Step-by-step gradual domain adaptation for rotating machinery fault diagnosis

Abstract

Deep-learning-based fault diagnosis (FD) methods have shown remarkable superiority in the field of fault prognostic and health management (PHM). However, the performance of a deep neural network relies heavily on a substantial labeled training dataset, which is rare in actual industrial scenarios. Moreover, the extracted fault features under different working conditions follow different joint distribution. As a result, a deep model trained under one condition cannot be extended to others. To address the existing problem, this paper proposes a step-by-step gradual domain adaptive neural network to conduct cross-domain FD, which can realize precise alignment between the source domain and the target domain. Firstly, the maximum mean discrepancy is used to perform primary domain adaptation. Furthermore, two classifiers are set up to limit the discrepancy of target domain data in the classification decision. Finally, the exact alignment of class-level features is achieved by category prototype alignment. The extensive experimental results show the superiority and stability of the proposed method when compared with other conventional approaches.