An on-edge Machine Learning model to estimate State-of-Health in Electric Vehicle Batteries

In recent years, the electric vehicle (EV) market has experienced a significant expansion, with the aim of meeting the needs of decarbonization and transitioning to sustainable mobility. With advancements in EV technology, there is a growing use of more sophisticated application methods, including better control systems and advanced sensing technologies. One of the most important components of an electric vehicle is represented by the high-voltage lithium-ion battery pack, which serves as the primary energy source for the entire system. While lithium-ion batteries are a good option for their efficiency and high energy storage capacity, their performance decreases over time due to ageing. A key challenge in battery management is the accurate estimation of the State of Health (SOH), which is essential to ensure reliability, safety and effective functionality. This thesis focuses on the development of an on-edge machine learning model for the SOH estimation of electric vehicle batteries. The principal target of the project is to initialize the prototyping of a system capable of operating onboard an electric vehicle while providing real-time SOH estimation. A data acquisition system was designed using a Raspberry Pi 4 Model B and an OBD-II scanner, enabling real-time extraction of battery parameters from the Battery Management System (BMS) of an electric vehicle. The system was tested under real-world driving conditions to collect time series data, including voltage, current, State of Charge (SOC), and temperature of the high-voltage battery pack. The acquired time series data is then processed using a machine learning model, executed locally on the embedded system, to estimate the SOH of the vehicle’s battery pack. To enable the execution of the machine learning algorithm on the Raspberry Pi, we designed a model exporting pipeline. First, the algorithm is trained in the cloud using data from electric vehicle battery packs; then, the resulting model is exported to the embedded system in TFLite format. For model training, we utilized different datasets, including synthetic data generated via a simulation environment and real data coming from drive sessions of electric vehicles provided by a private company. The proposed method shows the potential of using embedded systems in real-time diagnosis of batteries in electric vehicles, contributing to advancements in EV battery monitoring and predictive maintenance.