InfernoFusion: An Edge-Optimized Audio-Visual Deep Learning Framework for Early Fire Smoke Detection

The requirement for early, dependable fire detection is very high because it reduces damage and satisfies public safety requirements in large-scale complex environments. The paper presents an audio-visual early fire and smoke detection system based on deep learning for real-time deployment on edge devices. Besides, the system is not limited to vision alone, which can fail at low illumination or be compromised by occlusion due to visual ambiguity, but combines the visual features of fire and smoke with cracks and hissing as complementary acoustic cues to curb robustness. Convolutional neural networks are used to process visual data along with preprocessing and augmentation techniques aimed at enhancing generalization, while audio signals are analyzed using spectral features for the detection of fire-related acoustic patterns. A fusion multimodal strategy combines the predictions of both modalities, which can greatly reduce the number of false alarms and improve detection accuracy. It is also optimized to perform lightweight inference and assessed under different lighting conditions, thus sustaining performance across varied low-lux environments with particular reliance on audio cues in circumstances where visual clarity is compromised. The proposed framework is also scalable to multi-camera environments through parallel edge processing and could be eased into existing fire safety infrastructures involving means such as surveillance systems, alarms, and IoT-based alert mechanisms. The experimental results confirmed the proposed edge-optimized multimodal strategy would give a practical, accurate, and robust solution for early fire detection in real-world applications.