Embedded strategy for battery module states estimation using tiny machine learning models

Abstract

The growing adoption of battery powered systems demands advanced Battery Management System (BMS) solutions for accurate State of Charge (SoC) and State of Health (SoH) estimation, ensuring safe and efficient operation. To address this need, the authors propose a practical and scalable algorithm based on state-of-the-art TinyML models capable of handling variable-length input sequences without padding, for the SoH estimation. These models enable embedded SoC and SoH estimation at the module level, effectively functioning as smart sensors within a BMS. Four deep learning architectures were developed for battery state estimation, based on CNN, GRU, LSTM, and Dense layers. The models were converted into TinyML compatible formats using an innovative compression technique based on neural network projection. After projection, the models were fine-tuned for a few epochs to restore their accuracy, and then evaluated using metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), model size, and inference latency. Latency measurements were obtained on an RTX 3060 GPU and are reported only for relative model comparison. For SoC estimation, the models utilized a time window of 60 samples, each containing voltage, current, and temperature measurements, to produce accurate predictions throughout the first life phase of the battery (up to 70% SoH). The projected Pr_LSTM TinyML model was selected for deployment due to its minimal memory footprint (4.0 kB), low latency (0.391 ms), and low error rates (MAE: 0.0114, RMSE: 0.0286). For SoH estimation, the models were trained and tested using partially charging data of varying lengths, reflecting practical real-world conditions. Two input feature sets were considered: the Incremental Capacity Analysis (ICA) curve alone, and ICA combined with voltage data. The projected Pr_CNN_GRU TinyML model, using only ICA features, was selected for its high accuracy and compact size (3.4 kB), achieving an MAE of 0.0060, RMSE of 0.0130, and a latency of 1.607 ms. Finally, an embedded SoC/SoH estimation strategy was proposed to scale predictions from the cell to the module level, enabling potential real-time implementation on the Infineon Mobile Robot (IMR) BMS. • TinyML model optimization via neural projection for embedded SoC/SoH smart sensing. • Variable-length SoH models using partial-charge ICA curves and voltage data • End-to-end BMS framework scaling SoC/SoH from cell to module in real system