To address challenges in valve pose estimation based on point cloud registration—such as complex backgrounds, occluded or missing features, and noise interference—the article proposes a lightweight graph-spatial attention-hierarchical feature fusion network (LGSA-HFFNet) algorithm. This approach designs and employs a multi-scale parallel convolutional feature extraction layer to enhance feature extraction, prevent gradient explosion during training, and accelerate convergence. It further incorporates a lightweight graph-spatial attention (LGSA) module, a streamlined enhancement combining graph and spatial attention, to overcome neural network difficulties in extracting features from disordered point cloud information, enabling effective extraction of local point cloud features. Finally, the system is validated through pose estimation experiments and deployed in real-world valve pose estimation tasks. Experimental results demonstrate that the LGSA-HFFNet algorithm achieves an average relative translation error of 0.05 m and a rotation error of 0.984 degrees in valve point cloud registration experiments. It exhibits excellent robustness, with translation and rotation registration performance degrading by only 2% and 7.5%, respectively, under complex backgrounds. Registration time is reduced by 80.32% compared to the iterative closest point (ICP) algorithm, with performance significantly outperforming traditional methods like ICP and semidefinite-based randomized sample consensus (SDRSAC). In ModelNet40 benchmark tests, rotational and translational errors were reduced to 2.293 degrees and 0.006 m, respectively, with rotational accuracy reaching an advanced level and translational accuracy showing a significant advantage over existing models. In experiments using real-world valve pose estimation datasets with high noise interference, the proposed method achieves errors of 2.175 7 degrees and 0.036 m, representing reductions of at least 28.98% and 17.81% compared to existing models.