Tiny Machine Learning for Li-ion Battery State of Health Estimation

The high rise of Renewable Energy Sources (RES) and the stochastic nature of weather conditions cause a need for battery integration into energy storage systems. Sophisticated Battery Management Systems (BMS) are required to supervise and compute the key parameters among which many are not directly measurable using standard sensors. Many data-driven methods are frequently preferred for their estimation due to their superior accuracy, minimal energy consumption, and straightforward implementation on cloud platforms or Internet of Things (IoT) devices. The primary focus of this research is the development of compact, data-driven models tailored for battery State of Health (SoH) estimation prioritizing efficiency in terms of memory and power requirements, rendering them well-suited for integration into IoT devices. Specifically, the target device for implementation is the Infineon Cypress CY8CPROTO-062S3-4343W. The development process entailed offline training of a compact Convolutional Neural Network (CNN) and an Artificial Neural Network (ANN) utilizing a comprehensive dataset derived from five distinct batteries. The on-device deployment and quantization of the models was performed employing the ModusToolBox Machine Learning (MTB-ML) software provided by Infineon. The tests demonstrate satisfactory results for both selected models, with notably good accuracy, particularly in the initial phases of the battery life cycle. Regarding computational load, the ANN exhibits a distinct advantage over the more intricate CNN model.