Generalizable Fault Diagnosis Under Distribution Shifts Induced by Unseen Working Conditions via Synthetic and Adversarial Sample Learning

Fault diagnosis under distribution shifts induced by previously unseen operating conditions is of significant practical value. The main challenge lies in the lack of data from unseen conditions, which prevents diagnostic models from capturing the corresponding distribution patterns, resulting in degraded prediction accuracy. To address this issue, we propose a fault diagnosis framework aimed at enhancing generalization against distribution shifts arising from unseen operating conditions, with its effectiveness validated specifically on gearbox diagnostics. Our approach first employs a modified 1-D stable diffusion model to generate samples under unseen operating conditions. Concurrently, adversarial samples are supplied from monitoring data under known operating conditions using the fast gradient sign method to further enhance model robustness. Then, the monitoring samples, the synthetic samples, and the adversarial samples are jointly used to train an uncertainty-aware deep learning (UDL) model until convergence. Finally, both the classification accuracy and prediction uncertainty of the UDL model are assessed. To validate the effectiveness of the proposed approach, two planetary gearbox datasets were employed for testing. Experimental results demonstrate that the proposed method is capable of accurately performing fault diagnosis under unseen operating conditions, thereby verifying its robustness and generalization capability.