FIRE-EVSim: A BIM and 3D simulation-based framework for enhanced early fire detection in electric vehicle charging infrastructure

The expansion of electric vehicle (EV) charging infrastructure introduces unique fire safety challenges, particularly those associated with lithium-ion battery thermal runaway and high-voltage electrical faults, which differ from conventional fire scenarios. Current fire detection systems are limited by the scarcity of domain-specific training data in EV charging environments. To address this limitation, a simulation-driven data augmentation framework, Fire Incident Response Enhancement via Electric Vehicle Simulation (FIRE-EVSim), is proposed. This platform integrates real-world EV charging station configurations with advanced simulation technologies to generate large-scale synthetic fire datasets. Realistic 3D models are developed using Autodesk Revit for Building Information Modeling (BIM) and rendered in Unreal Engine 5 (UE5) to simulate diverse fire scenarios across varying intensities, environmental conditions, and temporal settings. The system produces 100 distinct fire scenarios and 10,000 annotated images capturing a wide range of fire behaviors and contextual factors. A YOLO11 deep learning model is trained using a hybrid dataset that combines the synthetic data with real fire imagery, demonstrating significantly improved detection performance. Evaluation results show a precision of 0.918, recall of 0.884, mAP@50 of 0.917, and F1-score of 0.901. The proposed vision-based approach leverages existing CCTV infrastructure to provide a low-cost and scalable platform for enhancing EV charging infrastructure safety through smart fire detection systems. • A novel FIRE-EVSim framework generates simulated fire data for EV infrastructure. • BIM and Unreal Engine were integrated to simulate diverse, real-world fire scenarios. • 10,000 annotated fire images were created across 100 controlled fire simulations. • The framework enables scalable, camera-based fire detection using existing CCTV. • YOLOv11 hybrid data-trained attained mAP@50 of 0.917 and F1-score of 0.901. • Contributes to safer EV infrastructure through simulation-driven fire detection.