【摘要】 船舶的管路系统是船舶可靠运行的保障，具有零部件种类多和结构复杂的特点。针对船舶管路系统在维护与替换过程中重建效率低和精度差的问题，对经典的点云深度学习网络PointNet++进行了改进，提出了一种船舶管路的重建方法。通过构建管路部件分割数据集，生成了高质量训练样本；对PointNet++模型采取多尺度采样半径调整和损失均衡化的改进策略，使其能够准确和均衡地捕捉各类部件的局部特征；根据分割结果求解出零部件特征参数和零部件之间的连接关系，实现了拓扑结构的恢复和管路的重建。实验结果显示，改进后的PointNet++模型在管路部件分割精度上取得了提升，对少数类部件的识别和分割更为精准；在管路重建方面实现了较高的精度，同时也保证了安装精度和可替换性，验证了所提方法可以满足实际船舶维修中高精度、自动化重建管路的需求，有效提升了管路替换的生产效率和可靠性。更多还原

【Abstract】 The piping system of a ship is crucial to ensuring reliable operation, characterized by a wide variety of components and structural complexity. Addressing the issues of low efficiency and accuracy in the reconstruction of piping systems during maintenance and replacement, this study refines the classical point-cloud deep-learning network PointNet++ and proposes a dedicated reconstruction method for ship piping systems. A specialized piping component segmentation dataset was constructed to generate high-quality training samples. Improvements to the PointNet++ model include multi-scale sampling radius adjustments and loss function balancing, enabling it to accurately and consistently capture local features across diverse component types. Using the segmentation results, component feature parameters and inter-component connections were determined, achieving topology restoration and complete piping reconstruction. The experimental results show that the improved PointNet++ model significantly enhances the accuracy of component segmentation, especially in recognizing and segmenting minority component classes. High precision was also achieved in the reconstructed piping, ensuring both installation accuracy and interchangeability. The proposed method meets the requirements for automated, high-precision reconstruction in real-world ship maintenance, effectively improving production efficiency and reliability in piping replacement.更多还原