Real-time CCTV-based deep learning for early detection of lithium-ion battery fires

Abstract

Lithium-ion battery (LIB) fires pose distinctive hazards—including toxic gas release, resistance to suppression, and potential re-ignition—making their early detection critical for effective mitigation. Current LIB fire detection approaches often rely on embedded sensors or wiring, limiting applicability in portable devices such as laptops, and e-scooters. Moreover, conventional fire detection systems cannot distinguish between LIB and non-LIB fires, limiting situational awareness and delaying appropriate response. This study presents a real-time deep learning framework for LIB fire recognition using CCTV footage, leveraging characteristic spatio-temporal combustion patterns such as jet-like flame projection, abrupt ignition bursts, temporary flame extinction, and re-ignition. A 3D convolutional neural network with a ResNet-based backbone was trained on a custom dataset comprising LIB fires, conventional fires, and non-fire scenes from diverse real-world environments. The model achieved ∼87 % accuracy, with balanced precision, recall, and F1-score, and processed video at 94 FPS with low false-alarm and miss rates. Temporal prediction analysis revealed that short-term classification fluctuations corresponded with actual combustion stages, providing interpretable insights into fire progression. Grad-CAM++ visualizations confirmed the network's focus on relevant LIB fire features. The proposed framework combines high accuracy, interpretability, and operational feasibility, offering deployment potential across residential, industrial, and transportation environments. Its adoption will enable early intervention, improve decision-making, and support safer integration of LIB-powered technologies in everyday life.