Point cloud registration is a fundamental task in three-dimensional computer vision,reverse engineering,and intelligent manufacturing,aiming to accurately align multiple point clouds into a unified coordinate system.However,in practical industrial scenarios,a vast majori-ty of scanned objects,such as mechanical gears and cylindrical components,exhibit significant rotational symmetry,which introduces severe correspondence ambiguity during the registration process.This inherent geometric ambiguity often misguides traditional optimization-based algo-rithms,such as the Iterative Closest Point and its variants,causing them to establish erroneous feature matches and inevitably become entrapped in incorrect local minima.The problem is fur-ther exacerbated in real-world environments because industrial point clouds typically suffer from high-intensity sensor noise,structural occlusions,and local geometric imperfections caused by manufacturing defects,making reliable alignment exceedingly difficult.Furthermore,while re-cent data-driven deep learning methods have shown promise,they heavily rely on massive a-mounts of identically distributed training data.To address these critical challenges,this paper proposes a novel,interpretable,and geometric prior-guided unsupervised point cloud registration framework specifically designed for objects of rotational symmetry.The framework consists of three stages.First,to overcome the instability of traditional principal component analysis in the presence of noise and missing regions,the algorithm estimates a robust rotational axis prior by leveraging the intrinsic spectral structure of the point cloud's covariance matrix.To further en-hance reliability,a geometric consistency constraint is introduced based on the observation that surface normals in the lateral regions of the object are approximately orthogonal to the true rota-tional axis.By combining this orthogonality constraint with a random sampling strategy and a point-to-axis distance histogram to filter out outliers and identify dominant symmetric structures,a highly robust rotational axis prior is successfully regressed.Second,based on this reliable pri-or,an innovative directional Branch-and-Bound global optimization strategy,termed Axial-BnB,is developed for coarse registration.Instead of exploring the entire SO(3)rotational space,the proposed method restricts the search for transformation parameters to a highly constrained coni-cal neighborhood centered around the estimated rotational axis.To accommodate potential minor deviations in the initial axis estimation,an adaptive expansion strategy is incorporated,allowing the search radius to gradually enlarge as the rotation angle increases,which significantly prunes the search space,boosts computational efficiency,and drastically reduces symmetry-induced am-biguity while strictly preserving the ability to approach the global optimum.Third,following the coarse alignment,a fine registration stage based on fuzzy clustering is employed to achieve high-precision alignment.By representing the dense point cloud with a sparse set of cluster centers and their corresponding covariance matrices,the method efficiently captures the local geometric trends and structural anisotropy,subsequently refining the pose parameters through a gradient descent optimization.Comprehensive experiments conducted on challenging public datasets and in-dustrial-scenario scanned point clouds validate that the proposed method significantly outperforms exist-ing traditional optimization algorithms and demonstrates vastly superior zero-shot generalization com-pared to contemporary deep learning models.This training-free,highly interpretable,and efficient framework successfully provides a robust solution for the precise registration of rotational symmetry point clouds,showcasing application potential in automated manufacturing.