Abstract: 3D Gaussian Splatting has emerged as a cutting-edge technology in radiation field modeling, offering a novel approach to the digitalization of underground coal mines. This advancement has become a pivotal component in developing autonomous navigation systems and digital twins for coal mine robots. However, existing SLAM methods for coal mine robots face challenges when facing factors such as dynamic illumination, lack of depth perception, and complex geometric structures. To address these issues, a high-precision 3D Gaussian Splatting SLAM method for multi-dimensional visual perception enhancement is proposed. First, a multi-stage image enhancement correction algorithm based on HIS space is proposed, which integrates MSRCR, SWF, and normalized gamma correction techniques, achieving RGB image detail enhancement through nonlinear illumination compensation and local contrast optimization. A color map-depth map feature-guided depth-completion network is constructed to fill depth data voids and suppress noise interference, thereby enhancing the quality of raw data. Then, a two-stage keyframe selection mechanism combining hybrid Euclidean distance with overlap ratio is designed to filter the minimum necessary keyframe set, achieving accurate and efficient keyframe selection under the premise of guaranteeing the integrity of the scene representation by using spatial disparity constraints and redundancy control of view coverage. Finally, a Gaussian management strategy with double constraints on transparency and spatiotemporally weighted observation frequency is designed to eliminate invalid Gaussian ellipsoids in low-transparency regions and strengthen the confidence of Gaussian ellipsoids observed stably over a long period, achieving globally consistent and high-quality mapping by combining loop closure detection and pose graph optimization. To verify the effectiveness of the proposed method, extensive comparative experiments are conducted in typical coal mine scenarios, such as mining roadways, relying on an independently designed and integrated mobile robot platform. The results show that the proposed method achieves centimeter-level trajectory tracking accuracy in complex underground coal mine scenarios, reducing the average tracking error by 12.4% compared to the current state-of-the-art baseline method. The Peak Signal-to-Noise Ratio of 3D mapping reaches 22.4 dB while maintaining high real-time performance, which is significantly better than mainstream algorithms, providing reliable technical support for the construction of digital twin mines.