Abstract: Aiming at the problems of low accuracy and stability of traditional coal pile identification methods for coal mine belt conveyors under complex working conditions and insufficient integrity of three-dimensional point cloud three-dimensional reconstruction of the coal pile surface, which leads to misjudgment of coal pile, this paper proposes an intelligent coal pile identification method based on geographically weighted principal component analysis (GWPCA) and adaptive α-shape three-dimensional point cloud reconstruction. Aiming at the problems of slow processing speed and low recognition rate caused by the large amount of point cloud data collected from the full view of the depth camera, the target point cloud data on the surface of the belt conveyor is determined by manually drawing the ROI(Region of Interest) local area, and the direct filtering is adopted to effectively filter out redundant points, noise points and isolated data. Realize the rapid and accurate acquisition of point cloud data of coal piles on the surface of belt conveyors. Aiming at the problem of missing point clouds on the side surface of coal piles, an improved Poisson surface reconstruction model based on geographically weighted principal component analysis (GWPCA) is constructed. Through the spatial topological relationship between the conveyor belt surface and the upper surface of the coal pile, the point cloud data completion is achieved by combining the legal direction constraint mechanism, and the minimum convex hull screening algorithm is adopted to eliminate the pseudo-plane interference. Aiming at the problem of surface segmentation of coal piles of different shapes and conveyor belts, the functional relationship between the compactness of the bottom boundary of the coal pile and the standard deviation of local curvature is established to realize the extraction of the bottom contour boundary of the coal pile by the adaptive α-shape algorithm, accurately segment the point cloud data of the coal pile and calculate its equivalent width. Aiming at the problem that the coal pile identification method using a single parameter is easily affected by local sensitive values, resulting in deviation in the identification results, a coal pile identification method based on the joint discrimination of the maximum width/height of the coal pile and the equivalent value is proposed. The expected value of the kernel density function of the local point cloud height value is used to calculate the equivalent height, and the minimum bounding box is used to calculate the maximum width and height values. An intelligent recognition experimental platform for coal pile on belt conveyors was established, and verification experiments for coal pile recognition under different experimental environments were carried out. The results show that under normal lighting conditions, the average errors of the maximum width and height detection are 0.40 cm and 0.35 cm respectively, and the correlation coefficient between the equivalent value of single frame width/height and the maximum value is 0.983, and the average detection speed is 1.15 seconds per frame, effectively reflecting the real state of the coal pile. In a dusty environment, the average errors of the maximum width and height detection are 0.46 cm and 0.47 cm respectively, and the average recognition speed of a single frame is 1.23 seconds. The equivalent values can accurately follow the overall shape changes of the coal pile, achieving accurate and reliable discrimination of the coal pile state.