Selective adversarial augmented network and system for rolling bearing fault diagnosis

Abstract

The application provides a selective adversarial enhancement network and system for rolling bearing fault diagnosis, a framework of the selective adversarial enhancement network is integrated with a balance enhancement module and a selective adversarial module, and is used for realizing fine-grained sub-domain alignment between a source domain and a target domain; the selective adversarial enhancement network comprises the balance enhancement module, the selective adversarial module and an uncertainty suppression module; the balance enhancement module is used for dynamically adjusting category distribution of the source domain and the target domain, so that the category distribution of the source domain and the target domain is balanced; the selective adversarial module is used for screening abnormal categories and accurately aligning shared categories, and realizes distribution alignment at a sub-domain level; and the uncertainty suppression module is used for suppressing uncertainty propagation caused by misclassification by optimizing a loss function, and improving alignment quality of samples close to a decision boundary.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, and in particular to a selective countermeasure enhancement network and system for rolling bearing fault diagnosis. Background Technology

[0002] Currently, domain adaptation methods are widely used in mechanical fault diagnosis. However, most existing methods assume that the source domain and the target domain share the same label space, making it difficult to handle the partial domain adaptation problem where the target domain label space is a subset of the source domain. They also face the challenge of uncertainty in classification boundaries. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a selective adversarial enhancement network and system for rolling bearing fault diagnosis. The balanced enhancement module effectively alleviates the negative transfer caused by class asymmetry by dynamically adjusting the class weight distribution between the source and target domains. The selective adversarial module accurately filters shared classes and removes abnormal class interference during adversarial training, achieving fine-grained feature alignment at the subdomain level. The uncertainty suppression module introduces a boundary uncertainty penalty term into the loss function to suppress the propagation of uncertainty in samples near the classification boundary, thereby improving the recognition accuracy and model robustness of samples near the classification boundary.

[0004] In a first aspect, embodiments of the present invention provide a selective adversarial enhancement network for rolling bearing fault diagnosis. The framework of the selective adversarial enhancement network integrates a balancing enhancement module and a selective adversarial module to achieve fine-grained sub-domain alignment between the source domain and the target domain. The selective adversarial enhancement network includes: a balancing enhancement module, a selective adversarial module, and an uncertainty suppression module. The balancing enhancement module is used to dynamically adjust the class distribution of the source domain and the target domain to achieve a balance between the class distributions of the source domain and the target domain. The selective adversarial module is used to filter out abnormal classes and accurately align shared classes to achieve sub-domain-level distribution alignment. The uncertainty suppression module is used to suppress the propagation of uncertainty caused by misclassification by optimizing the loss function, thereby improving the alignment quality of samples near the decision boundary.

[0005] In optional embodiments of this application, the selective adversarial enhancement network further includes: a feature extraction module and a classifier; the feature extraction module is used to extract domain-invariant features of the source domain and the target domain by sharing parameters; the balanced enhancement module is used to enhance the source domain samples by random selection, thereby increasing the diversity of the target domain samples; the selective adversarial module is used to selectively adversarially train the enhanced features of the source domain and the target domain, and to perform selective adversarial learning based on a strategy optimized by category and subdomain; the uncertainty suppression module is used to suppress uncertainty propagation by dynamically adjusting the loss function; and the classifier is used to classify fault features.

[0006] In an optional embodiment of this application, the feature extraction module is used to convert the one-dimensional vibration signal in the source and target domains into a two-dimensional image more suitable for feature extraction by means of Gram angular difference field (GADF); the feature extraction module uses a ResNet50 network as the basic architecture of the feature extraction network, and the ResNet50 network is used to extract deep features related to the bearing fault state.

[0007] In an optional embodiment of this application, the feature extraction module performs feature extraction, including: inputting two-dimensional image features transformed by GADF into a ResNet50 network; and progressively optimizing the extracted features through a convolutional block attention module (CBAM) and a simple attention module (SimAM).

[0008] In optional embodiments of this application, the aforementioned balancing enhancement module is used to enrich the class distribution of the target domain in the early stages of training, so that the class distribution of the target domain and the source domain are kept in balance; the selective adversarial module is used to align samples with the class prediction probability of the target domain using a class-level adversarial discriminator.

[0009] In an optional embodiment of this application, the above-mentioned balanced enhancement module is used to obtain the class prediction distribution of each sample through forward propagation using the data of the current batch in the target domain, and calculate the average probability of each class in the target domain; select source domain samples that meet a certain weight threshold to form an enhanced target domain batch; the balanced enhancement module is used to assign an enhancement weight to each sample from the source domain mini-batch according to the class probability of the target domain.

[0010] In an optional embodiment of this application, the selective adversarial module is used to configure an independent domain discriminator for each category in the source domain and to perform category-level adversarial training using the aforementioned enhanced target domain batches.

[0011] In an optional embodiment of this application, the uncertainty suppression module is used to dynamically suppress the negative impact of uncertain categories on the target domain classifier by calculating the uncertainty of the source domain samples.

[0012] In optional embodiments of this application, the training process of the selective adversarial enhancement network includes: a balancing enhancement stage, a selective adversarial stage, and an uncertainty suppression stage. The training process of the balancing enhancement stage includes: enhancing a portion of the source domain samples based on the source domain data to supplement the class distribution of the target domain, ensuring that the target domain has a more balanced class distribution in the early stages of training. The training process of the selective adversarial stage includes: after balancing enhancement, performing class-level adversarial training using a selective adversarial module, further reducing the interference of irrelevant classes in the source domain on the training of the target domain by judging the class matching degree between the source domain and target domain samples. The training process of the uncertainty suppression stage includes: using an uncertainty suppression module to perform weighted suppression on uncertain classes in the target domain, thereby improving the stability and accuracy of the target domain classifier. The model is optimized in the balancing enhancement stage, the selective adversarial stage, and the uncertainty suppression stage respectively to reduce the negative transfer effect of non-shared classes and enhance the cross-domain adaptability of the target domain.

[0013] Secondly, embodiments of the present invention also provide a selective countermeasure enhancement system for rolling bearing fault diagnosis, comprising: the aforementioned selective countermeasure enhancement network for rolling bearing fault diagnosis.

[0014] The embodiments of the present invention bring the following beneficial effects:

[0015] This invention provides a selective adversarial enhancement network and system for rolling bearing fault diagnosis. The balanced enhancement module effectively mitigates negative transfer caused by class asymmetry by dynamically adjusting the class weight distribution between the source and target domains. The selective adversarial module accurately filters shared classes and eliminates interference from abnormal classes during adversarial training, achieving fine-grained feature alignment at the subdomain level. The uncertainty suppression module introduces a boundary uncertainty penalty term into the loss function to suppress the propagation of uncertainty in samples near the classification boundary, thereby improving the recognition accuracy and model robustness of samples near the classification boundary.

[0016] Other features and advantages of this disclosure will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0017] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a selective adversarial enhancement network for rolling bearing fault diagnosis provided in an embodiment of the present invention;

[0020] Figure 2 A schematic diagram of a GRCS feature extraction module provided in an embodiment of the present invention;

[0021] Figure 3 A schematic diagram illustrating uncertainty suppression and decision boundary optimization provided in an embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram of a selective countermeasure enhancement system for rolling bearing fault diagnosis provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In recent years, with the rapid growth of manufacturing demand, industrial equipment is developing towards greater automation, complexity, and intelligence. This places higher demands on the fault diagnosis and health monitoring of industrial systems and their key components. Rotating machinery, as a crucial component of industrial systems, typically operates in unsteady and harsh environments, exhibiting complex and unpredictable failure modes that easily lead to various malfunctions. These failures not only severely impact production efficiency and product quality but can also trigger large-scale production accidents, even causing irreversible losses. Therefore, achieving efficient and accurate fault diagnosis for rotating machinery is not only crucial for ensuring production safety but also effectively reduces downtime, improves economic efficiency, and provides strong technical support for intelligent manufacturing.

[0025] With the rapid development of information technology, deep learning (DL), as a powerful data analysis and pattern recognition tool, has become one of the important technologies in the field of fault diagnosis due to its superior feature extraction and automatic learning capabilities. Compared with traditional machine learning methods, deep learning can automatically extract fault features from massive amounts of sensor data, reducing reliance on expert experience and thus significantly improving the efficiency and accuracy of fault diagnosis. Especially in the fault diagnosis of rotating machinery, deep learning has been widely used, capable of processing large-scale, multi-dimensional, and complex time-series data and quickly identifying fault modes. However, despite the significant progress made by deep learning technology in the field of fault diagnosis, it still faces the following two major challenges:

[0026] (1) High dependence on labeled data: Deep learning models usually require a large amount of high-quality labeled data. However, the randomness and complexity of fault events in industrial scenarios make it extremely difficult to obtain such data. Especially when the fault type changes frequently, the cost of collecting labeled data is extremely high, which seriously restricts the practical application of the model.

[0027] (2) Insufficient cross-domain generalization ability: Changes in the operating conditions of industrial equipment (such as speed and load) can lead to significant differences in the distribution of monitoring data. Deep learning-based diagnostic models are prone to performance degradation when faced with unknown operating conditions. Traditional training methods often assume that the training and test data distributions are consistent, but this assumption is difficult to adapt to the complex and ever-changing actual industrial environment. Therefore, improving the model's generalization ability under cross-domain conditions is one of the key issues that urgently needs to be addressed in current intelligent fault diagnosis.

[0028] To address the aforementioned issues, transfer learning (TL) offers an effective technique. Transfer learning aims to improve the transfer effect in the target domain (unlabeled data) by leveraging knowledge learned from the source domain (labeled data), thereby enhancing the model's generalization ability. Domain adaptation methods within transfer learning effectively address the problem of data distribution differences, improving the diagnostic performance of the target task through cross-domain knowledge transfer. In recent years, an increasing number of researchers have focused on applying domain adaptation methods to the field of fault diagnosis for industrial equipment, especially in scenarios involving multiple operating conditions and multiple domains.

[0029] However, traditional domain adaptation methods typically assume that the source and target domains have the same class space, which is not easily achieved in many practical applications. For example, in rotating machinery fault diagnosis, due to different operating conditions (such as changes in speed and load), the label space of the target domain is often only a subset of the label space of the source domain. That is, the target domain may lack samples of certain fault categories from the source domain. This situation is called Partial Domain Adaptation (PDA). In this case, directly applying traditional domain adaptation methods may lead to negative transfer problems because categories in the source domain that are not shared with the target domain (i.e., external classes) can interfere with network learning, thereby affecting the performance of the target task.

[0030] To address the partial domain adaptation problem, researchers have proposed various methods to screen and filter irrelevant categories in the source domain, reducing the interference of external class samples on model training. The main strategies for partial domain adaptation include sample weighting and class selection, subdomain alignment and local adversarial training, adversarial training and domain discriminators, dual-domain alignment, and the collaborative achievement of target domain distribution balance through feature consistency and pseudo-label generation. These methods identify and remove external class samples from the source domain, thereby enhancing the alignment effect of shared class samples and improving model performance in the target domain. While these methods alleviate the negative transfer problem to some extent, most do not fully exploit the potential of unlabeled data in the target domain and lack fine-grained screening of external class samples.

[0031] Based on this, this invention provides a Selective Adversarial Augmentation Network and System for Rolling Bearing Fault Diagnosis, specifically proposing a rolling bearing fault diagnosis method based on Selective Adversarial Augmentation Network (SAAN). First, by dynamically adjusting the class distributions of the source and target domains to achieve balance, the class asymmetry problem is resolved and the negative transfer effect is mitigated. Second, a selective domain adversarial module is designed to filter outomaly classes and accurately align shared classes, thereby alleviating the distribution mismatch problem caused by outliers and achieving subdomain-level distribution alignment to capture subtle differences in rolling bearing fault characteristics under different operating conditions. Finally, an uncertainty suppression module is proposed to optimize the loss function, effectively suppressing the propagation of uncertainty in erroneous classes and enhancing the alignment effect of samples near the classification boundary, thereby improving the classification performance and transfer effect of the target domain. Transfer diagnosis experiments using two rolling bearing datasets verify the superiority of this method.

[0032] To facilitate understanding of this embodiment, a selective adversarial enhancement network for rolling bearing fault diagnosis disclosed in this embodiment of the invention will first be described in detail.

[0033] Example 1:

[0034] This invention provides a selective adversarial enhancement network for rolling bearing fault diagnosis. The framework of the selective adversarial enhancement network integrates a balanced enhancement module and a selective adversarial module to achieve fine-grained subdomain alignment between the source domain and the target domain.

[0035] See Figure 1 The diagram shows a selective adversarial enhancement network for rolling bearing fault diagnosis. Figure 1 As shown in (a), the selective adversarial enhancement network includes: a balance enhancement module, a selective adversarial module, and an uncertainty suppression module; the balance enhancement module is used to dynamically adjust the class distribution of the source domain and the target domain so that the class distribution of the source domain and the target domain reaches a balance; the selective adversarial module is used to filter out abnormal classes and accurately align shared classes to achieve distribution alignment at the subdomain level; the uncertainty suppression module is used to suppress the propagation of uncertainty caused by misclassification by optimizing the loss function and improve the alignment quality of samples close to the decision boundary.

[0036] This embodiment proposes a novel partial domain adaptation method—Selective Adversarial Augmentation Network (SAAN). SAAN aims to address the shortcomings of existing methods in handling partial domain adaptation by combining a balanced augmentation module, a selective adversarial module, and an uncertainty suppression module. Specifically, it mainly includes:

[0037] (1) Propose the SAAN framework: integrate the balanced enhancement module and the selective adversarial module to achieve fine-grained subdomain alignment between the source domain and the target domain, effectively alleviating the negative migration problem in partial domain adaptation.

[0038] (2) Design an uncertainty suppression module: By optimizing the loss function, the uncertainty propagation caused by misclassification is suppressed, the alignment quality of samples close to the decision boundary is improved, and the accuracy and stability of the diagnostic results are enhanced.

[0039] (3) Verification of transfer learning ability and generalization performance: Cross-domain diagnostic tasks were performed in two rolling bearing experiments to verify the effectiveness of SAAN. In addition, the results of comparison with several other advanced TLFD methods demonstrate the superiority of SAAN.

[0040] This invention provides a selective adversarial enhancement network for rolling bearing fault diagnosis. The balanced enhancement module effectively mitigates negative transfer caused by class asymmetry by dynamically adjusting the class weight distribution between the source and target domains. The selective adversarial module accurately filters shared classes and eliminates interference from abnormal classes during adversarial training, achieving fine-grained feature alignment at the subdomain level. The uncertainty suppression module introduces a boundary uncertainty penalty term into the loss function to suppress the propagation of uncertainty in samples near the classification boundary, thereby improving the recognition accuracy and model robustness of samples near the classification boundary.

[0041] Example 2:

[0042] This invention provides another selective adversarial enhancement network for rolling bearing fault diagnosis, implemented based on the above embodiments. First, some concepts involved in this embodiment are introduced.

[0043] 1. Partial domain adaptation.

[0044] Domain adaptation (DA) enables models to learn transferable knowledge from labeled data in the source domain by reducing the differences in data distribution between the source and target domains. However, traditional domain adaptation methods typically assume that the source and target domains have the same label space (i.e., ...). This assumption does not hold true in practical applications. Partial Domain Adaptation (PDA) assumes that the target domain's tag space... It is the source domain tag space subset (i.e. However, even in shared tag spaces In this context, the distributions of the source and target domains still show significant differences (i.e., ).

[0045] In addition, there are class samples in the source domain that do not match the target domain. These samples may cause negative transfer problems, which further increases the complexity of cross-domain diagnosis.

[0046] In rolling bearing fault diagnosis, the actual operating environment of the target domain typically contains only a subset of categories, which is not entirely consistent with the category distribution of the source domain. This introduces a research problem related to PDA (Programmable Analytical Automation). and Represent the source domain and the target domain, respectively, where the source domain contains A labeled sample, the target domain contains Unlabeled samples. The source and target domain data come from different distributions. and Extracted from the target domain's label space. It is the source domain tag space subset (i.e. Furthermore, even in shared tag spaces Within this domain, the distributions of the source and target domains still differ significantly (i.e., Partial domain adaptation focuses on adapting only to categories common to both the source and target domains, thus avoiding the negative impact of mismatched categories in the source domain.

[0047] 2. Domain-based adversarial learning.

[0048] The core issue of transfer intelligence diagnostics is mapping the source and target domain data of bearings to a common feature space and aligning their distributions within that space. Domain adversarial learning is one of the main methods for feature distribution alignment, with the Domain Adversarial Training Neural Network (DANN) being a representative example. This network borrows ideas from generative adversarial learning, designing a domain discriminator that, through adversarial training with features as input, promotes the alignment of the source and target domain data distributions in the feature space. This prevents the domain discriminator from correctly identifying whether a feature belongs to the source or target domain, ultimately eliminating data distribution discrepancies.

[0049] Domain-Adversarial Networks (DANs) mainly consist of three parts: feature extractors. Domain discriminator and fault classifier The domain discriminator is primarily used to distinguish whether data features belong to the source or target domain. The feature extractor fine-tunes based on the error of the domain discriminator to confuse it, ultimately achieving transfer learning. The loss function of the domain adversarial network is as follows:

[0050] (1)

[0051] In the formula: Let m be the m-th sample in the union of the source and target domain samples; n represents the total number of samples in the source and target domains. This represents the domain label; a value of 1 indicates the source domain, and a value of 0 indicates the target domain. , and They are respectively , and Optimization parameters; and This represents the loss function of the classifier and the loss function of the domain discriminator; The weight parameters of the two loss functions were measured. The domain adversarial network is trained through the minimax optimization process in equation (2):

[0052] (2)

[0053] Through training, the feature extractor can eliminate differences in data distribution and extract domain-invariant features from the source and target domains. The classifier trained using source domain data can be directly used for classifying target domain data, thus transferring source domain diagnostic knowledge to the target domain.

[0054] 3. ResNet50.

[0055] ResNet (Residual Network) is a deep convolutional neural network that effectively solves the vanishing or exploding gradient problems commonly encountered in traditional deep networks during training by introducing a cross-layer connection mechanism. This mechanism allows some input data to bypass the convolutional layers and be directly passed to the output, thus preserving the original data features. Simultaneously, residual learning, by adding the input and output and combining it with the nonlinear mapping function ReLU, not only simplifies the model optimization process but also significantly improves the stability of gradient backpropagation, effectively addressing the challenge of training deep networks.

[0056] Thanks to this innovation, ResNet quickly became a classic architecture in the field of deep learning and has been widely applied in various fields such as image recognition, speech processing, and fault diagnosis. Within this framework, different versions of the ResNet model have been proposed, among which ResNet50, as a deep variant, has been widely adopted due to its superior performance. ResNet50 consists of 50 layers of neural networks, significantly improving feature extraction capabilities through a deeper network structure while maintaining low parameter count and computational complexity. Its architecture comprises two core modules: ConvBlock and IdentityBlock, used to change the network's dimensionality and deepen its structure, respectively. ConvBlock adapts to the input and output requirements of different layers by adjusting the number of channels, while IdentityBlock further learns higher-order features while maintaining the same input dimensionality. In the field of bearing fault diagnosis, ResNet50, with its deep convolutional neural network structure, can effectively capture subtle features in bearing operating data, extract key characteristics at multiple scales and dimensions, and demonstrates strong adaptability to data diversity and complexity. By deeply exploring the potential features in the operating state, ResNet50 can provide accurate and comprehensive diagnostic information, offering robust and reliable technical support for the condition monitoring and maintenance of industrial equipment, making it a highly promising tool in deep learning-based fault diagnosis research.

[0057] For rolling bearing fault diagnosis tasks in partial domain adaptation scenarios, this embodiment proposes an improved DANN-based model—Selective Adversarial Augmentation Network (SAAN). This model achieves accurate alignment of shared class features and effective alignment of subdomain distributions by balancing augmentation and selective domain adversarial strategies. Simultaneously, it dynamically adjusts the class distribution to mitigate the impact of negative transfer and optimizes the loss function to suppress uncertainty propagation, thereby improving the model's diagnostic performance in the target domain. A detailed framework of the SAAN model can be found in [link to documentation]. Figure 1 (b).

[0058] In some embodiments, the selective adversarial enhancement network further includes: a feature extraction module and a classifier; the feature extraction module is used to extract domain-invariant features of the source domain and the target domain by sharing parameters; the balanced enhancement module is used to enhance the source domain samples by random selection, thereby increasing the diversity of the target domain samples; the selective adversarial module is used to selectively adversarially train the enhanced features of the source domain and the target domain, and to perform selective adversarial learning based on a strategy optimized by category and subdomain; the uncertainty suppression module is used to suppress uncertainty propagation by dynamically adjusting the loss function; and the classifier is used to classify fault features.

[0059] First, in the data collection and preprocessing stage, one-dimensional vibration signals are acquired to obtain data from the source and target domains. Subsequently, the data is input into the feature extraction module (GRCS), which extracts domain-invariant features from the source and target domains using shared parameters. and To effectively mitigate negative transfer, SAAN introduces a balanced enhancement selective adversarial module. In this module, source domain samples are randomly selected and enhanced, increasing the diversity of target domain samples and thus improving the performance of the target domain. Next, the enhanced features of the source and target domains are trained using selective adversarial methods, employing a strategy based on category and subdomain optimization to avoid situations where certain categories in the source domain have no corresponding categories in the target domain, thereby reducing the negative transfer impact of irrelevant categories on the target domain. To further improve model performance, SAAN designs an adaptive uncertainty suppression module, which dynamically adjusts the loss function to suppress uncertainty propagation, ensuring the model's stability and robustness across multiple domains. Finally, a classifier is used to classify fault features, and the accuracy of fault diagnosis is improved by optimizing the loss function.

[0060] I. GRCS Feature Extraction Module

[0061] In some embodiments, the feature extraction module is used to convert one-dimensional vibration signals in the source and target domains into two-dimensional images more suitable for feature extraction using Gram angular difference field (GADF). The feature extraction module uses a ResNet50 network as the basic architecture of the feature extraction network, which is used to extract deep features related to bearing fault conditions.

[0062] This embodiment proposes a novel feature extraction module—GRCS (GADF-ResNet50-CBAM-SimAM)—aimed at effectively extracting features and enhancing the performance of the target domain in cross-domain tasks. This module improves the accuracy and generalization ability of feature extraction by combining various advanced network architectures and attention mechanisms.

[0063] First, the GADF (Grammian Angular Difference Field) method is introduced to transform one-dimensional vibration signals in the source and target domains into two-dimensional images more suitable for feature extraction. This captures more detailed spatial and angular information in the image, providing richer feature representations for subsequent feature extraction. This method enriches feature representations by enhancing the geometric relationships of the image. Subsequently, ResNet50 is adopted as the foundation of the feature extraction network. ResNet50 not only performs excellently in image classification but also effectively avoids the gradient vanishing problem common in training deep networks due to its residual connection design. In the GRCS module, ResNet50 is used as the core feature extractor to extract deep features related to bearing fault states. Furthermore, CBAM (Convolutional Block Attention Module) and SimAM (Simple Attention Module) are combined into the feature extraction process to further enhance feature expressiveness. Specifically, CBAM effectively extracts key region features in the image and suppresses the influence of irrelevant regions through channel attention and spatial attention mechanisms. SimAM, through its efficient spatial self-attention mechanism, helps to better capture common patterns between the source and target domains, avoiding interference from irrelevant class samples in the source domain on the training of the target domain. The combination of the two modules can optimize the feature extraction process at multiple levels, further enhancing cross-domain adaptability.

[0064] In some embodiments, the feature extraction module performs feature extraction, including: inputting two-dimensional image features transformed by GADF into a ResNet50 network; and progressively optimizing the extracted features through a convolutional block attention module (CBAM) and a simple attention module (SimAM).

[0065] The overall architecture of the GRCS module can be found in [reference needed]. Figure 2 The diagram illustrates a GRCS feature extraction module. The main steps of feature extraction include: first, inputting rich features from GADF into the ResNet50 network; then, progressively optimizing the extracted features through CBAM and SimAM modules, ultimately obtaining the feature representation extracted by the GRCS module. This process effectively improves the target domain diagnostic capability in PDA tasks, especially when there is a significant difference between the source and target domains. The GRCS module can help the model make more accurate classification predictions in the target domain.

[0066] II. Balanced Enhancement Selective Countermeasure Module (i.e., the general term for Balanced Enhancement Module and Selective Countermeasure Module).

[0067] In some embodiments, the balancing enhancement module is used to enrich the class distribution of the target domain in the early stages of training, so that the class distribution of the target domain is balanced with that of the source domain; the selective adversarial module is used to align samples with the class prediction probability of the target domain using a class-level adversarial discriminator.

[0068] In PDA tasks, the label spaces of the source and target domains typically only partially overlap, meaning the target domain labels are merely a subset of the source domain labels. When the target domain is small and the class distribution is uneven, directly aligning the overall distribution can easily lead to incorrect matching of target domain data to source domain classes that do not exist in its label space, resulting in negative transfer. Traditional approaches often use a "source domain filtering" strategy to remove irrelevant samples, but when the source domain is much larger than the target domain, over-filtering can discard potentially useful information. To address this contradiction, this embodiment proposes a Balanced Augmentation Selective Adversarial Module (BASAM). First, the target domain is appropriately "supplemented" in the early stages of training to enrich its class distribution and maintain a good balance with the source domain. Then, combined with the selective adversarial module, a class-level adversarial discriminator is used to align samples with the target domain's class prediction probabilities, thereby suppressing the interference of non-shared classes on the adversarial process. With the help of this module, the target domain can not only more fully mine useful samples that can be transferred from the source domain, but also automatically weaken those "abnormal classes" that have a very low probability of appearing in the target domain during adversarial alignment, thereby effectively reducing the risk of negative transfer and improving cross-domain recognition performance.

[0069] (1) Balanced Alignment Module.

[0070] In some embodiments, the balanced enhancement module is used to utilize the data of the current batch in the target domain to obtain the class prediction distribution of each sample through forward propagation and calculate the average probability of each class in the target domain; select source domain samples that meet a certain weight threshold to form an enhanced target domain batch; the balanced enhancement module can assign an enhancement weight to each sample from the source domain mini-batch according to the class probability of the target domain.

[0071] To achieve balanced enhancement, this embodiment first utilizes the data from the current batch in the target domain to obtain the predicted class distribution of each sample through forward propagation, and then calculates the average probability of each class in the target domain. Let the current target domain batch be... After classifier Obtain the predicted distribution Then define each category The average predicted probability is:

[0072] (3)

[0073] in, Indicates sample Predicted as category The probability. Because the target domain label only covers shared categories. Non-shared categories Typically lower.

[0074] Based on this, this embodiment designs a balanced enhancement module: from source domain mini-batch In this process, each sample is assigned an enhancement weight based on the class probability of the target domain. Specifically, for samples from the source domain... (in Let its enhanced weights be defined as:

[0075] (4)

[0076] in, This is the initial enhancement ratio, which gradually decreases as training progresses; Represents source domain samples Corresponding category Average predicted probability over mini-batch in the target domain; This represents the confidence level of the sample in the source domain (which can be represented by the maximum probability predicted by the classifier or other metrics); The adjustment function controls the upper bound of the enhancement weights, assigning greater weights to samples with higher prediction probabilities (i.e., shared class) in the target domain, while naturally suppressing samples of non-shared class.

[0077] Subsequently, source domain samples that meet certain weight thresholds are selected to form an enhanced target domain batch. ,Right now:

[0078] (5)

[0079] in, A preset threshold is used. In this way, only samples that share the same category as the target domain and have a high confidence level in the source domain can be "added" to the target domain, thereby effectively balancing the category distribution of the target domain batch.

[0080] (2) Selective Adversarial Module.

[0081] In some embodiments, the selective adversarial module is used to configure an independent domain discriminator for each category in the source domain and to perform category-level adversarial training using the aforementioned enhanced target domain batches.

[0082] While the balancing enhancement module enriches and balances the class distribution of the target domain to some extent, it cannot completely eliminate the interference of irrelevant classes from the source domain on the alignment process. This is especially true when there are "outlier classes" in the source domain that do not appear in the target domain, which can still lead to negative transfer. Therefore, this embodiment further proposes a selective adversarial module to more finely suppress the impact of these non-shared classes on cross-domain alignment, thereby further improving the overall adaptation performance. Unlike traditional single discriminator methods, this module provides a selective adversarial module for each class in the source domain. Configure an independent domain discriminator And utilize the aforementioned enhanced target domain batch Conduct category-level adversarial training.

[0083] Specifically, for each sample within a batch Record it as the predicted category The probability is In category Domain discriminator In this context, the adversarial loss for this sample can be expressed as:

[0084] (6)

[0085] in For feature extraction function, For the loss function with gradient inversion, The domain label is then used. Subsequently, this embodiment categorizes categories based on the target domain. average probability Weighting the discriminator loss, the overall selective adversarial loss is defined as:

[0086] (7)

[0087] Therefore, for shared categories with a high probability of occurrence in the target domain, their feature representations receive stronger adversarial correction, while for non-shared categories with a low probability of occurrence in the target domain, the adversarial correction is naturally reduced, thus effectively mitigating the negative transfer problem. Compared to traditional single-discriminator methods, multi-discriminator architectures can achieve more refined source domain sample selection at the category level, eliminating noisy data that does not match the distribution of the target domain, thereby significantly improving the accuracy and robustness of cross-domain adaptation.

[0088] III. Adaptive Uncertainty Suppression Module (AUSM).

[0089] In some embodiments, the uncertainty suppression module is used to dynamically and weightfully suppress the negative impact of uncertain categories on the target domain classifier by calculating the uncertainty of source domain samples.

[0090] In PDA tasks, the class differences between the source and target domains inevitably introduce classification uncertainty, especially when the target domain has few classes, which can lead to significant model uncertainty. To improve the robustness of the model, this embodiment proposes an Adaptive Uncertainty Suppression Module (AUSM). This module calculates the uncertainty of source domain samples and dynamically weights and suppresses the negative impact of uncertain classes on the target domain classifier, thereby reducing the interference of irrelevant class samples on target domain learning.

[0091] See Figure 3 The diagram shown illustrates uncertainty suppression and decision boundary optimization. Figure 3 This demonstrates the impact of uncertainty propagation, where the close boundaries between class 1 and class 2 in the source domain lead to misclassification during target domain prediction, thus reducing the classification accuracy of the target domain. To further mitigate this impact, AUSM employs a weighted complement entropy loss to reduce the negative influence of uncertain classes in the source domain. It weights source domain samples according to the magnitude of uncertainty to reduce the influence of irrelevant source domain samples that do not contribute to the target domain. The core loss function of AUSM is defined as follows:

[0092] (8)

[0093] in, It is a weighting factor for the source domain samples, which is dynamically adjusted based on the importance of the source domain samples in the target domain; It is the cross-entropy loss of the source domain samples, used to measure the classification accuracy of the source domain samples; and As balancing terms, the weights of the supplementary entropy loss term and the uncertainty penalty term are controlled respectively; This is an uncertainty penalty term used to suppress the interference of uncertain categories in the source domain on the target domain model. This term adjusts the contribution of source domain samples by quantifying the classification uncertainty of each sample.

[0094] AUSM is designed to dynamically adjust the impact of source domain samples on target domain training based on their uncertainty. In the target domain, source domain samples with lower uncertainty are assigned higher weights, while samples with higher uncertainty are suppressed, thus reducing their interference with the target domain learning process. In this way, AUSM effectively suppresses the negative impact of irrelevant categories in the source domain on the target domain, thereby improving the classification performance of the target domain.

[0095] IV. Algorithms and Training Strategies

[0096] In some embodiments, the training process of the selective adversarial augmentation network includes: a balanced augmentation phase, a selective adversarial phase, and an uncertainty suppression phase. The training process of the balanced augmentation phase includes: augmenting a portion of the source domain samples based on the source domain data to supplement the class distribution of the target domain, ensuring that the target domain has a more balanced class distribution in the early stages of training. The training process of the selective adversarial phase includes: after balanced augmentation, performing class-level adversarial training using a selective adversarial module, further reducing the interference of irrelevant classes in the source domain on the training of the target domain by judging the class matching degree between the source domain and target domain samples. The training process of the uncertainty suppression phase includes: using an uncertainty suppression module to perform weighted suppression on uncertain classes in the target domain, thereby improving the stability and accuracy of the target domain classifier.

[0097] In some embodiments, the model is optimized in the balance enhancement stage, the selective adversarial stage, and the uncertainty suppression stage, respectively, to reduce the negative migration impact of non-shared categories and enhance the cross-domain adaptability of the target domain.

[0098] The SAAN model addresses the label space differences between the source and target domains by combining Balanced Augmentation, Selective Adversarial, and Adaptive Uncertainty Suppression (AUSM) modules, thereby improving the target domain's performance in cross-domain fault diagnosis tasks.

[0099] The SAAN training process consists of three main stages: 1) Balanced Augmentation Stage: Based on the source domain data, a portion of the source domain samples are augmented to supplement the class distribution of the target domain, ensuring a more balanced class distribution in the early stages of training; 2) Selective Adversarial Stage: After balanced augmentation, a selective adversarial module is used for class-level adversarial training. By judging the class matching degree between the source and target domain samples, the interference of irrelevant classes in the source domain on the target domain training is further reduced; 3) Uncertainty Suppression Stage: Through an adaptive uncertainty suppression module, uncertain classes in the target domain are weighted and suppressed, thereby improving the stability and accuracy of the target domain classifier. By optimizing the model at different stages, the negative transfer effect of non-shared classes is reduced to enhance the cross-domain adaptability of the target domain.

[0100] Example 3:

[0101] This invention provides a simulation experiment analysis of a selective adversarial enhancement network for rolling bearing fault diagnosis, which is implemented based on the above embodiments.

[0102] To verify the effectiveness of the SAAN-based rolling bearing fault diagnosis model, this embodiment verifies the effectiveness of the proposed method through the fault diagnosis results of two bearing fault diagnosis cases. In Case 1, this embodiment utilizes variable speed operating condition data contained in the CWRU dataset to systematically analyze the fault characteristics of the bearing under different speed conditions. In Case 2, bearing data is collected using the Mechanical Fault Simulation (MFS) experimental platform in the Intelligent Rotating Machinery Maintenance Laboratory to more realistically reproduce the complexity of the bearing operating environment under actual conditions. This section details the data preprocessing process and provides detailed case studies. All experiments were conducted on the deep learning framework Torch.

[0103] In Case 1, the SAAN method provided in this embodiment maintained high classification accuracy in all tasks and scenarios, reaching or approaching 100% multiple times. It also remained stable as the missing percentage increased, which fully demonstrates its excellent robustness and generalization ability in a partial domain adaptation environment where the target domain label space is only a subset of the source domain. It can effectively suppress negative migration and significantly improve the classification performance of rolling bearing fault diagnosis.

[0104] In Case 2, the SAAN method provided in this embodiment achieved nearly 100% classification accuracy across all tasks and missing class settings with minimal fluctuations, demonstrating its high adaptability and robustness even when the target domain label space is only a subset of the source domain. In contrast, methods such as Baseline, M1, M2, SAN, SPDA, DANN, and CDAN generally experienced a decrease in accuracy after removing key failure modes, especially under conditions of multi-class missing values ​​(e.g., missing values ​​2, 3, 4, 5, and 6), further validating SAAN's significant advantages in suppressing negative transfer, accurately aligning shared classes, and optimizing classification boundaries.

[0105] To address the challenge that existing methods in mechanical fault diagnosis, where the target domain label space is only a subset of the source domain, neglect the fine-grained distribution differences between subdomains and the significant performance degradation caused by weight imbalances in uncertain samples near the classification boundary, this embodiment proposes a rolling bearing fault diagnosis model based on SAAN. SAAN abandons the dependence on completely overlapping label spaces and is specifically designed for partial domain adaptation scenarios. Specifically, the balancing enhancement module effectively mitigates negative transfer caused by class asymmetry by dynamically adjusting the class weight distribution between the source and target domains; the selective adversarial module accurately filters shared classes and eliminates abnormal class interference during adversarial training, achieving fine-grained feature alignment at the subdomain level; and the uncertainty suppression module introduces a boundary uncertainty penalty term into the loss function to suppress the propagation of uncertainty in samples near the classification boundary, thereby improving the recognition accuracy and model robustness of samples near the classification boundary. Experimental results show that SAAN can achieve effective feature alignment between the source and target domains under PDA conditions, and in the diagnostic task of Dataset 1, SAAN achieves an average accuracy of 92.15%, which is an improvement over existing methods.

[0106] Example 4:

[0107] This invention provides a selective countermeasure enhancement system for rolling bearing fault diagnosis, implemented based on the above embodiments. See also... Figure 4 The diagram shows a selective countermeasure enhancement system for rolling bearing fault diagnosis, which includes the selective countermeasure enhancement network for rolling bearing fault diagnosis provided in the foregoing embodiments.

[0108] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the selective countermeasure enhancement system for rolling bearing fault diagnosis described above can be referred to the corresponding process in the foregoing embodiments, and will not be repeated here.

[0109] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0110] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0112] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fault diagnosis of rolling bearings based on selective adversarial enhancement networks, characterized in that, The framework of the selective adversarial enhancement network integrates a balanced enhancement module and the selective adversarial module to achieve fine-grained subdomain alignment between the source domain and the target domain. The selective adversarial enhancement network includes: a balance enhancement module, a selective adversarial module, and an uncertainty suppression module; The balance enhancement module is used to dynamically adjust the class distribution of the source domain and the target domain so that the class distribution of the source domain and the target domain reaches a balance. The selective adversarial module is used to filter out anomaly categories and precisely align shared categories, achieving subdomain-level distribution alignment. The uncertainty suppression module is used to suppress the propagation of uncertainty caused by misclassification by optimizing the loss function, thereby improving the alignment quality of samples close to the decision boundary. The selective adversarial enhancement network further includes: a feature extraction module and a classifier; the feature extraction module is used to extract domain-invariant features of the source domain and the target domain by sharing parameters; the balanced enhancement module is used to enhance the source domain samples by random selection, thereby increasing the diversity of the target domain samples; the selective adversarial module is used to selectively adversarially train the enhanced features of the source domain and the target domain, and to perform selective adversarial learning based on a strategy optimized by category and subdomain; the uncertainty suppression module is used to suppress uncertainty propagation by dynamically adjusting the loss function; and the classifier is used to classify fault features. The feature extraction module is used to convert the one-dimensional vibration signals in the source domain and the target domain into two-dimensional images that are more suitable for feature extraction using the Gram angular difference field (GADF). The feature extraction module uses a ResNet50 network as the basic architecture of the feature extraction network, and the ResNet50 network is used to extract deep features related to the bearing fault state. The core loss function of the adaptive uncertainty suppression module is defined as follows: ; in, It is a weighting factor for the source domain samples, which is dynamically adjusted based on the importance of the source domain samples in the target domain; It is the cross-entropy loss of the source domain samples, used to measure the classification accuracy of the source domain samples; and As balancing terms, the weights of the supplementary entropy loss term and the uncertainty penalty term are controlled respectively; This is an uncertainty penalty term used to suppress the interference of uncertain categories in the source domain on the target domain model. This term adjusts the contribution of the source domain samples by quantifying the classification uncertainty of each sample.

2. The method according to claim 1, characterized in that, The feature extraction module performs feature extraction, including: The two-dimensional image features transformed by GADF are input into the ResNet50 network; The extracted features are progressively optimized using the Convolutional Block Attention (CBAM) module and the Simple Attention (SimAM) module.

3. The method according to claim 1, characterized in that, The balancing enhancement module is used to enrich the class distribution of the target domain in the early stages of training, so that the class distribution of the target domain is balanced with that of the source domain. The selective adversarial module is used to align samples with the class prediction probability of the target domain using a class-level adversarial discriminator.

4. The method according to claim 3, characterized in that, The balanced enhancement module is used to obtain the class prediction distribution of each sample through forward propagation using the data of the current batch in the target domain, and calculate the average probability of each class in the target domain; and select source domain samples that meet a certain weight threshold to form an enhanced target domain batch. The balanced enhancement module is used to assign an enhancement weight to each sample from the source domain mini-batch based on the class probability of the target domain.

5. The method according to claim 3, characterized in that, The selective adversarial module is used to configure an independent domain discriminator for each category in the source domain and to perform category-level adversarial training using the aforementioned enhanced target domain batches.

6. The method according to claim 1, characterized in that, The uncertainty suppression module is used to dynamically suppress the negative impact of uncertain categories on the target domain classifier by calculating the uncertainty of source domain samples.

7. The method according to any one of claims 1-6, characterized in that, The training process of the selective adversarial enhancement network includes: a balance enhancement phase, a selective adversarial phase, and an uncertainty suppression phase; The training process of the balanced enhancement phase includes: on the basis of the source domain data, taking a portion of the source domain samples for enhancement to supplement the class distribution of the target domain, ensuring that the target domain has a more balanced class distribution in the early stage of training; The training process of the selective adversarial phase includes: after the balance enhancement, class-level adversarial training is performed using the selective adversarial module, and the interference of irrelevant classes in the source domain on the training of the target domain is further reduced by judging the class matching degree between the source domain and the target domain samples. The training process of the uncertainty suppression stage includes: using the uncertainty suppression module to perform weighted suppression on uncertain categories in the target domain, thereby improving the stability and accuracy of the target domain classifier; The model is optimized in the balance enhancement stage, the selective adversarial stage, and the uncertainty suppression stage to reduce the negative migration impact of non-shared categories and enhance the cross-domain adaptability of the target domain.

8. A selective countermeasure enhancement system for rolling bearing fault diagnosis, characterized in that, Used to perform the rolling bearing fault diagnosis method based on selective adversarial enhancement network as described in any one of claims 1-7.

Images