Multi-source domain generalization for degradation monitoring of journal bearings under unseen conditions

Degradation detection and remaining useful life (RUL) prediction are the essential tasks of Prognostics and Health Management (PHM) designed to increase the reliability of key components and reduce unpredictable maintenance costs. Due to the lack of enough degradation monitoring data under unseen working conditions or equipment conditions in practical applications, the performance of most existing deep learning and transfer learning RUL prediction models will deteriorate. To address this problem, this paper combines the advantages of Gated Recurrence Unit (GRU) and Transformer structures to propose a multi-source domain generalization learning method. The proposed method can extract the generalized degradation feature representations from multiple available offline run-to-failure datasets under different known working conditions or equipment conditions to assist the prognosis tasks for practical application scenarios. The run-to-failure datasets of internal combustion engine journal bearings are used for case studies to validate the proposed method. The calculation results prove the superiority and effectiveness of the proposed method.