Abstract: Raw point clouds acquired by three-dimensional laser scanning in coal mine roadways contain a large amount of noise. Existing point cloud denoising methods cannot simultaneously remove multiple types of noisy points and incur high computational overhead, making it difficult to meet the requirements for real-time denoising of underground roadway point clouds. To address these problems, a multi-scale feature fusion network was proposed for coal mine roadway point cloud denoising. A multi-scale feature extraction module was used to fully preserve the fine-grained geometric details of point clouds and progressively fuse global features to construct hierarchical and robust point cloud features. A self-attention module was used to calculate the association scores of high-dimensional point cloud features after multi-scale feature fusion, establish dependencies among points in roadway point clouds, and implement adaptive weighting. A context enhancement module was used to enhance and fuse local and global contextual features and refine point cloud representations. Experimental results showed that the introduction of each of the three modules improved the denoising performance for roadway point clouds, with the self-attention module playing a key role in maintaining the balance between precision and recall. The multi-scale feature fusion network achieved a precision of 0.936 1, a recall of 0.928 7, and an F1 score of 0.932 4, efficiently retaining valid roadway point clouds while removing invalid noisy point clouds. Its number of floating-point operations per second, parameter count, and inference speed were 1.221×10⁹, 1.191×10⁶ parameters, and 2.25 m/s, respectively. The network achieves a good balance between roadway point cloud denoising accuracy and real-time performance.