A novel hierarchical transferable network for rolling bearing fault diagnosis under variable working conditions

The deterioration of bearing failure is a gradual process. Simultaneously identifying the fault pattern and severity is of great significance for understanding the failure evolution process and determining reasonable repair plans. However, the working conditions of rolling bearings change frequently, which further increases the difficulty of identifying fault pattern and severity. To overcome the above problem, this paper proposes a novel hierarchical transferable network (HTNet). Firstly, a two-layer hierarchical structure is designed to handle the fault pattern and severity recognition task, respectively. Then, in order to establish the correlation between these two layers, an adaptive subnet selection module is proposed to utilize the pseudo labels of fault patterns for separating the internal fault severity levels and ensuring the simultaneous diagnosis ability of the network. On this basis, a hierarchical domain adaptation method is presented to extract domain-invariant features from different classification tasks in different layers. Finally, two experimental cases of rolling bearings verify that the proposed method has better performance and transferability than the existing state-of-the-art methods under variable working conditions.