Generating Synthetic Sensor Data to Facilitate Machine Learning Paradigm for Prediction of Building Fire Hazard.

Area of Science:

• Fire Safety Engineering
• Computational Fluid Dynamics
• Machine Learning Applications

Background:

• Accurate fire localization is crucial for effective firefighting and building safety.
• Existing methods for fire simulation and data analysis can be complex and time-consuming.
• Machine learning offers potential for automated analysis of fire sensor data.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting fire location using simulated building sensor data.
• To introduce CData, an automated tool for generating CFAST fire simulation inputs and summarizing results.
• To assess the performance of Support Vector Machine, Decision Tree, and Random Forest algorithms in fire localization.

Main Methods:

• Utilized the CFAST (Consolidated Fires and Smoke Transport Model) fire simulation engine to generate time-series sensor data.
• Developed CData, an automated process for creating CFAST input files and analyzing simulation outputs.
• Trained and tested Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF) classification models on synthetic temperature data.

Main Results:

• Decision Tree and Random Forest models achieved high accuracy (93-96%) in predicting fire location.
• Support Vector Machine performance was found to be sensitive to the size of the training dataset.
• Feature importance analysis using DT and RF results identified key input variables for fire prediction.

Conclusions:

• Machine learning, particularly Decision Tree and Random Forest algorithms, shows significant promise for automated fire localization in buildings.
• The CData tool facilitates the generation of synthetic data for training and testing fire detection models.
• This learning-by-synthesis approach enhances situational awareness for firefighting operations.