Anomaly detection and fault prognosis for bearings

In this paper, a new bearing anomaly detection and fault prognosis method is proposed. The method detects bearing anomalies and then predicts its remaining useful life (RUL). To achieve these two goals, an autoregressive model, which is used to filter out fault-unrelated signals, is derived according to healthy bearing vibrational signals. A health index is developed to indicate bearing health conditions. Anomalies of bearings are detected by choosing an appropriate threshold with the aid of a Box-Cox transformation. A nonlinear model is built to track the bearings' degradation process and an extended Kalman filter is designed to model adaptation and RUL prediction. Finally, PRONOSTIA bearing data are used to demonstrate the effectiveness of the proposed method.