Task-structured prototype evolution with dual-stage reconstruction for multi-Source cross-domain few-shot learning

Abstract

Industrial fault diagnosis often faces the challenge of limited fault samples. Most existing closed-set diagnosis approaches focus on weighting and aligning source domains with the target domain. However, the domain shifts among multiple source domains are often overlooked. To address these issues, this paper proposes a Task-Structured Prototype Evolution with Dual-Stage Reconstruction (TSPE-DSR) framework for multi-source cross-domain few-shot closed-set fault diagnosis. First, a cross-domain task division mechanism is designed to construct meta-tasks from multiple source domains. The mechanism iteratively assigns pseudo-source and pseudo-target domains to form auxiliary tasks. These tasks exploit shared knowledge across domains and facilitate rapid adaptation to the target domain. Second, a dual-stage attention prototype reconstruction network is developed to model prototype evolution. A self-attention mechanism is introduced to achieve a progressive transition from embedding learning to prototype reconstruction. This process enables adaptive transfer of source-domain prototypes and effectively reduces cross-domain discrepancies. Finally, a joint training strategy is adopted by combining weighted prototype similarity loss and classification loss. This strategy improves classification performance, alleviates overfitting, and enhances model robustness. The effectiveness of TSPE-DSR is validated on the Tennessee Eastman Process(TEP), Case Western Reserve University(CWRU), and Aluminum Electrolysis Process(AEP) datasets. Experimental results demonstrate that TSPE-DSR outperforms existing methods in classification performance, achieving average accuracy improvements of 5.52%, 3.12%, and 4.79% over the strongest baseline on the TEP, CWRU, and AEP datasets, respectively. Our code is at https://github.com/Luce-hub/DAPRN.