Transfer fault prognostic for rolling bearings across different working conditions: a domain adversarial perspective

Modern machines generally operate under varying working conditions, which induces significant data distribution discrepancies in the gathered condition monitoring signals. However, most of the existing machine learning-based methods, especially deep learning (DL)-based fault prognostic methods, neglect the data distribution discrepancy between the training and testing data. As a result, most of the existing DL-based methods can only generalize well under identical working conditions, which is infeasible in real engineering practice. To solve this critical issue, a novel dual-branch neural network with a domain adversarial module is developed to achieve transfer fault prognostics across different operating conditions. A dual-branch-based DL model is first utilized to extract abundant degradation features from the heterogeneous inputs. Then, the domain adversarial technique is employed to solve the significant distribution discrepancy problem existing across different operating conditions. The proposed approach is validated experimentally through two rolling element bearing open-sourced datasets, i.e., the XJTU-SY bearing dataset and the PRONOSTIA bearing dataset. The experimental results demonstrate that the proposed method can accurately achieve the transfer fault prognostic task without any labelled data in the target domain, and performance comparisons with other state-of-the-art approaches are also presented.