Shapelet selection based on a genetic algorithm for remaining useful life prediction with supervised learning

Abstract

RUL (remaining useful life) shapelets were recently developed to overcome the shortcomings of similarity-based RUL prediction methods, such as high sensitivity to parameters. RUL shapelets are informative subsequences whose distances to a run-to-failure time series sample are very useful for predicting the RUL of the sample. However, the prediction performance and interpretability highly depend on the set of RUL shapelets, and it is very difficult to compose an optimized set. In this paper, we mathematically formalize the RUL shapelet composition problem with multiple objective functions. In addition, we analyze the characteristics of good RUL shapelet sets and develop a solution methodology based on a genetic algorithm. From the various experiments, we validate that the proposed method outperforms previous ones and suggest how to use the proposed method. The solution methodology developed in this paper can be applied to solve various RUL prediction problems. In addition, the findings on the RUL shapelets can help researchers develop their RUL shapelet-based solution.