Prototype‐attention domain adaptation for explainable bearing fault diagnosis

This study presents a prototype attention domain adaptation framework for explainable bearing fault diagnosis under varying operating conditions. The method aligns class centroids and dynamically weights target samples based on similarity and confidence. It achieves high accuracy with limited labeled data. The framework was evaluated on two benchmark datasets, Paderborn and CWRU, both reaching 100 percent classification accuracy. Interpretability is supported through prototype-to-instance similarity maps and visualizations of feature alignment.