Dual-domain generative adversarial network with cross-domain structural regularization for bearing fault diagnosis

Dual-branch self-guided network for single-domain generalization in sensor fault diagnosis

Most domain-generalization methods for sensor fault diagnosis require multiple labeled source domains. When only one source domain is available, inter-domain variations cannot be directly observed, making it difficult to identify informative samples and avoid overfitting to source-specific features. To address these challenges, a dual-branch self-guided network (DBSGN) is proposed. In the guidance branch, energy discrepancy and local feature distance identify class-representative samples, while style consistency selects domain-representative samples. In the decision branch, multi-scale feature extractors learn complementary representations, and distribution-uncertainty-guided interleaved learning enables the fault classifier to exploit different feature scales. Two-stage contrastive learning improves cross-scale distribution consistency, intra-class compactness, and inter-class separability. Scale-blurring adversarial training further suppresses scale-specific information and promotes transferable representation learning without target-domain access during model development. Experiments on real-world sensor data from a nickel flash smelting system show that DBSGN achieves an average accuracy of 94.89% across twelve cross-process-domain diagnostic tasks, 3.44 percentage points higher than the best comparison method under the same protocol. These results support its effectiveness in the studied single-source domain-generalization setting.

Heterogeneous dual-stream contrastive adversarial network for cross-domain mechanical fault diagnosis

In the cross-machine fault diagnosis task of bearings for rotating machinery, inherent physical differences among different mechanical devices cause severe domain distribution shift of fault features. Consequently, the diagnostic model trained on the source machine suffers a dramatic decline in generalization performance when applied to diagnosis tasks of target machines, making it difficult to meet the requirements of practical industrial diagnosis. To address this issue, this paper proposes a heterogeneous dual-stream contrastive adversarial network (HDCAN). Firstly, one-dimensional (1D) vibration signals are converted into two-dimensional images via the symmetric dot pattern, constructing a signal-image multimodal input system. Secondly, a parallel dual-stream network architecture is designed, and a cross-modal contrastive learning strategy is introduced simultaneously to effectively eliminate the negative transfer effect between heterogeneous modalities and realize the pre-alignment of multimodal features. On this basis, the domain adversarial learning mechanism is integrated to further extract generalized features with both fault discriminability and domain invariance, so as to improve the cross-machine adaptation capability of the model. To verify the effectiveness of the proposed method, experiments are conducted on six cross-machine diagnosis tasks from three public bearing fault datasets. The experimental results show that the average diagnostic accuracy of the HDCAN model reaches 92.10%, which significantly outperforms the current mainstream domain adaptive fault diagnosis methods. The experimental results fully demonstrate that the proposed HDCAN can effectively alleviate the distribution shift problem of cross-machine fault data and possesses great application potential in industrial practical scenarios.

An attention guided joint distribution domain adaptation method for fault diagnosis

Padé Approximant Neural Networks as Feature Extractors for Unsupervised Domain Adaptation in Bearing Fault Diagnosis

These results establish PadéNet-enhanced UDA as an accurate, broadly applicable approach for robust bearing fault diagnosis under varying operating conditions, with a reduced parameter count suited to resource-constrained embedded platforms.

Dynamic pseudo-label guided adversarial multi-scale graph convolutional network for cross-domain fault diagnosis.

Recently, domain adaptation, a form of transfer learning, has been extensively applied to mechanical fault diagnosis across diverse operating conditions to address the challenges of insufficient labeled data and frequent operational state transitions. However, most existing approaches focus solely on the spatial distribution of inter-domain categorical features, neglecting the clustering properties of feature clusters. Moreover, most pseudo-labeling approaches lack a rational screening mechanism based on the quality of the pseudo-labels. To address these challenges, this paper proposes a Dynamic Pseudo-Label Guided Adversarial Multi-Scale Graph Convolutional Network (DPAMGCN) for unsupervised cross-domain fault diagnosis. First, a network architecture is designed by cascading a multi-scale parallel Convolutional Neural Network (CNN) with a multi-receptive-field graph convolutional network (MRF-GCN) to extract features. Second, a multi-objective total-loss function is constructed that integrates geometric loss with three standard loss functions to jointly optimize feature clustering and spatial distribution. Finally, a dynamic threshold-based pseudo-label filtering strategy is proposed that, when combined with a geometric loss, enhances the model's generalization capability. Cross-domain transfer experiments conducted on the University of Ottawa (Ottawa) and Huazhong University of Science and Technology (HUST) benchmark datasets demonstrate that DPAMGCN achieves outstanding cross-domain diagnostic performance under the proposed optimization strategy and pseudo-label screening mechanism.

Source-free domain adaptation via robust pseudo-label optimization for cross-domain wind turbine bearing diagnosis

Newly built wind farms often face three practical barriers for bearing fault diagnosis: limited fault samples, expensive labeling, and restricted access to historical source-domain data due to privacy and security requirements. This paper presents a source-free domain adaptation framework named dual perturbation robust pseudo-label optimization (DPR-PLO) for cross-domain wind turbine bearing diagnostics using only a pre-trained source model and unlabeled target data. DPR-PLO combines noise-aware hybrid perturbation learning with dynamic prototype alignment to improve robustness against condition-induced distribution shifts. Specifically, it applies time–frequency perturbations and feature-space adversarial perturbations to enhance target diversity, constructs discriminative target representations via dynamic clustering and prototype updating, and adopts a two-stage pseudo-label refinement procedure to reduce noise propagation during self-training. Experiments on Paderborn University Dataset, Jiangnan University Dataset, and a self-collected dataset demonstrate that DPR-PLO achieves 97.56% average accuracy across 18 cross-condition tasks, surpassing benchmark methods by 6.95% and yielding more reliable pseudo-labels. These results indicate that DPR-PLO provides a feasible and robust solution for privacy-preserving bearing health management in new wind turbines.