Sharpness-Aware Gradient Alignment for Domain Generalization With Noisy Labels in Intelligent Fault Diagnosis

Ensuring the operational safety of rotating machinery through intelligent fault diagnosis is crucial for industrial profitability. However, the inherent variability in machine operations often leads to the generation of new datasets, compromising domain generalization. Additionally, the process of creating training data involves human intervention, introducing the risk of poor label auditing and resulting in noisy fault labels. To address these challenges, we present a novel approach called sharpness-aware gradient alignment (SAGA) for domain generalization with noisy labels (DGNLs) in fault diagnosis. The SAGA approach employs a dual network structure: a robust network leverages self-supervised learning and a gradient penalty to counteract the memorization of noisy labels, thus enhancing diagnostic robustness; and a generalizable network focuses on SAGAs to identify a suitable flat minimum, thereby improving diagnostic generalization capacity. The integration of individual networks in the dual structure is achieved through recursive parameter transfer, facilitating seamless communication of learned knowledge to realize both diagnostic robustness and generalization. Extensive bearing fault diagnostic experiments involving five domain generalization cases demonstrate the effectiveness of the proposed SAGA approach. Comparative results show that SAGA consistently outperforms state-of-the-art methods under diverse generalization gaps and varied noise rates.