AI-powered safe egress time assessment for complex building fire evacuation

Abstract

With the increasing height and complexity of modern buildings, a safe evacuation in fire emergencies becomes more challenging. The use of required safe egress time (RSET) helps evaluate the performance of building fire safety, but its quantification relies on costly computational simulations. This study proposes a deep-learning method to fast quantify RSET and furthermore support fire-evacuation assessment. Firstly, a database were built with 1068 evacuation simulations under buildings like stadiums and airport terminals with different occupancy distributions. A deep learning model based on generative adversarial networks is trained by inputs of building floor plans, initial occupant distribution, and exit capacity, as well as outputs of spatial-temporal occupant density fields. The trained model could effectively recognize the indoor spatial features, reproduce the evacuation process, and predict the RSET with an overall accuracy of 95 %. Additionally, the model can handle varying occupant densities across multiple regions, regardless of whether the regions are connected. The proposed smart design framework offers a novel prediction approach to rapidly estimate the required safe egress time, enables a fast performance-based fire safety analysis for a complex building, which can promote safer building design, also helps optimize fire evacuation process and enhance emergency response.