Abstract: A non-contact measurement method was proposed based on laser scanning， spatial point cloud data processing and depth learning to meet the practical requirements of on-line inspection for the surface defects of cross-rolling piercing plugs for seamless steel tubes. According to the characteristics of seamless steel tube production lines， the detection positions， system structures and data acquisition schemes of plug contours were determined， and the iterative closest point(ICP) registration method was introduced to achieve the registration of the measurement point cloud with the standard CAD model. The corresponding classification number set and gradual shape were designed for the head defects， and the point cloud depth learning method was used to realize the defect accurate classification and quantitative early warning. Aiming at the surface wear defects， the upper threshold of wear depth was set to monitor the wear degree accurately. In order to verify the reliability of the system， a physical simulation platform was built for plug inspection， and a plug model with different defects was customized by using 3D printing technology. The testing results show that the contour inspection errors are less than 0.06 mm， the classification accuracy of head defect may reach 97.7%， and the accuracy degree may reach 98.1%， which meets the requirements of on-line inspection.

Key words: cross-rolling piercing plug, on-line inspection for defect, deep learning of point cloud, defect classification

摘要： 针对无缝钢管斜轧穿孔顶头表面缺陷在线检测的现实需求，提出了一种基于激光扫描、空间点云数据处理及深度学习的非接触测量方法。根据无缝钢管产线特点确定了检测位置、系统构成和顶头轮廓数据采集方案，并引入迭代最近点（ICP）配准方法，实现了测量点云与标准CAD模型的配准。针对头部缺陷设计了相应的分类数集和渐变形态，使用点云深度学习方法实现了缺陷精确分类和量化预警。针对表面磨损缺陷，设置磨损深度上限阈值以实现磨损程度的精确监测。为验证系统可靠性，搭建了顶头检测物理模拟平台，并利用3D打印技术定制了含有不同缺陷特征的顶头实物模型。测试结果表明，表面轮廓检测误差在0.06 mm以内，头部缺陷分类精度可达97.7%、准确度可达98.1%，满足在线检测要求。

关键词: 斜轧穿孔顶头, 缺陷在线检测, 点云深度学习, 缺陷分类