Automated shaft-hole assembly is critical in aerospace and lunar infrastructure, demanding extreme precision. We present a monocular vision localization method that fuses semantic and geometric information for high-precision assembly on lunar power stations. Our approach combines deep learning-based detection with geometric analysis in an iterative refinement loop, enabling real-time correction of localization errors. Experimental results show that the proposed method achieves 3D localization of assembly holes with horizontal errors under 0.1 mm and surface normal deviations below 0.3°, even under complex lunar conditions. Compared with traditional methods, our approach greatly enhances detection accuracy and robustness, and exhibits stable convergence during optimization. These results provide a reliable technical foundation for high-precision assembly tasks in space exploration.