A feature fusion-based prognostics approach for rolling element bearings

Abstract

The emergence of prognostics and health management as a condition-based maintenance approach has greatly improved productivity, maintainability, and most essentially, reliability of systems. Invariably, a rolling-element bearing (REB) is the heart of rotating components; however, its failure can have daunting effects ranging from costly unexpected breakdown to catastrophic life-threatening situations. Consequently, the need for accurate condition monitoring and prognostics of REBs cannot be overemphasized. In view of achieving a more comprehensive condition assessment for prognostics of REBs, this study proposes a kernel principal component analysis (KPCA) feature fusion technique for degradation assessment and a deep learning model for prognostics. The deep learning method-deep long short-term memory (DLSTM) has shown an evident comparative advantage over the basic LSTM model and standard recurrent neural networks for time-series forecasting. Subsequently, the proposed prognostics model-KPCA-DLSTM performance was validated with a run-to-failure experiment on REBs and evaluated for accuracy against other prognostics methods reported in other works of literature using standard performance metrics. The proposed method was also used for REB remaining useful life (RUL) prediction and the results show that the KPCA-DLSTM does not only reflect a more monotonic bearing degradation trend but also yields better prognostics results.