关键词: 齿轮箱, 故障诊断, 迁移学习, 领域自适应, 动力学模型, 深度学习

Abstract: A fault diagnosis method was proposed for industrial gearboxes under small-sample and imbalanced data conditions. This approach integrated explainable dynamic simulation with deep domain adaptation learning. A rigid-flexible coupled dynamic model of the gearbox was constructed. High-fidelity vibration simulation data with clear physical meaning were generated. Scarce fault samples were effectively supplemented. Mechanistic explainability of fault-response relationships was provided. A feature extraction network was developed by combining multi-head self-attention and depthwise separable convolu-tion. Transient impact features in vibration signals were better captured. An improved T-Softmax loss function was designed. Learnable margins were introduced into the feature space. The spread of majority-class distributions was constrained. Minority-class feature cohesion was enhanced. Both normal samples and rare fault samples were accurately recognized. A refined joint distribution adaptation strategy was proposed based on Bayesian inference. Marginal and conditional distributions between simulation and measured data were aligned. Domain shift caused by model simplification and noise interference was reduced. Experiments were conducted under extreme imbalance with only 30 fault samples and a 1∶10 imbalance ratio. An average accuracy of 98.12 % was achieved. Performance was close to that of balanced data scenarios. On public datasets,the average accuracy exceeded 95 %. The method′s robustness and generalization capability were fully validated.

Key words: gearbox, fault diagnosis, transfer learning, domain adaptation, dynamic model, deep learning