Abstract: The digital 3D reconstruction process of coal mine fully mechanized working face roadways requires complete and dense coal wall point cloud data. Due to factors such as occlusion and limited viewing angle, the collected coal wall point cloud data of the fully mechanized working face is often incomplete and sparse, which affects downstream tasks and requires coal wall point cloud repair and completion. At present, there is a lack of datasets and network models for underground point cloud completion tasks. Existing models used for coal wall point cloud completion suffer from uneven distribution of point cloud density and loss of point cloud feature information. In order to solve the above problems, a coal wall point cloud completion network model based on residual optimization is designed. Supervised learning is used to learn point cloud feature information, and the complete point cloud is output by minimizing density sampling and iteratively optimizing the residual network. The method collects real coal wall point cloud data of fully mechanized working face underground, preprocesses and screens available data. The method creates a coal wall point cloud missing dataset by simulating random cavities. The missing dataset is used to train the residual optimization-based coal wall point cloud complementary network model. The experimental results show that compared with the classic FoldingNet, TopNet, AtlasNet, PCN, and 3D-Capsule point cloud completion network models, the residual optimization-based coal wall point cloud completion network model achieves the optimal level of chamfer distance, ground shift distance, and F₁ score for the constructed missing and sparse coal wall point clouds, with the best overall completion effect. It is able to achieve effective completion for the actual missing coal wall point clouds.