Human-AI interactive framework for smart evacuation safety analysis in large infrastructures

Abstract

The increasing scale and complexity of large urban infrastructures have led to greater pedestrian concentrations and high risks of crowd-related incidents in emergencies. This study develops an Intelligent Evacuation Prediction Tool (IEPTool) with a human-AI interactive framework for evacuation prediction and safety assessment in large infrastructures. The tool is equipped with a deep learning engine trained from a comprehensive evacuation-simulation database of 66 real-life architectural floor plans, including air terminals, exhibition centers, large stadiums, and various stations. By integrating long-short-term memory (LSTM) networks and generative adversarial networks (GAN), key metrics, including evacuation time, the pedestrian flow rate at each exit, and dynamic pedestrian density distribution, are predicted with a high accuracy of over 90%. Subsequently, a large language model (LLM) is incorporated for interactive risk analysis, enabling intelligent evacuation safety assessments and providing optimization guidance. The integrated graphical user interface allows fast and accurate evaluation of evacuation safety for complex floorplans. This intelligent framework provides practical and reliable support to fire safety design analysis and urban resilience management.