Abstract: In response to the complex challenges faced by coal mine safety production in the construction of smart mines, a set of artificial intelligence safety early warning systems for the entire process and multiple scenarios has been established. Through a closed-loop architecture of " perception-analysis-early warning-disposal" ,intelligent control of mine safety production is achieved. In the perception layer,an intelligent perception scheme for mine scenes based on deep learning models is designed to accurately identify complex mine environments. In the analysis layer,a collaborative early warning model for mining, digging transportation,and ventilation is developed to achieve comprehensive monitoring of production links. In the decisionmaking layer,machine learning algorithms and knowledge graph technologies are integrated to build a hybrid intelligent early warning system with cross-domain adaptability. In the application layer,a multi-dimensional perception early warning platform is built,including intelligent visualization of fully mechanized mining faces,dynamic monitoring of mine water and intelligent analysis of underground personnel behavior,achieving real-time and precise monitoring of the entire production process of the mine. Research shows that:① The AI model for mine scene recognition,which integrates DETR and DeepLabV3+,achieves a PA value of 0. 835 and an MIOU value of 0. 825 on high-resolution datasets. Combined with the SAM model,the recognition accuracy for four types of coal-based sites,including open-pit coal yards,underground coal yards,coal power sites and coal chemical sites,all exceed 0. 820. The verification in the Ordos and Ningdong bases achieved recognition rates of 0. 788 and 0. 838, respectively. ② The mine safety early warning system adopts a hierarchical design architecture,which can complete the entire process of intelligent perception and control from data collection and processing in mine production to business logic and application display. ③ The application verification of the system in a typical mine in the Ningdong base shows that the system has good practicability and reliability,providing a practical example for promoting the transformation of traditional mines to smart mines. 

Key words: smart mine,safety warning,deep learning,collaborative warning model,multi-scenario perception