Abstract: Aiming at the problems of incomplete spatial perception of fully mechanized caving face in thick coal seam, insufficient cutting control accuracy of shearer, difficult real-time identification of coal gangue state in caving process and insufficient coordination ability of shearer-hydraulic support-back scraper conveyor, taking 81203 intelligent fully mechanized caving face of Baode Coal Mine as the engineering background, the integration and application research of perception-decision-control closed-loop system is carried out. The spatial reconstruction method of working face based on three-dimensional inspection robot and laser simultaneous localization and mapping ( SLAM ) is constructed, and the digital coal cutting planning is carried out by integrating transparent geological model, inertial navigation elevation and actual cutting data. Based on the IU-Net semantic segmentation results, the gangue content on the image surface is calculated. Combined with the memory coal drawing parameter learning and the long-term and short-term memory ( LSTM ) network coal drawing time prediction, the “prediction time-gangue content threshold” double stop criterion is established. Relying on the ground sub-control center, the data of transparent geology, three-dimensional inspection, coal gangue identification, intelligent coal caving algorithm and support electro-hydraulic control system are connected. The industrial test shows that the average positioning error of the three-dimensional inspection robot is 0.046 m, and the maximum data transmission delay is 29 ms. The geological model can support the continuous propulsion of up to 10 coal mining cycles without updating. The automation rate of the shearer is 93.6%, the average control error is 66.92 mm, and the maximum error is 100 mm. The average accuracy of coal gangue recognition in the field image of 81203 working face is 86.62%. When the threshold of gangue content on the image surface is set to 5%, 20%, 30% and 50%, the system can trigger the corresponding hydraulic support to stop automatically. In the pre-working stage test of 81202, the average error between the learning time of memory coal caving and the actual coal caving time is 2.11 s. The results show that the constructed system realizes the closed-loop connection of space perception, cutting planning, coal gangue identification, coal drawing decision and support execution.