To address the limited 3D structural perception and insufficient feature discrimination in traditional welding joint classification and lack-of-fusion detection methods, this study proposed a 3D point cloud detection network that integrated geometric structure modeling with an attention mechanism, termed CA-PnPNet. First, the network was built upon the PointNet++ framework, in which a point neighborhood processing in 3D(PnP3D) was integrated into multiple feature extraction stages to strengthen the modeling of local spatial geometric relationships. In addition, a channel attention module (CAM) was incorporated to adaptively emphasize key features by capturing semantic dependencies across channels. Finally, the collaborative integration of these two modules at different feature layers enabled unified enhancement of both local point cloud geometric representation and semantic feature expression, resulting in more comprehensive 3D structural characterization. To validate the effectiveness of the method, multiple sets of comparative experiments were conducted. The results demonstrate that CA-PnPNet achieves an accuracy of 97.7% in the welding point cloud classification task, outperforming the baseline model by 1.9%, while improving the inference speed from 33.3 FPS to 36.1 FPS. These results validate the superior accuracy and real-time performance of the proposed method. Overall, CA-PnPNet provides an effective technical reference for intelligent detection and industrial quality monitoring of complex welded structures.