Similarity based remaining useful life prediction based on Gaussian Process with active learning

Abstract

Similarity based methods predict remaining useful life (RUL) by directly measuring the similarities between degradation trajectories, which can be applied when the exact time of equipment operation beginning is unknown and the degradation models cannot be established. However, the similarity measures are usually determined based on experience and experimental results. Establishing the appropriate similarity measures for different applications may be laborious. Besides, similarity based methods only provide deterministic RUL values without evaluating the associated prediction uncertainty. In this paper, a Gaussian process (GP) modeling framework is proposed for RUL prediction and establishing the relationship between similarity based methods and GP, which enriches the variety of similarity measures and improve their interpretability. The proposed framework can provide not only the predicted RUL value but also the associated prediction uncertainty. Besides, a novel active learning method is proposed to relieve the computational complexity of the GP model in consideration of uncertainty calibration, which enables application of the proposed framework in real time. The effectiveness of the proposed framework is illustrated through two case studies based on the C-MAPSS dataset and experimental GaAs degradation dataset.