Bearing fault data non-uniform resampling method based on attention generative adversarial network

Abstract

The invention relates to a method for non-uniform resampling of bearing fault data, which is characterized in that non-uniform resampling is carried out on an original fault signal based on a Generative Adversarial Network (GAN) on the premise of keeping the overall length of the fault signal unchanged. Through fault peak identification and setting of an adjustable up-sampling window, enhanced up-sampling processing is carried out on a fault peak area, and down-sampling compression is carried out on a signal stable area. According to the method, a trainable attention mechanism is introduced, attention weights are adaptively distributed in a generator coding stage, and the model is guided to more effectively reconstruct feature information of a fault peak region. In the training process, similarity matching with a high-sampling-rate reference signal is carried out in a sliding window mode, the alignment position of a peak area is positioned, reconstruction loss in a window is used as an optimization target, and the expression precision of a resampling result is improved. A finally output signal is formed by splicing an up-sampling enhanced peak segment and a down-sampling segment of a stable area, key fault information retention and redundant data compression are both considered, centralized distribution of sampling resources and enhanced reconstruction of fault features are achieved, and the method is suitable for fault data adaptive resampling and subsequent fault diagnosis tasks.

Description

Technical Field

[0001] This invention relates to the field of data resampling, and in particular to a method for non-uniform resampling of bearing fault data based on attention generative adversarial networks under resource-constrained conditions. Background Technology

[0002] With the development of intelligent manufacturing and equipment condition monitoring technologies, fault diagnosis is becoming increasingly important in industrial scenarios. Especially in rotating machinery such as bearing systems, timely extraction and analysis of fault signals are crucial for ensuring equipment safety and extending its service life. Building high-precision fault identification models typically requires a large amount of high-quality fault sample data. However, in practical industrial applications, the low probability of fault occurrence and high cost of data acquisition result in a limited and unevenly distributed amount of fault data available for model training. Particularly in fault spike regions, there is often insufficient sampling points, affecting the modeling and extraction of key features. In fault spike regions, due to their abrupt amplitude changes and short duration, they often contain crucial fault information; insufficient sampling can easily lead to the loss of key diagnostic features, thus affecting the model's recognition accuracy.

[0003] Furthermore, there are significant application contradictions in the selection of sampling frequency during fault signal acquisition and processing. On the one hand, too low a sampling frequency will result in the incomplete capture of fault spike features, leading to a decrease in diagnostic accuracy. On the other hand, while a high sampling frequency can obtain more fault details, it will significantly increase the hardware cost of the acquisition equipment, the pressure on data storage, and the computational burden of subsequent modeling. In addition, edge-cloud collaboration is becoming a trend in industrial intelligent fault diagnosis. Under this framework, it is necessary to send the device status data from the device end or edge end to the cloud. If the sampling frequency is too high, it will lead to high data transmission costs and a high risk of packet loss. In actual industrial scenarios, it is often desirable to maintain a high sampling frequency during the signal acquisition stage to retain key information, while reducing the amount of data during the storage or transmission stage to reduce system resource overhead. Therefore, how to reasonably allocate sampling resources within a limited sampling budget so that information in key fault areas can be fully captured has become one of the key issues in improving fault diagnosis performance.

[0004] To address the aforementioned issues, a common approach is to design appropriate data sampling methods to readjust the total amount or distribution characteristics of the experimental data after acquiring data using low-sampling-frequency sensors. Currently, common solutions in industry and academia mainly fall into three categories: 1) Upsampling method These methods increase the sampling point density in the original signal through interpolation and other means, thereby improving the ability to capture fault details. Common techniques include linear interpolation, spline interpolation, and zero interpolation, which can improve time-domain resolution to some extent. For example, CN 102170276B uses a programmable gate array to upsample ultrasonic signals. However, these methods often come at the cost of increasing the length of the original signal, bringing additional data storage pressure and computational burden; even with an increased number of data points, it is difficult to accurately reproduce the complex dynamic behavior of peak regions.

[0005] 2) Generative data completion method Generative methods based on deep learning, especially Generative Adversarial Networks (GANs), can learn the distribution characteristics of real data and complete sparse or missing parts of the original signal, thus improving data quality and distribution structure to some extent. For example, CN 114091615B established a method and system for completing electricity metering data based on GANs. However, this type of method focuses more on data generation and expansion, and does not necessarily guarantee that the enhancement is of the critical fault area, and its generation results have a certain degree of uncertainty.

[0006] 3) Signal reconstruction method Signal reconstruction methods mainly utilize techniques such as wavelet transform, compressed sensing, or Fourier analysis to extract potential high-frequency information from the original signal and reconstruct it into a more complete signal representation. For example, CN 113852979B reconstructs the original single-tone signal data from compressed sensing measurement data. This type of method has a positive effect on improving the frequency domain representation capability of the overall signal, but most methods require increased computational overhead, and their reconstruction process relies on a global strategy, lacking targeted processing for fault spike regions.

[0007] Although the above methods have alleviated the problems caused by low sampling frequency to some extent, there are still two key shortcomings that limit their practical application in high-precision fault diagnosis tasks.

[0008] 1) Some methods significantly increase the length of the original signal through interpolation or signal extension. While this increases the density of sampling points, it also brings additional data storage and transmission pressure, especially in large-scale industrial monitoring systems. Furthermore, the increased signal length directly increases the computational burden of subsequent models, hindering the deployment of lightweight, high-real-time diagnostic models.

[0009] 2) Even if some methods maintain the original length of the signal during sampling, most adopt a fixed step size sampling strategy, lacking the ability to adaptively perceive the signal content and failing to perform targeted sampling enhancement for fault peak regions. This "average treatment" sampling strategy often ignores non-stationary segments containing key fault information, making it difficult for the model to fully learn the most discriminative features during training, thus limiting the accuracy and robustness of fault identification.

[0010] Therefore, there is an urgent need for a method that, while maintaining the original signal length, can adaptively identify and enhance the sampling density of critical fault regions based on the inherent characteristics of the data. This method would involve targeted upsampling of critical information regions to increase the ability to actively perceive and enhance fault spikes in the time domain, and effective downsampling of redundant and stable regions to retain diagnostically sensitive information while compressing redundant data. This would achieve truly adaptive reconstruction of non-uniform sampling point distribution while balancing storage, computational efficiency, and diagnostic accuracy. Summary of the Invention

[0011] To address the aforementioned issues, this invention proposes a non-uniform resampling method for bearing fault data based on attention-based generative adversarial networks. While ensuring that the overall signal length remains unchanged after resampling, it achieves high-quality fault signal reconstruction and effective sampling point redistribution through fault peak region identification, upsampling window control, attention-guided upsampling of peak regions, and downsampling of stable regions. This provides more representative input data for fault diagnosis and subsequent analysis, while also considering key feature preservation and computational resource optimization.

[0012] The objective of this invention is achieved through the following technical solution: a method for non-uniform resampling of bearing fault data based on attention-based generative adversarial networks, the method comprising the following steps: (1) Collect the original bearing fault signal containing periodic impact spikes and a stable background; (2) Based on the bearing structural parameters and operating speed, calculate the theoretical fault characteristic frequency, and perform envelope processing on the original bearing fault signal to obtain the envelope spectrum. Search for the local maximum peak to obtain the position of the first fault peak, and deduce the positions of the remaining peaks according to the fault cycle step size. (3) Construct an upsampling window centered on each peak position, and change the length of each peak window based on the upsampling factor to improve the temporal resolution of the peak region; (4) The generative adversarial network based on the regional attention mechanism performs local upsampling on the identified peak regions. The position information of the peak regions and the stable regions are embedded into the network model structure as attention masks and assigned different initial weight values ​​respectively. The upsampled signal output by the generator and the original sampled signal are matched for similarity to achieve optimal alignment. The reconstruction loss between the generated window and the reference window and the discrimination loss of the model are calculated and iteratively trained. (5) Extract the upsampling window segment corresponding to each peak region in the output of the model after training, remove the peak regions, extract the main feature components of the stable regions, and reconstruct the signal; (6) The peak and non-peak regions are spliced ​​in sequence to restore the complete signal and obtain a resampled signal containing the peak enhancement structure and the stable compression structure.

[0013] Further, in step (2), the vibration signal is subjected to Hilbert transform to obtain its envelope signal, and then the envelope signal is subjected to fast Fourier transform to obtain its envelope spectrum; with the theoretical fault characteristic frequency as the center, a search window is set to search for the maximum response frequency in the envelope spectrum, the time-domain peak spacing is calculated according to the rate, and the peak search is used to obtain the first peak point with a distance greater than the threshold in the original signal, which is the initial fault peak position.

[0014] Furthermore, in step (3), the upsampling window can be customized with window length and upsampling factor control parameters to limit the upsampling range and factor of the peak region, thereby achieving controllable peak enhancement modeling.

[0015] Furthermore, in step (4), the attention mechanism employs a trainable attention weight vector, which guides the feature extractor to allocate more weights to the fault spike region during the generator's encoding stage, thereby enhancing the generator's feature representation capability in the spike region and containing more fault information for diagnosis.

[0016] Furthermore, in step (4), a one-dimensional attention weight vector with the same signal length is constructed based on the identified set of peak region locations to represent the importance of each position in the input sequence; the attention vector is embedded in the network structure of the generator as a modulation factor for feature channel weighting to continuously strengthen the back gradient propagation of fault peak regions during the training phase and improve the reconstruction capability of the region.

[0017] Furthermore, by using a sliding window, similarity matching and feature alignment are performed on the upsampled signal output by the generator and the original high-sampled signal at different window positions to determine the alignment position of the peak region. The window with the highest similarity is used as the reference region, and the upsampling loss between the generated signal and the reference signal within this window is calculated and used as the optimization objective to guide the training of the GAN generator.

[0018] Furthermore, for each fault peak region, under the condition of uniform upsampling ratio, the corresponding window length of the generator output is obtained according to the original sampling window length, realizing shape matching and alignment within each upsampling window; the segment that best matches the shape of the generated signal is searched in the original sampling window, and its normalized dot product similarity with the generated signal is calculated to determine the optimal matching position, obtain the corresponding alignment reference segment, and calculate the reconstruction error at the alignment position.

[0019] Furthermore, in step (5), the signal is downsampled in the stable region to remove high-frequency details and retain low-frequency trend information, thereby further reducing redundant sampling points.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1) The sampling area is more accurate, which can effectively concentrate and restore the key feature information of the fault.

[0021] Current mainstream fault signal resampling or enhancement methods in industry and academia often rely on uniform upsampling or equal-density processing of the entire signal. This fails to allocate sampling resources appropriately based on the periodicity and local high-information characteristics of fault-related impact signals, resulting in low model training efficiency and weak key feature extraction capabilities. This invention, based on the theoretical characteristic frequencies of rolling bearings, accurately identifies the starting position of fault spikes through spectral alignment analysis and calculates all spike regions periodically from this position. Furthermore, a fixed-length upsampling window is constructed with a uniform upsampling rate to achieve dense reconstruction of key regions. Simultaneously, downsampling compression in stable regions controls the overall data length. This strategy ensures complete coverage and enhancement of spike regions while avoiding unnecessary redundant sampling, improving the ability of the generated data to reconstruct structural diagnostic features.

[0022] 2) A lightweight attention mechanism guides GAN training, significantly improving model focusing efficiency and training effectiveness.

[0023] Existing attention mechanisms in signal modeling often rely on complex self-attention or cross-attention structures, requiring large-scale computation of correlation matrices or the introduction of positional encoding, which burdens training resources and model interpretability. This invention, however, constructs a binary region mask of "peak region - non-peak region" based on the pre-identification peaks. It uses only two trainable scalar parameters (corresponding to peak and stable regions respectively) to control the model's attention, guiding the generator to concentrate resources on feature modeling of peak regions during training. This region-level attention mechanism eliminates the need for point-by-point modeling, avoids the risk of parameter explosion, and allows for dynamic adjustment of weights during training. It combines lightweight structure with adaptive performance, significantly improving the quality of peak detail generation without altering the model structure.

[0024] 3) The dot product alignment loss improves the upsampling shape restoration quality and is more robust against positional errors.

[0025] Existing reconstruction loss functions typically calculate the point-by-point mean square error between the upsampled signal and the original reference signal, ignoring the potential micro-displacement between the two over time. This can easily lead to segments with "identical shape but incorrect position" being misclassified as incorrect samples. This invention introduces a dot product alignment mechanism into the loss function: a fixed window length (equal to the upsampled window length) slides through the original high-sampled signal, performing a dot product match with the generator's output window. The segment with the highest similarity is selected as the target, and the reconstruction loss is calculated at the aligned position. This method effectively eliminates the interference of displacement on the loss value, allowing the model optimization process to focus more on the shape consistency of the peak structure rather than its positional overlap, improving the accuracy of peak reconstruction and training robustness, and is particularly suitable for modeling periodic impact signals. Attached Figure Description

[0026] Figure 1 This is the overall system block diagram.

[0027] Figure 2 This is a workflow for non-uniform resampling.

[0028] Figure 3 This describes the workflow of upsampling fault spike regions in attention-based GANs.

[0029] Figure 4 This is a schematic diagram for searching for frequency spikes characteristic of faults.

[0030] Figure 5 This is a schematic diagram of optimal alignment of fault spikes based on similarity matching.

[0031] Figure 6 This is a schematic diagram for identifying fault spikes and fault cycles.

[0032] Figure 7 This is a comparison chart of the discrete points of the original signal and the resampled signal. Detailed Implementation

[0033] The implementation steps of the present invention will now be described in detail with reference to the accompanying drawings.

[0034] As shown in Figure 1, the overall system block diagram, based on the provided original low-sampling frequency fault signal data, presents a non-uniform resampling method for fault data based on attention-based generative adversarial networks. First, peak regions are identified based on fault characteristic frequencies, allowing for autonomous configuration of upsampling windows and scaling factors. Simultaneously, a region-level attention mechanism is introduced into the generator, guiding the model to focus on peak regions through trainable weights. Sliding window alignment matching is performed in the original high-sampling reference signal to extract the segment most similar to the generated window structure, and the reconstruction error is calculated accordingly for training optimization. Finally, the upsampled peak regions and downsampled stable regions are concatenated for output, achieving non-uniform signal resampling with enhanced information in key regions and optimized compression in non-critical regions while maintaining the overall signal length.

[0035] The specific implementation steps of this invention are as follows: Figure 2 As shown, it mainly includes 7 steps: Step 1: Input the original fault signal The acquired raw fault signal is typically an acceleration time-domain signal at a low sampling rate, containing periodic impact spikes and a stable background. The sampling frequency is usually insufficient to fully capture the spike details, and the overall signal sampling density is relatively limited. The raw signal serves as the input data for the non-uniform resampling method of this invention, providing the basic data source for subsequent fault spike enhancement.

[0036] Step 2: Fault Spike Identification Based on the known structural parameters of the bearing (number of rolling elements, rolling element diameter, pitch circle diameter, contact angle, etc.) and operating speed, the theoretical fault characteristic frequency is calculated using formulas (11)-(14). Subsequently, the Hilbert transform was used to process the envelope of the time-domain signal to obtain the envelope spectrum. And search for local maxima near the characteristic frequencies, such as Figure 4 As shown, the location of the first peak is determined. Then, using formula (4), based on the fault cycle step size... The remaining peak positions are derived recursively, and the position index is further fine-tuned near each expected position using a local sliding window extreme value search method. This allows for the identification of the set of index positions of all spikes in the entire signal. ,like Figure 6 As shown.

[0037] Step 3: Customize the upsampling window length and magnification for the fault spike region. Centered on each identified peak location, extend the construction length forward and backward by [length to be specified]. Peak window. Set a uniform upsampling factor. This makes the length of each spike window in the generator output become... This improves the temporal resolution of peak regions. To ensure the final total signal length remains unchanged, the downsampling ratio in stable regions is adaptively calculated based on the difference between the total number of points required in peak regions and the original number of points, thus completing the budget for the number of non-uniform resampling points. The upsampling configuration parameters include... and Users can flexibly configure the solution according to their custom needs, achieving adaptability and scalability.

[0038] In steps 2 and 3, adaptive identification and autonomous configuration of upsampling parameters are performed using peak regions based on fault characteristic frequencies. In fault diagnosis signals, peak regions often carry crucial transient information, but information folding or distortion easily occurs under low sampling rate conditions. To improve the expression density of peak regions, this invention first constructs an adaptive identification mechanism for fault peaks, combining time-domain envelope analysis and peak localization to achieve rapid calibration of peak positions, and further introduces configurable upsampling parameters to achieve automatic control of the upsampling range and sampling density.

[0039] Under ideal conditions, when a fault occurs, a periodic vibration signal will be generated from the faulty location during bearing operation. Starting from a fault spike in the vibration signal, with a fixed data length, a complete fault cycle of the bearing is formed. Therefore, identifying the first fault spike in the input signal can fully represent all the fault spikes contained in the entire signal.

[0040] First, the initial fault spike location is determined based on envelope analysis, and the sampling frequency is set to... The input vibration signal is a one-dimensional discrete sequence. The envelope signal is obtained by performing a Hilbert transform on the signal.

[0041] in, The input vibration signal is then subjected to a Fast Fourier Transform on the envelope signal. Obtain its envelope spectrum :

[0042] Based on theoretical fault characteristic frequency Set the search window to the center. Hz, search for the maximum response frequency in the envelope spectrum :

[0043] This observation frequency It more closely reflects the impact cycle under actual operating conditions, avoiding the impact of model errors on the calculation of the true time-domain peak value.

[0044] Calculate the time-domain peak spacing based on the observation frequency. :

[0045] Use peak search to obtain all distances greater than [a certain value] in the original signal. The first peak point is the initial peak position defined in this invention. Subsequent steps will be based on the cycle size. Constructing the ideal peak position sequence:

[0046] To improve the accuracy of peak positioning, in each Fine-tune the surrounding settings search window Search for the actual maximum amplitude point from the original signal:

[0047] Finally, the fault spike location sequence was obtained. These locations are used as the basis for subsequent non-uniform upsampling operations. For each peak location... Construct a symmetrical window :

[0048] in The window length is half the window length. Finally, a set of non-overlapping upsampling candidate regions is obtained:

[0049] Meanwhile, the remaining part of the original signal is a stable region. :

[0050] This invention allows users to independently set the upsampling factor for each peak region according to actual system requirements. To ensure that the resampling process does not exceed the original data length limit, the maximum allowable upsampling factor is defined as:

[0051] Based on the bearing's structure and rotation parameters, the theoretical fault characteristic frequency is calculated. , 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 in the acceleration envelope spectrum. For example, a fault in the outer race corresponds to the fault characteristic frequency BPFO; a fault in the inner race to BPFI; a fault in the cage to FTF; and a fault in the balls to BSF. The calculation formulas for these fault characteristic frequencies are shown below:

[0052] 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.

[0053] Step 4: Attention-based Generative Adversarial Network (GAN) upsamples the peak regions. In traditional fault signal upsampling modeling methods, GAN generators apply equal weight to the entire input signal, assigning the same level of attention to all regions during training without distinguishing between fault spikes and stationary regions. When faced with strongly non-stationary, information-concentrated fault signals, the model is easily diluted by numerous informationless stationary regions, making it difficult to focus on key spike responses. While fault spikes carry core diagnostic information, their proportion in the original signal is extremely small, making them easily overlooked during training. Furthermore, the output quality of the upsampling generator is limited, resulting in insufficient detail in the fault spike regions, failing to meet the requirements for high-fidelity data quality.

[0054] To address this, the present invention introduces a region guidance mechanism based on trainable attention weights. This mechanism aims to guide the model to focus on learning the features of fault spike regions through weight adjustment during training, thereby improving the modeling quality of key regions while ensuring modeling efficiency. This is achieved by using the set of spike region locations identified in the previous step. Construct a one-dimensional attention weight vector with the same length as the signal. This is used to represent the importance of each position in the input sequence. The initialization method is:

[0055] in, This represents the initial attention value for the peak region, set to 1.0; The initial value of interest for the stable region is usually set to 0 or a small constant. As trainable parameters in the generator network, they will be further optimized during model training to form a focus distribution that is more adapted to the current task.

[0056] The attention vector It can be embedded in the network structure of the generator as a modulation factor for feature channel weighting, which is used to continuously strengthen the back gradient propagation in fault spike regions during the training phase, thereby improving the reconstruction capability of these regions. Through this mechanism, the model can also automatically focus on important regions during training, ensuring high-quality restoration of key features and significantly improving the recognition and expressive power of upsampled reconstructed signals in diagnostically sensitive areas.

[0057] When upsampling to model fault spike regions, traditional methods typically use the mean square error (MSE) or mean absolute error (MAE) at fixed locations as the reconstruction loss. However, in real-world scenarios, there may be slight phase shifts or alignment errors between the upsampled signal and the original high-sampled signal. In such cases, even if the shapes are similar, conventional loss functions will still produce significant errors and fail to accurately reflect the generation quality.

[0058] To improve the fidelity of the generated signal in the fault spike region, this invention designs an optimal alignment reconstruction loss function based on dot product matching. This method only performs local alignment reconstruction evaluation on the upsampling results in the spike region, avoiding the error amplification problem caused by slight positional deviations, and ensuring that the shape of the spike signal is accurately learned and restored.

[0059] For each fault spike region, its original sampling window length is known to be... Under a uniform upsampling ratio Under the given conditions, the corresponding window length of the generator output is This structure is used to perform shape matching and alignment within each upsampling window.

[0060] In the original high-sampling signal The sliding window searches for the segment that best matches the shape of the generated signal.

[0061] Calculate each one with the generated signal Normalized dot product similarity:

[0062] Determine the optimal matching position:

[0063] Obtain the corresponding alignment reference fragment for:

[0064] At the alignment position The reconstruction error was calculated above:

[0065] in, It is the generated signal. This represents the count of the number of signal points contained in the window. 0 to Repeat this process for all peak windows, and the final total loss is defined as:

[0066] in The number of peak regions participating in upsampling training, where k represents the number of peak windows counted.

[0067] like Figure 3 As shown, this invention employs a Generative Adversarial Network (GAN) with a region attention mechanism to perform local upsampling processing on the identified peak regions. The overall upsampling process includes the following 8 steps: Step 4.1: Construct the initial attention mechanism generative adversarial network A Generative Adversarial Network (GAN) structure with a region attention mechanism is constructed, comprising two main modules: a generator and a discriminator. The generator receives the original low-sampled signal as input and simultaneously incorporates the location information of the peak region from step 2 as an attention mask embedded in the model structure. This mask marks peak and plateau regions, assigning them different initial weight values. Attention weight vector. Trainable parameters are embedded into the network to participate in the feature calculation of each layer of the generator, thereby achieving weighted guidance of local features.

[0068] Step 4.2: Attention mechanism guides generator to upsample in peak regions The original signal is input into the generator. During the generator training process, the attention weight vector is calculated using formula (15). The generator is guided to focus on reconstruction in peak window regions. Peak regions are optimized due to their higher weight, prompting the generator to perform more detailed and accurate feature reconstruction in these regions to enhance the representation of fault features; stable regions have lower weights to reduce the waste of computing resources.

[0069] Step 4.3: Calculate the position of maximum matching dot product using a sliding window. A sliding window is applied to the original high-sampled reference signal, and the generator outputs an upsampled signal. Compared with the original high-sampled signal Similarity matching is performed, and the dot product similarity is calculated position by position according to formula (17) to find the dot product. The largest position is selected as the segment that is closest to the shape of the generated window to achieve optimal alignment.

[0070] Step 4.4: Calculate the reconstruction loss After determining the optimal alignment position, the reconstruction loss between the generated window and the reference window is calculated according to formula (20). This loss effectively guides the generator to gradually optimize its ability to restore local details of the spikes during training, ensuring the consistency and integrity of the spike feature shape.

[0071] Step 4.5: Input discriminator to calculate discrimination loss To further improve the overall distribution authenticity and consistency of the generated data, the generated window and the real window are input into the discriminator together, and the discriminant loss is calculated to measure whether the generated sample has sufficient authenticity and discriminant confusion. This discriminant loss is used to evaluate the authenticity of the generated data in terms of global statistical characteristics, prompting the generator not only to locally restore shape features, but also to maintain consistency with the real signal in the overall signal distribution, thereby improving the naturalness of the generated data and its ability to discriminate confusion.

[0072] Step 4.6: Train the generator and discriminator of the GAN By using both reconstruction loss and discriminant loss as optimization objectives, the generator and discriminator model parameters are iteratively trained to improve the shape consistency and global discriminative ability of the generated signal. Through alternating optimization and iteration of the generator and discriminator, the realism of the generator's upsampled signal and the peak structure reconstruction effect are gradually improved.

[0073] Step 4.7: Determine if the generated signal meets the requirements. After each training round, the current generated result is confirmed to be acceptable by judging whether the loss function has converged or meets the accuracy requirements. If the requirements are met, the training is considered to have converged, and the training process is stopped to enter the next stage. If the requirements are not met, the next round of parameter optimization iteration of the generator and discriminator is carried out to ensure that the final output upsampled spike fragments reach the expected restoration accuracy and authenticity standards.

[0074] Step 4.8: Output upsampled segments of the fault spike region After training, upsampled window segments corresponding to each spike region are extracted from the generator output and used as key enhancement components in the non-uniform resampling structure of the fault signal. These upsampled segments offer significantly higher temporal resolution than the original low-sampled signal and highly reproduce the true spike characteristics in terms of shape, amplitude, and phase, providing high-quality input for subsequent splicing with downsampled stable regions. Ultimately, these upsampled spike region segments are integrated into the overall non-uniform resampling signal at the signal level, significantly improving the density of key fault information and diagnostic effectiveness of the overall signal.

[0075] Step 5: Wavelet transform is used to downsample the stationary region of the signal. In the signal segment after removing peak regions, the main feature components of the stable regions are extracted through wavelet multi-scale decomposition. Based on the downsampling ratio set in step 3, the signal is reconstructed, retaining low-frequency features and removing high-frequency redundancy, thereby effectively reducing the number of redundant data points and ensuring that the total length of the final signal remains unchanged while preserving the overall signal trend and energy characteristics.

[0076] Step 6: Segment upsampling segments from the fault spike region and downsampling segments from the stable region. After processing the spike and non-spike regions separately, they are spliced ​​together in sequence to restore the complete signal. At this point, the faulty spike region has a higher information density due to upsampling, the stable region remains compact due to downsampling, and the overall signal length remains consistent with the original input.

[0077] Step 7: Output the fault signal after non-uniform resampling The final output is a resampled signal containing both spike-enhanced and stationary compression structures. This signal has higher diagnostic effectiveness and structural expressive power, and can be used for subsequent feature extraction, fault classification, or prediction model training. It also significantly improves the performance of spike-type fault modeling without increasing the signal length, making it particularly suitable for real-world industrial applications where data is scarce or sampling resources are limited.

[0078] The above description is merely a preferred embodiment of the present invention and does not constitute a limitation on the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention. For example, wavelet filtering in the invention is only one possible implementation and can be replaced by other schemes such as empirical mode decomposition (EMD). Other possible alternatives can also be listed.

[0079] Figure 6This diagram illustrates the practical application of the spike identification algorithm of this invention on original low-sampling-rate fault signals. The figure shows the acceleration signal of a rolling bearing, where the periodically occurring high-amplitude impact signals are the fault spikes, reflecting the instantaneous response generated by the periodic impact of the rolling element on the defect location. This invention calculates the periodic information based on theoretical fault characteristic frequencies and identifies the first spike position based on spectral envelope analysis. Subsequently, it iteratively derives the remaining spikes using the theoretical period step size and further refines the calibration through a sliding window local extremum search method. The arrows in the figure indicate multiple automatically identified spike points, and the boxed areas show the periodic interval between two adjacent spikes, which is the fault period identified by this invention. This identification result provides a high-confidence spike localization basis for subsequent window construction and non-uniform upsampling strategies.

[0080] Figure 7 The implementation effects of this invention are presented. The upper half of the figure shows the distribution of sampling points in the original signal at a sampling rate of 10 kHz, while the lower half shows the non-uniformly resampled signal obtained after processing by the method of this invention. The shaded area in the figure is marked as the fault spike region, which contains the key information segment with the most concentrated impact features. It can be seen that in the original high-sampling signal, the number of points in the spike region is evenly distributed with other regions. However, in the non-uniformly sampled signal, the number of sampling points in the spike region is significantly increased, while the number of points in the stable region is compressed. This feature indicates that this invention upsamples and reconstructs the spike region through a generative adversarial network, while downsampling the stable region through wavelet transform, ultimately achieving enhanced sampling density of key information segments without changing the overall signal length. This comparison clearly reflects the significant advantages of the non-uniform resampling strategy in information focusing and structure preservation, providing higher quality data support for subsequent fault feature extraction and diagnostic modeling.

[0081] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for non-uniform resampling of bearing fault data based on attention-based generative adversarial networks, characterized in that, The method includes the following steps: (1) Collect the original bearing fault signal containing periodic impact spikes and a stable background; (2) Based on the bearing structural parameters and operating speed, calculate the theoretical fault characteristic frequency, and perform envelope processing on the original bearing fault signal to obtain the envelope spectrum. Search for the local maximum peak to obtain the position of the first fault peak, and deduce the positions of the remaining peaks according to the fault cycle step size. (3) Construct an upsampling window centered on each peak position, and change the length of each peak window based on the upsampling factor to improve the temporal resolution of the peak region; (4) The generative adversarial network based on the regional attention mechanism performs local upsampling on the identified peak regions. The position information of the peak regions and the stable regions are embedded into the network model structure as attention masks and assigned different initial weight values ​​respectively. The upsampled signal output by the generator and the original sampled signal are matched for similarity to achieve optimal alignment. The reconstruction loss between the generated window and the reference window and the discrimination loss of the model are calculated and iteratively trained. (5) Extract the upsampling window segment corresponding to each peak region in the output of the model after training, remove the peak regions, extract the main feature components of the stable regions, and reconstruct the signal; (6) The peak and non-peak regions are spliced ​​in sequence to restore the complete signal and obtain a resampled signal containing the peak enhancement structure and the stable compression structure.

2. The bearing fault data non-uniform resampling method according to claim 1, characterized in that: In step (2), the vibration signal is subjected to Hilbert transform to obtain its envelope signal, and then the envelope signal is subjected to fast Fourier transform to obtain its envelope spectrum. Centered on the theoretical fault characteristic frequency, a search window is set to search for the maximum response frequency in the envelope spectrum. The time-domain peak spacing is calculated based on the rate. Peak search is used to obtain the first peak point in the original signal whose distance is greater than the threshold, which is the initial fault peak position.

3. The bearing fault data non-uniform resampling method according to claim 1, characterized in that: In step (3), the upsampling window can be customized with window length and upsampling factor control parameters to limit the upsampling range and factor of the peak region, thereby achieving controllable peak enhancement modeling.

4. The bearing fault data non-uniform resampling method according to claim 1, characterized in that: In step (4), the attention mechanism uses a trainable attention weight vector, which guides the feature extractor to allocate more weights to the fault spike region during the encoding stage of the generator, so as to enhance the generator's feature representation ability in the spike region and contain more fault information for diagnosis.

5. The bearing fault data non-uniform resampling method according to claim 4, characterized in that: In step (4), a one-dimensional attention weight vector with the same length as the signal is constructed based on the identified set of peak region locations to represent the importance of each position in the input sequence. The attention vector is embedded in the network structure of the generator as a modulation factor for feature channel weighting to continuously strengthen the back gradient propagation of fault peak regions during the training phase and improve the reconstruction capability of the region.

6. The bearing fault data non-uniform resampling method according to claim 1, characterized in that: By using a sliding window, the upsampled signal output by the generator and the original high-sampled signal are similarly matched and feature aligned at different window positions to determine the alignment position of the peak region. The window with the highest similarity is used as the reference region, and the upsampling loss between the generated signal and the reference signal within this window is calculated and used as the optimization objective to guide the training of the GAN generator.

7. The bearing fault data non-uniform resampling method according to claim 6, characterized in that: For each fault spike region, under the condition of uniform upsampling ratio, the corresponding window length output by the generator is obtained according to the original sampling window length, so as to achieve shape matching and alignment within each upsampling window; Search for the segment in the original sampling window that best matches the shape of the generated signal, calculate its normalized dot product similarity with the generated signal, thereby determining the optimal matching position and obtaining the corresponding alignment reference segment. Calculate the reconstruction error at the alignment position.

8. The bearing fault data non-uniform resampling method according to claim 1, characterized in that: In step (5), downsampling is performed on the stable signal region to remove high-frequency details and retain low-frequency trend information, thereby further reducing redundant sampling points.

Images