Lightweight CNN-Based Embedded System for Li-Ion Battery Fault Diagnosis

Abstract

Real-time fault diagnosis in lithium-ion batteries is essential for ensuring the safety and reliability of modern energy systems, yet it remains challenging due to the strict resource constraints of embedded Battery Management Systems. While deep learning methods provide improved robustness under dynamic operating conditions compared to conventional approaches, their high computational complexity limits deployment on microcontrollers. To address these limitations, this paper proposes a lightweight TinyML-based fault diagnosis framework using a multi-scale Convolutional Neural Network optimized for on-device inference. The model processes multivariate time-series data comprising 31 battery features over 32-timestep windows to classify four operational states: Normal, High Resistance, Low Capacity, and Self-Discharge. The architecture employs parallel depthwise-separable convolutions to capture multi-scale temporal patterns, combined with a Squeeze-and-Excitation attention mechanism for efficient channel-wise feature refinement. Evaluated on the CH-BatteryGen dataset using cycle-level partitioning, the proposed model achieves 92.81% classification accuracy with a macro F1-score of 0.914, demonstrating strong and balanced performance across fault classes. Furthermore, the model is optimized through quantization to enable real-time, on-device inference on resource-constrained microcontrollers. These results demonstrate that the proposed framework effectively balances diagnostic accuracy and computational efficiency, providing a practical and scalable solution for intelligent battery fault diagnosis in embedded BMS applications.