Abstract: In view of the pain points of safety management and control such as lagging monitoring of gas, dust, and bracket force-bearing, difficulty in identifying personnel's "three violations" behaviors and others in fully mechanized mining faces of Houcun Coal Mine, an intelligent safety early warning platform with a four layer architecture of "perception - communication - analysis - application" is constructed. The platform integrates intelligent sensing terminals and multi-mode communication technology, integrates extreme value elimination - weighted sliding average filtering algorithm, improves YOLOv5 object detection and skeleton posture estimation model, and is equipped with XG-800 type gas sensor and DSM-5000 type dust sensor to achieve multi-source data collaborative early warning, and the identification accuracy rate of YOLOv5 model reaches 95%. The industrial operation shows that the response time for over limit early warning of typical parameters is shortened from 5.0 min to 1.9 min; Within half a year, accidents and near misses are decreased by 30% and 20% respectively; The problem closed-loop processing rate is improved to 94%, the processing cycle is shortened by 25%, and the annual direct economic benefit is approximately 1.5 million yuan. This research provides an engineering demonstration for the construction of intelligent safety early warning closed-loop management system for high-risk fully mechanized mining faces.