Zihao Deng

Rotating machinery fault diagnosis based on one-dimensional signals often suffers from performance degradation due to distribution shifts caused by varying operational conditions. Domain generalization (DG) methods aim to learn a model from labeled source domains that can generalize to unseen target domains. However, existing DG approaches typically face limitations in effectively simulating inter-domain shifts, imposing fine-grained consistency constraints, and handling ambiguous decision boundaries. To address these challenges, this paper proposes a novel DG framework. First, a Distribution-Aware Source Domain Feature Augmentation module is introduced, which perturbs channel-wise statistics of features to generate augmented samples that preserve semantics while exhibiting distributional diversity, thereby simulating potential domain variations. Second, an Uncertainty-Guided Feature Alignment Regularization is designed to dynamically weight the alignment loss based on feature uncertainty and the discrepancy between original-augmented feature pairs, enforcing robust feature consistency. Third, an Adaptive Angular Margin Loss is incorporated to enhance discriminative power by constructing adaptive decision boundaries in the angular space based on intra-class compactness and inter-class separation. Extensive experiments on the PHM2009 gearbox dataset and a proprietary bearing dataset under a leave-one-domain-out protocol demonstrate the superiority of the proposed method. It achieves average accuracies of 80.46% and 87.46% on the two cases, respectively, significantly outperforming state-of-the-art comparative methods. The ablation study further confirms the contribution of each core component.

Data augmentation and generation are popular solutions for the problem of imbalanced and insufficient data. Aiming to conduct cross-domain data generation, this article proposes a physics-guided data generation framework, termed “digital block printing (DBP).” Its core idea is to use the “Digital Printing Plate” constructed from the simulation data as a reference, and use the “Data Printer” trained with source-domain experimental data to generate synthetic data with a similar distribution to the target-domain experimental data. For a specific case, conditional-diffusion-based vibration signal generation is conducted in this article, with the target-domain data simulated by a phenomenological model as the “Digital Printing Plate,” and the diffusion model trained by source-domain experimental data as the “Data Printer.” The result shows that the method proposed in this article can generate experiment-like synthetic data, which processes both the frequency-domain characteristics similar to the target-domain simulation data, and time-domain characteristics similar to the source-domain experimental data.

Transfer learning methods have been widely applied in intelligent fault diagnosis in domain-shift scenarios. However, in gearbox fault diagnosis where domain shift and class imbalance coexist, traditional transfer learning methods exhibit bias toward behavioral shifts of the majority class. To address these challenges, dynamic decoupled reward meta-learning (DDRML) is proposed, which has a framework integrating dynamic reward adjustment, modular decoupling optimization, and meta-learning generalization capabilities. Initially, the first convolutional layer is replaced with a complex wavelet attention module (CWAM), which enhances standard wavelet convolution by decoupling amplitude and phase in time-frequency space and employs a multiscale weighted feature mechanism to select and correlate features across numerous frequency bands. Furthermore, a class-sensitive dynamic reward meta-learning (CDR-ML) component is designed, introducing the concept of the action-value function from reinforcement learning to guide the model toward minority samples via adaptive rewards. Building on this, a full-action loss is formulated to holistically optimize the model using reward values from all classes, significantly reducing training time. Experiments on two datasets demonstrate that DDRML significantly outperforms baselines under domain shift and class imbalance while exhibiting superior robustness to extreme imbalance ratios and noisy environments.