Open-set combustion-fingerprint classification with out-of-distribution alerting for environmental fire-hazard...

Area of Science:

• Environmental Science
• Chemical Engineering
• Data Science

Background:

• Hazardous fire effluent gases (CO, HCN, HCl, NOₓ) pose significant risks during structural and emerging fire incidents.
• Current fire-gas classifiers operate under a closed-set assumption, failing to identify novel combustion sources like wildland fires or EV battery fires.
• Rapid identification of fire-effluent composition is critical for environmental hazard assessment and respiratory protection.

Purpose of the Study:

• To develop and validate an open-set machine learning framework for fire-effluent hazard screening using combustion fingerprints.
• To integrate closed-set fire material classification with out-of-distribution (OOD) detection for novel hazard scenarios.
• To establish a reproducible, interpretable, and extensible screening layer for environmental fire-hazard monitoring.

Main Methods:

• Utilized multi-parameter combustion fingerprints (7 calorimetry features) from the NIST Fire Calorimetry Database (FCD).
• Employed machine learning models (Random Forest, LightGBM, XGBoost) for classifying four structural fire material classes.
• Integrated plug-in OOD detectors (LOF, Isolation Forest) and validated using leave-one-class-out (LOCO) and real-data tests on wildland combustion scans.

Main Results:

• Random Forest achieved a macro F₁ score of 0.679±0.067 for classifying structural fire materials.
• OOD detectors performed marginally above chance on structural classes but showed better detectability on novel wildland combustion data (AUROC 0.812 for LOF).
• SHAP analysis identified peak heat release rate (HRRmax), peak smoke extinction (Ksmoke,max), and integrated CO₂ (∫CO₂) as dominant features.

Conclusions:

• The developed open-set framework provides a valuable screening tool for environmental fire-hazard monitoring, mapping the capabilities and limitations of combustion fingerprints.
• OOD detectability is scenario-dependent, highlighting the need for robust validation across diverse fire types.
• The study offers an interpretable, extensible baseline for assessing acute environmental contaminant exposure during fire incidents.