Abstract: To address the challenge of balancing computational complexity and detection accuracy in point cloud-based defect detection for train body surfaces, a lightweight point cloud segmentation method, named PointNet-LSD, is proposed. Based on the PointNet backbone, a local geometric feature aggregation module is introduced in the feature extraction stage, where K-nearest neighbors are utilized to explicitly model the spatial relationships between points and their neighborhoods, thereby enhancing the representation capability for fine-grained geometric structures. Meanwhile, a channel attention mechanism is incorporated to adaptively recalibrate feature channels and suppress noise interference from large-area smooth backgrounds. In the model optimization stage, a teacher–student knowledge distillation strategy is adopted, where PointNet++ serves as the teacher network, and soft label distributions are used to guide the student network to learn more discriminative feature representations, thus improving the generalization ability of the lightweight model. Experimental results on a self-constructed train body point cloud dataset demonstrate that the proposed method achieves an inference speed of 378.59 FPS with only 1.67×10⁶parameters. Compared with the baseline PointNet model, the intersection over union, accuracy, and F1-score are improved by approximately 5.9%, 4.5%, and 2.9%, respectively. The results indicate that the proposed method can effectively enhance the detection performance of small defects while maintaining high computational efficiency, showing promising potential for practical engineering applications.