Evidential Deep Learning-Based Adversarial Network for Universal Cross-Domain Fault Diagnosis of Rotary Machinery

In recent years, domain adaption (DA) in fault diagnosis of rotary machinery has been attracting considerable attention. Recent advancements in closed-set, partial, and open-set DA fault diagnosis, have well addressed the label inconsistent problem where the relationship of label spaces between the source and target domains are assumed to be certain; however, previous information on fault types in the target domain is unavailable in applications, denoted as universal cross-domain fault diagnosis, where the above three kinds of DA methods are rendered ineffective. To address this issue, a novel evidential deep learning (DL)-based adversarial network is proposed for universal cross-domain fault diagnosis without making explicit assumptions on the relationship of label spaces between the source and target domains. First, the adversarial training strategy is used for domain-invariant feature extraction. Second, an evidence-based fault identifier is adopted for known fault identification by judging the confidence and uncertainty of predictions of fault samples. Third, exponential evidence score-based unknown estimation is developed for underlying unknown fault recognition. At last, experimental results on both the bearing fault dataset and gearbox fault dataset validated the superiority of the proposed method over other DA-based methods.