, but instead uses spectral normalization methods. This is because in motor fault diagnosis, abnormal amplitude is often a very critical feature, but in the presence of noise, normalization methods can dilute or even erase this feature.

[0047] III. Generating Signal Augmentation Datasets This invention constructs a signal quality filter based on three aspects: energy efficiency, physical stability, and fault significance. Each fault signal sample generated by the generator is first sent to the filter for screening to remove unqualified signals. The process continues until the number of generated signals reaches a specified number before the sample is judged by the discriminator.

[0048] Specifically: After the adversarial network completes training, the generator will generate samples according to the specified fault data type and quantity, expanding the sample set of the fault signal dataset. To ensure the quality of the generated samples, this invention introduces a sample quality filter to remove unqualified samples until the number of qualified samples generated reaches the specified requirement. The filtered samples are then input into a discriminator for discrimination.

[0049] The filter design is based on three aspects: energy efficiency, physical stability, and fault significance.

[0050] First, calculate each fault signal sample. mean By signal Subtract the mean This eliminates zero-point drift caused by objective factors such as measurement equipment errors. in, This is the revised new sample.

[0051] Regarding energy validity, the filter uses the root mean square (RMS) to measure the validity of a signal. If the RMS is below a certain threshold, the signal is considered too weak to be valid, and the signal sample is discarded. (RMS) The calculation method is as follows: in, This represents the total duration of the fault signal samples.

[0052] Regarding physical stability, the absolute value of the peak value is used as the criterion for evaluating the physical stability of the generated signal. If the absolute value of the peak value of the generated signal exceeds the set threshold, the signal is considered to be a numerical explosion signal generated by network instability, and the signal sample is removed.

[0053] Excess kurtosis is used to measure fault significance. Kurtosis is a measure of the fourth-order standard central moment of a probability distribution. The standard kurtosis of each corrected fault sample is calculated as follows: Analysis of the above formula reveals that when the signal follows a standard normal distribution, the corresponding standard kurtosis is 3. Therefore, excess kurtosis is defined as follows: When the excess kurtosis of the generated signal is less than a certain threshold, the signal is considered to be close to Gaussian noise and does not have fault significance, and the signal sample is removed.

[0054] If the generated signal passes the energy validity, physical stability, and fault significance tests of the filter, it is considered to have passed the quality test of the filter and is saved as the final generation result; if the generated signal fails the test, it is discarded.

[0055] Repeat steps two and three until fault signal samples that meet the quantity requirements and pass the screening are generated, and then form a signal augmentation dataset.

[0056] IV. Adversarial Network Training In this step, the signal augmentation dataset is loaded into the Generative Adversarial Network (GAN) using the data loading module. The signal data types are numbered by integer and then input into the coding layers MLP1 and MLP2 for type information extraction, thereby converting the signal type number information into a dense vector. Then, all parameters in the GAN are trained according to the pre-set hyperparameters until the network is stable, and the weight file when the network is stable is saved.

[0057] The network uses the Adam optimizer and the Hinge Loss function as the loss function. The adversarial generative network is updated in reverse based on the loss function until the loss function converges, resulting in the trained generator corresponding to the current fault type, and the weight file is saved.

[0058] The implementation details and parameter settings are as follows: First, set the relevant parameters for the global optimizer Adam, including momentum. Set it to 0 to prevent training oscillations caused by momentum accumulation, and correspondingly reduce momentum. Set the learning rate to 0.99; set the generator learning rate to 1e-4 and the discriminator learning rate to twice that of the generator, i.e., 2e-4. During the learning process, maintain dual time scale updates, i.e., update the discriminator twice for every update of the generator. This ensures that the discriminator's recognition ability is ahead of the generator's signal imitation ability and provides effective guidance to the generator during the update process.

[0059] Secondly, the number of initial mapping channels for high-dimensional features in the discriminator is set to 0.5 times the signal length, and the number of embedding channels for type information in the conditional scoring part of the discriminator is set to 0.5 times the signal length.

[0060] Finally, the generator input noise length is set to be the same as the training sample signal length, and the number of feature embedding channels used for class feature modulation in AdaINResBlock is set to 0.5 times the signal sample length.

[0061] During the inference phase, the generator is trained using the corresponding weight file based on the specified fault type and number of fault samples. Then, the filter selects the sample data generated by the generator, and the final signal augmentation dataset is the new fault signal sample. The fault type generated by the generator during the inference phase must not exceed the range of fault types in the training phase.

[0062] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening, characterized in that, include: S1 introduces spectral normalization to construct an adversarial generative network containing a generator and a discriminator; A filter was constructed based on three aspects: energy efficiency, physical stability, and fault significance. S2, collects and labels fault signals during motor operation to obtain real fault signal data containing fault type labels; S3, the generator randomly generates multiple pseudo-fault signals according to the specified fault type, and inputs all pseudo-fault signal data into the filter for filtering to obtain high-quality pseudo-fault signal data; S4. Input the fault signal dataset consisting of real fault signal data and high-quality pseudo fault signal data into the discriminator to obtain the final score used to distinguish between real fault signal data and high-quality pseudo fault signals. The discriminator further determines, based on the threshold and the final score, whether each fault signal in the fault signal dataset comes from real fault signal data or high-quality pseudo-fault signal data. S5. Based on the discriminator's discrimination result, the parameters of the adversarial generative network are updated in reverse and S3-S4 are repeated until the network converges. At this point, the trained generator and the finally generated high-quality pseudo-fault signal data are obtained. The finally generated high-quality pseudo-fault signal data is the new fault signal sample after data augmentation.

2. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening as described in claim 1, characterized in that, In step S1, the generator is used to receive a specified fault type and the number of pseudo fault signals generated, and outputs a specified number of pseudo fault signals of the corresponding fault type; the generator contains several cascaded AdaINResBlocks, which are used to control the fault type of the generated fault signals.

3. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening as described in claim 2, characterized in that, The specific process of AdaINResBlock is as follows: After receiving the input data, AdaINResBlock first encodes the fault type label to obtain the corresponding variance and mean, and then uses the variance, mean and input data to perform the first modulation and upsampling to obtain the upsampled output features. Secondly, the input data and the upsampled output features are concatenated through residuals and then normalized to obtain the input data for secondary modulation. The input data is modulated a second time to obtain the final output result.

4. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening as described in claim 3, characterized in that, The first modulation is specifically as follows: ； in, For input data, For the modulation output result, These are the variance and mean after the fault type labels are encoded, respectively. The method for the second modulation is the same as that for the first modulation, wherein The input data uses secondary modulation, and the variance and mean are obtained by re-encoding the fault type label.

5. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening as described in claim 1, characterized in that, The discriminator is specifically: After receiving the fault signal dataset, the discriminator first maps each fault signal data in the dataset to a feature vector. Then, the feature vector is sequentially processed through several cascaded ResDownBlocks for perception extraction. The results of the perception extraction are then sequentially processed through an activation layer and a global average pooling layer. The feature vector output by the global average pooling layer is used to calculate the final score of each fault signal data.

6. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening as described in claim 1, characterized in that, In step S3, for each pseudo-fault signal data generated by the generator, the filtering process is as follows: First, the filter receives pseudo-fault signal data and calculates the mean of each pseudo-fault signal data. Secondly, the corrected pseudo-fault signal is obtained by subtracting the corresponding mean from the pseudo-fault signal data. Finally, the filter sequentially screens the corrected pseudo-fault signals based on three aspects: energy efficiency, physical stability, and fault significance, ultimately obtaining high-quality pseudo-fault signal data.

7. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistics screening as described in claim 6, characterized in that, The specific process by which the filter selects the corrected pseudo-fault signals based on three aspects—energy efficiency, physical stability, and fault significance—is as follows: If the root mean square of the corrected pseudo-fault signal is less than the energy efficiency threshold, the corresponding pseudo-fault signal is deleted; otherwise, it enters the physical stability screening. If the absolute value of the peak value of the corrected pseudo-fault signal is greater than the physical stability threshold, the corresponding pseudo-fault signal is deleted; otherwise, it enters the fault significance screening. If the excess kurtosis of the corrected pseudo-fault signal is less than the fault significance threshold, the corresponding pseudo-fault signal is deleted; otherwise, it is retained.

8. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening as described in claim 7, characterized in that, The root mean square is specifically: ； in, Let be the signal value at time t, and L be the total duration of the pseudo-fault signal. It is the root mean square; The excess kurtosis is specifically as follows: ； Where D is the standard kurtosis of the spurious fault signal. This is an excessive kurtosis.

9. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening as described in claim 1, characterized in that, In S4, the final score for each fault signal data is obtained by weighted sum of the unconditional score and the conditional score; The unconditional score is obtained by calculating the confidence level of the fault signal data; the conditional score is obtained by calculating the similarity between the fault signal data and the corresponding fault type label.

10. The method for enhancing motor fault data based on spectral normalization conditional generative adversarial networks and high-order statistical screening as described in claim 1, characterized in that, In step S4, the discrimination process of the discriminator is as follows: The scoring thresholds for real fault signal data and spurious fault signal data are set as the first scoring threshold and the second scoring threshold, respectively. The discriminator uses a first score threshold to distinguish between real and fake fault signal data. If the final score of the real fault signal data is higher than the first score threshold, the fault signal data is determined to be real fault signal data. Otherwise, the fault signal data is determined to be fake fault signal data, indicating that the discriminator's judgment result is incorrect. The discriminator uses a second scoring threshold to distinguish between genuine and fake fault signal data. If the final score of the fake fault signal data is lower than the second scoring threshold, the fault signal data is determined to be fake fault signal data; otherwise, the fault signal data is determined to be genuine fault signal data, indicating that the discriminator's judgment result is incorrect.

Images