Bearing life prediction method based on the multiscale degradation fusion algorithm and an ensembled BIGRU network

Remaining useful life (RUL) prediction is a crucial way to prevent unexpected failures, reduce operating costs, and ensure the stable operation of bearings. However, the complete life cycle of bearings typically includes multiple degradation stages, each with a unique degradation trend. Single-scale data often only elucidate the characteristics of specific local stages and fail to provide a comprehensive representation of the entire degradation process. To address these issues and challenges, this study presents a bearing RUL prediction method based on the ensemble bidirectional gated recurrent unit (EBiGRU) model and a multiscale degradation fusion (MSDF) algorithm. First, a novel degradation stage division algorithm, called 3Sigma-Bottom-Up, was designed to divide the entire bearing life cycle into four distinct states: normal, mild, moderate, and severe degradations. Second, an ensemblelearning-based EBiGRU model, a local-scale analysis method, was proposed to obtain the RUL result for each bearing degradation stage. Third, a conservative RUL estimate, a global scale analysis index, was defined and calculated using the Weibull curve to fit the performance degradation feature. Finally, an MSDF algorithm was constructed to obtain the bearing RUL by fusing the local and global analysis results. Experiments were conducted with XITU-SY datasets and a bearing test platform. The results show that the proposed method is more effective than conventional life prediction methods, and that the EBiGRU model exhibits better accuracy in remaining-life prediction than the alternative models.