Selective adversarial enhancement network and system for rolling bearing fault diagnosis

Abstract

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

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a selective countermeasure enhancement network and system for rolling bearing fault diagnosis. Background Art

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

[0003] In view of this, the purpose of the present invention is to provide a selective adversarial enhancement network and system for rolling bearing fault diagnosis. The balance enhancement module effectively alleviates the negative migration caused by category asymmetry by dynamically adjusting the category weight distribution of the source domain and the target domain; the selective adversarial module accurately screens shared categories and eliminates abnormal category interference during adversarial training to achieve fine-grained feature alignment at the subdomain level; the uncertainty suppression module introduces a boundary uncertainty penalty term in the loss function to suppress the uncertainty propagation of samples near the classification boundary, thereby improving the recognition accuracy and model robustness of samples near the classification boundary.

[0004] In the first aspect, an embodiment of the present invention provides a selective adversarial enhancement network for rolling bearing fault diagnosis. The framework of the selective adversarial enhancement network integrates a balance enhancement module and a 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 category distribution of the source domain and the target domain to achieve a balance between the category distribution of the source domain and the target domain; the selective adversarial module is used to screen abnormal categories and accurately align shared categories to achieve distribution alignment at the subdomain level; the uncertainty suppression module is used to suppress the uncertainty propagation caused by misclassification by optimizing the loss function, thereby improving the alignment quality of samples close to the decision boundary.

[0005] In an optional embodiment of the present application, the above-mentioned selective adversarial enhancement network also 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 through shared parameters; the balanced enhancement module is used to randomly select and enhance the source domain samples to increase the diversity of the target domain samples; the selective adversarial module is used to perform selective adversarial training on the enhanced features of the source domain and the target domain, and perform selective adversarial learning based on the strategy of category and subdomain optimization; the uncertainty suppression module is used to suppress the propagation of uncertainty by dynamically adjusting the loss function; the classifier is used to classify fault features.

[0006] In an optional embodiment of the present application, the above-mentioned 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 through the Gram angle difference field GADF; the feature extraction module uses the 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 the present application, the above-mentioned feature extraction module performs feature extraction, including: inputting the two-dimensional image features converted by GADF into the ResNet50 network; and gradually optimizing the extracted features through the convolution block attention module CBAM and the simple attention module SimAM module.

[0008] In an optional embodiment of the present application, the above-mentioned balance enhancement module is used to enrich the category distribution of the target domain in the early stage of training so that the category distribution of the target domain and the source domain are balanced; the selective adversarial module is used to align samples with the category prediction probability of the target domain using a category-level adversarial discriminator.

[0009] In an optional embodiment of the present application, the above-mentioned balanced enhancement module is used to use the data of the current batch of the target domain to obtain the category prediction distribution of each sample through forward propagation, and calculate the average probability of each category in the target domain; select the 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 small batch according to the category probability of the target domain.

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

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

[0012] In an optional embodiment of the present application, the training process of the above-mentioned selective adversarial enhancement network includes: a balanced enhancement stage, a selective adversarial stage and an uncertainty suppression stage; the training process of the balanced enhancement stage includes: based on the source domain data, taking part of the source domain samples for enhancement, supplementing the category distribution of the target domain, and ensuring that the target domain has a more balanced category distribution in the early stage of training; the training process of the selective adversarial stage includes: after balanced enhancement, using the selective adversarial module to perform category-level adversarial training, and by judging the degree of category matching between the source domain and target domain samples, further reducing the interference of irrelevant categories in the source domain on the target domain training; the training process of the uncertainty suppression stage includes: through the uncertainty suppression module, weighted suppression of uncertain categories in the target domain, thereby improving the stability and accuracy of the target domain classifier; optimizing the models in the balanced enhancement stage, the selective adversarial stage and the uncertainty suppression stage respectively, reducing the negative transfer effect of non-shared categories to enhance the cross-domain adaptability of the target domain.

[0013] In a second aspect, an embodiment of the present invention further provides a selective countermeasure enhancement system for rolling bearing fault diagnosis, comprising: the above-mentioned selective countermeasure enhancement network for rolling bearing fault diagnosis.

[0014] The embodiments of the present invention bring the following beneficial effects: An embodiment of the present invention provides a selective adversarial enhancement network and system for rolling bearing fault diagnosis. The balance enhancement module effectively alleviates the negative migration caused by category asymmetry by dynamically adjusting the category weight distribution of the source domain and the target domain; the selective adversarial module accurately screens shared categories and eliminates abnormal category interference during adversarial training to achieve 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 uncertainty propagation of samples near the classification boundary, thereby improving the recognition accuracy and model robustness of samples near the classification boundary.

[0015] Other features and advantages of the present disclosure will be set forth in the following description, or some features and advantages may be inferred or unambiguously determined from the description, or may be learned by practicing the above-mentioned technology of the present disclosure.

[0016] In order to make the above-mentioned objectives, features and advantages of the present disclosure more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic diagram of the structure of a selective adversarial enhancement network for rolling bearing fault diagnosis provided by an embodiment of the present invention; Figure 2 A schematic diagram of a GRCS feature extraction module provided in an embodiment of the present invention; Figure 3 A schematic diagram of uncertainty suppression and decision boundary optimization provided by an embodiment of the present invention; Figure 4 A schematic structural diagram of a selective countermeasure enhancement system for rolling bearing fault diagnosis provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] In recent years, with the rapid growth of manufacturing demand, industrial equipment is developing towards greater automation, sophistication, and intelligence. This has placed higher demands on fault diagnosis and health monitoring of industrial systems and their key components. Rotating machinery, as a vital component of industrial systems, often operates in non-steady states and harsh working environments. Its failure modes are complex and difficult to predict, easily leading to the occurrence of multiple faults. These failures not only seriously affect production efficiency and product quality, but can also cause large-scale production accidents and even irreversible losses. Therefore, achieving efficient and accurate rotating machinery fault diagnosis is not only crucial for ensuring production safety, but also effectively reduces downtime, improves economic efficiency, and provides strong technical support for intelligent manufacturing.

[0021] 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 excellent feature extraction and automatic learning capabilities. Compared with traditional machine learning methods, deep learning can automatically extract fault features from massive sensor data, reducing reliance on expert experience, thereby significantly improving the efficiency and accuracy of fault diagnosis. In particular, deep learning has been widely used in rotating machinery fault diagnosis, capable of processing large-scale, multi-dimensional, and complex time series data and quickly identifying fault modes. However, despite the significant progress made in the field of fault diagnosis, deep learning technology still faces the following two major challenges: (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 types change frequently, the cost of collecting labeled data is extremely high, which seriously restricts the practical application of the model.

[0022] (2) Insufficient cross-domain generalization capability: 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 distribution of training and test data is consistent, an assumption that is difficult to adapt to the complex and changing actual industrial environment. Therefore, improving the model's generalization capability under cross-domain conditions is one of the key issues that need to be urgently addressed in current intelligent fault diagnosis.

[0023] To address these issues, transfer learning (TL) provides an effective technical approach. Transfer learning aims to improve the generalization of the model by leveraging knowledge learned in the source domain (labeled data) to improve the transfer effect in the target domain (unlabeled data). Domain adaptation methods in transfer learning can effectively address the problem of data distribution differences and improve the diagnostic performance of the target task through cross-domain knowledge transfer. In recent years, an increasing number of researchers have begun to focus on applying domain adaptation methods to the field of fault diagnosis of industrial equipment, especially in cross-operating and cross-domain scenarios.

[0024] However, traditional domain adaptation methods often assume that the source and target domains have the same category space, which is not easy to achieve in many practical applications. For example, in rotating machinery fault diagnosis, due to different operating conditions (such as speed and load changes), the label space of the target domain is often only a subset of the source domain label space. That is, the target domain may lack samples of certain fault categories in 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) will interfere with network learning, thereby affecting the performance of the target task.

[0025] To address the problem of partial domain adaptation, researchers have proposed a variety of methods to screen and filter irrelevant categories in the source domain to reduce the interference of external class samples on model training. The main strategies for partial domain adaptation include sample weighting and category selection, subdomain alignment and local adversarial training, adversarial training and domain discriminators, dual-domain alignment, and collaborative feature consistency and pseudo-label generation to achieve a balanced target domain distribution. These methods identify and remove external class samples in the source domain, thereby enhancing the alignment of shared class samples and improving model performance in the target domain. Although these methods have alleviated the negative transfer problem to some extent, most of them do not fully tap the potential of unlabeled data in the target domain and lack the fine-grained screening of external class samples.

[0026] Based on this, an embodiment of the present invention provides a selective adversarial augmentation network and system for rolling bearing fault diagnosis, and specifically proposes a rolling bearing fault diagnosis method based on a selective adversarial augmentation network (SAAN). First, by dynamically adjusting the category distribution of the source domain and the target domain to achieve a balance, the category asymmetry problem is solved and the negative transfer effect is reduced; secondly, by designing a selective domain adversarial module, abnormal categories are screened and shared categories are accurately aligned, thereby alleviating the distribution mismatch problem caused by outlier categories, and at the same time, subdomain-level distribution alignment is achieved to capture the subtle differences in rolling bearing fault characteristics under different operating conditions; finally, an uncertainty suppression module is proposed to optimize the loss function, effectively suppress the uncertainty propagation of the wrong category, and enhance the alignment effect of samples near the classification boundary, so as to improve the classification performance and migration effect of the target domain. The superiority of this method was verified by conducting migration diagnosis experiments on two rolling bearing datasets.

[0027] To facilitate understanding of this embodiment, a selective adversarial enhancement network for rolling bearing fault diagnosis disclosed in an embodiment of the present invention is first introduced in detail.

[0028] Example 1: An embodiment of the present invention provides a selective adversarial enhancement network for rolling bearing fault diagnosis. The framework of the selective adversarial enhancement network integrates a balance enhancement module and a selective adversarial module to achieve fine-grained subdomain alignment between the source domain and the target domain.

[0029] See also Figure 1 The schematic diagram of the structure of a selective adversarial enhancement network for rolling bearing fault diagnosis is shown in FIG. Figure 1 As shown in (a), the selective adversarial enhancement network includes: a balanced enhancement module, a selective adversarial module and an uncertainty suppression module; the balanced enhancement module is used to dynamically adjust the category distribution of the source domain and the target domain to balance the category distribution of the source domain and the target domain; the selective adversarial module is used to screen abnormal categories and accurately align shared categories to achieve distribution alignment at the subdomain level; the uncertainty suppression module is used to suppress the uncertainty propagation caused by misclassification by optimizing the loss function, thereby improving the alignment quality of samples close to the decision boundary.

[0030] This example proposes a new partial domain adaptation method, the Selective Adversarial Augmentation Network (SAAN). SAAN combines a balance enhancement module, a selective adversarial module, and an uncertainty suppression module to address the shortcomings of existing methods in partial domain adaptation. Specifically, it includes: (1) The SAAN framework is proposed: it integrates 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 transfer problem in partial domain adaptation.

[0031] (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 diagnosis results are enhanced.

[0032] (3) Verification of transfer learning and generalization performance: We conducted a series of cross-domain diagnostic tasks in two rolling bearing experiments to verify the effectiveness of SAAN. In addition, we compared it with several other advanced TLFD methods, and the results demonstrated the superiority of SAAN.

[0033] An embodiment of the present invention provides a selective adversarial enhancement network for rolling bearing fault diagnosis. The balance enhancement module effectively alleviates the negative migration caused by category asymmetry by dynamically adjusting the category weight distribution of the source domain and the target domain; the selective adversarial module accurately screens shared categories and eliminates abnormal category interference during adversarial training to achieve 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 uncertainty propagation of samples near the classification boundary, thereby improving the recognition accuracy and model robustness of samples near the classification boundary.

[0034] Example 2: The embodiment of the present invention provides another selective adversarial enhancement network for rolling bearing fault diagnosis, which is implemented on the basis of the above embodiment. First, some concepts involved in this embodiment are introduced.

[0035] 1. Partial domain adaptation.

[0036] Domain Adaptation (DA) enables the model to learn transferable knowledge from source domain labeled data by reducing the difference between the source domain and target domain data distribution. However, traditional domain adaptation methods usually assume that the source domain and target domain have the same label space (i.e. ), this assumption does not hold true in practical applications. Partial Domain Adaptation (PDA) assumes that the label space of the target domain is the source domain label space A subset of ), but even in the shared label space There is still a significant difference between the source and target domains (i.e. ).

[0037] In addition, there are category samples in the source domain that do not match the target domain. These samples may cause negative transfer problems, further increasing the complexity of cross-domain diagnosis.

[0038] In rolling bearing fault diagnosis, the actual operating environment of the target domain usually contains only some categories, which is not completely consistent with the category distribution of the source domain, which introduces the research problem of PDA. and Represent the source domain and target domain respectively, where the source domain contains labeled samples, the target domain contains unlabeled samples. The data of the source domain and the target domain are obtained from different distributions. and Extracted from the target domain, and the label space of the target domain is the source domain label space A subset of ). In addition, even in the shared label space There is still a significant difference between the source and target domains (i.e. ). Partial domain adaptation focuses on adapting only the categories shared by the source and target domains, avoiding the negative impact of mismatched categories in the source domain.

[0039] 2. Domain adversarial learning.

[0040] The core issue in transfer intelligent diagnosis is mapping the source and target domain data of bearings into a common feature space and aligning their distributions within that space. Domain adversarial learning is one of the main approaches for feature distribution alignment, and a representative example is the domain adversarial neural network (DANN). This network draws on the principles of generative adversarial learning to design a domain discriminator. Using features as input, it uses adversarial training to align the distributions of source and target domain data in the feature space. This prevents the domain discriminator from correctly identifying features as belonging to the source or target domain, ultimately eliminating data distribution discrepancies.

[0041] The domain adversarial network mainly consists of three parts: feature extractor , domain identifier and fault classifier Among them, the domain discriminator is mainly used to distinguish whether the data features belong to the source domain or the target domain. The feature extractor is fine-tuned according to the error of the domain discriminator to achieve the purpose of confusing the domain discriminator, and finally realize the transfer diagnosis. The loss function of the domain adversarial network is as follows: (1) Where: represents the mth sample in the union of source domain and target domain samples; n represents the total number of source domain and target domain samples; Indicates the domain label. When it is 1, it indicates the source domain, and when it is 0, it indicates the target domain. 、 and They are 、 and Optimization parameters of and represents the classifier loss function and the domain discriminator loss function; The weight parameters of the two loss functions are measured. The domain adversarial network is trained through the maximum and minimum optimization process in formula (2): (2) Through training, the feature extractor can eliminate the differences in data distribution and extract domain-invariant features in the source and target domains. The classifier trained using source domain data can be directly used to classify target domain data, thereby migrating source domain diagnostic knowledge to the target domain.

[0042] 3. ResNet50.

[0043] ResNet (Residual Network) is a deep convolutional neural network that effectively addresses the vanishing or exploding gradient problems common in traditional deep network training by introducing a cross-layer connection mechanism. This mechanism allows some input data to bypass the convolutional layers and pass directly to the output, preserving the original data characteristics. Furthermore, residual learning, by adding the input and output and combining them 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 difficulty of training deep networks.

[0044] Thanks to this innovation, ResNet quickly became a classic architecture in deep learning and has been widely used in fields such as image recognition, speech processing, and fault diagnosis. Within this framework, various versions of the ResNet model have been proposed, with ResNet50, a deeper variant, widely adopted for its superior performance. ResNet50 consists of a 50-layer neural network, significantly improving feature extraction capabilities through a deeper network structure while maintaining a low number of parameters and computational complexity. Its architecture consists of two core modules: ConvBlock and IdentityBlock, which are used to change the network dimension and deepen the network 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 input dimension unchanged. In the field of bearing fault diagnosis, ResNet50, with its deep convolutional neural network structure, is able to effectively capture subtle features in bearing operating data and extract key characteristics at multiple scales and dimensions, demonstrating its strong adaptability to data diversity and complexity. By deeply exploring the potential features in the operating status, 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.

[0045] For the rolling bearing fault diagnosis task in the partial domain adaptation scenario, this embodiment proposes an improved model based on DANN - Selective Adversarial Augmentation Network (SAAN). This model achieves accurate alignment of shared category features and effective alignment of subdomain distributions through balanced enhancement and selective domain adversarial strategies. It also dynamically adjusts the category distribution to mitigate the impact of negative transfer and optimizes the loss function to suppress uncertainty propagation, thereby improving the diagnostic performance of the model in the target domain. The detailed framework of the SAAN model can be found in Figure 1 (b).

[0046] In some embodiments, the selective adversarial enhancement network also 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 through shared parameters; the balanced enhancement module is used to randomly select and enhance source domain samples to increase the diversity of target domain samples; the selective adversarial module is used to perform selective adversarial training on the enhanced features of the source domain and the target domain, and perform selective adversarial learning based on the strategy of category and subdomain optimization; the uncertainty suppression module is used to suppress the propagation of uncertainty by dynamically adjusting the loss function; the classifier is used to classify fault features.

[0047] First, in the data collection and preprocessing stage, the source domain and target domain data are obtained by collecting one-dimensional vibration signals. Subsequently, the data is input into the feature extraction module (GRCS), which extracts domain-invariant features of the source domain and target domain by sharing parameters. and . In order to effectively reduce negative transfer, SAAN introduces a balanced enhancement selective adversarial module, in which source domain samples are randomly selected and enhanced to increase the diversity of target domain samples, thereby improving the performance of the target domain. Then, the enhanced features of the source and target domains are trained through selective adversarial training, and a strategy based on category and subdomain optimization is adopted to perform selective adversarial learning to avoid the situation where certain categories in the source domain have no corresponding categories in the target domain, thereby reducing the negative transfer effect of irrelevant categories on the target domain. In order to further improve the performance of the model, SAAN designs an adaptive uncertainty suppression module to suppress the propagation of uncertainty by dynamically adjusting the loss function to ensure the stability and robustness of the model across multiple domains. Finally, the fault features are classified by the classifier, and the accuracy of fault diagnosis is improved by optimizing the loss function.

[0048] 1. GRCS feature extraction module In some embodiments, 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 by means of the Gram angle difference field (GADF); the feature extraction module uses the 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.

[0049] This example proposes a novel feature extraction module, GRCS (GADF-ResNet50-CBAM-SimAM), designed to effectively extract features and enhance the performance of the target domain in cross-domain tasks. This module combines multiple advanced network architectures and attention mechanisms to improve feature extraction accuracy and generalization capabilities.

[0050] First, the Grammian Angular Difference Field (GADF) method was introduced to convert one-dimensional vibration signals in the source and target domains into two-dimensional images more suitable for feature extraction. This method captures more detailed spatial and angular information in the image, providing a richer feature representation for subsequent feature extraction. This method enriches the feature representation by enhancing the geometric relationships in the image. Subsequently, ResNet50 was adopted as the underlying architecture for the feature extraction network. ResNet50 not only performs well in image classification, but also, due to its residual connection design, effectively avoids the vanishing gradient problem commonly encountered when training deep networks. In the GRCS module, ResNet50 is used as the core feature extractor to extract deep features related to the bearing fault status. Furthermore, the Convolutional Block Attention Module (CBAM) and the Simple Attention Module (SimAM) are incorporated into the feature extraction process to further enhance the feature representation capability. Specifically, CBAM effectively extracts key image region features through channel-wise and spatial-attention mechanisms while suppressing the influence of irrelevant regions. SimAM, through its efficient spatial self-attention mechanism, helps better capture common patterns between the source and target domains, preventing irrelevant source domain samples from interfering with target domain training. The combination of these two modules optimizes the feature extraction process at multiple levels, further enhancing cross-domain adaptability.

[0051] In some embodiments, the feature extraction module performs feature extraction, including: inputting the two-dimensional image features converted by GADF into the ResNet50 network; and gradually optimizing the extracted features through the convolution block attention module CBAM and the simple attention module SimAM module.

[0052] The overall architecture of the GRCS module can be found in Figure 2Figure 1 shows a schematic diagram of a GRCS feature extraction module. The main steps of feature extraction include: first, inputting GADF-enriched features into the ResNet50 network; then, gradually optimizing the extracted features through the CBAM and SimAM modules, ultimately obtaining the feature representation extracted by the GRCS module. This process effectively improves the target domain diagnostic capabilities 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.

[0053] 2. Balance Enhancement Selective Countermeasures Module (i.e. the general term for the Balance Enhancement Module and the Selective Countermeasures Module).

[0054] In some embodiments, the balance enhancement module is used to enrich the category distribution of the target domain in the early stage of training to keep the category distribution of the target domain balanced with that of the source domain; the selective adversarial module is used to align samples with the category prediction probability of the target domain using a category-level adversarial discriminator.

[0055] In the PDA task, the label spaces of the source domain and the target domain usually only partially overlap, that is, the target domain labels are only a subset of the source domain labels. When the target domain is small and the category distribution is uneven, if the overall distribution is directly aligned, it is easy to mistakenly match the target domain data to the source domain category that does not exist in its label space, resulting in negative transfer. Traditional practices often use the "source domain filtering" strategy to eliminate irrelevant samples, but when the source domain is much larger than the target domain, excessive filtering will discard potentially useful information. To resolve this contradiction, this embodiment proposes a balanced augmentation selective adversarial module (BASAM). First, the target domain is moderately "supplemented" in the early stage of training to make it richer in category distribution and maintain a good balance with the source domain; then, combined with the selective adversarial module, the category-level adversarial discriminator is used to align the samples with the category prediction probability of the target domain, thereby suppressing the interference of non-shared categories on the adversarial process. With the help of this module, the target domain can not only more fully explore the beneficial samples that can be transferred from the source domain, but also automatically weaken those "abnormal classes" with extremely 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.

[0056] (1) Balanced Alignment Module

[0057] In some embodiments, the balanced enhancement module is used to use the data of the current batch of the target domain to obtain the category prediction distribution of each sample through forward propagation, and calculate the average probability of each category 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 category probability of the target domain.

[0058] In order to achieve balanced enhancement, this embodiment first uses the data of the current batch of the target domain to obtain the category prediction distribution of each sample through forward propagation, and calculates the average probability of each category in the target domain. Suppose the current target domain mini-batch is , through the classifier Get the predicted distribution , then define each category The average predicted probability is: (3) in, Representation sample Predicted as category Since the target domain labels only cover the shared categories , non-shared category Usually lower.

[0059] Based on this, this embodiment designs a balanced enhancement module: In , an enhancement weight is assigned to each sample according to the category probability of the target domain. Specifically, for the source domain sample (in ), let its enhanced weight be defined as: (4) in, is the initial enhancement ratio, which gradually decreases as the training progresses; Represents source domain samples Corresponding category The average predicted probability over the target domain mini-batch; Indicates the confidence of the sample in the source domain (which can be expressed by the maximum probability predicted by the classifier or other indicators); It is an adjustment function used to control the upper bound of the enhancement weight, so that samples with higher prediction probabilities (i.e., shared categories) in the target domain are given greater weights, while samples of non-shared categories are naturally suppressed.

[0060] Then, source domain samples that meet a certain weight threshold are selected to form an enhanced target domain batch ,Right now: (5) in, is a preset threshold. In this way, only those samples that share categories with the target domain and have high confidence in the source domain can be "supplemented" to the target domain, thereby effectively balancing the category distribution of the target domain batch.

[0061] (2) Selective Adversarial Module

[0062] In some embodiments, the selective adversarial module is used to configure an independent domain discriminator for each category in the source domain, and perform category-level adversarial training using the aforementioned enhanced target domain batches.

[0063] Although the balanced enhancement module enriches and balances the category distribution of the target domain to a certain extent, it still cannot completely eliminate the interference of irrelevant categories in the source domain on the alignment process, especially when there are "abnormal categories" in the source domain that do not appear in the target domain, it may still cause negative transfer. To this end, this embodiment further proposes a selective adversarial module to more finely suppress the impact of these non-shared categories on cross-domain alignment, thereby further improving the overall adaptation performance. Unlike the traditional single discriminator method, this module is used for each category in the source domain. Configure an independent domain discriminator , and use the aforementioned enhanced target domain batch Perform category-level adversarial training.

[0064] Specifically, for each sample in the batch , which is predicted to be the category The probability of In category Domain discriminator In the example, the adversarial loss can be expressed as: (6) in is the feature extraction function, is the loss function with gradient reversal, represents the domain label. Then, this embodiment classifies the category according to the target domain. The average probability of Weighting the discriminator loss, the overall selective adversarial loss is defined as: (7) Therefore, for shared categories with a high probability of appearing in the target domain, their feature representations are subject to stronger adversarial correction, while for non-shared categories with a low probability of appearing in the target domain, the adversarial correction strength is naturally reduced, effectively alleviating the negative transfer problem. Compared to traditional single-discriminator methods, the multi-discriminator architecture can achieve more refined source domain sample screening at the category level, eliminating noisy data that does not match the target domain distribution, thereby significantly improving the accuracy and robustness of cross-domain adaptation.

[0065] 3. Adaptive Uncertainty Suppression Module (AUSM)

[0066] In some embodiments, the uncertainty suppression module is configured to dynamically weight and suppress the negative impact of uncertain categories on the target domain classifier by calculating the uncertainty of source domain samples.

[0067] In the PDA task, the difference in categories between the source and target domains inevitably introduces classification uncertainty. This is especially true when the target domain has fewer categories, 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 them to suppress the negative impact of uncertain categories on the target domain classifier, thereby reducing the interference of irrelevant category samples on target domain learning.

[0068] See also Figure 3 A schematic diagram of uncertainty suppression and decision boundary optimization is shown in Figure 3 The impact of uncertainty propagation is demonstrated, where the boundaries between categories 1 and 2 in the source domain are close, leading to misclassification in the target domain prediction process, thereby reducing the classification accuracy of the target domain. To further reduce this effect, AUSM uses a weighted complement entropy loss to reduce the negative impact of uncertain categories in the source domain and weights source domain samples according to the 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: (8) in, It is the weighting factor of the source domain sample, which is dynamically adjusted based on the importance of the source domain sample in the target domain; is the cross entropy loss of the source domain samples, which is used to measure the classification accuracy of the source domain samples; and is a balance term, which controls the weights of the supplementary entropy loss term and the uncertainty penalty term respectively; 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.

[0069] AUSM is designed to dynamically adjust the impact of source domain samples' uncertainty on target domain training. In the target domain, lower-uncertainty source domain samples are given higher weights, while higher-uncertainty samples are suppressed, reducing their interference with the target domain learning process. In this way, AUSM effectively suppresses the negative impact of irrelevant source domain classes on the target domain, thereby improving target domain classification performance.

[0070] 4. Algorithms and training strategies.

[0071] In some embodiments, the training process of the selective adversarial enhancement network includes: a balanced enhancement stage, a selective adversarial stage, and an uncertainty suppression stage; the training process of the balanced enhancement stage includes: based on the source domain data, taking part of the source domain samples for enhancement to supplement the category distribution of the target domain, ensuring that the target domain has a more balanced category distribution in the early stage of training; the training process of the selective adversarial stage includes: after balanced enhancement, using the selective adversarial module to perform category-level adversarial training, and by determining the degree of category matching between the source domain and target domain samples, further reducing the interference of irrelevant categories in the source domain on the target domain training; the training process of the uncertainty suppression stage includes: through the uncertainty suppression module, weighted suppression of uncertain categories in the target domain, thereby improving the stability and accuracy of the target domain classifier.

[0072] In some embodiments, the models are optimized in the balance enhancement stage, the selective confrontation stage, and the uncertainty suppression stage respectively to reduce the negative transfer effect of non-shared categories to enhance the cross-domain adaptation capability of the target domain.

[0073] The SAAN model combines balanced augmentation, selective adversarial, and adaptive uncertainty suppression (AUSM) modules to handle the label space differences between the source and target domains, thereby improving the performance of the target domain in cross-domain fault diagnosis tasks.

[0074] The training process of SAAN is divided into three main stages: 1) Balanced enhancement stage: Based on the source domain data, some source domain samples are taken for enhancement to supplement the category distribution of the target domain, ensuring that the target domain has a more balanced category distribution in the early stages of training; 2) Selective adversarial stage: After balanced enhancement, the selective adversarial module is used to perform category-level adversarial training. By determining the degree of category matching between source and target domain samples, the interference of irrelevant categories in the source domain on the target domain training is further reduced; 3) Uncertainty suppression stage: Through the adaptive uncertainty suppression module, uncertain categories in the target domain are weightedly 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 categories is reduced to enhance the cross-domain adaptability of the target domain.

[0075] Example 3: An embodiment of the present invention provides a simulation experimental analysis of a selective adversarial enhancement network for rolling bearing fault diagnosis, which is implemented on the basis of the above embodiment.

[0076] To verify the effectiveness of the rolling bearing fault diagnosis model based on SAAN, 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 uses the variable speed operating condition data contained in the CWRU dataset to systematically analyze the fault characteristics of the bearing under different speed conditions; while in Case 2, the Mechanical Fault Simulation (MFS) experimental platform of the Rotating Machinery Intelligent Operation and Maintenance Laboratory is used to collect bearing data to more realistically reproduce the complexity of the bearing operating environment under actual working conditions. This section details the data preprocessing process and provides a detailed case study. All experiments are conducted on the deep learning framework Torch.

[0077] In Case 1, the SAAN method provided in this embodiment maintained a high classification accuracy in all tasks and scenarios, reaching or approaching 100% many times, and continued to operate stably as the missing ratio increased, fully demonstrating its excellent robustness and generalization ability in the partial domain adaptation environment where the target domain label space is only a subset of the source domain, and can effectively suppress negative transfer and significantly improve the classification performance of rolling bearing fault diagnosis.

[0078] In Case 2, the SAAN method provided in this example achieved nearly 100% classification accuracy with minimal fluctuation across all tasks and missing category settings, demonstrating its high adaptability and robustness even when the target domain label space is only a subset of the source domain. In contrast, the accuracy of methods such as Baseline, M1, M2, SAN, SPDA, DANN, and CDAN generally decreased after removing key failure modes, and was particularly unstable under conditions of multiple missing categories (e.g., missing 2, 3, 4, 5, 6). This further demonstrates SAAN's significant advantages in suppressing negative transfer, accurately aligning shared categories, and optimizing classification boundaries.

[0079] In mechanical fault diagnosis, the target domain label space is only a subset of the source domain. Existing methods ignore the fine-grained distribution differences between subdomains and the weight imbalance of uncertain samples near the classification boundary, which significantly degrades adaptation performance. This embodiment proposes a rolling bearing fault diagnosis model based on SAAN. SAAN abandons the reliance on complete label space overlap and is designed specifically for partial domain adaptation scenarios. Specifically, the balance enhancement module dynamically adjusts the class weight distribution between the source and target domains to effectively mitigate negative transfer caused by class asymmetry. The selective adversarial module accurately selects shared categories and removes interference from anomalous categories 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 recognition accuracy and model robustness. Experimental results demonstrate that SAAN effectively aligns source and target domain features under PDA conditions. On the diagnostic task of Dataset 1, SAAN achieves an average accuracy of 92.15%, significantly improving over existing methods.

[0080] Example 4: The embodiment of the present invention provides a selective anti-corrosion enhancement system for rolling bearing fault diagnosis, which is implemented on the basis of the above embodiment. Figure 4 The structure diagram of a selective countermeasure enhancement system for rolling bearing fault diagnosis is shown, and the selective countermeasure enhancement system includes: the selective countermeasure enhancement network for rolling bearing fault diagnosis provided by the above embodiment.

[0081] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working process of the selective countermeasure enhancement system for rolling bearing fault diagnosis described above can refer to the corresponding process in the aforementioned embodiment and will not be repeated here.

[0082] In addition, in the description of the embodiments of the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0083] 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 the present invention, or the portion that contributes to the prior art, or a portion 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 for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0084] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present 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.

[0085] Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. 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 above embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A selective adversarial enhancement network for rolling bearing fault diagnosis, 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 confrontation enhancement network includes: a balance enhancement module, a selective confrontation module and an uncertainty suppression module; The balance enhancement module is used to dynamically adjust the 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 to filter out abnormal categories and accurately align shared categories to achieve distribution alignment at the subdomain level; The uncertainty suppression module is used to suppress the uncertainty propagation caused by misclassification by optimizing the loss function, thereby improving the alignment quality of samples close to the decision boundary.

2. The selective adversarial enhancement network for rolling bearing fault diagnosis according to claim 1, characterized in that: The selective adversarial enhancement network also 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 through shared parameters; The balance enhancement module is used to randomly select and enhance the source domain samples to increase the diversity of the target domain samples; The selective adversarial module is used to perform selective adversarial training on the enhanced features of the source domain and the target domain, and perform selective adversarial learning based on a category and subdomain optimization strategy; The uncertainty suppression module is used to suppress uncertainty propagation by dynamically adjusting the loss function; The classifier is used to perform fault feature classification.

3. The selective adversarial enhancement network for rolling bearing fault diagnosis according to claim 2, characterized in that: 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 by using the Gram angle difference field (GADF); The feature extraction module uses the 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.

4. The selective adversarial enhancement network for rolling bearing fault diagnosis according to claim 3, characterized in that: The feature extraction module performs feature extraction, including: The two-dimensional image features converted by GADF are input into the ResNet50 network; The extracted features are gradually optimized through the convolutional block attention module CBAM and the simple attention module SimAM modules.

5. The selective adversarial enhancement network for rolling bearing fault diagnosis according to claim 2, characterized in that: The balance enhancement module is used to enrich the category distribution of the target domain in the early stage of training so that the category distribution of the target domain and the source domain are balanced; The selective adversarial module is used to align samples using a category-level adversarial discriminator with the category prediction probability of the target domain.

6. The selective adversarial enhancement network for rolling bearing fault diagnosis according to claim 5, characterized in that: The balancing enhancement module is used to use the data of the current batch of the target domain to obtain the category prediction distribution of each sample through forward propagation, and calculate the average probability of each category in the target domain; select the 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 category probability of the target domain.

7. The selective adversarial enhancement network for rolling bearing fault diagnosis according to claim 5, characterized in that: The selective adversarial module is used to configure an independent domain discriminator for each category in the source domain, and perform category-level adversarial training using the aforementioned enhanced target domain batches.

8. The selective adversarial enhancement network for rolling bearing fault diagnosis according to claim 2, characterized in that: The uncertainty suppression module is used to dynamically weight and suppress the negative impact of uncertain categories on the target domain classifier by calculating the uncertainty of source domain samples.

9. The selective adversarial enhancement network for rolling bearing fault diagnosis according to any one of claims 1 to 8, characterized in that: The training process of the selective adversarial enhancement network includes: a balance enhancement stage, a selective adversarial stage and an uncertainty suppression stage; The training process of the balance enhancement phase includes: based on the source domain data, taking some source domain samples for enhancement to supplement the category distribution of the target domain, ensuring that the target domain has a more balanced category distribution in the early stage of training; The training process of the selective adversarial stage includes: after the balance enhancement, using the selective adversarial module to perform category-level adversarial training, by determining the category matching degree of the source domain and the target domain samples, further reducing the interference of irrelevant categories in the source domain on the target domain training; The training process of the uncertainty suppression stage includes: performing weighted suppression on uncertain categories in the target domain through the uncertainty suppression module, thereby improving the stability and accuracy of the target domain classifier; Models are optimized in the balance enhancement stage, the selective confrontation stage, and the uncertainty suppression stage respectively to reduce the negative transfer effect of non-shared categories and enhance the cross-domain adaptability of the target domain.

10. A selective countermeasure enhancement system for rolling bearing fault diagnosis, characterized in that: include: The selective adversarial enhancement network for rolling bearing fault diagnosis according to any one of claims 1 to 9.

Images