IoT-Integrated Multimodal Large Language Model for Real-Time Fire Detection and Risk Mitigation

Rapid and accurate fire detection is essential for mitigating property damage and safeguarding lives. Traditional fire detection systems, typically reliant on centralized infrastructures, are hindered by latency, network dependency, and privacy vulnerabilities. This paper introduces an on-device large language model (LLM) framework that integrates multimodal sensor data with edge computing to enable real-time fire detection, prediction, and notification. Utilizing lightweight transformer architectures such as DistilBERT and MobileBERT, the proposed system achieves notable enhancements in detection responsiveness and predictive accuracy compared to conventional approaches. Extensive evaluations underscore the framework’s robustness, privacy preservation, and adaptability across varied environmental contexts, illustrating its potential to significantly strengthen community resilience to fire-related threats.