Unsupervised domain adaptation for bearing fault diagnosis using nonlinear impact dynamics model under limited supervision

Abstract

Rolling bearing is one of the crucial rotating parts of mechanical systems, which is usually exposed to high-load working conditions. The diagnosis of rolling bearing faults is significant for the health monitoring of the whole mechanical system. The deep learning method has been proven to be effective in many fault diagnosis occasions. However, sufficient labeled fault samples are unavailable in some practical industrial diagnosis tasks, which will lead to the serious performance degradation of traditional deep learning methods. Therefore, a rolling bearing dynamics model is established for generating sufficient simulation data for assisting the training process. Furthermore, to overcome the diagnostic performance degradation problem caused by the inconsistent feature distribution of simulation data and experimental data, adversarial learning is conducted to realize domain adaptation, thus capturing the generalized feature representation. The analysis results of an experimental rolling bearing dataset demonstrate the effectiveness of the proposed model, showing a potential industrial application value.