This thesis studies the battery state of charge (SOC) estimation and battery anomaly detection using machine learning technique. Classically, battery SOC is estimated using a model-based approach which requires extensive modeling, experimentation, and validation before accurate SOC estimation. The technique developed in this thesis for SOC estimation is unique and does not require battery temperature or capacity as inputs which are essential for model-based estimation.
Machine learning model is initially developed and validated using the simulated battery data. Later, the model is validated experimentally at different temperatures for NCM (Lithium nickel manganese cobalt oxide) battery. Finally, unsupervised machine learning algorithm is developed for the battery anomaly detection. Most of the battery anomalies reflect either change in battery resistance or diffusivity. Both types of changes are included in the simulated data. Anomaly detection algorithm can cluster out the faulty batteries. The methods developed in the thesis are suitable for on-line and real-time applications as it does not require a battery to be physically present in the laboratory.
