Early Flame and Smoke Detection in Valve Halls of Ultra-High-Voltage Converter Stations Using an Attention-Enhanced YOLOv5s Model

Early detection of incipient fire signs, such as dilute smoke emerging at valve hall penetration seals, remains a critical challenge for fire safety in ultra-high-voltage (UHV) converter stations. This study addresses the limitation through controlled valve hall sealing simulations that systematically reproduce the complete fire evolution process—from initial dilute smoke leakage, through dense smoke accumulation, to eventual flame overflow—thereby constructing a high-fidelity dataset tailored to complex converter station environments with low-contrast and varying illumination conditions. Building on this dataset, an attention-enhanced YOLOv5s model integrating multiple attention mechanisms (GAM, CBAM, CA, and ECA) is proposed as the core of an end-to-end visual detection framework. Rigorous experiments and ablation studies validate the framework’s superiority: dilute smoke detection precision improves substantially from a baseline of 57% to 81%, with recall increasing from 61% to 84%; dense smoke and flame detection achieve accuracies of 92% and 98%, respectively. Compared with the original model, false and missed alarm rates are significantly reduced, demonstrating strong robustness against background interference and lighting variations. The proposed method enables reliable, high-precision monitoring across all fire stages, providing a novel technical pathway for early fire warning in UHV converter stations and offering extensibility to other large-scale industrial fire monitoring scenarios.