Rolling bearing fault diagnosis method and system facing unbalanced semi-supervised domain generalization

Abstract

The application provides a rolling bearing fault diagnosis method and system for unbalanced semi-supervised domain generalization, and the unbalanced semi-supervised domain generalization network comprises a fault prior guided temporal enhancement module, a domain decoupled representation learning module and a frequency-aware hybrid expert network; the fault prior guided temporal enhancement module improves the diversity and distribution consistency of the minority class samples in combination with the physical mechanism of the fault mode; the domain decoupled representation learning module decouples the features into domain-invariant features and domain-specific features by constructing an auxiliary domain; the frequency-aware hybrid expert network realizes adaptive modeling of cross-domain features by using frequency domain decomposition and a dynamic gating mechanism; the unbalanced semi-supervised domain generalization network adopts a pseudo-label strategy combined with a confidence constraint and utilizes unlabeled data for joint training. The complex problems of unknown target domain, serious class imbalance and scarcity of labeled data in an industrial scene can be effectively solved, and the accuracy and robustness of rolling bearing fault diagnosis are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, and in particular to a method and system for diagnosing rolling bearing faults generalized to unbalanced semi-supervised domains. Background Technology

[0002] The booming development of intelligent manufacturing has made Prognostics and Health Management (PHM) of rotating machinery play a central role in ensuring equipment reliability, improving production efficiency, and operational safety. As a key component of modern industrial systems, rotating machinery, such as rolling bearings, increasingly demands early fault diagnosis and predictive maintenance to effectively avoid unplanned downtime, extend equipment lifespan, reduce maintenance costs, and ensure production continuity. Given these urgent needs, researchers have adopted Deep Learning-based Fault Diagnosis (DLFD) methods to achieve more accurate fault classification. While DLFD methods have achieved significant results under fixed operating conditions, rotating machinery typically operates under highly dynamic and complex conditions, such as continuous changes in load, speed, temperature, and humidity, leading to a significant shift in the distribution of monitoring data, known as "domain shift." This distributional variability severely challenges the accuracy and reliability of fault diagnosis models, causing a sharp decline in the performance of DLFD methods when faced with domain shift. What is particularly serious is that existing fault diagnosis methods generally face the "cold start" problem, that is, when the target domain lacks labeled data or the target domain is unknown, the model is difficult to initialize effectively, resulting in low diagnostic accuracy and failing to meet the dynamic adaptation needs of actual industrial scenarios. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a rolling bearing fault diagnosis method and system for unbalanced semi-supervised domain generalization, in order to propose an unbalanced semi-supervised domain generalization network of Fault-PriorGuided Temporal Augmentation-Domain Decoupling Representation Learning (FPGTA-DDRL) for rolling bearing fault classification.

[0004] In a first aspect, embodiments of the present invention provide an imbalanced semi-supervised domain generalization network, comprising: a fault prior-guided temporal enhancement module, a domain decoupling representation learning module, and a frequency-aware hybrid expert network; the fault prior-guided temporal enhancement module is used to improve the diversity and distribution consistency of minority class samples by combining the physical mechanism of fault modes; the domain decoupling representation learning module is used to decouple features into domain-invariant features and domain-specific features by constructing an auxiliary domain; the frequency-aware hybrid expert network is used to achieve adaptive modeling of cross-domain features by utilizing frequency domain decomposition and dynamic gating mechanisms; the imbalanced semi-supervised domain generalization network is also used to perform joint training using unlabeled data by combining a pseudo-label strategy with confidence constraints.

[0005] In optional embodiments of this application, the aforementioned unbalanced semi-supervised domain generalization network is further used to decouple domain invariance and domain-specific features through a dual-branch collaborative mechanism during the unknown working condition generalization stage, and to improve the discriminativeness and generalization of feature representation by relying on a frequency-aware hybrid expert network, so as to achieve adaptive recognition and identification of rolling bearing fault types under unknown working conditions.

[0006] In optional embodiments of this application, the aforementioned fault prior guidance timing enhancement module is used to construct an enhancement function by explicitly maintaining the time structure of the signal in conjunction with the bearing fault mode; the fault prior guidance timing enhancement module is also used to achieve structured resampling by combining KMeans clustering and SMOTE oversampling.

[0007] In optional embodiments of this application, the aforementioned fault-prior-guided timing enhancement module is used to segment the original signal through a sliding window based on a fault-prior timing enhancement strategy, and to set corresponding timing enhancement strategies for the physical mechanisms and signal characteristics of different fault modes; the fault-prior-guided timing enhancement module is also used to generate synthetic samples by combining clustering and interpolation on minority class samples after enhancement based on a structured oversampling method that combines KMeans clustering and SMOTE oversampling.

[0008] In optional embodiments of this application, the aforementioned domain decoupling representation learning module is used to aggregate the representations of the source domain into a unified auxiliary domain, and to achieve systematic decoupling of domain-invariant features and domain-specific features by jointly optimizing the alignment loss of domain-invariant features and the decoupling loss of domain-specific features; wherein, the source domain includes: labeled source domain and unlabeled source domain; the domain decoupling representation learning module is used to promote knowledge transfer and cross-domain generalization ability between source domains by minimizing the average sum of squared Euclidean distances of the domain-invariant features between each source domain and the auxiliary domain; the domain decoupling representation learning module is also used to constrain the domain-specific features of each source domain and the auxiliary domain to be nearly orthogonal based on cosine similarity (i.e., maximizing the angular diversity of domain-specific features between different domains), reduce the correlation between domain-invariant features and domain-specific features, and enhance the discriminative ability and independence of the learned features.

[0009] In optional embodiments of this application, the frequency-aware hybrid expert network incorporates frequency domain decomposition, expert dynamic selection, and fusion mechanisms to achieve adaptive modeling of multi-source features.

[0010] In optional embodiments of this application, the aforementioned imbalanced semi-supervised domain generalization network is used to apply a high-confidence pseudo-label mechanism to the cross-domain joint training process through a semi-supervised optimization strategy; the imbalanced semi-supervised domain generalization network is used to perform forward inference on unlabeled samples using the current model in each training cycle to obtain the category prediction probability; for the target sample with the highest prediction probability and higher than a set threshold, the category corresponding to the target sample is selected as the pseudo-label, and the target sample and the labeled sample participate in the training process together.

[0011] In optional embodiments of this application, the aforementioned imbalanced semi-supervised domain generalization network achieves robust generalization of the rolling bearing fault diagnosis model under imbalanced and heterogeneous distribution conditions by jointly optimizing multiple loss functions; wherein, the optimization objectives of the imbalanced semi-supervised domain generalization network include: minimizing the classification loss of labeled samples, jointly optimizing the feature decoupling loss, and minimizing the pseudo-label loss of unlabeled samples by adopting a high-confidence pseudo-label mechanism.

[0012] Secondly, embodiments of the present invention also provide a rolling bearing fault diagnosis method for unbalanced semi-supervised domain generalization, applied to the aforementioned unbalanced semi-supervised domain generalization network. The unbalanced semi-supervised domain generalization network includes: a fault prior-guided temporal enhancement module, a domain decoupling representation learning module, and a frequency-aware hybrid expert network. The method includes: the fault prior-guided temporal enhancement module combining fault mode physical mechanisms to improve the diversity and distribution consistency of minority class samples; the domain decoupling representation learning module decoupling features into domain-invariant features and domain-specific features by constructing an auxiliary domain; the frequency-aware hybrid expert network using frequency domain decomposition and dynamic gating mechanisms to achieve adaptive modeling of cross-domain features; the unbalanced semi-supervised domain generalization network combining a confidence-constrained pseudo-label strategy using unlabeled data for joint training; and the rolling bearing fault diagnosis based on the trained unbalanced semi-supervised domain generalization network.

[0013] The embodiments of the present invention bring the following beneficial effects: This invention provides a rolling bearing fault diagnosis method and system for unbalanced semi-supervised domain generalization. The network includes an unbalanced semi-supervised domain generalization network framework based on FPGTA-DDRL. By collaboratively integrating a fault prior-guided temporal enhancement module, domain decoupling representation learning, and a frequency-aware mixture-of-experts (FAMoE) network, it organically combines cross-domain knowledge transfer, multi-level feature fusion, and semi-supervised learning. This enables FPGTA-DDRL to achieve efficient adaptation in dynamic industrial environments with unbalanced samples, heterogeneous distributions, and scarce labels, thus overcoming the limitations of isolated optimization in traditional methods.

[0014] 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.

[0015] 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

[0016] 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.

[0017] Figure 1 This is a schematic diagram of the structure of an unbalanced semi-supervised domain generalization network provided in an embodiment of the present invention; Figure 2 A schematic diagram of an unbalanced semi-supervised domain generalization network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a fault prior guidance timing enhancement module provided in an embodiment of the present invention; Figure 4 A schematic diagram of a frequency-domain sensing expert hybrid expert network provided in an embodiment of the present invention; Figure 5 A schematic diagram of a prior art provided for an embodiment of the present invention; Figure 6 A flowchart of a rolling bearing fault diagnosis method for unbalanced semi-supervised domain generalization provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of a rolling bearing fault diagnosis system for unbalanced semi-supervised domain generalization provided in an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] Currently, existing research on domain offset in rotating machinery fault diagnosis can be mainly summarized into methods such as Domain Adaptation Fault Diagnosis (DAFD), Domain Generalization Fault Diagnosis (DGFD), Imbalanced Domain Generalization Fault Diagnosis (IDGFD), and Semi-supervised Domain Generalization Fault Diagnosis (SDGFD). While these methods have made progress in their respective specific fields, most focus on addressing single challenges and struggle to comprehensively address the complex interplay of multiple problems in real-world industrial scenarios. Specifically, the DAFD method emphasizes aligning the data distributions of the source and target domains to mitigate offset.

[0020] However, the DAFD method is highly dependent on the availability of labeled data for the target domain. When the target domain is unknown, its generalization ability is limited, making it difficult to adequately address dynamic industrial environments. This limitation has prompted researchers to shift their focus to strategies that do not rely on labeled target domains, namely the DGFD method. This method aims to build robust models independent of specific target domain data, improving the model's cross-domain adaptability by mining domain-invariant features, thereby better addressing the problems of changing operating conditions and domain shifts in industrial scenarios.

[0021] While DGFD performs well in scenarios with abundant labeled data, it tends to overfit to a large number of classes in the source domain when faced with severe class imbalance (scarce fault samples), thus severely impairing its generalization ability to the minority of fault classes. To address the limitations of DGFD in diagnosing minority classes in imbalanced scenarios, Imbalanced Domain Generalization (IDGFD) methods have emerged. These methods focus on solving the problem of scarce fault samples and often use strategies such as resampling, weighted loss, or synthetic sample generation to balance the class distribution.

[0022] However, while IDGFD effectively mitigates class bias, its core bottleneck lies in its heavy reliance on labeled data, making it ill-suited to the prevalent problem of extremely scarce labeled data in real-world industrial settings. Therefore, to effectively utilize the massive amounts of unlabeled data from industrial environments to compensate for label scarcity, researchers have proposed the Semi-supervised Domain Generalization (SDGFD) method. SDGFD integrates limited labeled data with a large amount of unlabeled data, using mechanisms such as pseudo-labels or consistency regularization to mine latent information and improve the model's robustness.

[0023] In summary, although existing research has made progress in single dimensions such as "alignment domain offset", "utilizing unlabeled data" and "mitigating class imbalance", most of them focus on local optimization and are difficult to comprehensively address the complex challenges of "unknown target domain, severe class imbalance and scarce labels" in real industrial fault diagnosis scenarios.

[0024] Based on this, this invention provides a rolling bearing fault diagnosis method and system for unbalanced semi-supervised domain generalization. The network includes an unbalanced semi-supervised domain generalization network framework based on FPGTA-DDRL. By collaboratively integrating a fault prior-guided temporal enhancement module, domain decoupling representation learning, and a frequency-aware Mixture-of-Experts (FAMoE) network, it organically combines cross-domain knowledge transfer, multi-level feature fusion, and semi-supervised learning. This enables FPGTA-DDRL to achieve efficient adaptation in dynamic industrial environments characterized by imbalanced samples, heterogeneous distributions, and scarce labels, overcoming the limitations of isolated optimization in traditional methods. The main contributions of this embodiment include: (1) The FPGTA-DDRL framework is proposed and defined as an advanced paradigm to address the challenge of imbalanced semi-supervised domain generalization. This framework improves the diversity and distribution consistency of minority class samples by constructing a fault prior-guided temporal enhancement module; and designs a domain decoupled representation learning module to effectively decouple domain-invariant features from domain-specific features, thereby improving the discriminativeness and robustness of cross-domain knowledge transfer. (2) FAMoE is constructed. This network innovatively designs a frequency domain decomposition and dynamic gating mechanism, which can adaptively allocate expert weights and fully explore the complementarity of high and low frequency information. Combined with a high-confidence pseudo-label semi-supervised strategy, FPGTA-DDRL can efficiently utilize massive amounts of unlabeled data, thereby effectively alleviating the domain offset and cold start problems; (3) Extensive experimental verification and comparative analysis were conducted on multiple datasets. The experimental results fully demonstrate the significant superiority of FPGTA-DDRL in highly imbalanced and unknown operating conditions, providing a more robust fault diagnosis solution for demanding industrial application scenarios.

[0025] To facilitate understanding of this embodiment, a detailed description of an unbalanced semi-supervised domain generalization network disclosed in this embodiment of the invention will be provided first.

[0026] Example 1: This invention provides an imbalanced semi-supervised domain generalization network. Addressing the common problems of sample imbalance, heterogeneous distribution, and scarce labels in rolling bearing fault diagnosis under unknown operating conditions, this embodiment proposes an imbalanced semi-supervised domain generalization network that can perform intelligent diagnosis based on FPGTA-DDRL.

[0027] See Figure 1 The diagram shows the structure of an imbalanced semi-supervised domain generalization network, which includes: a fault prior-guided timing enhancement module, a domain decoupling representation learning module, and a frequency-aware hybrid expert network. The fault prior-guided temporal enhancement module is used to improve the diversity and distribution consistency of minority class samples by combining the physical mechanism of fault modes; the domain decoupling representation learning module is used to decouple features into domain-invariant features and domain-specific features by constructing an auxiliary domain; the frequency-aware hybrid expert network is used to achieve adaptive modeling of cross-domain features by utilizing frequency domain decomposition and dynamic gating mechanism; and the imbalanced semi-supervised domain generalization network is also used to perform joint training using unlabeled data by combining a pseudo-label strategy with confidence constraints.

[0028] In this embodiment, the fault prior guidance timing enhancement module can use the KMeansSMOTE algorithm (a data augmentation method combining K-Means clustering and SMOTE oversampling techniques) for structured resampling. The KMeansSMOTE method combines K-means clustering with SMOTE, selectively oversampling in a minority of dense and imbalanced clusters to preserve the local manifold and reduce noise amplification.

[0029] Specifically, first, K-means clustering is applied to the input dataset to reveal substructures. Second, clusters are filtered based on imbalance ratios, with clusters exceeding a threshold being oversampled, and samples are allocated inversely proportional to minority class density to prioritize sparse regions. Finally, SMOTE is invoked within the selected clusters to generate synthetic samples, ensuring adaptive balance.

[0030] The hybrid expert network in this embodiment can incorporate a router module to dynamically distribute input to multiple parallel subnetworks (i.e., experts) based on feature patterns, thereby enhancing the model's adaptability across different subspaces. Its overall structure mainly consists of two parts: a router and an expert library.

[0031] See Figure 2 The diagram illustrates an imbalanced semi-supervised domain generalization network, covering the entire process from data preprocessing and feature extraction to generalized diagnosis under unknown conditions. First, addressing the uneven distribution of fault categories in the data, a fault prior-guided temporal enhancement module is designed to improve data balance and the diversity and representativeness of minority class samples. Building upon this, a domain decoupling representation learning module is constructed. Through cross-domain data integration, it enriches the model's knowledge representation space and achieves efficient feature decoupling, separating features into domain-invariant and domain-specific features. Then, to handle different feature flows, FAMoE is proposed to achieve accurate modeling and efficient fusion of cross-domain features. This network effectively captures and integrates the discriminative and complementary nature of features across frequency bands through dynamic adaptive allocation of expert weights, thereby improving feature diversity and the model's cross-domain robustness. To fully utilize unlabeled samples, this embodiment employs semi-supervised optimization. A high-confidence pseudo-label mechanism is used to include unlabeled samples in joint training, alleviating the knowledge bottleneck caused by label scarcity and enhancing the model's adaptability to unknown distributions.

[0032] In some embodiments, the unbalanced semi-supervised domain generalization network is also used in the unknown working condition generalization stage to decouple domain invariance and domain-specific features through a dual-branch collaborative mechanism, and rely on a frequency-aware hybrid expert network to improve the discriminativeness and generalization of feature representation, so as to achieve adaptive recognition and identification of rolling bearing fault types under unknown working conditions.

[0033] Finally, in the generalization stage of unknown working conditions, a dual-branch collaborative mechanism is used to decouple domain invariant and domain-specific features, and FAMoE is used to improve the discriminativeness and generalization of feature representation, so as to achieve adaptive and accurate identification of rolling bearing fault types under unknown working conditions.

[0034] Example 2: This invention provides another unbalanced semi-supervised domain generalization network, implemented based on the above embodiments. The specific implementation methods of each module of the unbalanced semi-supervised domain generalization network are described in detail.

[0035] I. Fault Prior Guidance Timing Enhancement Module: In some embodiments, the fault prior-guided timing enhancement module is used to construct an enhancement function by explicitly maintaining the time structure of the signal in conjunction with the bearing fault mode; the fault prior-guided timing enhancement module is also used to achieve structured resampling by combining KMeans clustering and SMOTE oversampling.

[0036] Current mainstream fault diagnosis data augmentation methods (such as SMOTE and Mixup) mostly rely on interpolation or perturbation within the sample space, ignoring the inherent temporal dynamic structure of vibration signals and failing to fully incorporate domain prior knowledge of common rolling bearing fault modes. This leads to a significant decrease in the model's representation and discrimination capabilities for minority classes when samples are scarce or class distributions are extremely imbalanced.

[0037] To address the aforementioned shortcomings, this embodiment proposes an FPGTA module. Its core lies in explicitly preserving the signal's temporal structure and constructing a customized enhancement function based on typical bearing fault modes (such as periodic faults, impact faults, and random disturbances). Simultaneously, it incorporates KMeansSMOTE to achieve structured resampling, effectively improving the diversity and distribution consistency of minority class samples. (See [link to relevant documentation]). Figure 3 The diagram shows a structural schematic of a fault prior guidance timing enhancement module.

[0038] In some embodiments, the fault prior-guided timing enhancement module can specifically include two collaborative stages in the unbalanced fault diagnosis framework: (1) designing a fault prior-based timing enhancement strategy: segmenting the original signal through a sliding window, and injecting features into minority class samples from a physical perspective based on the physical mechanisms of different fault modes to improve the representativeness of the samples; (2) performing structured oversampling: further generating synthetic samples from the minority class samples after the above timing enhancement processing by combining clustering and interpolation, improving the coverage of the feature space from a distribution perspective, and assisting in improving the class discrimination ability. The specific implementation process is as follows: First, a sliding segmentation method with a data window of length 300 is used to segment the original signal, ensuring the temporal correlation and statistical characteristics between segments. Then, each signal segment is normalized. (1) in, The original signal, and These are the mean and standard deviation of the signal, respectively. This is the standardized signal.

[0039] Next, to enhance sample diversity, a fault prior-driven enhancement strategy is constructed. The enhancement strategy is dynamically adjusted based on the statistical characteristics (such as mean, standard deviation, etc.) of each signal segment to simulate different fault modes. Different fault modes (such as inner race fault, outer race fault, rolling element fault, etc.) correspond to different enhancement methods. In each signal segment... The generated enhanced signal It can be represented as: (2) in, These enhancements are generated based on the fault mode and signal characteristics. Specifically, for inner-circle faults (periodic faults), the enhancement is... ,in For amplitude, For frequency, For time; for outer race failures (impact failures), the enhancement term is... ,in Let be the impulse function. For impact time; and for rolling element failure (random soft failure), the enhancement term is... This means adding noise within a random time period.

[0040] Relying solely on the aforementioned enhancements is insufficient to completely resolve the discrimination degradation problem caused by class imbalance. Therefore, this embodiment further introduces the KMeansSMOTE algorithm to perform distribution-consistent resampling for minority class samples. It uses KMeans clustering analysis to resample the enhanced signal samples... The samples are divided into several clusters with similar intra-class structures. Then, neighboring samples within each cluster are selected for linear interpolation. The formula for synthesizing new samples is as follows: (3) in, and These are two augmented samples belonging to the same fault category within the same KMeans cluster. These are the generated synthetic samples. It is a randomly generated weighting factor that determines the generation location of the synthetic sample.

[0041] II. Domain Decoupling Representation Learning Module: In some embodiments, the domain decoupling representation learning module is used to aggregate the representations of the source domain into a unified auxiliary domain, and to achieve systematic decoupling of domain-invariant features and domain-specific features by jointly optimizing the alignment loss of domain-invariant features and the decoupling loss of domain-specific features. The source domain includes labeled source domains and unlabeled source domains. The domain decoupling representation learning module is used to promote knowledge transfer and cross-domain generalization ability between source domains by minimizing the average sum of squared Euclidean distances between the domain-invariant features of each source domain and the auxiliary domain. The domain decoupling representation learning module is also used to constrain the domain-specific features of each source domain and the auxiliary domain to be nearly orthogonal (i.e., maximizing the angular diversity of domain-specific features between different domains) based on cosine similarity, thereby reducing the correlation between domain-invariant features and domain-specific features and enhancing the discriminative ability and independence of the learned features.

[0042] To achieve efficient knowledge transfer across multiple source domains and improve the discriminative power of feature representations, this embodiment proposes a Domain Decoupling Representation Learning (DDRL) framework. This framework first aggregates the representations of all source domains (including labeled and unlabeled source domains) into a unified auxiliary domain. Then, it uses a cosine similarity loss function to regularize the feature components in the auxiliary domain, explicitly reducing the correlation between domain-invariant features and domain-specific features, thereby achieving system decoupling between the two.

[0043] like Figure 2 As shown, the specific calculation process of this module is as follows: First, a unified auxiliary domain is constructed as a global feature benchmark. This auxiliary domain is a new domain generated by aggregating the feature representations of all source domains (including labeled and unlabeled source domains). Here, the total number of domains in this embodiment is defined as N. s The first N s-1 The Nth is the original source domain. s Each of these is an auxiliary field generated.

[0044] Step 1: Calculate the domain-invariant feature loss function L inv It is defined as the sum of squared average Euclidean distances between the domain-invariant features of each source domain and the auxiliary domain, aiming to map the features of all source domains into a unified, domain-invariant auxiliary space: (4) in, Indicates the first Domain-invariant eigenvectors of each source domain N represents the field-invariant eigenvector of the auxiliary field. s To assist in calculating the total number of source domain sets, Representing the Euclidean norm. This is achieved by introducing a denominator... Normalization of the loss ensures that the domain-invariant loss does not fluctuate drastically with changes in the number of source domains, laying a foundation for subsequent multi-objective collaborative optimization.

[0045] Step 2: Calculate the domain-specific feature loss function L spec This loss function, based on cosine similarity, aims to explicitly constrain the orthogonality of domain-specific features between each source and auxiliary domain, thereby achieving decoupling between them. The specific expression is: (5) in, For the domain-specific features of the i-th source domain, For domain-specific features of auxiliary domains, Represents cosine similarity, | | indicates the absolute value operator. When calculating the domain-specific feature loss function, the logic is to calculate the feature correlation between each original source domain and the auxiliary domain. Since the number of source domains involved in the comparison is... Therefore, when calculating the average loss, the denominator is taken as... This ensures that the loss function is mathematically consistent with the actual number of comparison terms, thus accurately representing the average decoupling degree between each source domain and auxiliary domain. By minimizing this loss (i.e., minimizing the absolute value of the similarity between features of each source domain and auxiliary domain), specific features of different domains can be forced to be independent of each other in space, effectively separating domain-invariant features from domain-specific features.

[0046] Step 3: Joint Optimization. The two losses mentioned above are jointly optimized and combined into a total feature decoupling loss, which is then weighted by coefficients. The balance between the two is adjusted as follows: (6) in, To balance the hyperparameters, the relative importance of domain-invariant feature alignment and domain-specific feature decoupling is adjusted. Because... and Both definitions adopt a method based on the number of comparison terms. The mean-based calculation method ensures that both are on the same mathematical order. This equivalence ensures that each loss term has a substantial impact on model parameter updates during gradient descent, thus avoiding the problem of any loss term being overwhelmed and achieving true collaborative optimization. Based on this, by adjusting... This allows for flexible control over the trade-off between generalization and discriminative abilities in the model: when there are significant differences in operating conditions across different domains, the generalization ability should be reduced. Prioritize ensuring generalization; when it is necessary to enhance category differentiation, it is advisable to increase... To improve the ability to make judgments. By adjusting It allows for flexible control over the trade-off between generalization and discriminative abilities in the model.

[0047] III. Frequency-Aware Hybrid Expert Networks: In some embodiments, the frequency-aware hybrid expert network incorporates frequency domain decomposition, dynamic expert selection, and fusion mechanisms to achieve adaptive modeling of multi-source features.

[0048] To enhance the model's adaptive feature modeling capability under complex and variable signal conditions, this embodiment proposes a Frequency-Aware Mixture-of-Experts (FAMoE) network. This structure integrates frequency domain decomposition, dynamic expert selection, and fusion mechanisms to achieve deep adaptive modeling of multi-source features. (See also...) Figure 4 The diagram shows a frequency-domain sensing expert hybrid expert network.

[0049] First, the input signal is processed by an FFT frequency domain decomposition module to obtain high-frequency components F. high With low-frequency component F low This process effectively distinguishes different frequency domain information contained in the signal, making subsequent feature processing more targeted. For each frequency band component, a multi-expert ensemble (MEE) is used for deep feature extraction to obtain discriminative feature representations in each frequency domain.

[0050] Next, all frequency band features will be fed into a frequency-aware gating network to achieve dynamic expert selection and fusion. This gating network includes modules such as Conv, GELU activation, pooling, fully connected layers, Softmax, and Top-K sparse selection. Given the input features F, the expression for calculating the gating weights is: (7) in, , For convolution and fully connected parameters, This indicates global average pooling, GELU is the activation function, Softmax is used for weight normalization, and Top-K only retains the K experts with the largest weights.

[0051] Subsequently, all frequency band features are fed into a frequency-domain-aware gating network to enable dynamic selection of experts and weight fusion. This gating network calculates the matching scores between the input features and each expert, and implements a "sparse screening first, then renormalization" strategy to ensure the scale stability of the output features.

[0052] Specifically, the gating weights are calculated as shown in formula (7): First, through convolution (parameters...) Global average pooling ( ) and fully connected layer (parameters) The process involves calculating the original score vector, then using a Top-K operation to retain only the K highest-scoring expert indices and setting the scores of the remaining experts to negative infinity. Finally, the retained K scores are renormalized using the Softmax function. This mechanism ensures that the sum of the weights of the selected K experts is strictly equal to 1, effectively avoiding feature output scale collapse caused by sparse expert selection and guaranteeing the numerical stability of deep feature representations. In this embodiment, the MoE (Mixture-of-Experts) consists of a frequency-aware gating network and N parallel expert networks. Specifically, each expert network is composed of N parallel lightweight convolutional sub-networks. Each expert network, through different convolutional kernel sizes or receptive fields, focuses on extracting specific frequency domain components (such as the aforementioned high-frequency component F). high or low-frequency component F low The local pattern features in the ) are obtained by weighting and fusing the outputs of all expert networks according to the sparse weights calculated by the gating network to obtain the final fused feature representation F. fused The calculation formula is as follows: (8) Among them, W n Indicates the first The gating weight of each expert, Indicates the first The output features of an expert network. Due to the sparse selection mechanism in formula (7), the weight vector W has only K non-zero values ​​(corresponding to the selected experts), and satisfies Through this adaptive weighted fusion method, the model can dynamically allocate computational resources to the best-matching expert sub-network based on the frequency characteristics of the input signal, thus achieving efficient modeling of multi-source features.

[0053] IV. Semi-supervised optimization strategy: In some embodiments, an imbalanced semi-supervised domain generalization network is used to apply a high-confidence pseudo-label mechanism to the cross-domain joint training process through a semi-supervised optimization strategy.

[0054] To fully explore the potential information of unlabeled samples and improve the model's generalization ability in the target domain, this embodiment introduces a semi-supervised optimization strategy, applying a high-confidence pseudo-label mechanism to the cross-domain joint training process.

[0055] In some embodiments, an imbalanced semi-supervised domain generalization network is used to perform forward inference on unlabeled samples using the current model in each training cycle to obtain the class prediction probability; for the target sample with the highest prediction probability and higher than a set threshold, the class corresponding to the target sample is selected as the pseudo-label, and the target sample and the labeled sample participate in the training process together.

[0056] Specifically, in each training cycle, the current model is first used to perform forward inference on unlabeled samples to obtain class prediction probabilities. For samples with the highest prediction probabilities that are higher than a set threshold, their corresponding classes are selected as pseudo-labels, and these high-confidence unlabeled samples are used together with the labeled samples in the subsequent training process.

[0057] This strategy effectively alleviates the knowledge bottleneck caused by label scarcity in engineering practice. On the one hand, by assigning supervisory signals to unlabeled samples through pseudo-labels, it promotes the model's adaptive representation and alignment with the target domain data distribution. On the other hand, joint training helps improve the discriminativeness and robustness of the feature space, reducing performance degradation caused by distribution differences. The loss term for pseudo-labeled samples can be expressed as: (9) in, Let I be the number of unlabeled samples adopted, and let I(·) be the indicator function. I(·) is only used when the first... The maximum class probability of each sample is greater than the confidence threshold. Only then will the sample be adopted. This is a high-confidence pseudo-label. Let be the predicted probability of the model that the m-th sample belongs to class c. This represents the total number of categories.

[0058] V. Objective Function Optimization: In some embodiments, the imbalanced semi-supervised domain generalization network achieves robust generalization of the rolling bearing fault diagnosis model under imbalanced and heterogeneous distribution conditions by jointly optimizing multiple loss functions; wherein, the optimization objectives of the imbalanced semi-supervised domain generalization network include: minimizing the classification loss of labeled samples, jointly optimizing the feature decoupling loss, and minimizing the pseudo-label loss of unlabeled samples by adopting a high-confidence pseudo-label mechanism.

[0059] In summary, this embodiment achieves robust generalization of the rolling bearing fault diagnosis model under conditions of imbalance and heterogeneous distribution by jointly optimizing multiple loss functions. During training, the total loss L needs to be minimized. total While ensuring the classification accuracy of labeled samples, it effectively improves the decoupling ability of features between domains and makes full use of the potential information of unlabeled samples, thus alleviating the knowledge bottleneck caused by label scarcity.

[0060] Specifically, the optimization objectives mainly include three aspects: (1) minimizing the classification loss L of labeled samples. cls (2) Jointly optimize feature decoupling loss L decouple (2) Improve the generalization ability and discriminative power of cross-domain feature representation; (3) Adopt a high-confidence pseudo-label mechanism to minimize the pseudo-label loss L of unlabeled samples. pseudoThis involves fully exploring the information in unlabeled data. To achieve the above objectives, the network optimization problem can be formally expressed as follows: (10) in, For model parameters, These are the weighting coefficients. Where, the classification loss L... cls The definition is as follows: (11) in, This indicates the number of labeled samples. For the number of categories, For the sample The true label, For the model to sample Category The predicted probability. This loss is used to measure the model's accuracy in classifying labeled samples and is fundamental to improving its discriminative ability.

[0061] Example 3: This invention provides an experimental verification, comparative analysis, and discussion of an unbalanced semi-supervised domain generalization network, based on the above embodiments.

[0062] See Figure 5 The diagram shown is a schematic representation of an existing study. Figure 5 (a) Figure 5 (c) Figure 5 (c) The main bottlenecks of DGFD, IDGFD, and SDGFD methods can be summarized as follows: (1) Domain generalization (DGFD) builds a model by learning "domain-invariant" features from multiple labeled source domains, aiming to generalize to an unknown target domain. However, this strategy lacks awareness of the potential characteristics of the target domain (such as subtle distributional differences), such as Figure 5 (a) shows that (2) although IDGFD attempts to balance the source domain data in multi-domain imbalanced generalization, its decision boundary still has a "misdiagnosis" problem in the target domain for identifying minority class faults, indicating that its generalization ability is still unstable. Figure 5 (b) shows. (3) Under the complex background of "domain offset + label missing", SDGFD is prone to generating pseudo-labels with class bias. These pseudo-label errors are amplified during the consistency regularization process, further pulling the model alignment direction incorrectly, inducing boundary oscillations and unstable alignment phenomena, such as Figure 5 As shown in (c).

[0063] To verify the effectiveness of this invention, this embodiment was tested on the CWRU and MFS datasets. Experimental results show that the imbalanced semi-supervised domain generalization network proposed in this embodiment achieves an accuracy of over 93% across various task conditions, significantly outperforming existing technologies and demonstrating the effectiveness of the FPGTA, DDRL, and FAMoE modules.

[0064] Example 4: This invention provides a rolling bearing fault diagnosis method for unbalanced semi-supervised domain generalization, which is implemented based on the above embodiments and applied to the unbalanced semi-supervised domain generalization network provided in the aforementioned embodiments. The unbalanced semi-supervised domain generalization network includes: a fault prior-guided timing enhancement module, a domain decoupling representation learning module, and a frequency-aware hybrid expert network.

[0065] Based on the above description, see Figure 6 The flowchart shown illustrates a rolling bearing fault diagnosis method generalized to unbalanced semi-supervised domains. This method includes the following steps: Step S602: The fault prior guidance timing enhancement module combines the physical mechanism of fault modes to improve the diversity and distribution consistency of minority class samples; Step S604: The domain decoupling representation learning module decouples the features into domain-invariant features and domain-specific features by constructing an auxiliary domain; Step S606: The frequency-aware hybrid expert network uses frequency domain decomposition and dynamic gating mechanism to achieve adaptive modeling of cross-domain features; Step S608: The imbalanced semi-supervised domain generalization network is jointly trained using unlabeled data, combined with a pseudo-label strategy based on confidence constraints. Step S610: Perform fault diagnosis of rolling bearings based on the trained unbalanced semi-supervised domain generalization network.

[0066] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the rolling bearing fault diagnosis method for unbalanced semi-supervised domain generalization described above can be referred to the corresponding process in the aforementioned embodiments of the unbalanced semi-supervised domain generalization network, and will not be repeated here.

[0067] Rolling bearing fault diagnosis in real industrial environments often faces multiple challenges, including imbalanced class distribution, limited labeled data, and the absence of target domains, posing a severe test to the model's generalization ability and diagnostic reliability. To address this, this embodiment proposes an imbalanced semi-supervised domain generalization network combining fault prior-guided temporal enhancement and domain decoupling representation learning for rolling bearing fault classification. First, a fault prior-guided temporal enhancement module is proposed to alleviate inter-class distribution bias. Second, a domain decoupling representation learning mechanism is designed to effectively separate and coordinate domain-invariant and domain-specific representations. Furthermore, a frequency-aware hybrid expert structure is constructed to fully exploit the complementarity of multi-frequency domain information. Finally, a pseudo-label strategy with confidence constraints is combined to effectively utilize unlabeled samples to improve the model's adaptability under label-constrained conditions. Experimental results across operating conditions using the CWRU and MFS datasets demonstrate that this proposed method significantly outperforms existing methods in both diagnostic accuracy and stability, validating its robustness in complex industrial scenarios.

[0068] The unbalanced semi-supervised domain generalization framework based on FPGTA-DDRL proposed in this embodiment addresses the challenges of unknown target domain, class imbalance, and missing labels in rolling bearing diagnosis under complex industrial environments. Through three innovative approaches—fault prior-driven data augmentation, decoupled representation learning of auxiliary domain features, and frequency-aware expert hybrid feature extraction—it achieves high-precision, balanced fault diagnosis under cross-domain and extreme imbalance conditions. Experimental results show that the method achieves an accuracy exceeding 93% in both CWRU and MFS cross-condition tasks. Ablation studies confirm the irreplaceable role of each core module; feature visualization analysis demonstrates its advantages in feature alignment and inter-class differentiation across different domains. In summary, FPGTA-DDRL exhibits outstanding performance in robustness, generalization ability, and balanced recognition, demonstrating its potential for widespread application in practical industrial environments. Future research will further explore multimodal sensor information fusion, online adaptive domain generalization, and lightweight deployment to support a wider range of equipment condition monitoring and predictive maintenance applications.

[0069] Example 5: This invention also provides a rolling bearing fault diagnosis system generalized to unbalanced semi-supervised domains, used to run the aforementioned rolling bearing fault diagnosis method generalized to unbalanced semi-supervised domains; see also Figure 7 The diagram shows a structural schematic of a rolling bearing fault diagnosis system generalized to unbalanced semi-supervised domains. The rolling bearing fault diagnosis system generalized to unbalanced semi-supervised domains includes a memory 100 and a processor 101. The memory 100 is used to store one or more computer instructions, which are executed by the processor 101 to implement the above-mentioned rolling bearing fault diagnosis method generalized to unbalanced semi-supervised domains.

[0070] Furthermore, Figure 7The rolling bearing fault diagnosis system shown for unbalanced semi-supervised domain generalization also includes a bus 102 and a communication interface 103. The processor 101, the communication interface 103 and the memory 100 are connected through the bus 102.

[0071] The memory 100 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 102 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0072] Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 100, and processor 101 reads information from memory 100 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0073] 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.

[0074] 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.

[0075] 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.

[0076] 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. An unbalanced semi-supervised domain generalization network, characterized in that, The unbalanced semi-supervised domain generalization network includes: a fault prior-guided timing enhancement module, a domain decoupling representation learning module, and a frequency-aware hybrid expert network. The fault prior guidance timing enhancement module is used to improve the diversity and distribution consistency of minority class samples by combining the physical mechanism of fault modes. The domain decoupled representation learning module is used to decouple features into domain-invariant features and domain-specific features by constructing an auxiliary domain. The frequency-aware hybrid expert network is used to achieve adaptive modeling of cross-domain features by utilizing frequency domain decomposition and dynamic gating mechanisms. The imbalanced semi-supervised domain generalization network is also used to perform joint training using unlabeled data, combined with a pseudo-label strategy that incorporates confidence constraints.

2. The unbalanced semi-supervised domain generalization network according to claim 1, characterized in that, The unbalanced semi-supervised domain generalization network is also used in the unknown working condition generalization stage to decouple domain invariance and domain-specific features through a dual-branch collaborative mechanism, and rely on the frequency-aware hybrid expert network to improve the discriminativeness and generalization of feature expression, so as to achieve adaptive recognition and identification of rolling bearing fault types under unknown working conditions.

3. The unbalanced semi-supervised domain generalization network according to claim 1, characterized in that, The fault prior guidance timing enhancement module is used to construct an enhancement function by combining the time structure of the explicit hold signal with the bearing fault mode; The fault prior guidance timing enhancement module is also used to combine KMeans clustering and SMOTE oversampling to achieve structured resampling.

4. The unbalanced semi-supervised domain generalization network according to claim 3, characterized in that, The fault prior-guided timing enhancement module is used to segment the original signal through a sliding window based on the fault prior timing enhancement strategy, and to set corresponding timing enhancement strategies for different fault modes based on their physical mechanisms and signal characteristics. The fault prior guidance timing enhancement module is also used to generate synthetic samples by combining clustering and interpolation on minority class samples after enhancement, based on a structured oversampling method that combines KMeans clustering and SMOTE oversampling.

5. The unbalanced semi-supervised domain generalization network according to claim 1, characterized in that, The domain decoupling representation learning module is used to aggregate the representations of the source domain into a unified auxiliary domain. By jointly optimizing the domain-invariant feature alignment loss and the domain-specific feature decoupling loss, the system decoupling of domain-invariant features and domain-specific features is achieved. The source domain includes labeled source domains and unlabeled source domains. The domain decoupled representation learning module is used to promote knowledge transfer and cross-domain generalization ability between source domains by minimizing the average sum of squared Euclidean distances of the domain-invariant features between each source domain and the auxiliary domain. The domain decoupled representation learning module is also used to constrain the domain-specific features of each source domain and the auxiliary domain to be nearly orthogonal based on cosine similarity, thereby reducing the correlation between domain-invariant features and domain-specific features and enhancing the discriminative ability and independence of the learned features.

6. The unbalanced semi-supervised domain generalization network according to claim 1, characterized in that, The frequency-aware hybrid expert network incorporates frequency domain decomposition, dynamic expert selection, and fusion mechanisms to achieve adaptive modeling of multi-source features.

7. The unbalanced semi-supervised domain generalization network according to claim 1, characterized in that, The imbalanced semi-supervised domain generalization network is used to apply a high-confidence pseudo-label mechanism to the cross-domain joint training process through a semi-supervised optimization strategy. The imbalanced semi-supervised domain generalization network is used to perform forward inference on unlabeled samples using the current model in each training cycle to obtain the category prediction probability; for the target sample with the highest prediction probability and higher than a set threshold, the category corresponding to the target sample is selected as the pseudo-label, and the target sample and the labeled sample participate in the training process together.

8. The unbalanced semi-supervised domain generalization network according to any one of claims 1-7, characterized in that, The unbalanced semi-supervised domain generalization network achieves robust generalization of the rolling bearing fault diagnosis model under unbalanced and heterogeneous distribution conditions by jointly optimizing multiple loss functions. The optimization objectives of the imbalanced semi-supervised domain generalization network include: minimizing the classification loss of labeled samples, jointly optimizing the feature decoupling loss, and minimizing the pseudo-label loss of unlabeled samples by using a high-confidence pseudo-label mechanism.

9. A method for fault diagnosis of rolling bearings generalized to unbalanced semi-supervised domains, characterized in that, The method is applied to the imbalanced semi-supervised domain generalization network according to any one of claims 1-8, wherein the imbalanced semi-supervised domain generalization network comprises: a fault prior-guided timing enhancement module, a domain decoupling representation learning module, and a frequency-aware hybrid expert network; the method comprises: The fault prior guidance timing enhancement module combines the physical mechanism of fault modes to improve the diversity and distribution consistency of minority class samples; The domain decoupled representation learning module decouples features into domain-invariant features and domain-specific features by constructing an auxiliary domain. The frequency-aware hybrid expert network utilizes frequency domain decomposition and dynamic gating mechanisms to achieve adaptive modeling of cross-domain features; The imbalanced semi-supervised domain generalization network, combined with a pseudo-label strategy based on confidence constraints, is jointly trained using unlabeled data. Fault diagnosis of rolling bearings is performed based on the trained unbalanced semi-supervised domain generalization network.

10. A rolling bearing fault diagnosis system for unbalanced semi-supervised domain generalization, characterized in that, The rolling bearing fault diagnosis system for unbalanced semi-supervised domain generalization is used to execute the rolling bearing fault diagnosis method for unbalanced semi-supervised domain generalization as described in claim 9.

Images