Deep meta-domain-adversarial neural network for machinery fault diagnosis under multiple operating conditions

Abstract

This paper investigates the machinery fault diagnosis problem in the case of multiple operating conditions. Considering that the mechanical equipment may experience different operating conditions during operation, a meta-domain-adversarial neural network (MDANN) is established for improving the adaptability of transfer models in multi-target-domain fault diagnosis, which utilizes meta-learning technique to view the multi-target-domain fault diagnosis as multiple tasks. In the data preprocessing stage, the continuous deep belief network is selected for outliers removal. Furthermore, to make the MDANN model available for the partial domain multi-target-domain fault diagnosis problem, a new group-wise comparison approach is proposed. Compared with the existing results, the proposed MDANN allows one trained fault diagnosis model to cope with different operating conditions, and it can be extended to address scenarios where the label spaces in the source and target domains are different. Finally, the proposed fault diagnosis method is experimented on the bearing datasets and compared with several state-of-the-art approaches, which proves its superiority and effectiveness.