Mitigating Catastrophic Forgetting in Cross-Domain Fault Diagnosis: An Unsupervised Class Incremental Learning Network Approach

Abstract

While deep learning has found widespread application in fault diagnosis, it continues to face three primary challenges. First, it assumes that training and test datasets adhere to the same distribution, which is often not the case in industries with varying conditions. Second, it relies heavily on the availability of abundant labeled data for training, overlooking the reality that newly collected data are frequently unlabeled. Third, neural networks frequently encounter catastrophic forgetting, a critical concern in dynamic industrial settings with emerging faults. Therefore, this article proposes an unsupervised class incremental learning network (UCILN), to mitigate catastrophic forgetting in cross-domain fault diagnosis, particularly in situations where the target domain lacks labeled data. A memory module and a semifrozen and semiupdated incremental strategy are designed to balance the retention of old knowledge with the acquisition of new information. Test results obtained from the Case Western Reserve University (CWRU) and Paderborn University (PU) datasets demonstrate the exceptional performance of UCILN.