Construction and Validation of an AI-Based Fire Detection Model Using Fire Simulation Images

This study developed and validated an artificial intelligence (AI)-based fire detection model integrating Fire Dynamics Simulator imagery with the YOLOv11 algorithm to enhance the rapid response in large-scale complex buildings. To overcome the physical limitations of real-world experiments, numerical simulations were conducted for a five-story building to establish a high-purity dataset of 4,315 sets that reflected the dynamic behavior of flames and smoke. Based on this dataset, optimized training was performed on the YOLOv11 model, achieving a high reliability of 95.0% mAP@50 in internal evaluations. Notably, the model validated robust spatial generalization by recording a detection rate of 94.8% in tests using emergency operation center data not included in the training phase. Furthermore, cross-domain validation using 35 real-world fire images identified fires with an average accuracy of 80.2% while revealing limitations regarding atypical smoke and environmental noise. This study establishes an engineering foundation for practical disaster response systems by combining high-precision simulation reliability with the real-time capabilities of AI. The identified limitations serve as a basis for further model refinement through the acquisition of diverse industrial field data.