Physics-Guided GAN With Manhattan Attention: A Novel Approach for Imbalanced Bearing Vibration Fault Diagnosis

This study introduces a novel Physics-Guided Generative Adversarial Network (PGAN) tailored explicitly for diagnosing rolling bearing faults under severely imbalanced datasets. PGAN integrates domain-specific physical constraints into the generative process, thereby enhancing both the physical realism and interpretability of the generated data. Key innovations include a generator conditioned on physics-based features and random noise, the introduction of dedicated physics-guided loss functions, and the incorporation of Manhattan Attention to improve the extraction of essential vibration features. Extensive experiments conducted using the CWRU and HUST datasets demonstrate PGAN’s superior performance compared to several state-of-the-art methods. Results indicate that PGAN significantly mitigates the impact of dataset imbalance, achieving robust diagnostic accuracy even in extreme conditions. Further visualization and comparative analyses underscore the method’s capability to deliver highly discriminative and interpretable features.