Research on Fire Source Recognition and Fire Extinguishing Algorithms Based on Multimodal Fusion and Lightweight...

Area of Science:

• Embedded Systems
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Conventional fire monitoring systems suffer from high false alarm rates and delayed responses.
• Lack of closed-loop control limits the effectiveness of current fire suppression technologies.
• Complex environments necessitate robust and autonomous fire detection and extinguishing solutions.

Purpose of the Study:

• To develop an embedded fire detection, tracking, and extinguishing system.
• To enhance system reliability by reducing false alarms through multimodal fusion.
• To achieve real-time autonomous fire suppression with high accuracy and efficiency.

Main Methods:

• Implemented a heterogeneous cooperative computing architecture with OpenMV4 H7 Plus and STM32F103C8T6 microcontrollers.
• Utilized a decision-level RGB-infrared fusion mechanism with a pruned, INT8-quantized lightweight FOMO model for perception.
• Employed a Bayesian inference-based multimodal fusion framework for decision-making and a PID-PI dual-loop controller for execution.

Main Results:

• Achieved real-time fire detection with 210 ms inference latency and a 1.8 MB model size.
• Reduced system-level false alarm rate to 2.0% after infrared fusion verification.
• Demonstrated an average fire recognition accuracy of 95.6% and extinguishing accuracy within ±5 cm.
• Completed the perception-to-extinguishing cycle within 8.5 seconds under varying illumination.

Conclusions:

• The proposed system offers a robust, real-time, and energy-efficient solution for autonomous embedded fire-fighting.
• Multimodal information fusion and lightweight neural models are effective in overcoming limitations of conventional systems.
• The system provides a practical and scalable solution for diverse environments including household, industrial, and warehouse settings.