Multiple-Model Estimation-based Prognostics for Rotating Machinery

Data-driven based prognostics approaches are currently attracting unprecedented attention. Considering flexibility, statistical model based prognostics, which present more transparency and usually incorporate statistical models with stochastic filtering, are studied in this paper. Traditionally, versions of Kalman filter and Particle filter are combined with a single statistical model. However, the service life of a bearing could undergo several health stages, such as a normal stage, a slight degradation stage, and a severe degradation stage. Thus, a single model cannot represent well the full degradation process of a bearing. As such, the concept of Multi-Model Estimation (MME) has been introduced in the field of CM of rotating machinery. Initially the Switching Kalman Filter (SKF) has been proposed combining three linear statistical models, but the Classic Kalman filter (CKF) might be insufficient for nonlinearity estimations. In this paper, we extend MME from CKF to nonlinear system estimation, for instance with extended Kalman filter, unscented Kalman filter and Particle filter, which can well overcome the drawbacks of CKF. Besides, twelve statistical models are studied and used for the automatic switch in the proposed multi-model methodology. The methodology is tested and evaluated on fifteen experimental datasets and it can be concluded that the extended MME outperforms the classic switching Kalman filter.