Hodges–Lehmann and Vision‐Language Model Based Intelligent Fire Detection in Complex Environments

Fires in industrial and environmental settings pose threats to process safety. Current vision‐based fire detection models face challenges in balancing high accuracy with lightweight deployment, suffer from high false alarm rates, and lack a closed‐loop inspection workflow. To enhance fire prevention capabilities and practical applicability, it is necessary to develop a methodology encompassing robust detection models and intelligent agents for fire safety inspection. Therefore, this study proposes a two‐stage robust fire detection methodology. Particularly, a novel Hodges–Lehmann‐based Robust Attention (HLRA) module is developed, utilising statistical estimators to achieve parameter‐free spatial and channel attention. Based on this, a real‐time detection model named RobustFireDet is proposed, integrating a triangular masked convolution and a robust HL‐IoU loss to improve accuracy while reducing parameters. Subsequently, a classification model named RobustFireNet is designed to mitigate false alarms by focusing on small‐target flames using the CoAlign Loss function. Experiments compare the proposed models with mainstream counterparts. RobustFireDet attains a detection mAP@0.5 of 74.3%, surpassing the second‐best YOLOv11 by 2.2 percentage points with 22.7% fewer parameters. RobustFireNet achieves a small‐flame classification accuracy of 96.9%, surpassing the second‐best EfficientNetV2‐S by 2.5 percentage points. The two‐stage pipeline further suppresses spurious fire boxes by 29.2% and false‐alarm images by 25.4% in an end‐to‐end evaluation. On top of the two‐stage detection results, a Vision‐Language Model‐based intelligent agent called FirePatrolAgent is further developed for scene analysis and recommendation generation, forming a closed‐loop inspection system that completes detection, verification, analysis, and response within 25.43 s per risk event.