Reinforced subdomain-adversarial generalization network for multi-source fault diagnosis in industrial robotic joint reducers

Abstract To address the issue of reduced fault diagnosis accuracy for industrial robot joint reducers under unknown operating conditions, a multi-source domain adaptive fault diagnosis method based on a Reinforced Subdomain-Adversarial Adaptation Network (RSAAN) is proposed. This method effectively extracts domain-invariant features, enhancing the robustness of fault diagnosis under unfamiliar conditions. First, intrinsic distribution clustering is utilized to extract features that emphasize fault-specific differences while minimizing the influence of operating condition variations, thereby improving the generalizability of the domain-invariant features. Next, local adversarial training is introduced on top of global adversarial training to achieve better feature alignment when the distribution discrepancy between domain-invariant features and generalized source domain features is small. Finally, an adaptive adversarial factor is employed to adjust the weight balance between global and local adversarial training, enabling the model to extract and refine domain-invariant features from multiple source domains to adapt to various unseen operating conditions. The proposed method achieved over 95% diagnostic accuracy in multiple cross-domain diagnostic scenarios using publicly available gearbox and robot bearing datasets, demonstrating its effectiveness and reliability in diagnosing faults under unknown conditions. This approach provides critical technical support and decision-making guidance for fault diagnosis in industrial robots.