A Method of Rolling Bearing Fault Diagnosis Using Transfer Learning

Rolling bearing fault diagnosis is an important task that is critical for realizing predictive maintenance in industry 4.0. However, due to the lack of labeled data of rolling bearing faults and the complexity of working conditions, it is expected to construct fault diagnosis model with good generalization through leveraging a small number of labeled data collected from limited source domains and make it perform fault diagnosis well in other target domains. In this paper, we propose a method of rolling bearing fault diagnosis using transfer learning. We first construct a group of Domain-Adversarial Neural Networks (DANNs), each of them does its best to identify the faults of target domain via making the distribution of each source domain data and the distribution of the target domain data become similar as much as possible. After that, the group of DANNs are combined through the Mixture of Experts (MoE) to build the fault diagnosis model MoE-DANN, which achieves satisfactory fault diagnosis results in the target domain by unleashing the potential and balancing the effect of each expert. Extensive experiments demonstrate that MoE-DANN yields better diagnosis accuracy on open benchmark datasets than the baseline models. Additionally, it generalizes well even it is trained by a limited number of labeled data.