Exploratory study of battery aging analysis with machine learning models to complete multi-physical ones for more adaptable systems

Electric vehicles represent one of the most future proof means of transport to face the energy challenge, as they are less polluting and economically viable. Their batteries are based on lithium-ion technology. Indeed, in addition to its high energy to mass ratio, self-discharge is relatively low compared to other batteries. However, significant work hasn't been done to understand the aging sequences of batteries related to their use. Several solutions have been proposed for batteries State Of Health (SOH) diagnosis but most of them are based on multi-physics methods that do not consider the system's operational condition. Our contribution is to make an exploratory study by applying Machine Learning models for battery SOH prediction from readily available values. We decided to test several supervised and unsupervised approaches in order to train a binary classifier. A battery is considered used when its SOH drops below 80%, and normal otherwise. Features consists of voltage and current measured at the battery output. We used a wide range of algorithms, ranging from neuron networks, probabilistic models, ensemblist methods and clustering algorithms. We concluded that ensemblist algorithms present the best performances: up to a precision of 84 %.