Rolling bearing defect detection method integrating wavelet denoising and generative adversarial network

Abstract

The application relates to the technical field of bearing defect detection, and discloses a rolling bearing defect detection method integrating wavelet denoising and a generative adversarial network. The method comprises the following steps: firstly, an optimal threshold of a Garrote wavelet is adaptively determined by using a particle swarm optimization algorithm to effectively denoise original vibration signals of the rolling bearing; then, a generative adversarial network guided by multiple conditions is constructed, high-quality synthetic waveform data are generated through double-domain feature fusion and multiple loss constraints; then, a multi-scale feature aggregation network of a discriminator is transferred to a classifier through knowledge transfer, and fine-tuning is performed by using a small sample training set; finally, a pseudo-label self-training strategy is adopted, an extended training set is constructed by using high-confidence synthetic data, the classifier is intensively trained, the test set is input into the intensively trained classifier, and a diagnosis result of a rolling bearing defect type is output, so that the accuracy and robustness of defect recognition are improved.

Description

Technical Field

[0001] This application relates to the field of bearing defect detection technology, and in particular to a method for detecting rolling bearing defects that integrates wavelet denoising and generative adversarial networks. Background Technology

[0002] Rolling bearings are core components in precision machinery, and their manufacturing quality directly determines the final performance, lifespan, and reliability of the equipment. In automated, high-speed production processes, various factors such as processes, materials, and equipment can lead to various manufacturing defects in bearings, including cracks, scratches, and corrosion. Failure to detect these defects in a timely manner can result in serious safety hazards and economic losses. Therefore, accurate and efficient quality inspection of bearings off the production line is crucial.

[0003] In recent years, deep learning has made significant progress in the field of intelligent quality inspection due to its powerful automatic feature extraction capabilities, providing a new approach to accurately identify defects in bearing manufacturing. However, this technology still faces two major challenges in practical production line applications: small sample size and high noise levels. On the one hand, under modern high-quality production management systems, defective samples are inherently low-probability events, making data acquisition difficult and resulting in extremely scarce defect sample data suitable for model training. This leads to the model being prone to overfitting, making it difficult to fully learn the characteristics of various defects, severely restricting the accuracy and generalization ability of the inspection model. On the other hand, the production workshop environment is complex, and the signal-to-noise ratio of the signals collected by acquisition equipment (such as acoustic or vibration sensors) is low. Background noise from the production line often masks weak defect feature signals, which not only increases the difficulty of model learning but also directly affects the accuracy of inspection results.

[0004] Therefore, how to achieve high-precision intelligent defect inspection under conditions of scarce defect samples and strong noise has become a key problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a rolling bearing defect detection method that integrates wavelet denoising and generative adversarial networks, enabling high-precision intelligent defect inspection in the context of scarce defect samples and strong noise.

[0006] In a first aspect, this application provides a method for detecting defects in rolling bearings that integrates wavelet denoising and generative adversarial networks, the method comprising:

[0007] Step S1: Obtain the original vibration signal of the rolling bearing, use the particle swarm optimization algorithm to find the optimal Garrote wavelet threshold, perform noise reduction processing on the original vibration signal, segment and label the denoised signal, and construct the initial training set and test set.

[0008] Step S2: Establish a conditional generative adversarial network. Its generator receives the denoised signal waveform, corresponding category label, physical features and random noise from the initial training set, and outputs synthetic waveform data. The synthetic waveform data is then input into the discriminator for training.

[0009] Step S3: After the conditional generative adversarial network is trained, the backbone network of multi-scale feature aggregation in its discriminator is transferred to the classifier as the feature extraction backbone, and the classifier is fine-tuned using the denoised signal in the initial training set.

[0010] Step S4: Input the synthesized waveform data into the fine-tuned classifier to obtain pseudo-label data. Construct an extended training set based on the pseudo-label data to reinforce the classifier. Input the test set into the reinforced classifier and output the diagnostic results of the rolling bearing defect type.

[0011] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0012] This application proposes a rolling bearing defect detection method integrating wavelet denoising and generative adversarial networks (GANs), achieving significant beneficial effects in the field of small-sample rolling bearing defect detection. First, the particle swarm optimization Garrote wavelet threshold denoising method overcomes the limitations of traditional threshold selection, achieving an optimal balance between noise suppression and feature preservation through an adaptive optimization mechanism, effectively improving the quality of the input signal and providing a high-fidelity data foundation for subsequent deep learning models. Second, the designed multi-condition jointly guided GAN successfully solves the data scarcity problem under small-sample conditions. Through dual-domain feature fusion and multiple loss constraints, it ensures that the generated synthetic data not only has visual realism but also maintains a high degree of consistency with the real signal in multiple physical dimensions such as time domain, frequency domain, envelope spectrum, and statistical features.

[0013] Furthermore, the knowledge transfer mechanism proposed in this application transfers the multi-scale feature aggregation network trained in the discriminator to the classifier, enabling the classifier to have strong feature extraction capabilities in the early stages of training, significantly reducing the model's dependence on a large amount of labeled data. The pseudo-label self-training strategy further expands the scale of training data, and by selecting high-confidence generated samples to construct an extended training set, it effectively improves the model's ability to identify rare defect categories. In summary, this application not only solves the core pain point of difficulty in obtaining defect samples in industrial sites to a certain extent, but also ensures the accuracy and reliability of defect detection through a systematic technical architecture. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of an embodiment of the rolling bearing defect detection method integrating wavelet denoising and generative adversarial network in this application.

[0016] Figure 2 This is a schematic diagram of a noisy vibration signal in an embodiment of this application;

[0017] Figure 3 This is a path diagram of fitness value changes in the embodiments of this application;

[0018] Figure 4 To find the optimal threshold in the embodiments of this application;

[0019] Figure 5 This is a schematic diagram of the noise-reduced vibration signal in an embodiment of this application;

[0020] Figure 6 This is a detailed structural diagram of the dual-domain feature fusion unit in the generator of the embodiments of this application;

[0021] Figure 7 This is a detailed structural diagram of the upsampling block in the generator in the embodiments of this application;

[0022] Figure 8 This is a core architecture diagram of the discriminator in the embodiments of this application;

[0023] Figure 9 This is a schematic diagram of the confusion matrix of the final model in the embodiments of this application;

[0024] Figure 10 This is a schematic diagram of the confusion matrix of a random single sample in an embodiment of this application. Detailed Implementation

[0025] This application provides a rolling bearing defect detection method integrating wavelet denoising and generative adversarial networks. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0026] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the rolling bearing defect detection method integrating wavelet denoising and generative adversarial networks in this application includes:

[0027] Step S1: Obtain the original vibration signal of the rolling bearing, use the particle swarm optimization algorithm to find the optimal Garrote wavelet threshold, perform noise reduction processing on the original vibration signal, segment and label the denoised signal, and construct the initial training set and test set.

[0028] Step S1 includes: acquiring the original vibration signals of the rolling bearing under multiple operating conditions; performing wavelet decomposition on the original vibration signals to obtain a series of original wavelet coefficients; generating a set of threshold candidate solutions based on the particle swarm optimization algorithm; inputting each threshold candidate solution into the Garrote threshold function; and performing noise reduction processing on the original wavelet coefficients to obtain the processed wavelet coefficients. The Garrote threshold function is: ,in, These are the original wavelet coefficients. For threshold candidate solutions, Parameters for controlling smoothness, For the processed wavelet coefficients, wavelet reconstruction is performed on all processed wavelet coefficients to obtain the denoised vibration signal; the kurtosis and mean square error of the denoised vibration signal are calculated, and a fitness function is constructed based on the kurtosis and mean square error. ,in, For ravine, Mean square error, For fitness, and The threshold candidate solutions are preset with weights; the threshold candidate solutions are iteratively updated with the goal of maximizing the fitness function. The threshold candidate solution that maximizes the fitness function is defined as the optimal Garrote wavelet threshold, and the final denoised signal is obtained based on the optimal Garrote wavelet threshold. The final denoised signal is segmented and labeled, and combined to generate a sample dataset. The sample dataset is then divided into an initial training set and a test set.

[0029] Specifically, in order to address the problems of low signal-to-noise ratio and weak fault features in the detection of rolling bearing defects, especially how to effectively extract early fault features under strong noise background, this application adopts an adaptive wavelet threshold denoising method based on particle swarm optimization. This method can automatically find the optimal denoising parameters according to the signal characteristics, effectively suppressing noise while fully preserving fault features, thus laying the foundation for subsequent small-sample defect identification.

[0030] Specifically, during the data acquisition phase, raw vibration signals of rolling bearings under different operating conditions are collected from industrial sites or standard testing platforms. These signals manifest as time-series data containing rich fault information. For example, normal bearing signals exhibit stable random vibration characteristics, while inner ring fault signals show periodic impact responses. These raw signals are often interfered with by background noise, making early, subtle fault characteristics difficult to observe directly. Figure 2 The diagram shows a vibration signal containing noise. In the signal preprocessing stage, wavelet transform is used to decompose the original vibration signal into multiple scales. The purpose of this step is to convert the time-domain signal to the time-frequency domain and expand the signal to different scale spaces through wavelet basis functions to obtain wavelet coefficients corresponding to different frequency bands. Among them, the low-frequency coefficients mainly reflect the macroscopic features of the signal, while the high-frequency coefficients contain the detailed information and noise components of the signal. This multi-resolution analysis characteristic lays the foundation for subsequent noise separation and feature extraction.

[0031] To effectively remove noise and preserve fault characteristics, this application utilizes a particle swarm optimization algorithm, using the composite function of the signal as the fitness function, to find the optimal Garrote wavelet threshold. The specific process is as follows: First, a set of particles is initialized in the parameter space, where the position of each particle represents a candidate threshold for the Garrote threshold function. For example, candidate values ​​are randomly generated within a reasonable range of [2.0, 2.1], and these candidate thresholds will serve as the initial solutions for the PSO search. Next, a composite fitness function is constructed as the optimization objective. The design of this function reflects a comprehensive evaluation of the noise reduction effect: kurtosis is used to quantify the significance of periodic impact features in the signal, ensuring that fault features are enhanced; mean square error is used to control the degree of signal distortion, ensuring the integrity of the overall waveform structure; and weighting coefficients are used to further refine the function. and Adjustments, such as settings =0.7, =0.3, which makes the optimization process focus more on the preservation of fault characteristics.

[0032] During the iterative optimization process, each candidate threshold All of these need to be evaluated through the following process: Candidate thresholds Substituting the Garrote threshold function into the wavelet coefficients, this function utilizes a unique nonlinear shrinkage property, i.e., when... Smooth contraction occurs when... Noise separation is achieved by setting the time to zero. The processed wavelet coefficients are reconstructed to obtain the denoised signal. The kurtosis and mean square error of this signal are calculated and substituted into the fitness function to obtain a score. Figure 3 As shown in the graph, the fitness value changes steadily with the number of iterations, gradually increasing from an initial 6.712 to 6.743. This indicates that the algorithm effectively improves the noise reduction effect. Figure 4 As shown, the optimal threshold is found by adjusting the threshold within the range of 2.02 to 2.07. After 75 iterations, the optimal threshold is obtained and converges, proving that the PSO algorithm effectively finds the optimal threshold. After finding the optimal threshold, it is applied to the detail coefficients obtained by wavelet decomposition and wavelet reconstruction is performed to obtain a high-quality denoised signal, as shown. Figure 5 The diagram shows the vibration signal after noise reduction. By comparing the individual historical optimal solution with the global optimal solution of the group, the velocity and position of the particles are continuously updated. This collaborative search mechanism causes the candidate threshold to continuously evolve towards the optimal solution. After multiple iterations, such as 100 iterations, the algorithm converges to the optimal threshold that maximizes the fitness function. The final optimal threshold achieves the optimal balance between noise suppression and feature preservation: it can effectively filter out background noise while completely preserving the fault impact characteristics, providing a high-quality input signal for subsequent data augmentation and defect identification. The denoised signal is standardized, and a fixed-length sliding window is used for sample segmentation. Each sample segment is labeled as "normal", "inner race fault", "outer race fault" or "rolling element fault" according to the actual state of the corresponding bearing. The labeled samples are randomly divided proportionally, with 70% used as the initial training set for model training and 30% used as the test set for performance evaluation, finally constructing a standardized dataset that can be used by deep learning models.

[0033] Step S2: Establish a conditional generative adversarial network. Its generator receives the denoised signal waveform, corresponding category label, physical features and random noise from the initial training set, and outputs the synthesized waveform data. The synthesized waveform data is then input into the discriminator for training.

[0034] In step S2, establishing a conditional generative adversarial network includes: constructing a multi-condition jointly guided generative adversarial network, which includes a generator and a discriminator. The generator consists of a dual-domain feature fusion layer, a fully connected layer, an upsampling block, and a final output layer connected sequentially. The discriminator consists of a multi-scale feature aggregation backbone network, a label embedding layer, and multiple fully connected layers connected sequentially. The denoised signal waveform, the corresponding defect category label, the physical feature vector extracted from the waveform, and the random noise vector in the initial training set are input into the generator, and the generator outputs synthesized waveform data. The synthesized waveform data, the real denoised waveform in the initial training set, and the corresponding defect category label are input into the discriminator.

[0035] Specifically, a conditional generative adversarial network (GAN) with multi-condition joint guidance is constructed and trained. The GAN includes a generator and a discriminator. The generator employs a multi-path encoding strategy and comprises three main components: an input encoding and fusion module, a feature transformation module, and a progressive upsampling block. In the input encoding stage, random noise, classification labels, and physical features are encoded through fully connected layers, while the original waveform input is fed into a dual-domain feature fusion module for deep feature extraction. At this stage, the generator is a standard generator. The specific implementation of the dual-domain feature fusion module within the generator is as follows: Figure 6 As shown, this module employs a parallel dual-path structure to process the input waveform. The time-domain feature path directly extracts the local temporal features of the waveform through a one-dimensional convolutional layer (64 channels). The frequency-domain path transforms the signal to the frequency domain through an FFT layer, and then extracts spectral features through a series of one-dimensional convolutional layers (128 channels). The features from both paths are dimensionality-reduced by a two-dimensional global average pooling layer, and then adaptively weighted and fused through a gated fusion mechanism. The final output is a dual-domain feature representation containing complete time-frequency information. These features are fused in the concatenation layer, and then their dimensions are adjusted through a fully connected layer. Subsequently, four upsampling blocks progressively convert the fused features into a complete waveform sequence. The specific structure of the upsampling blocks is shown below. Figure 7 As shown, each upsampling block sequentially contains a linear interpolation layer, a one-dimensional convolutional layer, a Gaussian noise injection layer, a batch normalization layer, and a LeakyReLU activation layer. The linear interpolation layer is responsible for improving the temporal resolution of the feature sequence, the one-dimensional convolutional layer performs feature transformation, the Gaussian noise injection increases the diversity of generated samples, the batch normalization ensures training stability, and the LeakyReLU activation function introduces non-linear transformation. By cascading four such upsampling blocks, low-dimensional features are gradually converted into complete waveform sequences.

[0036] The core architecture of the discriminator is as follows Figure 8As shown, the core architecture is a multi-scale feature aggregation architecture, which is an important guarantee for ensuring the quality of generated samples. The discriminator simultaneously receives real waveforms and generated waveforms, as well as corresponding category labels. The waveform data is processed by a multi-scale feature extraction and fusion module, which contains five cascaded residual blocks. Each residual block is followed by a one-dimensional global average pooling layer, which outputs feature vectors of 512, 256, 128, 64, and 32 dimensions, respectively. After these multi-scale features are merged in the concatenation layer, they are recalibrated by the SE (Squeeze-and-Excitation) module to enhance the weight of important features. At the same time, the label information is converted into feature vectors through the embedding layer and fused with the waveform features in the feature concatenation layer. Finally, the discrimination result is output through a fully connected layer.

[0037] The dual-domain feature fusion mechanism in the generator and the multi-scale feature aggregation structure in the discriminator form a good correspondence. The former ensures that the generated samples have physical rationality in both the time and frequency domains, while the latter performs detailed evaluation of the generated samples from multiple scales. The combined effect of the two enables the generative adversarial network to effectively learn the essential feature distribution of real vibration signals. At the same time, the generator can fully integrate multi-source conditional information to generate vibration signals with physical authenticity, while the discriminator achieves accurate identification of signal details through multi-scale feature aggregation. This adversarial training mechanism ensures that the generated samples not only conform to the statistical distribution of real data but also maintain the physical characteristics of fault features, providing a high-quality data foundation for subsequent small-sample defect identification.

[0038] Step S2 further includes: during the training of the conditional generative adversarial network, the specific expression for the total generator loss is: ,in, To combat the losses, For the loss of experts, The generator's parameters are optimized based on the total generator loss, which is used as the feature matching loss. The discriminator's parameters are optimized based on the total discriminator loss. The parameters of the generator and discriminator are iteratively updated through the backpropagation algorithm until the model converges, resulting in a trained generative adversarial network.

[0039] This application also sets up multiple loss functions to guide the training process of generative adversarial networks. The total loss function of the generator consists of three key parts: adversarial loss, expert loss, and feature matching loss.

[0040] The expression for adversarial loss is: , This indicates that the discriminator evaluates the generated samples. The rating.

[0041] Specifically, minimizing the adversarial loss essentially requires the generator to maximize the discriminator's score for the generated samples. During training, this loss function drives the generator to continuously optimize its network parameters, enabling the synthesized waveforms it produces to achieve higher realism scores in the discriminator's evaluation system. This effectively deceives the discriminator and counteracts the loss. Ensure that the generated samples score highly in the discriminator, thereby driving the generator to produce more realistic waveforms.

[0042] Among them, the expert loss consists of the temporal reconstruction loss. Frequency domain amplitude loss Envelope spectrum loss and statistical losses Composition, the expression is: ,in, , , and These represent the weights corresponding to the time-domain reconstruction loss, frequency-domain amplitude loss, envelope spectrum loss, and statistical loss, respectively; the time-domain reconstruction loss is: , For real data, To generate data; the frequency domain amplitude loss is: = , This is a Fourier transform operation; the envelope spectrum loss is: , For real data obtained through Hilbert transform The envelope, The generated data obtained through Hilbert transform The envelope; the statistical loss is: , This is a statistical feature vector extracted from real vibration data. The statistical feature vector extracted from the synthetic data output by the generator, where, This represents the L1 normal form for each prediction error vector.

[0043] Specifically, the expert loss function is designed to ensure that the generated waveform remains consistent with the real signal across multiple physical dimensions. Constraints are imposed on four aspects: time-domain reconstruction, frequency-domain amplitude, envelope spectrum, and statistical loss. The weighting coefficients for these four losses are set based on experimental optimization. In practice, the calculation method for each loss term is as follows: Time-domain reconstruction loss. The L1 norm is used to measure the difference between the generated waveform and the real waveform in the time domain. This loss term ensures the basic waveform characteristics of the generated signal; frequency domain amplitude loss. By comparing the spectral characteristics of the signals using Fast Fourier Transform (FFT) and calculating the L1 norm difference between the two amplitude spectra, the generated signal not only closely approximates the real signal in the time domain but, more importantly, maintains a high degree of consistency in its frequency domain characteristics. This is particularly evident in the accurate reproduction of characteristic frequency components crucial for bearing fault diagnosis. Envelope spectrum loss is achieved by extracting the real data x and the generated data using Hilbert transform. The envelope of the signal is obtained by FFT transformation, and the envelope spectrum is obtained by calculating the L1 norm difference between the two envelope spectra at bearing characteristic frequencies, such as the inner race fault frequency of 132Hz. This ensures that the generated waveform accurately reproduces the periodic characteristics of the fault impact. This loss term can reduce the harmonic energy error of the generated signal at the fault frequency; statistical feature loss. By calculating the real data x and the generated data The L1 norm difference of the multidimensional statistical feature vector is used to achieve this. The statistical feature vector includes at least the root mean square, kurtosis, skewness, and peak factor of the time domain indicators, as well as the spectral centroid and mean square frequency of the frequency domain indicators, thereby ensuring that the generated waveform is consistent with the distribution of the real fault signal in terms of key statistical characteristics.

[0044] The expression for the feature matching loss is: ,in, It is the feature extraction function that the discriminator extracts from the intermediate layer. It is the square of the Euclidean distance.

[0045] Specifically, feature matching loss This is used to constrain the consistency of the distribution of generated samples and real samples in the discriminator's feature space. The specific implementation process is as follows: During the forward propagation of the discriminator, the end of the multi-scale feature aggregation module is selected as the feature extraction point, and the feature extraction function F(·) is defined as the output feature vector of this module. For real data x and generated data... Calculate the deep feature representations obtained by the discriminator from the intermediate layer, respectively. and During training, the mean of the feature vectors of the real samples and the mean of the feature vectors of the generated samples are calculated for each batch. Then, the squared Euclidean distance between these two mean vectors is calculated. This loss term ensures that the generator not only needs to deceive the final output layer of the discriminator but also needs to simulate the feature distribution of the real data on the intermediate feature layers of the discriminator. Through this constraint, the data distribution learned by the generator remains highly consistent with the distribution of the real data in the feature space, effectively improving the feature quality of the generated samples and the stability of training.

[0046] The three loss terms—adversarial loss, expert loss, and feature matching loss—are weighted and summed to form a complete generator optimization objective, which jointly drives the generator parameter update. At the same time, this application also uses WGAN-GP loss to drive the discriminator parameter optimization, and updates the parameters of the generator and discriminator through the backpropagation algorithm until the model converges, resulting in a trained conditional generative adversarial network.

[0047] The specific expression for the total loss of the discriminator is as follows: ,in, For WGAN-GP loss, It is the mathematical expectation. It is the actual data distribution. It generates data distribution. For real data, The synthesized data output by the generator is defined as the generated data. It is the discriminator's score of the real data. It is the discriminator's score of the generated data. These are points randomly interpolated between real and generated data. These are the weights of the gradient penalty.

[0048] Specifically, in constructing the loss function of the discriminator, this application adopts WGAN-GP loss as the optimization objective. This loss function consists of three terms with clear physical meaning: the first term: The first term represents the discriminator's expected score for the generated sample; this term needs to be minimized during training to achieve higher scores for the generated samples. The second term: This represents the discriminator's expected score for real samples; maximizing this term is necessary to ensure higher ratings for real samples. The third term: It is a gradient penalty term. The sampling points are obtained through random linear interpolation between real and generated samples. In the specific implementation, the gradient penalty coefficient is set to 10, which is the optimal value verified by extensive experiments. The calculation of the gradient penalty term needs to be performed throughout the entire sample space. In practice, the corresponding interpolation point is calculated for each real-generated sample pair in each training batch, and the gradient norm of the discriminator at these interpolation points is required to be as close to 1 as possible. This design ensures that the discriminator satisfies the Lipschitz continuity condition, avoiding the capacity reduction problem caused by weight pruning in traditional WGAN, and ensuring the stability of the training process. Through this loss function design, the discriminator can accurately distinguish between real and generated samples during training while maintaining good function smoothness. This stable training environment provides the generator with a continuous and effective gradient signal, enabling it to gradually learn the distribution characteristics of real vibration signals and ultimately generate high-quality samples with physical authenticity.

[0049] At the training execution level, this application also employs an alternating optimization strategy to iteratively update the generator and discriminator parameters. In each training round, the generator parameters are first fixed, and the discriminator loss is calculated based on the currently generated samples and real samples. The discriminator parameters are then updated through backpropagation. Subsequently, the discriminator parameters are fixed again, and the generator's multiple loss function is calculated. The generator parameters are also updated through backpropagation. This alternating training process allows the generator and discriminator to improve together in adversarial training. The generator gradually learns the distribution characteristics of real vibration signals, while the discriminator continuously improves its ability to distinguish between real and fake samples. In particular, a dynamic weight adjustment mechanism is introduced during training. As the number of training rounds increases, the weight coefficients of each loss term in the generator are gradually adjusted. Initially, the focus is on adversarial loss to quickly establish basic generation capabilities. In the middle stage, the expert loss is strengthened to ensure the consistency of physical characteristics. In the later stage, the various losses are balanced to optimize overall performance. This adaptive training strategy effectively avoids the model getting trapped in local optima and ensures the quality of the final generated samples.

[0050] Furthermore, during the training process in step S2, this application also introduces an EMA (Exponential Moving Average Generator) mechanism. Specifically, the system generates a "shadow" model that maintains the standard generator, namely the EMA generator. This EMA generator is not trained directly through backpropagation, but rather its weights are updated each time the standard generator updates. Then, update its weights using an exponential moving average method. Its update rules can be conceptualized as follows:

[0051] ,in It is an attenuation coefficient close to 1 (e.g., 0.999). For the weights of the standard generator, The weights of the EMA generator are defined by the above mechanism, which aims to eliminate the weight oscillations of the standard generator during training, thereby obtaining a more stable model with generally higher quality generated samples. This EMA generator will be used specifically for the synthetic data generation task in subsequent steps.

[0052] Through the above training process, the generator can eventually produce synthetic samples that are highly consistent with real fault vibration signals in terms of time-domain waveform, frequency-domain features and statistical characteristics. These high-quality generated samples provide reliable data support for the subsequent training of defect identification models under small sample conditions.

[0053] Step S3: After the conditional generative adversarial network is trained, the backbone network of multi-scale feature aggregation in its discriminator is transferred to the classifier as the feature extraction backbone, and the classifier is fine-tuned using the denoised signal in the initial training set.

[0054] Step S3 further includes: after the conditional generative adversarial network (GAN) is trained, transferring the network structure and weight parameters of the multi-scale feature aggregation backbone network in its discriminator to the classifier as the feature extraction backbone network of the classifier; inputting the denoised signal and its class labels from the initial training set into the classifier, freezing the weights of the feature extraction backbone network, and training only the classification head at the top of the classifier; unfreezing the weights of the entire classifier network, using a preset learning rate, and fine-tuning the classifier end-to-end using the initial training set to obtain the fine-tuned classifier; the total loss of the classifier is the cross-entropy loss, expressed as: ,in, For the number of categories, For batch size, For sample index, For category indexing, For real labels, This represents the class probabilities predicted by the model.

[0055] Specifically, the classifier, as the core module for ultimately determining the bearing defect type, is essentially a deep learning-based multi-class pattern recognition model. This model consists of a feature extraction backbone network and a classification head. The feature extraction backbone network directly transfers from the multi-scale feature aggregation module in the trained discriminator. This module, through a cascaded residual block structure, can extract discriminative multi-scale depth features from the vibration signal. The classification head consists of fully connected layers and a Softmax activation function, responsible for mapping the extracted depth features to specific defect category probabilities. In the implementation process, after receiving the preprocessed vibration signal, the classifier first performs multi-level feature abstraction on the input signal through the feature extraction backbone network, generating a feature vector containing key fault information. This feature vector is then fed into the classification head, where it undergoes feature transformation through fully connected layers. The Softmax function then calculates the probability distribution of each defect category (including normal, inner race fault, outer race fault, and rolling element fault). The entire training process uses the cross-entropy loss function as the optimization objective. For example, the classifier workflow is illustrated using a denoised inner race fault vibration signal: 1024 After the original signal is input into the classifier, the multi-scale feature aggregation backbone network transferred from the discriminator automatically extracts a 256-dimensional depth feature vector. The feature pattern corresponds to the periodic impact characteristics of the inner ring fault. After the feature vector is transformed by the fully connected layer of the classification head, the four-category probability distribution is output by the Softmax function as [normal: 0.02, outer ring fault: 0.05, inner ring fault: 0.90, rolling element fault: 0.03]. The system finally selects the "inner ring fault" with the highest probability value as the diagnosis result. The whole process shows how the classifier transforms the original vibration signal into an accurate defect category judgment.

[0056] Step S4: Input the synthesized waveform data into the fine-tuned classifier to obtain pseudo-label data. Construct an extended training set based on the pseudo-label data to reinforce the classifier. Input the test set into the reinforced classifier and output the diagnostic results of the rolling bearing defect type.

[0057] In step S4, constructing an expanded training set based on pseudo-label data includes: inputting synthetic waveform data into the fine-tuned classifier, generating pseudo-labels for each synthetic waveform data, selecting pseudo-label data with confidence levels higher than a preset threshold, and merging them with the initial training set to generate an expanded training set.

[0058] Specifically, pseudo-label data is generated through the following process: First, a large amount of synthetic waveform data is generated using an EMA generator. This data itself does not have a real defect category label. Then, this batch of synthetic data is input into a fine-tuned initial classifier, which predicts the defect category of each synthetic sample and outputs the corresponding confidence score. Taking a synthetic vibration signal as an example, when the signal is input into the classifier, the model outputs its probability distribution in the four defect categories as follows: [Normal: 0.02, Inner ring fault: 0.06, Outer ring fault: 0.07, Rolling element fault: 0.85]. Since the predicted probability of the "Rolling element fault" category reaches 0.85, exceeding the set confidence threshold of 0.8, the system labels the synthetic sample as the "Rolling element fault" category, and together with its waveform data, it constitutes a pseudo-label sample. In this way, the system can select samples with high prediction confidence from a large amount of unlabeled synthetic data generated by the generator as pseudo-label training data. These pseudo-label data are merged with the original labeled data and used together for the reinforcement training of the classifier, which effectively expands the size and diversity of the training set and improves the model's ability to identify rare defect categories.

[0059] Using the final trained model, 100 samples were randomly selected from each of the four types of samples (normal, inner race fault, outer race fault, and rolling element fault) in an independent test set for diagnosis. Figure 9 As shown, the confusion matrix of the final model reveals that 399 out of 400 samples were accurately classified, achieving a diagnostic success rate of 99.75%. This demonstrates the model's excellent performance in classification tasks and its high robustness. Then, random inner race fault signals are fed into the final model for rolling bearing manufacturing defect detection, as shown... Figure 10 The image shows the confusion matrix of a random single sample. Figure 10 It can be seen that the fault signal was successfully diagnosed, and the diagnosis result was obtained, that is, the diagnostic signal is an inner ring fault signal, which proves the accuracy of the small sample rolling bearing manufacturing defect inspection model.

[0060] In summary, the rolling bearing defect detection method integrating wavelet denoising and generative adversarial networks proposed in this application has achieved significant beneficial effects in the field of small-sample rolling bearing defect detection. First, the particle swarm optimization Garrote wavelet threshold denoising method overcomes the limitations of traditional threshold selection, achieving an optimal balance between noise suppression and feature preservation through an adaptive optimization mechanism, effectively improving the quality of the input signal and providing a high-fidelity data foundation for subsequent deep learning models. Second, the designed multi-condition jointly guided generative adversarial network successfully solves the data scarcity problem under small-sample conditions. Through dual-domain feature fusion and multiple loss constraints, it ensures that the generated synthetic data not only has visual realism but also maintains a high degree of consistency with the real signal in multiple physical dimensions such as time domain, frequency domain, envelope spectrum, and statistical features.

[0061] Furthermore, the knowledge transfer mechanism proposed in this application transfers the multi-scale feature aggregation network trained in the discriminator to the classifier, enabling the classifier to have strong feature extraction capabilities in the early stages of training, significantly reducing the model's dependence on a large amount of labeled data. The pseudo-label self-training strategy further expands the scale of training data, and by selecting high-confidence generated samples to construct an extended training set, it effectively improves the model's ability to identify rare defect categories. In summary, this application not only solves the core pain point of difficulty in obtaining defect samples in industrial sites to a certain extent, but also ensures the accuracy and reliability of defect detection through a systematic technical architecture.

[0062] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0063] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0064] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications 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 this application.

Claims

1. A rolling bearing defect detection method integrating wavelet denoising and a generative adversarial network, characterized in that, The method comprises: Step S1: obtaining an original vibration signal of a rolling bearing, using a particle swarm optimization algorithm to find an optimal Garrote wavelet threshold, performing noise reduction processing on the original vibration signal, segmenting and labeling the denoised signal, and constructing an initial training set and a test set; The step S1 comprises: acquiring original vibration signals of the rolling bearing under multiple working conditions, wavelet-decomposing the original vibration signals to obtain a series of original wavelet coefficients; generating a set of threshold candidate solutions based on a particle swarm optimization algorithm, inputting each threshold candidate solution into a Garrote threshold function, denoising the original wavelet coefficients to obtain processed wavelet coefficients, the Garrote threshold function being: wherein, is the original wavelet coefficient, is the threshold candidate solution, is a parameter for controlling smoothness, is the processed wavelet coefficient, wavelet-reconstructing all the processed wavelet coefficients to obtain a denoised vibration signal; calculating kurtosis and mean square error of the denoised vibration signal, and constructing a fitness function based on the kurtosis and the mean square error: wherein, is the kurtosis, is the mean square error, is the fitness, and is a preset weight; iteratively updating the threshold candidate solutions, taking maximizing the fitness function as an objective, defining the threshold candidate solution that makes the fitness function maximum as an optimal Garrote wavelet threshold, and obtaining a final denoised signal based on the optimal Garrote wavelet threshold; segmenting and labeling the final denoised signal, combining to generate a sample data set, and dividing the sample data set into the initial training set and the test set; Step S2: establishing a conditional generative adversarial network, wherein a generator of the conditional generative adversarial network receives a denoised signal waveform in the initial training set, a corresponding category label, a physical feature, and random noise, and outputs synthetic waveform data, and the synthetic waveform data is input into a discriminator for training; Step S3: after the conditional generative adversarial network is trained, a multi-scale feature aggregation backbone network in the discriminator is migrated to a classifier as a feature extraction backbone, and the denoised signal in the initial training set is used to fine-tune the classifier; Step S4: the synthetic waveform data is input into the fine-tuned classifier to obtain pseudo-label data, an expanded training set is constructed based on the pseudo-label data, the classifier is intensively trained, and the test set is input into the intensively trained classifier to output a diagnosis result of a rolling bearing defect type.

2. The method of claim 1, wherein, The conditional generative adversarial network established in step S2 comprises: A multi-condition joint guided generative adversarial network is constructed, the generative adversarial network comprises a generator and a discriminator, the generator is sequentially connected by a dual-domain feature fusion layer, a full connection layer, an up-sampling block, and a final output layer, and the discriminator is sequentially connected by a multi-scale feature aggregation backbone network, a label embedding layer, and a multi-layer full connection layer; The denoised signal waveform in the initial training set, the corresponding defect category label, the physical feature vector extracted from the waveform, and the random noise vector are jointly input into the generator, and the generator outputs synthetic waveform data; The synthetic waveform data, the real denoised waveform in the initial training set, and the corresponding defect category label are jointly input into the discriminator.

3. The method of claim 2, wherein, Step S2 further comprises: In the process of training the conditional generative adversarial network, the generator total loss is specifically expressed as: wherein, is the adversarial loss, is the expert loss, is the feature matching loss, and the generator optimization parameter is driven based on the generator total loss; The WGAN-GP loss is taken as the total loss of the discriminator, and the discriminator optimization parameters are driven based on the discriminator total loss; The parameters of the generator and the discriminator are iteratively updated through a back propagation algorithm until the model converges, and a trained generative adversarial network is obtained.

4. The method of claim 3, wherein, The expert loss is composed of a time-domain reconstruction loss , a frequency-domain amplitude loss , an envelope spectrum loss , and a statistical loss , and an expression is: wherein, , , and are weights corresponding to the time-domain reconstruction loss, the frequency-domain amplitude loss, the envelope spectrum loss, and the statistical loss, respectively. The time domain reconstruction loss is: , is real data, is generated data; The frequency domain amplitude loss is: = , is a Fourier transform operation; The envelope spectrum loss is: , is an envelope of real data obtained by a Hilbert transform, is an envelope of generated data obtained by a Hilbert transform, ​​ The statistical loss is: , is a statistical feature vector extracted from real vibration data, is a statistical feature vector extracted from synthetic data output from the generator, wherein, denotes the L1 norm of each prediction error vector.

5. The method of claim 4, wherein, The expression of the adversarial loss is: , represents the score of the discriminator on the generated data .

6. The method of claim 4, wherein, The expression of the feature matching loss is: wherein, is a feature extraction function of the discriminator from the intermediate layer, is the square of the Euclidean distance.

7. The method of claim 3, wherein, The total loss of the discriminator is specifically expressed as: wherein, is the WGAN-GP loss, is the mathematical expectation, is the real data distribution, is the generated data distribution, is the real data, is the synthetic data output by the generator, defined as the generated data, is the score of the discriminator on the real data, is the score of the discriminator on the generated data, is a point randomly interpolated between the real data and the generated data, is the weight of the gradient penalty.

8. The method of claim 1, wherein, Step S3 further comprises: After the conditional generative adversarial network is trained, the network structure and weight parameters of the multi-scale feature aggregation backbone network in the discriminator are migrated to the classifier as a feature extraction backbone network of the classifier; The denoised signal and the category label thereof in the initial training set are input into the classifier, the weight of the feature extraction backbone network is frozen, and only the classification head at the top of the classifier is trained; The weights of the entire classifier network are unfrozen, a preset learning rate is used, the initial training set is used to fine-tune the classifier in an end-to-end manner, and a fine-tuned classifier is obtained; The total loss of the classifier is cross-entropy loss, expressed as: wherein, is the number of classes, is the batch size, is the sample index, is the class index, is the true label, is the class probability predicted by the model.

9. The method of claim 1, wherein, In step S4, the expanded training set is constructed based on the pseudo-label data, which comprises: The synthetic waveform data is input into the fine-tuned classifier to generate pseudo labels for each synthetic waveform data, and pseudo label data with a confidence higher than a preset threshold is screened out and merged with the initial training set to generate an expanded training set.

Images