FireMAS: Improving Fire Detection System by Leveraging Vision Language Models Using Multi-Agent System Framework

In industrial fire detection scenarios characterized by high-ceiling environments, deep learning-based methods exhibit superior efficacy by relying on visual data rather than on smoke density or thermal gradients. However, these models are prone to generating a high rate of false alarms, a problem that is difficult to mitigate due to their inherent black-box nature. To address this limitation, this paper introduces the FireMAS system, which utilizes state-of-the-art Vision-Language Models to incorporate environmental context into model predictions. The approach employs a multi-agent mechanism where independent agents analyze the scene from diverse global and local perspectives and collaboratively validate fire events, thereby reducing false alarms and improving robustness. This system achieves enhanced detection performance by decreasing false positives, resulting in a more reliable detection framework. To the best of our knowledge, FireMAS is the first work to integrate a multi-agent system for incorporating semantic contexts with a deep learning model at the inference stage in the industrial fire detection setting. The integration of our proposed system with a detection model improves the Area Under the Receiver Operating Characteristic Curve (AUROC) by an average of 0.18 points and, in low false alarm regions, by a margin of 11.24 points on industrial datasets. A detailed analysis of the system’s effectiveness confirms that our method can be effectively applied in industrial fire detection use-cases.