Switching Unscented Kalman Filters With Unknown Transition Probabilities for Remaining Useful Life Prediction of Bearings

Since bearings are critical components of a mechanical equipment, timely fault detection and accurate prediction of remaining useful life (RUL) are essential for ensuring sufficient time for maintenance and replacement. Switching multimodel prognostics have been extensively studied to better describe the entire degradation process. However, existing studies typically assume that model transition probabilities are known and constant, which is unrealistic in practical applications. To address this research gap, this study proposes the switching unscented Kalman filter-expectation maximization (SUKF-EM) algorithm. This algorithm first introduces a novel method for integrating health indicators (HIs). Subsequently, to resolve the issue of real-time estimation of model transition probabilities, an objective function for evaluating these probabilities is constructed based on the EM algorithm. Following this, an algorithmic framework for real-time identification using stochastic approximation is derived. Finally, the switching unscented Kalman filtering algorithm is integrated to achieve RUL prediction for bearings. The effectiveness of the proposed method has been validated using run-to-failure experimental data from the Intelligent System Maintenance Center at the University of Cincinnati, as well as bearing datasets from Xi'an Jiaotong University and the Changxing Sumyoung Technology (XJTU-SY).