Abstract
To address the challenges of distribution shift and limited fault samples in cross-condition bearing fault diagnosis, this paper proposes a cross-condition few-shot diagnosis method based on a learnable diagonal metric (DMT) and structured optimization, termed DMT-ML. First, to reduce the dependence of conventional methods on time-frequency image preprocessing and improve end-to-end modeling capability, a metric-based meta-learning framework is developed directly on raw one-dimensional vibration signals, in which a 1D tokenizer and a transformer encoder are introduced to capture long-range dependencies and more stable discriminative patterns across operating conditions. Second, to overcome the limitation of fixed Euclidean distance in prototypical inference under anisotropic cross-condition feature distributions and unequal dimensional importance, a learnable DMT module is designed to adaptively reweight feature dimensions, thereby improving the robustness and discriminative reliability of prototype matching. Furthermore, to address the unstable decision boundaries caused by intra-class dispersion and inter-class overlap in the embedding space under few-shot settings, a joint objective combining cross-entropy loss with intra-class and inter-class optimization is introduced to enhance intra-class compactness and inter-class separability. Finally, experiments are conducted on the Case Western Reserve University, Jiangnan University, and Paderborn University datasets under a strict condition-wise episodic evaluation protocol. The results show that the proposed method achieves strong diagnostic accuracy and stability under most cross-condition few-shot task settings. These findings confirm the effectiveness of the proposed method for cross-condition few-shot bearing fault diagnosis.
Funding Information
  • New Generation Information Technology Innovation Project of the Ministry of Education of China (No. 2020ITA02057)
  • National Natural Science Foundation of China (No. 62076215)

This publication has 31 references indexed in Scilit: