A New Incremental Remaining Useful Life Prediction Framework Based on Online Knowledge Distillation and Stochastic Process

Remaining useful life (RUL) prediction in practical engineers usually has the following challenges: (1) the RUL prediction with sequentially-collected online data easily suffers from varying working conditions with no historically-accumulated whole-life data available, (2) degradation randomness, as well as noise interference, will unavoidably result in the accumulation of prognostic error. To address these problems, this paper introduces stochastic process model and proposes an incremental RUL prediction framework based on online knowledge distillation. Using offline whole-life data and early part of online degradation data, the teacher model is first initialized by means of time series recursive prediction. For each online data block that sequentially arrives, the soft-labels are calculated by the teacher model, while the pseudo-labels are further determined by a Bayesian updated Wiener process model. By minimizing the loss between the student model's output and these two kinds of labels, an online knowledge distillation is realized for the incremental RUL prediction. A set of experiments is conducted on the two widely-used bearing datasets IEEE PHM Challenge 2012 and XJTU-SY. The results demonstrate that the proposed framework can periodically update the prediction model and then obtain accurate prediction results on sequentially-collected online data, which offers a feasible solution to actual prognostic tasks.