Domain adaptive deep belief network for rolling bearing fault diagnosis

Abstract

As the essential components of rotating machines, rolling bearings always operate in variable working conditions and suffer from different failure modes. To address the issue of lacking substantial labeled samples in new working conditions, a domain adaptive deep belief network (DA-DBN) is proposed for rolling bearing fault diagnosis. Firstly, the DBN model is pre-trained by the labeled samples which are composed of raw vibration signals and their corresponding time domain and frequency domain indicators. Secondly, the domain adaption method in transfer learning is applied to calculate the multi-kernel maximum mean discrepancies (MK-MMD) between the known working condition data and new working condition data in multiple layers. Thus, the loss function composed of MK-MMD and classification error can be obtained, and back propagation algorithm is used to fine-tune model parameters. Finally, the datasets with five fault patterns are collected to evaluate the performance of the DA-DBN. The results demonstrate that the proposed DA-DBN can achieve more than 92% fault classification accuracy under three noise levels; the average accuracy of fault classification under variable working conditions is 93.5%, which is the highest compared with other models.