Reinforcement Learning based Classification of Anomalous Degradation Behavior in EV Battery Pack

Abstract

Despite extensive progress in electric vehicle fault modelling, detection and classification of anomalous behavior in unforeseen circumstances is still a huge challenge in battery safety. The surety of billions of automobiles safety is still a question. Due to insufficient data and unpredictability of the behavior of the batteries, deep learning methods have shown limited success in anomaly detection and classification. Moreover, deep learning models tend to learn the anomalies during training while on testing they try to fit in the anomalies that leads to false negative instances. In recent times, unsupervised and semi-supervised learning techniques are in great demand in this domain. Reinforcement Learning has previously been used in robotics and autonomous driving for interaction, decision making, manipulation and control. Reinforcement Learning can self-learn the internal distribution of the normal and abnormal data and create anomalous instances. In this research, an algorithm is proposed for detecting and classifying anomaly in EV Battery packs using Deep Q-Network and its variant. Synthetic dataset is generated to train the model and testing has been done on lab generated data to ensure the robustness of the proposed model.