Time Series Feature extraction for Lithium-Ion batteries State-Of-Health prediction

Lithium-Ion batteries are a core component of many devices recently designed. Despite their very high performances, their use in electric vehicles involves sometimes harsh conditions, which results in a limited driving range and a variable cycle life from one vehicle to another. The goal of prognostics and health management applied to Lithium-Ion batteries in electric vehicles is to better understand the ageing mechanisms that take place during the whole cycle life of a battery through the observation of operating data. Predicting with accuracy the State-Of-Health of a battery, in relationship with usage data, is a key step in the development of electric vehicles and of the improvement of any battery powered devices. This paper proposes an approach based on the extraction of features from current, voltage and temperature curves during charge and discharge to predict the evolution of the State-Of-Health of a battery. A sliding window of cycles is used as input to a Long Short Term Memory neural network that outputs a multi-step ahead prediction of the State-Of-Health. Our study is based on several datasets, namely two major benchmarks published by the MIT and the NASA Prognostics Centre of Excellence.