Cross-Domain Open-Set Fault Diagnosis Based on Target Domain Slanted Adversarial Network for Rotating Machinery
Zuqiang Su, Weilong Jiang, Yang Zhao, Song Feng, Shuxian Wang, Maolin Luo
Abstract
Unsupervised domain adaption (DA) is a well-established technique for fault diagnosis of rotating machinery, which has attracted considerable attention in recent years. However, existing DA methods assume that the label spaces of the source domain and target domain are consistent, and this assumption is not always satisfied in industrial settings as new fault types would inevitably occur during operation. Aiming at this issue, an open-set fault diagnosis (OSFD) network is proposed for rotating machinery, denoted as the target domain slanted adversarial network (TDSAN). Specifically, two significant innovations are incorporated. Firstly, a target domain slanted classifier (TDSC) is developed to tackle the biased learning problem by leveraging the target domain data distribution. Secondly, an adaptive threshold for unknown fault identification is designed to enhance the distinguishability between known and unknown faults in the target domain. Finally, to evaluate the effectiveness and robustness of the proposed TDSAN, extensive experiments were conducted on two fault datasets: a bearing fault dataset and a gearbox fault dataset. The ablation study was also performed to validate the contributions of each innovation of the proposed TDSAN, and the experimental results demonstrated the superiority of the proposed framework.