Multi-level weighted dynamic adversarial adaptation network for partial set cross-domain fault diagnosis

Abstract

Domain adaptation (DA)-based methods have been successfully applied in fault diagnosis, but their effectiveness relies on the assumption that the source and target label spaces are identical, which does not always hold in industrial scenarios. A more realistic scenario, named partial transfer diagnosis where the target label space is a subset of the source label space, is explored in this study. To this end, a multi-level weighted dynamic adversarial adaptation network (MWDAAN) is proposed. To prevent the negative effects of outlier data, a parallel sample filtering scheme based on the multi-level weighting strategy is designed, which enhances the effect of correcting noise weights and adequately evaluates the transferability of source samples. On this basis, a dynamic domain adversarial framework embedded with the sample filtering scheme is constructed to maximize the positive transfer of shared classes. Extensive experimental results on four cases demonstrate the effectiveness and superiority of the proposed method.