A modified domain adversarial approach based on model and data-driven for bearing fault diagnosis

Abstract

Bearing fault diagnosis is essential for maintaining the reliability and stability of mechanical equipment. However, obtaining sufficient labeled data in real scenarios is high-risk and challenging, which limits the further application of traditional data-driven approaches. In this research, a novel model and data-driven approach called the modified domain adversarial neural network (MDANN) is developed for bearing fault identification. Specifically, a bearing dynamic model is established, so that the priori information on bearing failures can be acquired by finite element simulation. A data-driven MDANN model is then developed for feature extraction and cross domain transfer from simulated data to measured data. The attention module is introduced for feature weight reassignment, so that the priority of domain-invariant features is increased and negative transfer is suppressed. The improved loss function incorporating adaptive CORAL is designed to align both marginal and conditional distributions. Finally, the validity of the proposed MDANN is validated through two cases. The results demonstrate that the domain transfer capability of MDANN outperforms other methods in cross-domain tasks.