Real-time detection and temperature forecasting of large-space building fires using machine learning

In firefighting, timely knowledge of fire behavior is essential but often lacking. This study integrates building design and recorded gas temperatures to determine fire conditions and forecast temperature changes, employing a machine learning system that merges long short-term memory (LSTM) networks with transfer learning. The model is initially trained on datasets comprising 1000 samples from parametric fire models and 200 samples from field simulations, facilitating real-time predictions based on on-site data. Simulations in portal frame buildings show the model achieves over 95% accuracy in fire detection and 90% in gas temperature forecasting. A technique using correlation coefficients and standard deviations effectively identifies damaged thermocouples with over 96% accuracy, assuming a damage ratio under 30%. Validation in two real fire incidents demonstrated over 92% accuracy in fire location and over 89% accuracy in predicting gas temperatures 20 minutes ahead, with processing times of 2.14 seconds and 1.83 seconds, respectively. The machine learning framework also proves resilient to variations in ventilation conditions, making temperature predictions using reliability theory. This framework provides critical insights into fire status and progression for firefighters, contributing to advanced firefighting strategies.