Data enhancement method and device for bearing fault signal

Abstract

The invention discloses a data enhancement method for a bearing fault signal, and the method comprises the steps: obtaining a fault signal, carrying out the labeling of a real fault signal according to a fault type, and enabling the fault signal and a label to form a data set; constructing an initial model, wherein the initial model comprises a layered GAN generator and a discriminator; training the initial model by using the data set to obtain a data enhancement model for expanding the fault data sample; and inputting fault information in the to-be-enhanced data set into the data enhancement model to obtain a sample signal and a fault type, and adding the sample signal and the fault type into the to-be-enhanced data set for data enhancement. The invention further provides a data enhancement device. The method provided by the invention can comprehensively reflect fault data of time domain fault features and frequency domain fault features in real samples, thereby providing high-quality data support for data-driven fault diagnosis under small samples and unbalanced working conditions.

Description

Technical Field

[0001] This invention relates to the field of data enhancement, and more particularly to a data enhancement method and apparatus for bearing fault signals. Background Technology

[0002] Bearings, as crucial support components in rotating machinery, often operate under extreme conditions, making them one of the most frequent failure points in rotating equipment. Bearing fault prediction and remaining life estimation have become very popular research topics in recent years. Compared with traditional model-based and signal analysis-based fault diagnosis methods, data-driven fault diagnosis algorithms offer higher accuracy, stronger scalability, and do not rely on precise mathematical models.

[0003] However, current data-driven fault diagnosis algorithms often suffer from two main problems in practical applications: (1) Data-driven fault diagnosis algorithms often require a large amount of fault data for training, while practical applications typically only provide a small amount of fault experimental data; (2) The size of experimental data for various faults in practical applications is not uniform. For example, bearings may experience various faults during use, such as outer ring faults, inner ring faults, ball faults, and cage faults. The number of outer ring faults may be far greater than the number of ball faults. This imbalanced fault dataset reduces the fault classification accuracy of data-driven fault diagnosis algorithms.

[0004] Patent document CN120705635A discloses a method and system for diagnosing rolling bearing faults, including the following steps: preprocessing the original vibration signals of rolling bearings from multiple source domains to obtain multiple time-domain datasets and multiple spectrum datasets; constructing a fault diagnosis model, which includes a first feature extractor, a second feature extractor, a first classifier, a second classifier, a fusion feature extractor, and a joint classifier; performing data augmentation on similar samples in the multiple time-domain datasets to generate multiple augmented datasets; and training the fault diagnosis model using the time-domain datasets, augmented datasets, and spectrum datasets.

[0005] Patent document CN109544468A discloses an image data augmentation method, which includes: determining the category of the image to be augmented, extracting the category bounding boxes in the image using annotation software, determining the target region and noise interference region, changing the pixel values ​​of the noise interference region, and generating new samples. However, these methods have certain limitations: on the one hand, the transformation parameters usually rely on empirical settings and lack specificity; on the other hand, noise processing is not targeted and may even introduce more interference, leading to a decrease in image quality or weakening of features, especially in low signal-to-noise ratio scenarios, where the enhancement effect may be counterproductive. Summary of the Invention

[0006] The purpose of this invention is to provide a data augmentation method and apparatus for bearing faults. This method can comprehensively reflect the fault data of time-domain fault characteristics and frequency-domain fault characteristics in real samples, thereby providing high-quality data support for data-driven fault diagnosis under small sample and unbalanced operating conditions. To achieve the first objective of this invention, the following technical solution is provided: a data augmentation method for bearing faults, comprising the following steps: Acquire fault signals and label real fault signals according to fault type, and combine fault signals and labels to form a dataset; Build the initial model, including a hierarchical GAN ​​generator and a discriminator; The hierarchical GAN ​​generator includes a time-domain generator and a frequency-domain generator. The time-domain generator is used to extract time-domain features from the fault signal, and the frequency-domain generator is used to extract frequency-domain features from the fault signal. The frequency-domain features are processed by inverse Fourier transform to form corresponding transformed time-domain features. The time-domain features and transformed time-domain features are fused using a weighted fusion method to output fused features. The discriminator judges the generated signal based on the input sample and the loss function to estimate the probability that the sample comes from the real data distribution; The initial model is trained using the dataset to obtain a data augmentation model for expanding fault data samples; Input the signal samples in the dataset to be augmented into the data augmentation model, and output the generated sample signals corresponding to the fault; The obtained sample signals are added to the dataset to be augmented for data augmentation.

[0007] This invention reconstructs the generator module and discriminator loss function of GAN based on the features of different types of fault signals, and uses inverse Fourier transform to fuse the generator output signal, thereby obtaining a dataset with richer features.

[0008] Specifically, the fault signal needs to be preprocessed before input, including noise reduction, data normalization, and sample segmentation.

[0009] Specifically, the fault signal includes vibration signal or sound signal.

[0010] Specifically, the denoising employs one or more of the following: low-pass filter, band-pass filter, wavelet transform, or variable mode decomposition, to ensure that the core features of the signal are preserved.

[0011] Specifically, during the training process, multiple loss functions are used to train the initial model, including the generation loss function of the hierarchical GAN ​​generator and the composite loss function of the discriminator.

[0012] Specifically, the generation loss function includes a time-domain generation loss function for measuring the difference between the generated signal and the real signal in the time domain, and a frequency-domain generation loss function for measuring the difference between the generated signal and the real signal in the frequency domain.

[0013] Specifically, the expression for the time-domain generation loss function is as follows: ;in, Indicates the number of samples. Represents a time-domain generator. This indicates the generation of sample signals. Indicates the first A real sample signal, Indicates the first Each input random noise, This represents the L2 norm of the square.

[0014] Specifically, the expression for the frequency domain generation loss function is as follows: ;in, Indicates the number of samples. Represents a frequency domain generator. Indicates the first A real sample signal, Indicates the first Each input random noise, This represents the magnitude after performing a Fourier transform on the real sample data. This represents the modulus of the generated frequency domain sample.

[0015] Specifically, the fusion features are generated through a weighted fusion function, the expression of which is as follows: ;in, This represents the weighted fusion function. Indicates the input noise. This is the inverse Fourier transform function.

[0016] Specifically, the expression for the inverse Fourier transform is as follows: ;in, For frequency domain sampling point index, For time-domain sampling point index, The output of the frequency domain generator is the first... One frequency component, This indicates the number of sampling points for the spectrum sample.

[0017] Specifically, the composite loss function is constructed by introducing a sample quality evaluation index on top of the discriminator loss function, and its expression is as follows: ;in, This indicates the generation of sample quality evaluation indicators. This represents the original loss term of the discriminator loss function. This represents the weight parameters of the original loss term in the discriminator's loss function.

[0018] Specifically, the expression for the generated sample quality evaluation index is as follows: ;in, Represents the frequency domain characteristic error term. Represents the time-domain characteristic error term. This represents the weight calculation function.

[0019] Specifically, the frequency domain feature error term is constructed by summing the amplitude and frequency errors of the envelope spectrum of the generated signal and the original signal at the fault characteristic frequency.

[0020] Specifically, the time-domain feature error term is constructed based on cosine similarity for vibration signals.

[0021] To achieve the second objective of the present invention, the following technical solution is provided: a data augmentation device for performing the steps of the above-described data augmentation method for bearing faults.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: The hierarchical generator architecture, consisting of a time-domain generator and a frequency-domain generator, learns the feature information in their respective domains. By decoupling the modeling, it improves the ability to express complex signals. This structure has modular characteristics and is suitable for various industrial time-series signal modeling tasks.

[0023] A mechanism based on inverse Fourier transform and subsequent fusion is proposed. The frequency domain output of the frequency domain generator is restored to the time domain signal and then weighted and fused with the output of the time domain generator to obtain the fused generated sample. This method significantly improves the performance of the generated data in terms of time-frequency consistency and periodic structure. Attached Figure Description

[0024] Figure 1 This is a flowchart of a data augmentation method for bearing faults provided in this embodiment; Figure 2 This is a flowchart of the preprocessing provided in this embodiment; Figure 3 This embodiment provides the structural design and parameter optimization process for the hierarchical GAN ​​module. Figure 4 This is the logical framework of the hierarchical GAN ​​module provided in this embodiment; Figure 5 This describes the training process of the model provided in this embodiment; Figure 6 The evaluation and analysis process provided in this embodiment; Figure 7 This is a time-domain comparison diagram of the generated data and the original data provided in this embodiment; Figure 8 This is a frequency domain comparison diagram of the generated data and the original data provided in this embodiment. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] like Figure 1 As shown, this embodiment provides a data augmentation method for bearing faults, which includes the following steps: Fault signals are acquired and labeled with fault types, forming a dataset. An initial model is constructed, including a hierarchical GAN ​​generator and a discriminator. The hierarchical GAN ​​generator includes a time-domain generator and a frequency-domain generator. The time-domain generator extracts time-domain features from the fault signals, and the frequency-domain generator extracts frequency-domain features. The frequency-domain features are processed using an inverse Fourier transform to form corresponding transformed time-domain features. A weighted fusion method is used to fuse the time-domain features and the transformed time-domain features to output fused features. The discriminator estimates the probability that the samples come from the real data distribution. Signal samples from the dataset to be augmented are input into the discriminator in the data augmentation model as control samples during training. Random noise is input into the generator to train the initial model, obtaining a data augmentation model for expanding fault data samples and corresponding generated sample signals for the faults. The obtained sample signals are added to the dataset to be augmented for data augmentation.

[0027] More specifically, in this embodiment, the fault signal is selected as the vibration signal, and data preprocessing is performed on the fault signal. In this embodiment, the role of data preprocessing is to reduce noise, normalize, and construct samples from the original vibration signal data to meet the input requirements of the GAN network. The input to data preprocessing is the collected original signal, and the output is the training sample set. High-quality data preprocessing can significantly improve the stability of subsequent model training and the quality of sample generation, and is a fundamental step in the accurate modeling of the entire system. Figure 2 The flowchart shown illustrates that the data preprocessing process is divided into 5 sub-steps: The input algorithm requires real data from failure experiments.

[0028] Noise in signals is removed by using filtering methods such as low-pass filters, band-pass filters, wavelet transforms, and variable mode decomposition to ensure that the core features of the signal are preserved. For example, the cutoff frequency of the low-pass filter can be set according to the needs of fault diagnosis, and it can usually be set to 1000Hz.

[0029] Min-Max normalization is applied to the data to ensure it falls within the range [0, 1], thus preventing data with different units or magnitudes from affecting subsequent processing. The normalization process can be represented as: ;in, This represents taking the maximum value. This represents taking the minimum value. This represents the data obtained after normalization. This represents the original data.

[0030] The normalized data is divided into samples of appropriate length to provide a sufficient number of samples for the generative adversarial network, and to ensure that each sample contains complete feature information within its length.

[0031] Output the processed real samples for subsequent training tasks.

[0032] The initial model proposed in this embodiment includes a hierarchical generator structure based on hierarchical GAN, which enables the generator to learn time-domain and frequency-domain features simultaneously, and sets its network parameters, discriminator loss function structure and fusion mechanism.

[0033] like Figure 3 The diagram shows the design flow for the above-mentioned layered GAN structure, including: The model is built using Generative Adversarial Networks (GANs), where a generator transforms input noise into a signal, and a discriminator compares the generated signal with the real signal. They are trained adversarially using the optimization functions listed below. The discriminator prompts the generator to produce more realistic signals, and the generator improves the discriminator's ability to distinguish between real and generated signals.

[0034] This embodiment proposes a hierarchical GAN ​​module, such as Figure 4 As shown: A hierarchical generator is designed based on the time-domain and frequency-domain characteristics of vibration signals. This includes a time-domain generator. and frequency domain generator Their inputs are the same, from randomly sampled noise. The output of the time-domain generator is a time-domain signal. The output of the frequency domain generator is a frequency domain signal. This allows the generated samples to learn the temporal characteristics of the original signal in the time domain and the spectral characteristics in the frequency domain, respectively.

[0035] Time domain generator and frequency domain generator This forms a hierarchical generator architecture. It consists of a weighted fusion function. Generate fusion signal The process can be represented by the formula: ;in, This is the inverse Fourier transform function.

[0036] A time-domain generator can capture the temporal characteristics of input noise. The time-domain generation loss function in a time-domain generator measures the difference between the generated signal and the real signal in the time domain; its expression is as follows: ;in, Indicates the number of samples. Represents a time-domain generator. This indicates the generation of sample signals. Indicates the first A real sample signal, Indicates the first Each input random noise, This represents the L2 norm of the square.

[0037] The design concept of the frequency domain generator originates from frequency domain analysis methods in signal processing. By using frequency domain loss as the loss function of the frequency domain generator, complex data that conforms to the frequency characteristics of the signal can be generated. Frequency domain loss measures the difference between the generated signal and the real signal in the frequency domain, and its expression is as follows: ;in, Indicates the number of samples. Represents a frequency domain generator. Indicates the first A real sample signal, Indicates the first Each input random noise, This represents the magnitude after performing a Fourier transform on the real sample data. This represents the modulus of the generated frequency domain sample.

[0038] The hierarchical generator is initialized with parameters and its structure set to ensure stable network training. Parameters that need to be configured for both the time-domain and frequency-domain generators include: input and output dimensions, number of convolutional layers, number of channels, kernel size, stride, activation function, normalization method, and upsampling method. The hierarchical generator structure and parameters used in this invention are shown in Tables 1 and 2. Specifically, the first fully connected layer is used to expand latent features, the second fully connected layer is used to output the real part of the spectrum, and the third fully connected layer is used to output the imaginary part of the spectrum, together forming a complete complex spectrum. These parameters can be adjusted according to different application scenarios.

[0039] ; .

[0040] like Figure 4 As shown, time-domain generated samples Frequency domain sample generation It requires a hierarchical sample fusion function. Transform into fusion sample In this embodiment, linear weighting is used, and its expression is as follows: ;in, This represents the signal weighting factor, used to adjust the proportion of the time-domain generated signal and the frequency-domain generated signal in the fused generated signal.

[0041] Indicates taking The real part of the middle.

[0042] In the hierarchical sample fusion function, the frequency domain signal output by the frequency domain generator is... The signal is returned to the time domain by inverse Fourier transform and then linearly weighted with the signal from the time-domain generator. The input to the generated signal fusion function is the time-domain signal generated by the time-domain generator and the frequency-domain signal generated by the frequency-domain generator; the output is the fused generated signal.

[0043] To evaluate the fused generated signal and implement adversarial training, a discriminator needs to be constructed. The discriminator structure and parameters used in this embodiment are shown in Table 3. These parameters can be adjusted according to different application scenarios.

[0044] .

[0045] In this embodiment, a quality evaluation index for generated samples based on time-frequency domain feature fusion is proposed. This indicator can not only be used to evaluate the quality of generated samples, but can also be incorporated into the discriminator's loss function as an additional correction term for the fusion of time-domain and frequency-domain errors. This process can be expressed as the formula: ;in, This indicates the generation of sample quality evaluation indicators. This represents the original loss term of the discriminator loss function. This represents the weight parameters of the original loss term in the discriminator's loss function.

[0046] Evaluation indicators From the time-domain characteristic error term Frequency domain characteristic error term and indicator weighting coefficients Composition, its expression is: .

[0047] exist The full length of each sample is selected for calculating cosine similarity. The first five order fault characteristic frequencies and their amplitudes are selected as frequency domain features. Furthermore, the selection of fault characteristic frequencies depends on the fault category of the original signal. Faults occurring at different locations in the bearing will result in corresponding fault characteristic frequencies and their harmonics appearing in the acceleration envelope spectrum. For example, the fault characteristic frequency for the outer ring of the bearing is BPFO (Ball Pass Frequency of Outer ring); the fault characteristic frequency for the inner ring of the bearing is BPFI (Ball Pass Frequency of Inner ring); the fault characteristic frequency for the bearing cage is FTF (Fundamental Train Frequency); and the fault characteristic frequency for the bearing balls is BSF (Ball Spin Frequency).

[0048] In this embodiment, the weight calculation function The linear weighted algorithm is selected, and its specific expression is as follows: ;in, This represents the weighting coefficient of the indicator.

[0049] In this embodiment, the expression for the time-domain feature error term is as follows: ;in, The maximum order of the characteristic frequencies can be selected based on the application scenario (usually...). This can meet engineering application requirements. To generate the signal at the first Amplitude at the characteristic frequency of the first-order fault. For the true fault signal in the first Amplitude at the characteristic frequency of the first-order fault. For the true fault signal in the first Amplitude at the characteristic frequency of the first-order fault. For the generation of the signal Fault characteristic frequencies, The first of the real fault signals Fault characteristic frequencies.

[0050] The temporal feature error term is constructed based on cosine similarity, and its expression is as follows: ;in, A represents the vector of the generated signal, and B represents the vector of the actual fault signal. This represents the total number of components in a vector. Represents the first in the vector Each component value. Therefore, the cosine similarity for vibration signals. Taking the absolute value allows for a better characterization of the similarity between two vibration signals.

[0051] In this embodiment, the calculation formulas for the above-mentioned fault characteristic frequencies are as follows: ; ; ; ;in, The frequency of the rotating shaft, The diameter of the bearing balls. The bearing's mean diameter, The initial contact angle of the bearing. This refers to the number of balls in the bearing.

[0052] Evaluation indicators A discriminator loss function is added to reconstruct the discriminator's optimization objective and further promote the generation of time-frequency and frequency domain features of the entire network parameters to more closely approximate the real signal.

[0053] A hierarchical GAN ​​is trained using an adversarial training strategy, enabling the generator to produce samples with realistic fault characteristics from random noise. An evaluation phase is used to assess the quality of the samples generated by the generator.

[0054] The inputs to this step are the GAN structure and parameter configuration; real training samples; and initial weights for the discriminator and generator. The outputs are the converged hierarchical GAN ​​model and fault signal samples that approximate the real data distribution. Adversarial training is the core of GANs; high-quality training enables the generator to construct realistic signals, which is the fundamental guarantee of the usability of the generated samples. The specific steps are as follows: Figure 5 The following is stated: Real samples serve as positive samples for the discriminator, ensuring that the discriminator can learn the distribution of real data and use it to improve its ability to distinguish between real samples and generated samples.

[0055] The training parameters for hierarchical GANs include learning rate, optimizer, batch size, number of training epochs, and discriminator update frequency (usually 1:1, meaning the generator is trained once and the discriminator is trained once).

[0056] Layered GANs are trained alternately based on adversarial mechanisms and the training parameters set in the previous steps. The neural network parameters in the generator and discriminator are updated alternately, so that the generated signal approximates the time-frequency characteristics of the real signal.

[0057] Set a maximum number of training rounds. When the discriminator cannot distinguish between real samples and generated samples, or when the maximum number of training rounds is reached, the generated sample is used as the final output sample.

[0058] The similarity between the generated samples and the real samples is compared by outputting their time-domain and frequency-domain signals. The frequency-domain feature error term is calculated. (If the first 5 order fault characteristic frequencies and their amplitudes are used), if If the frequency domain features of the generated sample meet expectations, then the time domain feature error term is also calculated. ,if If so, the temporal characteristics of the generated sample are considered to have met expectations. Comprehensive evaluation index. Selection of expected value and index weighting coefficient Related, can be adjusted as needed.

[0059] Based on the application scenarios and requirements of data augmentation techniques, the quality of the generated data should be comprehensively evaluated and analyzed from multiple dimensions. This step quantifies the quality of the generated samples, verifies whether the GAN generation effect meets the application requirements, and provides a basis for subsequent model optimization. The process is as follows: Figure 6 As shown: To comprehensively evaluate the quality of samples generated by the hierarchical GAN ​​proposed in this invention, this invention defines corresponding indicators from four dimensions: similarity, diversity, non-repetition, and consistency of physical features.

[0060] To ensure that the generated samples closely resemble real data in terms of overall shape and signal trend, this invention introduces two complementary metrics in the similarity evaluation sub-step: cosine similarity (CS) and mean square error. This allows for a dual assessment of amplitude-level errors and shape trend differences.

[0061] Mean square error The calculation process can be expressed as a formula: ;in, Represents the sample length. and These represent the generated sample and the real sample respectively in the 1st... The values ​​of each sampling point.

[0062] This embodiment also includes a diversity evaluation step to measure the distribution coverage of generated samples in the feature space. This step uses Kullback-Leibler divergence as the diversity index of the generated samples to quantify their distribution. Compared with the true sample distribution The difference between them is defined as follows: ;in, and Both represent KL divergence, used to indicate the degree of deviation between the generated sample distribution and the real sample distribution.

[0063] and The real sample and the generated sample are respectively in the 1st... Probability distribution values ​​for each category or feature dimension.

[0064] This indicator can capture the degree of overlap in sample distributions; the smaller the value, the better the diversity of the generated samples.

[0065] To ensure that the generated samples retain key fault characteristics in the frequency domain, a physical characteristic consistency evaluation index is established, with fault characteristic frequencies and their amplitudes as the core. This index is based on the analysis of typical fault frequencies of a specified order in the envelope spectrum, calculating the deviations of the generated samples from the reference samples in terms of frequency position and amplitude.

[0066] Using the frequency domain characteristic error term, the first Fault characteristic frequency error of the order and fault characteristic frequency amplitude error The following technical solution is constructed: ; This indicator not only quantifies the consistency between frequency and amplitude, but also reflects the degree to which the generated samples retain diagnostic features at the physical level, which is an important foundation for ensuring the usability of the generated data engineering.

[0067] This embodiment also provides a data enhancement device for performing the steps of the data enhancement method for bearing fault signals provided in the above embodiments.

[0068] To better illustrate the technical effects of the solution provided in this embodiment, outer ring fault data will be used as an example.

[0069] in, Figure 7 To compare the original measured signal samples and the samples generated by the hierarchical GAN ​​in the time domain, it can be found that the samples generated by the hierarchical GAN ​​possess the key time domain characteristics of the real samples, such as amplitudes close to [-1,1], similar fluctuation patterns, and similar noise levels.

[0070] Figure 8To compare the frequency domain features of real samples and samples generated by hierarchical GAN, it can be seen that the main frequency components and peak values ​​of the two signals are similar, and the amplitudes are also similar in the high-frequency and low-frequency regions. This indicates that hierarchical GAN ​​has learned the frequency domain features of real signals quite well. Furthermore, the calculated metrics of the generated data are superior to those of ordinary GAN. For example, compared to ordinary GAN, the cosine similarity of the inner and outer loop data of hierarchical GAN ​​shows improvements of up to 40% and 70%, respectively.

[0071] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0072] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

[0073] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A data augmentation method for bearing fault signals, characterized in that, Includes the following steps: Acquire fault signals and label real fault signals according to fault type, and combine fault signals and labels to form a dataset; Build the initial model, including a hierarchical GAN ​​generator and a discriminator; The hierarchical GAN ​​generator includes a time-domain generator and a frequency-domain generator. The time-domain generator is used to extract time-domain features from the fault signal, and the frequency-domain generator is used to extract frequency-domain features from the fault signal. The frequency-domain features are processed by inverse Fourier transform to form corresponding transformed time-domain features. The time-domain features and transformed time-domain features are fused using a weighted fusion method to output fused features. The discriminator judges the generated signal based on the input sample and the loss function to estimate the probability that the sample comes from the real data distribution; The initial model is trained using the dataset to obtain a data augmentation model for expanding fault data samples; Input the signal samples in the dataset to be augmented into the data augmentation model, and output the generated sample signals corresponding to the fault; The obtained sample signals are added to the dataset to be augmented for data augmentation.

2. The data enhancement method for bearing fault signals according to claim 1, characterized in that, The fault signal needs to be preprocessed before input. The preprocessing includes noise reduction, data normalization, and sample segmentation.

3. The data enhancement method for bearing fault signals according to claim 1, characterized in that, During training, multiple loss functions are used to train the initial model, including the generation loss function of the hierarchical GAN ​​generator and the composite loss function of the discriminator.

4. The data enhancement method for bearing fault signals according to claim 3, characterized in that, The generation loss function includes a time-domain generation loss function for measuring the difference between the generated signal and the real signal in the time domain, and a frequency-domain generation loss function for measuring the difference between the generated signal and the real signal in the frequency domain.

5. The data enhancement method for bearing fault signals according to claim 4, characterized in that, The expression for the temporal generation loss function is as follows: ;in, Indicates the number of samples. Represents a time-domain generator. This indicates the generation of sample signals. Indicates the first A real sample signal, Indicates the first Each input random noise, This represents the L2 norm of the square.

6. The data enhancement method for bearing fault signals according to claim 4, characterized in that, The expression for the frequency domain generation loss function is as follows: ;in, Indicates the number of samples. Represents a frequency domain generator. Indicates the first A real sample signal, Indicates the first Each input random noise, This represents the magnitude after performing a Fourier transform on the real sample data. This represents the modulus of the generated frequency domain sample.

7. The data enhancement method for bearing fault signals according to claim 3, characterized in that, The composite loss function is constructed by introducing a sample quality evaluation index based on the discriminator loss function, and is expressed as follows: ;in, This indicates the generation of sample quality evaluation indicators. This represents the original loss term of the discriminator loss function. This represents the weight parameters of the original loss term in the discriminator's loss function.

8. The data enhancement method for bearing fault signals according to claim 7, characterized in that, The expression for the generated sample quality evaluation index is as follows: ;in, Represents the frequency domain characteristic error term. Represents the time-domain characteristic error term. This represents the weight calculation function.

9. The data enhancement method for bearing fault signals according to claim 8, characterized in that, The frequency domain feature error term is constructed by summing the amplitude and frequency errors of the envelope spectrum of the generated signal and the original signal at the fault characteristic frequency.

10. The data enhancement method for bearing fault signals according to claim 8, characterized in that, The time-domain feature error term is constructed based on cosine similarity for vibration signals.

Images