A lightweight fire hazard recognition model for urban subterranean buildings suitable for resource-constrained embedded systems

To enhance spatial utilization, architectural facilities are gradually expanding into underground spaces. However, underground structures pose challenges due to their enclosed nature and poor ventilation, resulting in high fire concealment rapid propagation and significant hazards in underground building facilities. In recent years, fire detection methods utilizing deep-learning visual technology have been widely adopted. Nevertheless, existing detection models are generally tailored for outdoor environments such as forests, and these models often exhibit issues such as large parameter counts and slow inference speeds, which impede the deployment of target detection algorithms on mobile embedded devices. To address these issues, this paper proposes a lightweight fire risk identification algorithm for underground building facilities based on an improved YOLOv8n model, named YOLOv8n-Risk. On a self-established dataset comprising flames, smoke, and electrical sparks, the YOLOv8n-Risk model achieves a recall rate of 83.1%, mAP50 of 89.4%, and mAP50-95 of 64.1%. These metrics represent improvements of 2.2%, 1.8%, and 2.0%, respectively, compared to the original YOLOv8n model. Furthermore, the proposed model has a parameter size of only 2.2 M, reducing the original model’s parameters by 30%. Experimental results demonstrate that the model exhibits strong generalization, high accuracy, and a small parameter size, facilitating its deployment on embedded inspection platforms with resource constraints. This provides robust technological support for advancing the development of smart cities.