Cascaded deep learning for flame detection and heat release rate quantification in fire safety

Area of Science:

• Fire Safety Engineering
• Artificial Intelligence
• Thermodynamics

Background:

• Accurate real-time heat release rate (HRR) measurement is crucial for effective fire safety engineering.
• Existing methods may lack efficiency or accuracy in dynamic fire scenarios.

Purpose of the Study:

• To develop and validate a cascaded deep learning framework for non-contact, real-time HRR measurement.
• To integrate visual flame detection with thermodynamic analysis for improved fire risk assessment.

Main Methods:

• A cascaded framework combining an enhanced YOLOv8n (with ECA and BiFPN) for flame detection and a dual-branch CNN for HRR quantification.
• The detection module achieved 95.2% precision and 88.3% recall.
• The HRR module processed spatial and frequency-domain flame features, demonstrating high accuracy (R²=0.976) with significantly fewer parameters than VGG16/ResNet50.

Main Results:

• The integrated framework showed robust performance across various flame phases, validated using the NIST fire database.
• The HRR quantification module outperformed Vision Transformers in terms of Mean Absolute Error (MAE).
• The cascaded approach optimizes efficiency by activating HRR analysis post-detection, reducing false alarms.

Conclusions:

• The proposed deep learning framework offers an efficient and accurate non-contact method for real-time HRR measurement.
• This approach establishes a new paradigm for non-contact fire risk assessment.
• Further validation is needed, particularly during peak HRR phases.