Bearing fault diagnosis method of speed robust feature learning and minority class generation in data imbalance and speed change environment

Abstract

The invention relates to a bearing fault diagnosis method for speed robust feature learning and minority class generation in a data imbalance and variable speed environment, and belongs to the technical field of bearing fault diagnosis, and the method comprises the following steps: S1, mapping an original vibration signal to a potential space with an unchanged state through employing an encoder based on variational mutual information VIB, realizing fault key feature extraction; s2, on the basis of a feature-level sample enhancement mechanism, generating samples with discrimination capability through potential space re-parameterization; s3, constructing a multi-task loss function, and cooperatively realizing fault feature discernibility, consistency of generated features and original features, and decoupling of operation state information; and S4, bearing faults are diagnosed.

Description

Technical Field

[0001] This invention belongs to the field of bearing fault diagnosis technology, and relates to a bearing fault diagnosis method based on speed robust feature learning and minority class generation in data imbalance and variable speed environments. Background Technology

[0002] As a critical component of rotating machinery, the health of rolling bearings directly affects the operational reliability of the equipment. Therefore, timely and accurate condition monitoring and fault diagnosis are of great significance for ensuring safe operation and extending the service life of equipment. Vibration signals contain rich information about the health status of equipment, and vibration-based intelligent fault diagnosis methods have achieved significant results.

[0003] In practical engineering applications, equipment often operates under variable speed conditions, and the distribution characteristics of vibration signals may change significantly under different operating conditions, which increases the difficulty of fault diagnosis. At the same time, fault samples in actual production often suffer from data imbalance, with data for certain fault types being relatively scarce, making it difficult for diagnostic models to fully learn the characteristics of a few categories. Therefore, studying the data imbalance problem under variable speed conditions has significant theoretical and practical implications.

[0004] Methods for addressing data imbalance can be broadly categorized into three types: data-level methods, model-level methods, and reweighting strategies. Among data-level techniques, Synthetic Minority Oversampling Technique (SMOTE) is the most representative, mitigating class imbalance by interpolating minority class instances in the feature space to generate new samples. Additionally, data augmentation can be performed based on simulated dynamic models or Generative Adversarial Networks (GANs). However, these methods are primarily applicable to constant rotational speeds or fixed operating conditions. At the model level, Jia et al. proposed Deep Normalized Convolutional Neural Networks (DNCNN) to address fault diagnosis under unbalanced rolling bearing sample conditions. Wu et al. introduced a Model Independent Framework (MAF) specifically designed for class imbalance fault diagnosis. Reweighting strategies, such as Focus Loss (FCL) and Label Distribution Aware Marginal Loss (LDL), also frequently mitigate class imbalance by adjusting the loss function.

[0005] While the aforementioned methods have achieved some success in mitigating data imbalance, most studies still focus on constant speed or single operating conditions. Under variable speed conditions, vibration signals often exhibit significant non-stationarity and cross-condition distribution shifts, making methods relying on sample generation or static distribution assumptions difficult to adapt to different operating states. Furthermore, under variable speed conditions, the sample distributions of different fault types and their corresponding operating conditions are typically more imbalanced. The sample size for minority classes is limited, and their characteristic distributions shift significantly with changes in speed. Existing methods still face significant challenges in simultaneously addressing operating condition variations and class imbalances, making it difficult to guarantee diagnostic performance. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a bearing fault diagnosis method based on speed robust feature learning and minority class generation in a data imbalance and variable speed environment.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A bearing fault diagnosis method based on speed-robust feature learning and minority class generation in imbalanced data and variable speed environments includes the following steps: S1: The original vibration signal is mapped to the state-invariant latent space using an encoder based on variational mutual information (VIB) to achieve the extraction of key fault features. S2: Based on the feature-level sample enhancement mechanism, discriminative samples are generated through latent space reparameterization. S3: Construct a multi-task loss function to collaboratively achieve fault feature discriminability, consistency between generated features and original features, and decoupling of operational status information; S4: Diagnose bearing failures.

[0008] Furthermore, in step S1, VIB introduces variational approximation to transform the mutual information objective function into a form that can be optimized by a deep neural network; In the VIB method, conditional distribution Neural network parameterization typically assumes that it follows a Gaussian distribution:

[0009] in and These represent the mean and standard deviation of the encoder network output, respectively. The mutual information between the input and the features is represented as:

[0010] Due to edge distribution Due to the difficulty in obtaining the necessary priors, VIB employs a simple prior distribution. By making an approximation, we obtain a variational estimate of the upper bound of the mutual information:

[0011] The mutual information between features and labels is represented as follows:

[0012] in Label entropy; by introducing a discriminative model To approximate the conditional distribution Obtain the lower bound of the mutual information variation:

[0013] Finally, the optimization objective of variational mutual information (VIB) is expressed as: .

[0014] Furthermore, step S1, which involves using an encoder based on variational mutual information (VIB) to map the original vibration signal to a state-invariant latent space to extract key fault features, includes: Using a VIB-based encoder, the original vibration signal Mapping to the latent space, the encoder extracts a compact representation and generates VIB parameters, including the mean. With variance Characterizes the Gaussian distribution in the state-invariant latent space; Furthermore, step S1 specifically includes the following steps: S11: Obtain a compact global vector through three consecutive convolutional layers. :

[0015] S12: Transfer vector Projected onto the mean through a linear mapping layer With log variance :

[0016] in , , and These are the weights and biases of the linear layer, respectively. express The dimension; Furthermore, step S2, which uses a feature-level sample enhancement mechanism to generate discriminative samples through latent space reparameterization, includes: Fault-related features were obtained by sampling from the Gaussian distribution using VIB reparameterization. ; Simultaneously generate redundant feature related representations It is used to capture interference information and also to apply decoupling constraints to prevent operational status information from infiltrating. ; VIB parameters Store the data to build a potential feature library for each fault category; Based on the potential feature library, new discriminative features are generated through a reparameterization process.

[0017] Furthermore, step S2 specifically includes the following steps: S21: From Gaussian distribution using VIB reparameterization techniques Medium sampling:

[0018] S22: Classify each fault category All training samples For features stored in the feature library:

[0019] in This represents the total number of fault types. For faults The number of training samples; S23: Calculate the number of samples in each class and compare it with the largest class size. Perform comparisons to determine the minority classes; for each minority class Calculate the number of missing samples. By using the corresponding class library Random sampling Generate synthetic features and apply VIB reparameterization techniques:

[0020] in Indicates from category Feature library The sample index is randomly selected from the data; in addition, Indicates the current category Sample set The synthetic feature index generated in the process; S24: Synthesize features Compared with the original features in the current batch The features are concatenated to form an enhanced feature set. :

[0021] in Indicates the total number of augmented samples; S25: Enhanced conditionally invariant features The data is then fed into a fault classifier, which is implemented using a fully connected network. The classifier outputs a predicted logit value for each sample. :

[0022] S26: At the same time, through Another linear mapping yields redundant features related to rotational speed or external disturbances. :

[0023] By explicitly modeling state-related components, the training process in and Apply decoupling constraints to prevent state information from infiltrating. .

[0024] Furthermore, the multi-task loss function includes: First, apply orthogonality constraints to ensure that and Effective decoupling between:

[0025] The orthogonality loss is:

[0026] exist The introduction of supervised comparison loss for a batch sample ,in To determine the corresponding category labels, first calculate the pairwise Euclidean distances between all samples:

[0027] The sets of positive and negative sample pairs are defined as follows:

[0028] set up and Let represent the number of positive sample pairs and negative sample pairs in the batch, respectively. Then, the supervised comparison loss is expressed as:

[0029] in For interval parameters; Applying multi-scale maximum mean difference (MMD) loss ensures the latent features of the reconstruction. Features related to the original fault Keep aligned:

[0030] Kernel function Defined as a multi-scale radial basis function:

[0031] Finally, the cross-entropy classification loss function is used to supervise the tomographic class prediction:

[0032] The overall multi-task loss function consists of four terms: .

[0033] The beneficial effects of this invention are as follows: The VIB encoder of this invention maps the original vibration signal to a state-invariant latent space, achieving high-fidelity extraction of key fault features. The feature-level sample enhancement mechanism generates discriminative samples through latent space reparameterization, effectively alleviating the class imbalance problem. Furthermore, the multi-task loss function, as the core driving mechanism for network training, collaboratively achieves fault feature discriminability, consistency between generated and original features, and decoupling of operational state information. This achieves comprehensive optimization of the latent representation, significantly improving diagnostic performance and robustness in variable speed and imbalance scenarios. Through extensive experimental verification on a self-built variable speed imbalance dataset and publicly available benchmark datasets, VIB-FGNet surpasses many cutting-edge methods in both diagnostic accuracy and robustness, demonstrating its effectiveness in complex variable speed imbalance conditions.

[0034] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0035] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the VIB architecture; Figure 2 This is a schematic diagram of the characteristic distribution under variable speed conditions; Figure 3 This is a diagram of the overall framework of VIB-FGNet; Figure 4 Results for different models under data imbalance conditions; Figure 5 The confusion matrix analysis is for Task C, where (a) is WDCNN; (b) is LapNet; (c) is ResNet18; (d) is ClassBD; (e) is TFN; and (f) is VIB-FGNet. Figure 6 The distributions of real and generated features for multiple classes of samples are shown, where (a) is VIB-FGNet and (b) is SMWDC. Figure 7 The feature distributions of real samples and generated samples in category 1 are shown, where (a) is VIB-FGNet and (b) is SMWDC. Figure 8 The results of the velocity suppression index assessment; Figure 9 The results of different models on imbalanced data. Detailed Implementation

[0036] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0037] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0038] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0039] Example 1: like Figure 1 As shown, this invention provides a bearing fault diagnosis method based on speed robust feature learning and minority class generation under data imbalance and variable speed environments. This method treats speed variations as redundant disturbances and achieves latent feature decoupling through multiple designed constraints, effectively suppressing condition-related disturbances and forming a consistent latent representation across speed ranges. Based on this, a category-level feature generation strategy based on latent feature distribution is constructed in the shared cross-speed latent space. Minority class fault features are dynamically enhanced through reparameterized sampling, effectively alleviating the sample imbalance problem under variable speed conditions. Experimental results show that this method has significant advantages over existing state-of-the-art methods under different speed variation and data imbalance ratios, verifying its effectiveness in variable speed and imbalance fault diagnosis tasks.

[0040] This invention re-examines data augmentation and fault representation modeling from a feature perspective. To improve fault diagnosis performance, it proposes a feature generation network based on variational information bottleneck (VIB-FGNet), aiming to learn a latent feature space robust to changes in operating state. This network suppresses redundant operating condition information through information constraints and extracts the most discriminative fault classification features by combining supervised learning with an explicit feature decoupling mechanism. Class-level controllable sampling is implemented in the latent feature space to enhance the representational ability of minority class samples, thereby mitigating the impact of data imbalance on diagnostic performance. Experimental results show that this method significantly improves the fault diagnosis accuracy of imbalanced data under variable speed conditions. The main contributions of this invention are as follows: (1) The VIB-FGNet framework is proposed to address the challenges of non-stationary fault features and highly unbalanced samples under variable speed conditions. This framework can alleviate feature drift caused by speed changes and enhance the discriminativeness and stability of minority class fault features.

[0041] (2) A potential feature decoupling mechanism is proposed, which separates the operating condition-related variation from the fault discrimination information to achieve rate invariant fault characterization across operating conditions.

[0042] (3) A class-level feature generation strategy was developed based on the latent feature distribution. By controlling the sampling, minority class features are stably generated, thereby alleviating the sample imbalance problem under variable speed conditions.

[0043] The IB theory explains the core objective of representation learning in supervised learning from an information theory perspective. The input sample is denoted as... The corresponding class tag is denoted as The learned representation is denoted as According to IB theory, ideal representation The output should be preserved as much as possible. Relevant information, while compressing its contained input. Information. Therefore, To become the minimum sufficient representation of this task. This objective can be formulated as the following optimization problem: (1) in Indicates mutual information, Let be a preset constant. Introducing Lagrange multipliers, the above objective can be transformed into an unconstrained form: (2) in It is a trade-off parameter between the degree of feature compression and the discriminative performance.

[0044] Although mutual information theory provides clear conceptual guidance, directly computing mutual information terms in deep neural networks is a challenging task. and This is typically challenging. To address the problem of mutual information computation, a variational mutual information (VIB) method is proposed, the structure of which is illustrated in the diagram below. Figure 1 As shown, VIB introduces a variational approximation, transforming the mutual information objective function into a form that can be optimized by a deep neural network.

[0045] In the VIB method, conditional distribution Neural network parameterization typically assumes that it follows a Gaussian distribution: (3) in and These represent the mean and standard deviation of the encoder network output, respectively.

[0046] Therefore, the mutual information between the input and the features can be expressed as: (4) Due to edge distribution Due to the difficulty in obtaining the necessary priors, VIB employs a simple prior distribution. (Usually, a standard normal distribution is chosen) We then approximate the mutual information to obtain an upper bound variational estimate: (5) The mutual information between features and labels can be represented as: (6) in This is the label entropy, independent of model parameters. It is achieved by introducing a discriminative model. To approximate the conditional distribution The variational lower bound of mutual information can be obtained: (7) Ultimately, the optimization objective of variational mutual information (VIB) can be expressed as: (8) The VIB-FGNet proposed in this invention mainly includes an encoder, a fault classifier, a sample imbalance enhancement module, and a multi-task loss function, aiming to achieve high-precision fault diagnosis under variable speed and imbalance conditions.

[0047] 1. VIB-based encoder and feature-level sample generation Under variable speed operation conditions, vibration signals exhibit significant non-stationary characteristics, with their amplitude, frequency structure, and time-domain waveform changing with rotational speed. Therefore, the signal patterns of the same fault differ significantly at different speeds. Furthermore, fault samples collected at different speeds are often unevenly distributed. These problems cause the features extracted by the encoder to drift with rotational speed and exhibit an unbalanced distribution in the feature space. An ideal fault feature distribution should be able to resist changes in rotational speed to achieve high diagnostic accuracy, while maintaining relative balance at different speeds, such as… Figure 2 As shown.

[0048] To address the aforementioned problems, this invention maps the variable-speed vibration signal to a latent feature space and employs multiple constraints to suppress velocity-related information. Furthermore, it alleviates the sample imbalance problem from the perspective of feature enhancement. Firstly, a VIB-based encoder is designed to process the original vibration signal... Mapping to the latent space allows for the preservation of the most fault-discriminating information while suppressing redundant state-induced mutations. Specifically, the encoder extracts a compact representation and generates VIB parameters—means. With variance These parameters characterize the Gaussian distribution in the state-invariant hidden space. Fault-related features. It is not directly output by the encoder, but is sampled from this distribution using the VIB reparameterization technique. This sampling mechanism ensures It can capture inherent fault attributes and is robust to changes in operating state. In parallel processing, the encoder generates redundant feature-related representations. This is used to capture interference information (such as velocity-induced amplitude and frequency fluctuations). This representation is further used to apply decoupling constraints to prevent the infiltration of operational status information. In addition, the VIB parameter The data is stored to construct a potential feature library for each fault category. Based on this library, new discriminative features can be generated through the same reparameterization process, thereby mitigating the class imbalance problem. The specific steps are as follows: First, a compact global vector is obtained through three consecutive convolutional layers. : (9) Then, the vector Projected onto the mean through a linear mapping layer With log variance : (10) in , , and These represent the weights and biases of the linear layer. Finally, express Dimensions.

[0049] Next, From Gaussian distribution using VIB reparameterization techniques Medium sampling: (11) In addition, each fault category All training samples For features stored in the feature library: (12) in This represents the total number of fault types. For faults The number of training samples.

[0050] To address the class imbalance problem, the framework automatically calculates the number of samples in each class and compares it with the largest class size. A comparison is performed to determine the minority class. For each minority class... Calculate the number of missing samples. Then, by using the corresponding class library... Random sampling Generate synthetic features and apply VIB reparameterization techniques: (13) in Indicates from category Feature library The sample index is randomly selected from the data. Furthermore, Indicates the current category Sample set The synthetic feature index generated in the process.

[0051] Finally, synthesize the features Compared with the original features in the current batch splicing to form an enhanced feature set : (14) in This indicates the total number of samples to be augmented.

[0052] Enhanced conditionally invariant features The data is then fed into a fault classifier, which is implemented using a fully connected network. The classifier outputs a predicted logit value for each sample. : (15) In addition, through Another linear mapping can obtain redundant features related to rotational speed or external disturbances. : (16) extract The purpose is not for direct fault classification, but rather as an auxiliary representation for subsequent constraint functions. By explicitly modeling state-related components, the training process can be... and Apply decoupling constraints to prevent state information from infiltrating. .

[0053] 2. Design of Multi-task Loss Function To achieve feature decoupling between fault-related and redundant components and to effectively generate fault features, a multi-task loss function was designed. First, orthogonality constraints were applied to ensure... and Effective decoupling between: (17) The orthogonality loss is: (18) To enhance intra-class compactness and inter-class separability, in The method incorporates supervised comparison loss.

[0054] For a batch sample ,in To determine the corresponding category labels, first calculate the pairwise Euclidean distances between all samples: (19) The set of positive sample pairs and negative sample pairs is defined as follows: (20) set up and Let represent the number of positive and negative sample pairs in the batch, respectively. Then, the supervised comparison loss can be expressed as: (twenty one) in This is the interval parameter.

[0055] Subsequently, multi-scale maximum mean difference (MMD) loss is applied to ensure the latent features of the reconstruction. Features related to the original fault Keep aligned: ,(twenty three) Kernel function Defined as a multi-scale radial basis function: (twenty four) Finally, the cross-entropy classification loss function is used to supervise the tomographic class prediction: (25) Therefore, the total multi-task loss function consists of four terms: (26) The overall framework of the proposed VIB-FGNet is as follows: Figure 3 As shown.

[0056] To quantitatively evaluate the model's ability to eliminate the influence of speed at different rotational speeds, this invention designs three evaluation indicators from three dimensions: statistical consistency, geometric similarity, and class separability: Speed ​​Variance Index (SVI), Cross-Speed ​​Cosine Similarity (CSC), and Same Fault / Different Fault Ratio (SDR). The feature representation is denoted as... The fault category is denoted as The rotational speed condition is denoted as The feature set corresponding to the c-th type of fault under the k-th operating condition is denoted as . .

[0057] The SVI metric is used to measure the difference in feature means under different speed conditions for the same fault category. First, the feature means under different speed conditions are calculated for each fault category and each feature dimension: (27) Then the mean and variance of this feature dimension at different rotational speeds are calculated: (28) Finally, the average value is taken for all fault categories and feature dimensions: (29) The smaller the value of this index, the weaker the influence of speed. If the characteristics of the same fault at different speeds are such that the SVI approaches 0 when the speeds are almost the same, it indicates that the influence of speed has been eliminated.

[0058] The CSC index calculates the average cosine similarity of feature vectors from samples at different speeds within the same fault category. If the influence of speed is suppressed, samples from different speeds should cluster into the same subspace, thus significantly improving the similarity. Given a fault category... With any two different speeds The pairwise cosine similarity between samples is calculated as follows: (30) Then, the average similarity across velocities was calculated: (31) The closer this value is to 1, the more consistent the behavior of the same fault is at different speeds, that is, the smaller the speed-related interference.

[0059] The SDR (Speed ​​Distance Ratio) index reflects the ratio of intra-class cross-speed distance to inter-class distance. It is calculated by first determining the average distance between samples of the same fault category at different speeds (representing the degree of speed influence), and then calculating the average distance between samples of different fault categories (representing fault distinguishability). The ratio reflects the importance of speed interference relative to fault differences. A lower ratio indicates that the intra-class cross-speed difference is much smaller than the inter-class difference, implying stronger fault distinguishability and weaker speed influence. The intra-class cross-speed distance representing speed influence is defined as follows: (32) Inter-class distance (representing fault separability) is defined as follows: (33) Ultimately, SDR is defined as follows: (34) In this embodiment, the effectiveness of the proposed VIB-FGNet under variable speed and data imbalance conditions was verified using the Spectra Quest Variable (SQV) dataset and the collected experimental dataset. Comparative experiments and model performance analysis were also conducted.

[0060] For comparison, five advanced methods designed to address data imbalance and five cutting-edge intelligent diagnostic models were used: (1) First, five common imbalance fault diagnosis methods are adopted, covering three categories: sample synthesis method (SMOTE), model-driven method (DNCNN, MAF, SMOTECNN) and weight adjustment method (FCL, (LDL)).

[0061] (2) Five advanced intelligent diagnostic models—ResNet18, WDCNN, LapNet, TFN and ClassBD—were compared and evaluated to conduct a comprehensive analysis of the proposed VIB-FGNet model.

[0062] Case 1: The dataset used in this invention was obtained through experimental acquisition. The test bench mainly consists of a load motor, a torque and speed measuring device, a reducer, a rolling bearing test module, and a drive motor. The bearing used in the experiment was SKF-61908, the vibration sensor was B&K 8347-C, the reducer was YT803-320, and the data acquisition card was NI-9234. Experimental data were acquired under a rated load of 2 Nm and an acceleration of 200 rpm / s, with the speed uniformly varying from 100 rpm to 1500 rpm, and the sampling frequency set to 12 kHz. During the experiment, vibration signals were measured for eight different health states of the rolling bearing using the vibration sensor, including: inner ring fault (IF), outer ring fault (OF), ball fault (BF), normal state (NC), inner ring-ball combined fault (IB), inner and outer ring combined fault (IO), outer ring-ball combined fault (OB), and inner and outer ring-ball combined fault (IOB). The number of balanced samples for each fault type is set to 400, the sample length is 3072, and the number of test samples is 50. The specific settings for the number of unbalanced samples are shown in Table 1.

[0063] Table 1

[0064] Based on the assumptions shown in Table 1, the number of samples for each type of imbalance fault in the training set is set to construct an imbalance dataset. Figure 4 The results of different models on various combinations of imbalance fault types are presented. It can be observed that the proposed VIB-FGNet achieves the highest diagnostic accuracy across all tasks. Especially in complex imbalance scenarios, this method consistently maintains a high accuracy of approximately 0.99, while the performance of other models significantly declines. In contrast, ResNet18, WDCNN, LaplaceAlexNet, TFN, and ClassBD generally achieve lower accuracy than this method in multi-type imbalance tasks, with some models even dropping to the 0.7-0.85 range in more demanding scenarios. Experimental results demonstrate that this method exhibits the best robustness and diagnostic performance when handling different combinations of imbalance fault types.

[0065] In addition, to clearly demonstrate the performance of each model in different fault categories, Figure 5 Tables (a)-(f) present the confusion matrices for each model under task C. For example... Figure 5As shown, even under imbalanced sample conditions, the proposed method still achieves high diagnostic accuracy, with predictions for almost all categories approaching the ideal diagonal. In contrast, ResNet18, WDCNN, LaplaceAlexNet, TFN, and ClassBD all exhibit varying degrees of classification errors in categories with fewer samples. Particularly for categories with small sample sizes, such as IF, OF, and BF, the comparative models often confuse them with neighboring categories, indicating a lack of clear inter-class separability.

[0066] Using IOB fault categories as imbalanced samples, different imbalance ratios were set to evaluate the diagnostic performance of each model under imbalance conditions. The results are shown in Table 2. Under high imbalance conditions (1:10), the accuracies of FCL and LDL were 0.880 and 0.885, respectively. When the classes were perfectly balanced (1:1), the accuracies improved to 0.953 and 0.948. MAF and DNCNN performed relatively poorly under high imbalance conditions (1:10, 1:8). However, as the class distribution tended to be balanced, their accuracies improved to approximately 0.883 and 0.885. Conversely, VIB-FGNet maintained high accuracy under all imbalance ratios, demonstrating robustness to class imbalance and highlighting its diagnostic performance under complex imbalance conditions. 。

[0067] Table 2

[0068] Taking a 1:2 imbalance ratio as an example, we simulated the sample missing situation of different fault categories to verify the feasibility of the proposed feature enhancement method. Figure 6 Figure (a) shows the distribution of the true features extracted by the proposed model and the corresponding generated features in the feature space. In the comparative experiment, the original imbalanced data was oversampled and augmented using the SMOTE method. The augmented samples were then input into WDCNN for feature learning (called SMWDC), and their feature distribution is shown in Figure (a). Figure 6 As shown in (b).

[0069] like Figure 6 As shown in (a), the proposed method achieves effective feature enhancement while preserving the original geometric structure by adaptively modeling the distribution of minority class samples in the feature space. Therefore, the generated features are highly consistent with the true features in terms of distribution and spatial location. In contrast, Figure 6 Although the features generated by SMWDC in (b) are generally close to the real features, there is still a certain degree of feature bias. This is because the feature enhancement process is not jointly optimized with the feature learning stage.

[0070] also, Figure 7Figures (a) and (b) show the feature distributions of real and generated samples in Category 1, further confirming the above observations. Jensen-Shannon divergence (JS), MMD, and Euclidean distance (ED) are used as effective metrics for quantitative evaluation of the differences between real and generated features. The results are summarized in Table 3. Experiments show that the proposed method achieves lower values ​​on all evaluation metrics, indicating that the generated features are closer to the real feature distribution in both statistical distribution and geometric structure.

[0071] Table 3

[0072] To verify the features extracted by the encoder It can effectively eliminate the influence of rotational speed and improve the accuracy of model fault diagnosis. The influence of rotational speed on feature extraction is quantified by designing indicators. Figure 8 Evaluation results for SVI, CSC, and SDR are presented. The proposed VIB-FGNet model outperforms the comparison models in all three metrics. The SVI value of the features extracted by VIB-FGNet is close to zero and much lower than that of the baseline model, indicating that the method effectively suppresses the interference of velocity changes. In addition, its CSC value reaches 0.935 (higher than ResNet's 0.869), proving that it has better clustering effect on samples of the same class in the feature space and significantly reduces intra-class differences. The SDR value is 0.2714, lower than ResNet's 0.4836, indicating that velocity interference has less impact on samples of different classes and the features have stronger discriminative power.

[0073] Case 2: This benchmark dataset originates from the publicly released variable speed bearing dataset from Xi'an Jiaotong University. The motor bearing used in the experiment is model NSK6203, and the sampling frequency is set to 25.6 kHz. Simulations were conducted for different damage levels under operating conditions of inner race faults (minor, moderate, and severe, labeled 0, 1, and 2 respectively) and outer race faults (minor, moderate, and severe, labeled 4, 5, and 6 respectively), while also including a healthy operating condition (labeled 3), resulting in a total of seven healthy states. During the testing phase, data segments collected during the gradual acceleration from a standstill to 3000 rpm were analyzed. The sample length and sampling quantity configuration remained consistent with Case Study 1.

[0074] An imbalanced dataset was constructed (see Table 4), and different models were used to analyze the data. The results are as follows: Figure 9 As shown, the proposed VIB-FGNet model exhibits significant performance advantages over other models, achieving more accurate and stable diagnostic results. In contrast, the comparative models are more significantly affected by data imbalance, leading to reduced diagnostic accuracy and greater performance fluctuations. This phenomenon indicates a limitation in robustness under imbalanced data conditions.

[0075] Table 4

[0076] The algorithm proposed in this invention was also compared with other schemes designed for imbalanced datasets, and the results are shown in Table 5. The diagnostic performance of all models generally improves as the imbalance ratio gradually decreases. The proposed VIB-FGNet shows a significant advantage under all imbalance ratios, with a diagnostic accuracy close to or reaching 1. In contrast, traditional methods such as FCL, LDL, and MAF are severely limited under extreme imbalance conditions (1:10, 1:8), with accuracy generally below 0.87. DNCNN and SMCNN perform even worse under imbalance conditions, with SMCNN showing a significant performance drop under extreme imbalance conditions.

[0077] Table 5

[0078] The quality of generated samples was evaluated using Class 1, with an imbalance ratio of 1:2, as an example. JS divergence, MMD, and ED were used to measure the difference between generated sample features and real sample features. For comparison, the SMOTE method was used to enhance the imbalanced samples before inputting them into different intelligent diagnostic models. The distance between the features generated by SMOTE and the features of the real samples was calculated, and the results are summarized in Table 6. According to Table 6, the proposed method generates features with lower metric values ​​compared to the real features. These results indicate that under imbalanced data conditions, the proposed method can achieve more effective feature enhancement, thereby further improving the accuracy of fault diagnosis.

[0079] Table 6

[0080] Example 2: An electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method described in Embodiment 1 when executing the computer program.

[0081] Example 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0082] Example 4: A computer program product includes a computer program that, when executed by a processor, implements the method described in Example 1.

[0083] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0084] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0085] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0086] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0087] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0088] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0089] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0090] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A bearing fault diagnosis method based on speed-robust feature learning and minority class generation under imbalanced data and variable speed environments, characterized in that: Includes the following steps: S1: The original vibration signal is mapped to the state-invariant latent space using an encoder based on variational mutual information (VIB) to achieve the extraction of key fault features. S2: Based on the feature-level sample enhancement mechanism, discriminative samples are generated through latent space reparameterization. S3: Construct a multi-task loss function to collaboratively achieve fault feature discriminability, consistency between generated features and original features, and decoupling of operational status information; S4: Diagnose bearing failures.

2. The bearing fault diagnosis method based on speed robust feature learning and minority class generation under imbalanced data and variable speed environments as described in claim 1, characterized in that: In step S1, VIB introduces variational approximation to transform the mutual information objective function into a form that can be optimized by a deep neural network. In the VIB method, conditional distribution Neural network parameterization typically assumes that it follows a Gaussian distribution: in and These represent the mean and standard deviation of the encoder network output, respectively. The mutual information between the input and the features is represented as: Due to edge distribution Due to the difficulty in obtaining the necessary priors, VIB employs a simple prior distribution. By making an approximation, we obtain a variational estimate of the upper bound of the mutual information: The mutual information between features and labels is represented as follows: in Label entropy; by introducing a discriminative model To approximate the conditional distribution Obtain the lower bound of the mutual information variation: Finally, the optimization objective of variational mutual information (VIB) is expressed as: 。 3. The bearing fault diagnosis method based on speed robust feature learning and minority class generation under imbalanced data and variable speed environments as described in claim 1, characterized in that: Step S1, which involves using an encoder based on variational mutual information (VIB) to map the original vibration signal to a state-invariant latent space to extract key fault features, includes: Using a VIB-based encoder, the original vibration signal Mapping to the latent space, the encoder extracts a compact representation and generates VIB parameters, including the mean. With variance It describes the Gaussian distribution in the state-invariant latent space.

4. The bearing fault diagnosis method based on speed robust feature learning and minority class generation under imbalanced data and variable speed environments as described in claim 3, characterized in that: Step S1 specifically includes the following steps: S11: Obtain a compact global vector through three consecutive convolutional layers. : S12: Transfer vector Projected onto the mean through a linear mapping layer With log variance : in , , and These are the weights and biases of the linear layer, respectively. express Dimensions.

5. The bearing fault diagnosis method based on speed robust feature learning and minority class generation under imbalanced data and variable speed environments as described in claim 1, characterized in that: Step S2, based on the feature-level sample enhancement mechanism, generates discriminative samples through latent space reparameterization, including: Fault-related features were obtained by sampling from the Gaussian distribution using VIB reparameterization. ; Simultaneously generate redundant feature related representations It is used to capture interference information and also to apply decoupling constraints to prevent operational status information from infiltrating. ; VIB parameters Store the data to build a potential feature library for each fault category; Based on the potential feature library, new discriminative features are generated through a reparameterization process.

6. The bearing fault diagnosis method based on speed robust feature learning and minority class generation under imbalanced data and variable speed environments as described in claim 5, characterized in that: Step S2 specifically includes the following steps: S21: From Gaussian distribution using VIB reparameterization techniques Medium sampling: S22: Classify each fault category All training samples For features stored in the feature library: in This represents the total number of fault types. For faults The number of training samples; S23: Calculate the number of samples in each class and compare it with the largest class size. Perform comparisons to determine the minority classes; for each minority class Calculate the number of missing samples. By using the corresponding class library Random sampling Generate synthetic features and apply VIB reparameterization techniques: in Indicates from category Feature library The sample index is randomly selected from the data; in addition, Indicates the current category Sample set The synthetic feature index generated in the process; S24: Synthesize features Compared with the original features in the current batch The features are concatenated to form an enhanced feature set. : in Indicates the total number of augmented samples; S25: Enhanced conditionally invariant features The data is then fed into a fault classifier, which is implemented using a fully connected network. The classifier outputs a predicted logit value for each sample. : S26: At the same time, through Another linear mapping yields redundant features related to rotational speed or external disturbances. : By explicitly modeling state-related components, the training process in and Apply decoupling constraints to prevent state information from infiltrating. .

7. The bearing fault diagnosis method based on speed robust feature learning and minority class generation under imbalanced data and variable speed environments as described in claim 1, characterized in that: The multi-task loss function includes: Apply orthogonality constraints to ensure and Effective decoupling between: The orthogonality loss is: exist The introduction of supervised comparison loss for a batch sample ,in To determine the corresponding category labels, first calculate the pairwise Euclidean distances between all samples: The sets of positive and negative sample pairs are defined as follows: set up and Let represent the number of positive sample pairs and negative sample pairs in the batch, respectively. Then, the supervised comparison loss is expressed as: in For interval parameters; Applying multi-scale maximum mean difference (MMD) loss ensures the latent features of the reconstruction. Features related to the original fault Keep aligned: Kernel function Defined as a multi-scale radial basis function: Finally, the cross-entropy classification loss function is used to supervise the tomographic class prediction: The overall multi-task loss function consists of four terms: 。 8. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the bearing fault diagnosis method as described in any one of claims 1-7, which involves speed robust feature learning and minority class generation under data imbalance and variable speed environments.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the bearing fault diagnosis method as described in any one of claims 1-7, which involves speed robust feature learning and minority class generation under data imbalance and variable speed environments.

10. A computer program product, characterized in that: The invention includes a computer program that, when executed by a processor, implements the bearing fault diagnosis method as described in any one of claims 1-7, which involves speed-robust feature learning and minority class generation under data imbalance and variable speed environments.

Images