Prototype-oriented domain adaptation for rotating machinery fault diagnosis

Deep learning-based fault diagnosis has received considerable attention for its potential to enhance the reliability and safety of rotating machinery. However, traditional approaches often assume identical data distributions between training and testing sets and rely on large, labeled datasets, which restricts their effectiveness in practical applications. To overcome these barriers, we propose a novel prototype-oriented domain adaptation (PODA) framework for unsupervised cross-domain fault diagnosis, featuring three key innovations. First, we establish dual-domain class prototypes to explicitly model category characteristics in both source and target domains, overcoming the constraints of conventional single-domain adaptation. Second, a bidirectional prototype alignment mechanism is introduced to minimize intra-class discrepancies while maximizing inter-class separability across domains through prototype attraction–repulsion mechanism. Third, a discriminative retraining phase is incorporated to enhance classifier decision boundaries by aligning target samples with their corresponding prototypes. Evaluations on three public datasets demonstrate that PODA surpasses existing algorithms, significantly improving diagnostic accuracy and addressing the generalization challenges posed by distributional shifts.