Fire type is a critical determinant of hazard escalation, suppression effectiveness, and emergency response strategy. However, current fire detection systems reduce this complexity to binary fire/no-fire outputs, treating fundamentally different fire phenomena equivalently at the detection stage. This limitation is particularly significant given the recent introduction of Class L fire for lithium-ion battery fires (LIB). Despite advances in deep learning, no established framework enables operational multi-class fire-type differentiation using spatiotemporal video analytics. To address this gap, we present a real-time classification framework leveraging flame motion, smoke evolution, and flicker dynamics. A curated CCTV-based dataset was developed comprising four categories: conventional fires (Classes A–F), LIB fires, benign fires, and non-fire scenes. Among evaluated architectures, the SlowFast network achieved the most consistent performance (accuracy 86.5%, precision 88.5%, recall 86.5%, F1-score 87.0). Temporal analysis and activation heatmaps further support model interpretability. By advancing detection beyond binary alarms toward fire-type differentiation, this work provides a basis for risk-informed response strategies, reduced false alarms, and improved operational handling of LIB-related fire hazards. • Real-time system detects and classifies fire types from surveillance video • Fire type classification leverages fire dynamics and spatiotemporal cues • Categorizes LIB, Conventional (A–F), benign fires and non-fire scenes • SlowFast network achieved 86.5% accuracy with strong generalization • Enabling next-generation adaptive, context-aware fire detection systems • Temporal analysis and heatmaps explain model decisions
Ali et al. (Sun,) studied this question.