Bearing fault diagnosis method integrating generation screening and adaptive feature enhancement

Abstract

The invention discloses a bearing fault diagnosis method integrating generative screening and adaptive feature enhancement. The method comprises the steps of S1, collecting original data and expanding a data scale by using a generative adversarial network; s2, extracting a high-quality sample by adopting a correlation screening mechanism; s3, noise is suppressed and key features are strengthened through a feature denoising module; and S4, realizing fault identification by using the bidirectional convolutional network. According to the method, the WGAN-OP and the SCC are combined to realize high-quality data expansion, and the distribution consistency and availability of the generated samples are improved. Through a feature denoising and enhancement integrated module (AFME), in combination with modal decomposition and a multi-scale entropy screening strategy, the signal-to-noise ratio is remarkably improved, key features are highlighted, an adaptive bidirectional time sequence convolutional network (ABiTCN) is constructed, multi-direction dependent capture and feature dynamic fusion are realized, and the diagnosis stability and generalization ability of the model under complex working conditions can be improved.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, specifically a bearing fault diagnosis method that integrates generative screening and adaptive feature enhancement. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent fault diagnosis models based on machine learning and deep learning have gradually replaced traditional rule-based reasoning methods and become the mainstream direction for industry applications. However, the performance of traditional intelligent fault diagnosis models is highly dependent on large-scale, high-quality labeled samples. Obtaining labeled samples requires multiple steps, including accurate determination of equipment fault types by professional technicians, collection and preprocessing of fault signals, and manual labeling of sample categories. This not only consumes a lot of manpower and material resources but is also limited by the randomness and scarcity of fault occurrences. For some faults under extreme working conditions, due to their extremely low probability of occurrence, it is almost impossible to collect sufficient samples. At the same time, there are natural differences in the frequency of occurrence of different fault categories, resulting in an uneven distribution of categories in the sample set, characterized by "redundancy in the majority class and scarcity in the minority class." When the model is trained on a limited or unevenly distributed dataset, it is prone to overfitting (i.e., the model overlearns noise and local features in the training samples, resulting in a significant decrease in its generalization ability to new samples). Furthermore, insufficient learning of minority class samples can lead to classification bias, causing missed or incorrect diagnosis of critical faults, posing a serious threat to production safety.

[0003] To address the aforementioned issues, small-sample fault diagnosis technology has emerged, aiming to achieve high-precision fault identification and classification using a small number of labeled samples. However, current small-sample fault diagnosis technology is still in its developmental stage. Limited by factors such as sample characteristics, data generation quality, signal processing capabilities, and model building mechanisms, a mature and stable technical system has not yet been formed, and the main problems are as follows: (1) Limited number of real samples and uneven distribution of categories: In actual industrial scenarios, normal equipment operation data dominates, and fault data itself is a scarce resource. Some key faults (such as insulation aging faults of high-voltage equipment and micro-deformation faults of precision instruments) have long occurrence cycles and strong concealment, so the number of samples that can be collected is often less than 10, which is far from meeting the training requirements of traditional models. At the same time, the problem of uneven sample distribution has a dual nature. On the one hand, the distribution is uneven among different fault categories; on the other hand, the sample distribution is uneven among different working conditions under the same fault category, resulting in extremely poor adaptability of the model in complex and ever-changing actual working conditions.

[0004] (2) Insufficient authenticity and diversity of generated samples: To alleviate the problem of scarce real samples, sample expansion methods based on generative models such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) have been widely used, but current generation techniques have obvious defects. In terms of authenticity, the generated fault signals often suffer from feature distortion; in terms of diversity, the generated samples are mostly limited to slight variations of existing real samples, making it difficult to cover the sample features under different working conditions and different fault severity levels. As a result, the expanded sample library still has a large number of feature gaps, which cannot effectively improve the generalization ability of the model.

[0005] (3) Significant noise interference in the signal, easily masking key fault features: During the fault signal acquisition process in industrial sites, multi-source noise interference is inevitable. When this noise is superimposed on the fault signal, it will cause a significant reduction in the signal-to-noise ratio, especially for early weak faults, whose characteristic amplitude is usually much smaller than the noise amplitude, and key fault features are severely masked. Traditional signal processing methods are difficult to effectively separate noise from fault features, causing the model to fail to capture early fault information and miss the best maintenance opportunity.

[0006] (4) Insufficient time-dependent modeling capability, making it difficult to fully represent the dynamic characteristics under complex operating conditions: The fault evolution process of industrial equipment has significant temporal characteristics. Fault features accumulate and change gradually over time, and fault signals at different times have strong dependencies. Although most current small-sample fault diagnosis models can extract the spatial features of signals, they lack the ability to model temporal dependencies and cannot capture the dynamic characteristics in the fault evolution process. Especially under complex operating conditions, the temporal correlation of fault signals is more complex, and the model cannot fully represent this dynamic change, resulting in a significant decrease in the accuracy of fault diagnosis under complex operating conditions, which cannot meet the actual operation and maintenance needs.

[0007] In summary, the label sample dependency problem of traditional intelligent fault diagnosis models and the four major challenges faced by current small-sample fault diagnosis technologies severely restrict the in-depth application of intelligent fault diagnosis technologies in real industrial scenarios. There is an urgent need to propose a fault diagnosis scheme that can overcome sample limitations and improve the robustness of feature extraction and temporal modeling capabilities. Summary of the Invention

[0008] This invention aims to provide a bearing fault diagnosis method that integrates generative screening and adaptive feature enhancement, enabling high-precision and robust intelligent fault diagnosis under limited sample conditions.

[0009] To address the above technical problems, the specific solution adopted by this invention is as follows: a bearing fault diagnosis method that integrates generative screening and adaptive feature enhancement, comprising the following steps: S1. Collect raw data and expand the data scale using generative adversarial networks; S2. Use a correlation screening mechanism to extract high-quality samples; S3. Suppress noise and enhance key features through the feature denoising module; S4. Fault identification is achieved using bidirectional convolutional networks.

[0010] As a further optimization of the bearing fault diagnosis method integrating generative filtering and adaptive feature enhancement of the present invention, step S1 specifically includes the following steps: S1-1 Data Acquisition and Preliminary Processing Raw bearing vibration data was collected, and outlier removal, missing value imputation, detrending and normalization were performed on the raw data. The resulting dataset was then divided into training and testing sets. S1-2, High-fidelity data generation A one-sided gradient penalty mechanism is introduced into the WGAN model to construct the WGAN-OP data generation model. The training set is used as the input of the WGAN-OP data generation model to generate data.

[0011] As a further optimization of the bearing fault diagnosis method integrating generative filtering and adaptive feature enhancement of the present invention, the WGAN-OP model uses different loss functions L for D and G. D and L G They are respectively: The formula for calculating the gradient penalty term GP is: in, It is the gradient norm used to enforce Lipschitz constraints, obtained by applying P data and P z (z) is used to calculate the weighted average to obtain x; This represents the gradient operator used to calculate the gradient of the D function as its input; The L2 norm of the gradient; In the WGAN-OP model, the formula for calculating x and the loss functions for D and G are as follows: .

[0012] As a further optimization of the bearing fault diagnosis method integrating generative filtering and adaptive feature enhancement of the present invention, step S2 specifically includes the following: The Spearman rank correlation coefficient is used to measure the rank correlation between generated samples and real samples. A threshold is set to filter the generated data, and only high-quality samples with high correlation are retained to supplement the original dataset.

[0013] As a further optimization of the bearing fault diagnosis method integrating generative filtering and adaptive feature enhancement of the present invention, step S3 specifically includes the following: A feature denoising module, AFME, is constructed. This module first performs multimodal decomposition of the signal using an adaptive FIR filter bank, and extracts impact-sensitive modes with correlation kurtosis (CK) as the optimization objective. Then, it performs a two-layer screening using inter-modal Pearson correlation coefficient and multi-scale entropy (MSE) to remove redundant and high-noise modes. Finally, it uses the modal importance index S... k Weighted fusion of each modality is performed.

[0014] Step S3 specifically includes the following steps: S3-1, Modal Decomposition Mode decomposition of the original signal is performed based on an adaptive finite impulse response filter bank. Let the coefficients of the k-th filter be f. k (l), the input signal is x(n), and the output of the k-th mode is expressed as: Where L is the filter length, u k (n) represents the k-th component of the filter bank output, with the correlation kurtosis CK as the optimization objective function, defined as: Where T is the time delay parameter, and the symbol * denotes autocorrelation calculation, which is performed by maximizing... To achieve adaptive extraction of fault-featured modes; S3-2, Two-layer modal screening First, similarity screening is performed, and then the Pearson correlation r between any two modes is calculated. ij : When |r ij | When the threshold θ is exceeded, it is considered a highly similar mode pair, and only the mode with higher kurtosis is retained; Secondly, multi-scale entropy screening is performed. By calculating the MSE value of each mode and sorting them in ascending order, modes with concentrated information and good noise suppression effect are retained first. S3-3, Feature-weighted reconstruction Define the modal importance index S K for: Among them, Ek K k MSE k C k β represents the energy percentage, kurtosis, and multiscale entropy of the k-th mode, respectively; i The weighting coefficients for each indicator are based on S. k Calculate the modal weights w k And perform weighted reconstruction: Where, x r (n) represents the final reconstructed signal.

[0015] As a further optimization of the bearing fault diagnosis method integrating generative filtering and adaptive feature enhancement of the present invention, step S4 specifically includes the following: An adaptive bidirectional temporal convolutional network ABiTCN is constructed. This model consists of two symmetrical TCN branches, one forward and one backward. An adaptive feature fusion layer (AFF) is introduced, and the fusion weights are learned through a lightweight channel attention mechanism. The contribution ratio of the bidirectional features is adaptively adjusted according to the signal characteristics. The fused features are mapped to a fully connected layer (FC) through global average pooling (GAP). The fault category probability is output by a Softmax classifier to complete the intelligent diagnosis of bearing faults under multiple working conditions.

[0016] As a further optimization of the bearing fault diagnosis method integrating generative filtering and adaptive feature enhancement of the present invention, the two symmetrical TCN branches of ABiTCN are used to extract feature information in different time directions. The forward TCN is used to model the dependency relationship from the past to the future, and the backward TCN is used to model the dependency relationship from the future to the past. Its calculation form is as follows: In the formula, X is the input signal sequence, Reverse(.) represents the time reversal operation, and Y... f and Y b These represent the forward and backward feature outputs, respectively. An adaptive weighted fusion layer (AFF) is introduced, which automatically allocates the fusion ratio of forward and backward features through learnable parameters. The fusion method is defined as follows: Where α∈[0,1] is the fusion weight coefficient, used to control the proportion of forward and backward features in the final representation. This weight is learned through a lightweight channel attention mechanism. Among them, Y f and Y b These represent the channel representations of the forward and backward features after global average pooling, respectively. The fused feature matrix FFF is then subjected to global average pooling and input into a fully connected layer for fault classification, specifically as follows: Where GAP(.) is the global average pooling operation, W c With b c These are the classification layer weights and bias terms, respectively. This represents the probability distribution of fault categories output by the model.

[0017] On the other hand, the present invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the above-described bearing fault diagnosis method.

[0018] On the other hand, the present invention also provides a storage medium including a stored computer program, wherein the computer program, when running, controls the device where the storage medium is located to execute the above-mentioned bearing fault diagnosis method.

[0019] The present invention has the following beneficial effects: 1. This invention effectively improves the stability and sample authenticity of adversarial training by introducing a one-sided gradient penalty mechanism in WGAN-OP. It can generate high-fidelity fault samples that are highly consistent with the real distribution, fundamentally alleviating the problems of small sample size and data imbalance. Combined with the SCC (Spearman Correlation Coefficient) correlation screening strategy, it achieves high correlation screening and redundant sample suppression of the generated data, significantly improving the structural quality and representativeness of the dataset. 2. The Adaptive Feature Mode Enhancement (AFME) module of this invention introduces adaptive FIR filter bank decomposition, two-layer mode selection, and feature weighted reconstruction strategies based on the traditional mode decomposition idea, achieving multi-scale noise suppression and key feature enhancement. This module significantly improves the signal-to-noise ratio and feature separability of signals in high-noise environments, making fault features clearer in the envelope spectrum and providing high-quality input for subsequent classification modeling. 3. In the classification stage, this invention constructs an Adaptive Bidirectional Temporal Convolutional Network (ABiTCN) that fuses long-term and short-term temporal information through forward and backward convolutional branches, and combines this with an Adaptive Feature Fusion Layer (AFF) to achieve dynamic weighting of bidirectional features. This mechanism takes into account both historical and future dependencies, enabling more stable and robust feature representation under complex temporal conditions. Compared with traditional TCNs and recurrent networks, ABiTCN performs better in temporal dependency capture and feature robustness.

[0020] 4. Experimental results on the Southeast University SEU-AE signal dataset and the Xi'an Jiaotong University XJTU-SY accelerated life test dataset show that the fault diagnosis method of this invention achieves significant performance improvements under both finite and imbalanced sample conditions, with average recognition accuracies reaching 99.95% and 99.02%, respectively. Compared with mainstream deep diagnostic models, the framework of this invention exhibits significant advantages in diagnostic accuracy, generalization ability, and stability, verifying its universality and engineering application potential under complex working conditions. Attached Figure Description

[0021] Figure 1 This is a logic block diagram of the fault diagnosis method of the present invention; Figure 2 This is a flowchart illustrating the fault diagnosis method of the present invention; Figure 3 This is an AE signal rolling bearing fault test bench for Southeast University. Figure 4 Generate sample SCC correlation coefficients for AE signal fault data; Figure 5 Envelope spectra of the AE signal at a thickness of 0.5 mm for both the original and generated samples; Figure 6 Envelope spectra before and after AFME decomposition and reconstruction of 0.5mm fault data; Figure 7 Accuracy for different imbalance ratios on the AE dataset; Figure 8 The iterative graph, T-SNE graph, and confusion matrix graph of the fault diagnosis results of the ABiTCN model are shown. Figure 9 This serves as a testbed for the XJTU-SY dataset. Figure 10 Generate sample SCC correlation coefficients for XJTU-SY fault data; Figure 11 Envelope spectra of the original and generated samples of Bearing2_1; Figure 12 Figure showing the ablation experiment results for the XJTU-SY dataset; Figure 13 The iterative graph, T-SNE graph, and confusion matrix graph of the fault diagnosis results of the ABiTCN model are shown. Figure 14 The figure shows the results of the comparative experiment on the XJTU-SY dataset. Detailed Implementation

[0022] To better understand the present invention, the following embodiments further illustrate the content of the present invention, but the content of the present invention is not limited to the following embodiments.

[0023] This invention provides a bearing fault diagnosis method that integrates generative filtering and adaptive feature enhancement, comprising the following steps: S1. Collect raw data and expand the data scale using generative adversarial networks; S2. Use a correlation screening mechanism to extract high-quality samples; S3. Suppress noise and enhance key features through the feature denoising module; S4. Fault identification is achieved using bidirectional convolutional networks.

[0024] Step S1 specifically includes the following steps: S1-1 Data Acquisition and Preliminary Processing Raw bearing vibration data was collected, and outlier removal, missing value imputation, detrending and normalization were performed on the raw data. The resulting dataset was then divided into training and testing sets. S1-2, High-fidelity data generation A one-sided gradient penalty mechanism is introduced into the WGAN model to construct the WGAN-OP data generation model. The training set is used as the input of the WGAN-OP data generation model to generate data.

[0025] To address the common problems of pattern collapse and gradient vanishing during training in traditional GANs, this invention employs an improved generative adversarial network, WGAN-OP, which introduces a one-sided gradient penalty (OGP) mechanism to achieve high-fidelity generation of fault data under complex working conditions.

[0026] GANs primarily consist of a Generator (G) and a Discriminator (D). G aims to make the generated fake data more realistic by inputting noise and feedback into D. D compares the data generated by G with the original data, determining whether the data is real or fake, and then passes the feedback back to G. The interaction between G and D achieves a maximum and minimum game with a cross-entropy objective. Where E represents the expected value, which helps the average D or G to perform on the distribution of real or generated data; Z is random noise; Pdata(x) is the distribution of real samples; Pz(z) is the distribution of generated data samples; D(x) represents the probability that x is a real sample; and G(z) is a generated sample.

[0027] Training instability and model collapse are common problems in GAN models, often hindering their effectiveness. To address these issues, WGAN was introduced, a novel approach that utilizes one-sided gradient-penalized distance as a loss function. One-sided gradient-penalized distance is an advanced metric for measuring distributional dissimilarity, representing a substantial advancement in overcoming the limitations of traditional GAN ​​models. To improve stability and efficiency during training, the WGAN model employs different loss functions L for D and G. D and L G They are respectively: The WGAN model is further refined by incorporating Lipschitz constraints, which limit the maximum gradient of the function and play a crucial role in enhancing the model's stability. In WGAN-OP, D strengthens the Lipschitz continuity to ensure its stability, which could otherwise lead to instability during training. The gradient penalty term (GP) is calculated as follows: in It is the gradient norm used to enforce the Lipschitz constraint; x is obtained by weighted averaging of Pdata and Pz(z); This represents the gradient operator used to calculate the gradient of the D function as its input; The L2 norm of the gradient is used, and this modification significantly alleviates problems associated with unstable training and model collapse. In the WGAN-OP model, the formula for calculating x and the loss functions for D and G are: The Feature-Mode Enhancement (AFME) module addresses the issue that in industrial equipment operating under complex conditions, vibration signals are often affected by strong noise, random interference, and structural resonance, leading to the masking or distortion of fault features and severely impacting diagnostic accuracy. To improve signal distinguishability and feature purity, this invention designs an Adaptive Feature-Mode Enhancement (AFME) module. This module achieves multi-scale noise suppression and fault impact feature enhancement through adaptive FIR filter bank mode decomposition and feature weighted reconstruction, providing high-fidelity input for subsequent classification models.

[0028] AFME first performs mode decomposition on the original signal based on an adaptive finite impulse response (FIR) filter bank. Let the coefficients of the k-th filter be fk(l), and the input signal be x(n), then the output of the k-th mode can be expressed as: Where L is the filter length, and uk(n) represents the k-th component of the filter bank output. To effectively highlight the potential periodic impact characteristics in the vibration signal, AFME uses correlation kurtosis (CK) as the optimization objective function. Correlation kurtosis comprehensively reflects the impulsiveness and periodicity of the signal, and is defined as: Where T is the time delay parameter, and the symbol * denotes autocorrelation calculation. By maximizing This method makes the modes output by the filter more sensitive to impulse components, thereby enabling adaptive extraction of modes containing fault characteristics. Compared with traditional empirical mode decomposition (EMD) and variational mode decomposition (VMD), this method does not require setting the number of modes and penalty parameters, and has stronger adaptability and stability.

[0029] After filter bank decomposition, multiple modal components are typically obtained. However, not all modes carry valid fault information; some components exhibit strong noise or are highly similar to other modes. To address this, AFME designed a two-layer modal screening strategy to extract high-value modes: First, similarity screening is performed. This is done by calculating the Pearson correlation r between any two modes. ij : When |r ij | When the threshold θ is exceeded, it is considered a highly similar mode pair, and only the mode with higher kurtosis is retained to reduce redundancy and retain the component with more significant impact characteristics.

[0030] Secondly, multiscale entropy (MSE) screening is performed. MSE measures the complexity and randomness of modal signals; a smaller value indicates a more ordered signal and lower noise content. By calculating the MSE value of each mode and sorting it in ascending order, modes with concentrated information and good noise suppression can be preferentially retained. After these two screening steps, AFME obtains a set of effective modes containing key fault information, laying the foundation for subsequent signal reconstruction.

[0031] Traditional modal reconstruction often employs equal-weighted superposition, failing to consider the differences in fault feature representation capabilities among different modes. Therefore, this invention proposes a feature-sensitivity-based weighted reconstruction mechanism that adaptively adjusts the reconstruction weights according to the importance of each mode, thereby achieving "targeted enhancement" of fault features. The modal importance index S is defined. K for: Among them, E k Kk MSE k C k β represents the energy percentage, kurtosis, and multiscale entropy of the k-th mode, respectively; i These are the weighting coefficients for each indicator. Based on S... k Calculate the modal weights w k And perform weighted reconstruction: Where xr(n) is the final reconstructed signal. This mechanism enhances high-contribution modes during reconstruction while weakening noise modes, thereby significantly improving the signal-to-noise ratio and making the impulse features in the envelope spectrum more prominent.

[0032] In complex industrial vibration signals, fault features often exhibit symmetrical or mutually modulated relationships. Relying solely on unidirectional convolution for modeling can easily lead to the loss of temporal discriminative information. To address this, this invention constructs an Adaptive Bidirectional Temporal Convolutional Network (ABiTCN), which simultaneously considers the forward and backward dependencies of the time series in its network structure, thereby achieving complete modeling of temporal information and joint expression of contextual features.

[0033] ABiTCN consists of two symmetrical TCN branches, used to extract feature information in different time directions. The forward TCN models dependencies from the past to the future, while the backward TCN models dependencies from the future to the past. Its computational form is as follows: In the formula, X is the input signal sequence, Reverse(.) represents the time reversal operation, and Y... f and Y b These represent the forward and backward feature outputs, respectively. After inverse recovery, the backward output Y... b Rearranged to be with Y f The alignment on the time axis allows the model to utilize both preceding and following temporal features, resulting in a strong response to both periodic components and non-stationary features in the signal.

[0034] Traditional BiTCNs typically fuse bidirectional features using feature concatenation or average weighting, but fixed fusion strategies cannot dynamically adjust feature weights based on signal characteristics. Therefore, this invention introduces an Adaptive Feature Fusion (AFF) layer, which automatically allocates the fusion ratio of forward and backward features using learnable parameters. The fusion method is defined as follows: Where α∈[0,1] are the fusion weight coefficients, used to control the proportion of forward and backward features in the final representation. These weights are learned through a lightweight channel attention mechanism. Among them, Y f and Y b These represent the channel representations of the forward and backward features after global average pooling, respectively. In this way, the model can adaptively adjust the fusion ratio of the bidirectional features according to the energy distribution and feature intensity of the signal in different time periods, thereby maintaining stable feature representation capabilities under different operating conditions.

[0035] The fused feature matrix FFF is then subjected to global average pooling and input into a fully connected layer for fault classification, specifically as follows: Where GAP(.) is the global average pooling operation, W c With b c These are the classification layer weights and bias terms, respectively. This represents the probability distribution of fault categories output by the model.

[0036] Compared with traditional unidirectional TCN, ABiTCN significantly enhances the contextual modeling capability of temporal features while maintaining parallel computing efficiency. Its main innovations include: (1) introducing a bidirectional convolutional structure, which can capture historical and future dependency information simultaneously, improving the discrimination capability of complex signals; (2) designing an adaptive fusion mechanism, realizing dynamic weighting of bidirectional features, avoiding the limitations of fixed fusion strategies; and (3) adopting a lightweight structure, maintaining low computational complexity while ensuring feature transfer stability.

[0037] The specific process is as follows: (1) Data Acquisition and Preliminary Processing: Vibration signals of rolling bearings under normal conditions and various typical fault conditions were collected through the experimental platform to construct an original sample set. To ensure the validity and consistency of the input data, outlier removal, missing value imputation, detrending, and normalization were performed on the original signals. Subsequently, the dataset was divided into training and test sets in a 7:3 ratio.

[0038] (2) High-fidelity data generation (WGAN-OP): To address the issues of scarce fault samples and uneven class distribution, this invention introduces the WGAN-OP (Wasserstein GAN with One-sided Gradient Penalty) generation model to achieve high-fidelity sample expansion. This model strengthens the Lipschitz continuity of the discriminator by adding a one-sided gradient penalty mechanism to the traditional WGAN, effectively suppressing gradient vanishing and mode collapse phenomena, thereby improving the stability and realism of the generated samples. The samples generated by WGAN-OP are highly consistent with real data in terms of distribution characteristics, providing rich and high-quality data support for subsequent feature extraction and classification modeling.

[0039] (3) Relevance Screening (SCC): The generated faulty samples need further screening to ensure their representativeness and quality. This invention uses the Spearman Rank Correlation Coefficient (SCC) to measure the rank correlation between the generated samples and the real samples, and sets a threshold to screen the generated data. Only high-quality samples with high correlation are retained to supplement the original dataset, thereby improving the statistical consistency and effectiveness of the expanded dataset.

[0040] (4) Feature Denoising and Enhancement (AFME): Under complex operating conditions, vibration signals are often mixed with strong noise and random interference, affecting fault feature extraction. Therefore, this invention designs a feature mode extraction module to decompose and denoise the signal. AFME first performs multimodal decomposition of the signal using an adaptive finite impulse response (FIR) filter bank, and uses correlation kurtosis (CK) as the optimization target to extract impact-sensitive modes. Subsequently, a two-layer screening is performed using intermodal Pearson correlation coefficient and multiscale entropy (MSE) to remove redundant and high-noise modes. In the signal reconstruction stage, based on the modal importance index S... k Weighted fusion of various modes enhances high-energy, low-entropy modes while suppressing noise modes. After AFME processing, the signal-to-noise ratio is significantly improved, and the fault impact characteristics in the envelope spectrum are more prominent.

[0041] (5) Bidirectional Temporal Modeling and Classification (ABiTCN): Based on the enhanced signal, this invention constructs an Adaptive Bidirectional Temporal Convolutional Network (ABiTCN) for fault identification. This model consists of two symmetrical TCN branches, forward and backward, responsible for modeling the historical and future dependencies of the time series, respectively, thus achieving complete capture of temporal information. To achieve dynamic fusion of forward and backward features, ABiTCN introduces an Adaptive Feature Fusion (AFF) layer, which learns fusion weights through a lightweight channel attention mechanism, adaptively adjusting the contribution ratio of bidirectional features according to signal characteristics. Finally, the fused features are mapped through Global Average Pooling (GAP) and a fully connected layer (FC), and the Softmax classifier outputs the fault category probability, achieving intelligent diagnosis under multiple operating conditions.

[0042] Dataset Introduction: In the data acquisition experiment, the spindle speed was set to 30 r / min, and the rolling bearing was UC210. ​​The AE acquisition system consisted of a UT-1000 wideband AE sensor, a PCI-2 acoustic emission acquisition card, a 40 dB preamplifier, and an industrial computer. The sampling frequency was 500kHz. Three types of AE signals were acquired under normal, fault, and fault conditions: a normal rolling bearing, a rolling bearing with a 0.5mm inner ring crack, and a rolling bearing with a 0.8mm inner ring crack. The acquisition platform was as follows: Figure 3 As shown.

[0043] As mentioned above, the magnitude of the imbalance ratio affects the accuracy of fault diagnosis. This invention provides experimental verification. The bearing fault diagnosis experiment is set with imbalance ratios of 2:1, 3:1, 4:1, and 5:1, with each sample length set to 200. Under an imbalance ratio of 2:1, the dataset contains 1000 samples. The experimental task design for the AE dataset is shown in the table below.

[0044] Based on the table above, four innovative experiments with different imbalance ratios were obtained. In subsequent experiments, the samples generated by WGAN-OP were added to the datasets of the four imbalanced experimental tasks, forming four new balanced datasets.

[0045] WGAN-OP generates samples for evaluation. In this example, WGAN-OP generates new samples for each category and evaluates their quality using SCC and envelope spectrum comparisons. A higher Spearman correlation coefficient indicates a closer resemblance to the real sample. Taking Southeast University's AE signal fault data as an example, the WGAN-OP model generates 50 new samples for each fault category. Since a higher SCC correlation coefficient indicates a closer resemblance to the real sample, the SCC correlation coefficient between the generated sample and the original sample for each fault category is calculated, and the two most similar samples are selected. The calculation results are as follows: Figure 4 As shown.

[0046] Depend on Figure 4 It can be seen that the 10th and 31st generated samples of the 0.5mm fault data for the AE signal are most similar to the original samples, and the 7th and 24th generated samples of the 0.8mm fault data for the AE signal are most similar to the original samples, indicating that these two sets of generated fault samples are closest to the real samples. To further verify the similarity between the generated samples and the real samples, envelope spectra of the real and generated samples of the 0.5mm fault data are plotted, as shown below. Figure 5 As shown.

[0047] Depend on Figure 5 As can be seen, the generated samples filtered by SCC have the same fault characteristic frequencies and their harmonics as the original samples. The figure clearly shows that the generated samples can basically reflect the regular distribution of the original samples, which is beneficial for the fault diagnosis model to grasp the characteristics of the data. In summary, the above two evaluation indicators verify that WGAN-OP has excellent data generation capabilities and can generate high-quality signals that are highly similar to the original signals.

[0048] from Figure 5 It can be seen that the fault data contains a large amount of noise, which seriously affects the fault diagnosis of the neural network. This invention uses AFME for decomposition and reconstruction. Taking the 0.5mm inner circle fault data of the AE signal from Southeast University as an example, the generated samples are concatenated with the original samples to form a large sample dataset. AFME decomposition and reconstruction are then applied to the large sample dataset, and the envelope spectrum is plotted on the reconstructed data. The results are as follows. Figure 6 As shown.

[0049] from Figure 6 It can be seen that the amplitude of the fault data significantly increases after AFME decomposition and reconstruction, and the fault characteristics are more prominent, which is beneficial to the neural network fault learning and diagnosis. Although the fault frequency doubling does not change significantly compared to the original data samples, the fault characteristics are more prominent, thereby improving the fault diagnosis performance of the neural network. Subsequent work will also verify the role of AFME in this invention through neural networks.

[0050] The specific method for verifying fault diagnosis under limited data conditions is as follows: The impact of different imbalance ratios on fault diagnosis The purpose of this experiment is to study the diagnostic performance of the ABiTCN model under different imbalance ratios. The fault diagnosis experiments were set with imbalance ratios of 2:1, 3:1, 4:1, and 5:1. Simultaneously, this invention considers that in practical engineering problems, the amount of normal bearing data is much greater than the amount of fault data. A high-quality signal highly similar to the fault data is generated using a WGAN-OP generative adversarial network model. Based on the SCC correlation coefficient, the high-quality signal most similar to the original signal is selected to address the problem of insufficient fault data. The amount of fault data is increased for different imbalance ratios to achieve the ideal state of an imbalance ratio of 1:1. In subsequent work, the effectiveness of the proposed fault diagnosis method will be verified by comparing different experimental methods. Method 1: Verify the diagnostic performance of four sets of unbalanced original data in the ABiTCN model. Experiment 2: Increase fault sample data using the WGAN-OP model to achieve the ideal state of 1:1 for all four sets of data, and input them into the ABiTCN model to verify its diagnostic performance. Experiment 3: Use AFME to decompose and reconstruct the four sets of unbalanced original data, and input them into the ABiTCN model to verify its diagnostic performance. Experiment 4: Under the ideal condition that all four sets of imbalanced original data reach a 1:1 ratio, AFME decomposition and reconstruction were applied, and the data were input into the TCN model to verify its diagnostic performance. Experiment 5: Under the ideal condition that all four sets of imbalanced original data reach a 1:1 ratio, AFME decomposition and reconstruction were applied, and the data were input into the ABiTCN model to verify its diagnostic performance. All data samples in the above five experiments were divided into 70% training set and 30% test set. To consider diagnostic efficiency, the model training cycle was set to 10 times, and each experiment was performed 5 times. The average accuracy was taken. The experimental results are as follows: Figure 7 As shown.

[0051] like Figure 7As shown, in Method 1, as the imbalance ratio increases from 2:1 to 5:1, the diagnostic accuracy of the ABiTCN model improves from 71.4% to 83.32%. During this process, the amount of fault data for inner races of 0.5mm and 0.8mm remains constant, while the amount of normal bearing data gradually increases, leading to an increase in the number of training samples. The ABiTCN model can learn more feature information from these training samples, thus improving diagnostic accuracy. In Method 2, the amount of fault data for different imbalance ratios is increased using a WGAN-OP generative adversarial network model to achieve the ideal imbalance ratio of 1:1. Fault diagnosis using the ABiTCN model shows that with the increase in fault data, good results are achieved under all four different imbalance ratios, with accuracy improvements of 7.39%, 10.09%, 7.7%, and 6.24% respectively compared to when the fault data amount was not increased. This demonstrates that the number of fault samples is crucial for deep learning-based fault diagnosis models. In Method 3, after using AFME decomposition and reconstruction on the four sets of imbalanced raw data, the fault diagnosis accuracy was significantly improved. Under the four different imbalance ratios of the raw data, the accuracy was above 97%, fully demonstrating the effectiveness of AFME decomposition and reconstruction. In Method 4, the amount of fault data with different imbalance ratios was increased using the WGAN-OP generative adversarial network model to achieve the ideal state of a 1:1 imbalance ratio. After applying AFME decomposition and reconstruction to the processed data, it was input into the TCN network. Under the four sets of data, the average accuracy reached 98.80%, 99.12%, 100.00%, and 86.03%.

[0052] Based on four sets of imbalanced raw data, five experimental methods were tested to... Figure 5 As can be seen, under four different imbalance ratios of original data, the original data + WGAN-OP + AFME + ABiTCN method proposed in this invention significantly outperforms the other four methods. Moreover, under the original data condition of an imbalance ratio of 4:1, the diagnostic accuracy of the method proposed in this invention reaches an astonishing 100%. It should be emphasized that the ABiTCN model of this invention is trained only 10 times, fully demonstrating the excellent fault diagnosis performance of the method proposed in this invention.

[0053] ABiTCN Model Fault Diagnosis Visualization With the original data having an imbalance ratio of 5:1, WGAN-OP+AFME was used to process the data, which was then input into the ABiTCN model to verify the model's stability and convergence speed. The model iteration graph was plotted, and the verification accuracy curves for training accuracy, test accuracy, training loss, and test loss are shown below. Figure 8 As shown in Figure a, to visually demonstrate the feature extraction performance of the model, the features extracted by the fault classification model are visualized using T-SNE. The visualization results are as follows: Figure 8 As shown in b. The detailed performance of the model is represented in the form of a confusion matrix, as shown in... Figure 8 As shown in c.

[0054] from Figure 8 As can be seen from a, the prediction accuracy of this model increases almost linearly, while the loss function decreases linearly. After approximately 10 iterations, the model's test accuracy remains around 100%. This is why the training cycle of the ABiTCN model in this invention is only set to 10 iterations; the test accuracy is around 100%, the test loss is at a low level, and the advantages of model computational efficiency are retained. Figure 8 As can be seen from b, the feature clusters of each category are different and non-overlapping, indicating that the method proposed in this invention can improve the fault diagnosis capability of the ABiTCN model by generating high-quality data and highlighting fault features through AFME decomposition and reconstruction, thereby accurately identifying different types of faults. Figure 8 As shown in Figure c, the horizontal axis of the confusion matrix represents the predicted label, and the vertical axis represents the true label of the input data. The elements on the main diagonal of the confusion matrix represent the percentage of correctly classified samples, while the other elements represent the percentage of misclassified samples. Figure 8 As shown in c, the classification accuracy for both normal and faulty data reached 100%. Therefore, it can be concluded that this method not only has satisfactory results in terms of accuracy, but also exhibits excellent performance in terms of stability and convergence speed.

[0055] This embodiment uses the Xi'an Jiaotong University XJTU-SY rolling bearing accelerated life test dataset to illustrate the fault diagnosis capability of the invention method. The test bench used for the XJTU-SY dataset is as follows: Figure 9 As shown.

[0056] The platform consists of an AC motor, a motor speed controller, a shaft, support bearings, a hydraulic loading system, and test bearings. The test bearings are LDK UER204 rolling bearings. Two PCB 352C33 unidirectional accelerometers are fixed to the horizontal and vertical directions of the test bearings via magnetic bases, respectively. A DT9837 portable dynamic signal acquisition device was used to collect vibration signals during the experiment. The sampling frequency was set to 25.6 kHz, the sampling interval was 1 min, and each sampling duration was 1.28 s. Three types of operating conditions were designed for the experiment, with five bearings in each condition. The dataset contains the full life cycle vibration signals of 15 rolling bearings. This embodiment selects seven of these datasets as the experimental dataset. The information of the selected datasets is shown in the table below.

[0057] WGAN-OP Sample Evaluation Similar to Example 1, using the Xi'an Jiaotong University XJTU-SY rolling bearing accelerated life test dataset as an example, the WGAN-OP model generates 50 new samples for each fault category. Since a higher SCC correlation coefficient indicates a closer resemblance to the real sample, the SCC correlation coefficient between the generated sample and the original sample for each fault category is calculated. The two samples most similar to the original sample are then selected. The calculation results are as follows: Figure 10 As shown.

[0058] Taking Bearing2_1 fault data as an example, from Figure 10 It can be seen that the 3rd and 10th generated samples have the highest SCC correlation coefficients with the original samples, indicating that these two fault-generated samples are closest to the real samples. Due to space limitations, the SCC correlation coefficients of the remaining 6 fault data generated samples are not analyzed one by one. To further verify the similarity between the generated samples and the real samples, envelope spectra of the original Bearing2_1 fault sample and its 3rd and 10th generated samples are plotted, as shown below. Figure 11 As shown.

[0059] Depend on Figure 11 As can be seen, the generated samples filtered by SCC have the same fault feature frequencies and their harmonics as the original samples. The figure clearly shows that the generated samples can basically reflect the regular distribution of the original samples. In summary, the above two evaluation metrics verify that the WGAN-OP model also has excellent data generation capabilities on the Xi'an Jiaotong University XJTU-SY rolling bearing accelerated life test dataset, and can generate high-quality signals that are highly similar to the original signals.

[0060] Fault diagnosis under limited data conditions a: The impact of different bearing fault data volumes on fault diagnosis Because the XJTU-SY rolling bearing accelerated life test dataset from Xi'an Jiaotong University lacks normal bearing data, it's impossible to consider the impact of different imbalance ratios on fault diagnosis. This invention treats this as small-sample fault data and utilizes the WGAN-OP model to generate fake fault samples to address the low accuracy of fault diagnosis with small-sample fault data. The purpose of this experiment is to study the diagnostic performance of the ABiTCN model under different numbers of fault samples. Each sample length was set to 500, and 455 samples were used for training and testing the model in the original small-sample fault data. The diagnostic results are shown in the table below.

[0061] As shown in the table above, when the number of fault samples was 455, the diagnostic accuracy of the ABiTCN model was only 55.65%. However, as the number of fault samples gradually increased, the diagnostic accuracy improved from 55.65% to 81.27%. The highest diagnostic accuracy was achieved at 1365 fault samples, reaching 81.27%. When the number of fault samples reached 1813, the diagnostic accuracy decreased to 79.70%. Literature review revealed that too many fake samples can interfere with the neural network's learning of fault features and fault classification. In summary, compared to training and testing using only the original fault samples, increasing the number of fault samples by 910 improved the diagnostic accuracy of the ABiTCN model by 25.62%, fully demonstrating that using the WGAN-OP model to generate high-quality fault data and increasing the number of small-sample fault data can improve fault diagnosis accuracy.

[0062] b: Ablation experiment Ablation experiments were conducted using the XJTU-SY rolling bearing accelerated life test dataset from Xi'an Jiaotong University. The proposed method was compared with four other methods, with diagnostic accuracy as the metric. This verified the superiority of the proposed method on the XJTU-SY rolling bearing accelerated life test dataset. The four fault diagnosis methods were: original data + ABiTCN, original data + FMD + ABiTCN, original data + WGAN-OP + ABiTCN, and original data + WGAN-OP + FMD + TCN. The number of fault samples was increased using the WGAN-OP model. Since the impact of adding 455, 910, and 1358 fault samples on the diagnostic performance of the ABiTCN model has already been verified above, this experiment will not be repeated here. In the ablation experiment, the number of samples increased by the WGAN-OP model was assumed to be 455. That is, the original data (Data) refers to 455 fault samples, and Data + WGAN-OP refers to 910 fault samples. The diagnostic accuracy was recorded in [the relevant section]. Figure 12 middle.

[0063] like Figure 12As shown, the result of using the original data + ABiTCN was the least ideal, with an accuracy of only 55.65%. For the original data + WGAN-OP + ABiTCN method, the accuracy reached 67.70%, showing an improvement compared to not using WGAN-OP to increase the number of fault samples. Furthermore, compared to the original data + FMD + ABiTCN method, the fault diagnosis accuracy significantly improved after using FME decomposition and reconstruction, reaching 98.88%, fully demonstrating the effectiveness of FME decomposition and reconstruction. To illustrate the superiority of the proposed ABiTCN model in fault diagnosis, a comparison with the TCN model was conducted, and the results fully verified the excellent performance of the ABiTCN model. The proposed method achieved a diagnostic accuracy of 99.02%. In summary, the proposed method also exhibits good fault diagnosis performance on the Xi'an Jiaotong University XJTU-SY rolling bearing accelerated life test dataset.

[0064] c: Visualization of ABiTCN model fault diagnosis Taking 1813 fault sample data as an example, the iterative graph, T-SNE graph, and confusion matrix graph of the fault diagnosis results of the ABiTCN model are plotted, as follows: Figure 13 As shown.

[0065] from Figure 13 (a) It can be seen that the prediction accuracy of the model increases almost linearly, while the loss function decreases linearly. After approximately 10 iterations, the model's test accuracy remains around 100%. Figure 13 (b) It can be seen that the feature clusters of each category are different and non-overlapping, indicating that the method proposed in this invention can accurately identify different types of faults. From Figure 13 (c) It can be seen that the classification accuracy of the fault data labeled 0, 3, 4, 5, and 6 all reached 100%, and the classification accuracy of the fault data labeled 1 and 2 also reached 97.65% and 96.59% respectively, which are basically at an ideal level. Therefore, it can be concluded that the method proposed in this invention not only has satisfactory results on the Southeast University AE signal rolling bearing fault dataset, but also has excellent performance on the Xi'an Jiaotong University XJTU-SY rolling bearing accelerated life test dataset.

[0066] By comparing the method of this invention with four other advanced methods on the Xi'an Jiaotong University XJTU-SY rolling bearing accelerated life test dataset, using diagnostic accuracy as the metric, the superiority of the method based on original data + WGAN-OP + FME + ABiTCN is demonstrated. The four fault diagnosis methods are original data + WGAN + CNN, original data + VMD + SCC + TCN, original data + SMOTE + Attention + LSTM, and original data + ADASYN + ABiTCN. The diagnostic accuracy of each experimental method is shown below. Figure 14As shown.

[0067] like Figure 14 As shown, the results of the original data + WGAN + CNN method are the least ideal, with an average accuracy of only 15.47%. For the original data + VMD + SCC + TCN method, the average accuracy reaches 82.77%, showing an improvement in diagnostic accuracy compared to the previous method. Specifically, compared to the original data + SMOTE + Attention + LSTM method, the LSTM network with the added attention mechanism achieves an average accuracy of 47.84%, while the method proposed in this invention achieves an average accuracy of 99.02%. Compared to the original data + ADASYN + ABiTCN method, the ABiTCN network achieves an average accuracy of 64.71%, while the method proposed in this invention achieves an average accuracy of 99.02%.

[0068] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A bearing fault diagnosis method that fuses the generation of screening and adaptive feature enhancement, characterized in that, The bearing fault diagnosis method comprises the following steps: S1, collecting original data and expanding the data scale by using a generative adversarial network; S2, extracting high-quality samples by using a correlation screening mechanism; S3, suppressing noise and strengthening key features by using a feature denoising module; S4, realizing fault identification by using a bidirectional convolutional network.

2. The method of claim 1, wherein the method is a fusion-generated screening and adaptive feature-enhanced bearing fault diagnostic method, characterized by, The step S1 specifically comprises the following steps: S1-1, data acquisition and preliminary processing Collect original bearing vibration data, remove outliers, fill in missing values, and perform detrending and normalization processing on the original data, and divide the processed data set into a training set and a test set; S1-2, high-fidelity data generation Introduce a one-sided gradient penalty mechanism in the WGAN model to build a WGAN-OP data generation model, and use the training set as the input of the WGAN-OP data generation model to generate data.

3. The method of claim 2, wherein the method is a fusion-generated screening and adaptive feature-enhanced bearing fault diagnostic method, characterized by: D and G of the WGAN-OP model adopt different loss functions L D and L G , respectively: The calculation formula of the gradient penalty term GP is: wherein, is the gradient norm for enforcing Lipschitz constraint by weighting P data and P z (z) to obtain x; denotes the gradient operator that computes the gradient of the D function at its input; is the L2 norm of the gradient; In the WGAN-OP model, the formula for calculating x and the loss function of D and G are: 。 4. The method of claim 2, wherein the method is a fusion-generated screening and adaptive feature-enhanced bearing fault diagnostic method. The step S2 specifically comprises the following contents: Use the Spearman rank correlation coefficient to measure the rank correlation between the generated samples and the real samples, and set a threshold to filter the generated data, and only keep the high-quality samples with high correlation to supplement to the original data set.

5. The method and system of claim 2, wherein, The step S3 specifically comprises the following contents: A feature denoising module AFME is constructed. Firstly, the module performs multi-modal decomposition on the signal through an adaptive FIR filter bank, and extracts the sensitive modal of impact features with the correlation kurtosis CK as the optimization target. Then, the dual-layer screening is performed by using the Pearson correlation coefficient and the multi-scale entropy MSE between the modes to remove the redundant and high-noise modes. Finally, the modes are fused by weighting according to the mode importance index S k The modes are fused by weighting.

6. The method of claim 5, wherein the method is a fusion-generated screening and adaptive feature-enhanced bearing fault diagnostic method, characterized by, The step S3 specifically comprises the following steps: S3-1, modal decomposition Based on adaptive finite impulse response filter bank, the original signal is modal decomposed, and the coefficient of the kth filter is f k (l), the input signal is x(n), and the kth modal output is represented as: where L is the filter length, u k (n) denotes the kth component of the filter bank output, with the correlation kurtosis CK as the optimization objective function, which is defined as: where T is the time delay parameter, the symbol * denotes the autocorrelation operation, and the maximum is taken over the time delay parameter , the adaptive extraction of the fault feature modal is completed. S3-2, double-layer modal screening First, similarity screening is performed, and the Pearson correlation coefficient r between any two modalities is calculated ij : When |r ij When |r exceeds a set threshold θ, it is considered a high similarity modal pair, and only the modal with higher kurtosis is retained. Secondly, multi-scale entropy screening is performed, the MSE value of each modal is calculated and sorted in ascending order, and the modal with good information and noise suppression effect is preferentially retained; S3-3, feature weighting reconstruction Defining a modality significance indicator S K is: wherein E k , K k , MSE k , C k respectively represent the energy proportion, kurtosis and multi-scale entropy of the kth modality; β i is the weighting coefficient of each index, and the modal weight w k is calculated according to S k , and weighted reconstruction is performed: where x r (n) is the final reconstructed signal.

7. The method of claim 2, wherein the method is a fusion-generated screening and adaptive feature-enhanced bearing fault diagnostic method, characterized by: The step S4 specifically comprises the following contents: An adaptive bidirectional time series convolutional network ABiTCN is constructed, the model is composed of two symmetrical TCN branches in the forward and backward directions, an adaptive feature fusion layer AFF is introduced, a lightweight channel attention mechanism is used to learn the fusion weight, the contribution ratio of bidirectional features is adaptively adjusted according to the signal characteristics, the fused features are mapped by global average pooling GAP and fully connected layer FC, and the fault category probability is output by the Softmax classifier, completing the intelligent diagnosis of bearing faults under multiple working conditions.

8. The bearing fault diagnosis method of claim 7, wherein: The two symmetrical TCN branches of the ABiTCN are used to extract feature information in different time directions, the forward TCN is used to model the dependence from the past to the future, and the backward TCN is used to model the dependence from the future to the past, and the calculation form is: where X is the input signal sequence, Reverse(.) denotes the time reversal operation, Y f and Y b denote the forward and backward feature outputs, respectively. An adaptive weighted fusion layer AFF is introduced, the fusion ratio of forward and backward features is automatically assigned through learnable parameters, and the fusion method is defined as: Wherein, alpha is a fusion weight coefficient in [0, 1], used to control the proportion of forward and backward features in the final representation, and the weight is obtained through a lightweight channel attention mechanism: where Y f and Y b respectively represent the channel representation of the forward and backward features after global average pooling. The fused feature matrix FFF is input into the fully connected layer after global average pooling to realize fault classification, and the specific representation is: where GAP(.) is a global average pooling operation, W c and b c are the classification layer weights and bias terms, respectively, is the probability distribution of failure classes output by the model.

9. A terminal device, comprising: The computer program is configured to be executed by the processor, and the processor executes the computer program to implement the bearing fault diagnosis method according to any one of claims 1 to 8.

10. A storage medium, characterized by The storage medium comprises a stored computer program, and the computer program controls a device where the storage medium is located to execute the bearing fault diagnosis method according to any one of claims 1 to 8.

Images