Advanced battery diagnostics for electric vehicles using CAN based BMS data with EKF and data driven predictive models

Abstract

Accurate evaluations of battery State of Health (SoH) and State of Charge (SoC) are critical for Electric Vehicles (EVs) safety, performance, and durability. This study proposes a novel hybrid diagnostic framework that combines statistical analysis, machine learning, and model-based estimation to improve battery monitoring capabilities. The cell-level voltage, temperature, current, and SoC were measured on 15S2P LiFePO4 battery pack utilizing a CAN interface. By implemting the Extended Kalman Filter (EKF), the impact of sensor noise and model errors was reduced in-comarision with the convetional Coulomb Counting techniques. The Random Forest regression model was used to train for SoH assessment with duty cycles, cycle duration, temperature gradient, and voltage spread, the model performed better than linear regression techniques with projected accuracy. The k-means clustering is used to group the cells with comparable behaviors, and the cells with outlier behaviors were identified using dynamic time warping (DTW) based on temporal deviation. The principal component analysis (PCA) is used for voltage imbalance trends identification and analysis. With the data-driven algorithms, the battery health assessments can be made reliable and interpretable by combining physical estimating techniques. This is a practical and expandable solution for improved EV battery diagnostics and life-cycle management.