Fault Prognosis of Roller Bearings Using the Adaptive Auto-Step Reinforcement Learning Technique

This paper presents the implementation of a new adaptation algorithm to model the crack propagation of roller bearings and to predict their Remaining Useful Life (RUL). The developed algorithm is designed based on the adaptive auto-step reinforcement-learning method combined with a crack propagation model. The advantage of this algorithm is that it is now able to not only estimate the defect growth rate online, but also to predict the RUL of a roller bearing element. The presented defect propagation model incorporated in this work is an extension to the Paris’s formula that is well known in the fracture mechanics community. Further, a new adaptive filtering technique, referred to as the auto-step, is presented in this paper and is used to estimate the parameters of the crack propagation model in real-time. The prognosis structure is first compares values of both the predicted and the measured defect sizes, and then, tunes the parameters of the crack propagation model. Simulation results obtained by the auto-step method are then compared with results obtained by the Recursive Least Square (RLS) adaptive filter. The proposed prognosis strategy is distinct itself from other approaches in terms of obtaining higher accuracy as well as faster convergence rate.