Collaborative Kernel-based Nonlinear Degradation Modeling with Transfer Learning for Remaining Useful Life Prediction

Abstract

A novel nonlinear collaborative modeling method for remaining useful life (RUL) prediction is proposed. This method uses a kernel-based Wiener process (KWP) model, which formulates a nonlinear drift function with the weighted combination of kernel functions. Compared with the existing Wiener process models, this kind of modeling allows characterizing the nonlinearity of degradation more accurately and flexibly. To address the problem of error accumulation and lack of data in long-term prediction, a transfer learning scheme, based on the KWP models, is developed by leveraging multiple historical degradation trends from different units to collaboratively describe the degradation process of interest, even with limited data. The positive transfer learning is realized by introducing cross-correlations into the drift functions for obtaining more robust and accurate results than those obtained by constructing multiple individual models, one for each unit. The unknown model parameters are estimated by a spare estimation algorithm. Then, based on the KWP model, a close-form expression of the RUL distribution is derived for uncertainty quantification. An online framework is also proposed to iteratively predict the RUL. Finally, the proposed method is verified on lithium-ion battery datasets and compared to existing methods. The outcomes demonstrate the effectiveness and superiority of the proposed method for RUL prediction.