Survival Analysis using Deep Learning for Predictive Aging Models of Batteries in Electric Vehicles

Abstract: With the growing EV market, predictive maintenance of batteries is one of the key challenges faced by the car industry. Since this technology is still nascent, most of the battery data obtained is through simulations which may not give an accurate estimate of the battery behavior in the real world. The goal of this thesis is to predict the probability of occurrence of an event in the future when we have incomplete information about the life cycle of the battery parameters from the entire fleet of cars in operation in real world conditions. We achieve this through Survival Analysis. This technique has proven to be a reliable solution for time to failure analysis. In this thesis we estimate the level of degradation of a PHEV car battery through time based on the data collected till now. Through our software toolchain we have seen well defined estimates till 4 months into the future through statistical analysis. The toolchain parameters are flexible to adapt it on any kind of raw battery data. It has been demonstrated that some of these methods can also be substituted with neural networks as a proof of concept. The results of these deep learning algorithms have a large scope of improvement as we collect more data for the entire fleet in the future. Some critical observations and future research strategies have been discussed at the end.