考虑结合面特征的多材质部件工业CT伪影机理及高精度边缘提取方法

Multi-material components have broad application in the industrial fields, such as the automobile, the aviation and the aerospace. It needs precision quality control to make sure that the machines work reliably. It is a challenging and promising research topic to measure it precisely by use of industrial computed tomography (CT). However, the complex joint surface of multi-material components and the unclear distribution mechanism of artifacts lead to some difficulties in edge detection, which causes low detection accuracy. Aiming at these problems, this project adopts the strategy of combining simulation and experiment to carry out the following researches. Based on the multi-energy CT theory, the characteristics of multi-material components projection images are analyzed in order to establish precise mathematical model of image segmentation and obtain the artifact of joint surface. Considering the characteristics of the joint surface, the artifact distribution is analyzed and the artifact correction method is established. A high-precision edge detection method is established based on the combination of the single-material volume data, the artifact characteristics and the probability and statistics theory. Based on researches above, a high precision measurement method with industrial CT for multi-material components structure will be established. The research results have an important significance to the development of precision measurement of complex parts and dimensional CT metrology.

由多个零件组成的多材质部件，广泛应用在航空、航天、汽车等工业领域，精密的整体质量检测是其可靠运行的基础，利用工业计算机断层扫描(CT)技术对其进行高精度检测是极富挑战且有前景的研究方向，然而多材质部件结合面结构复杂且伪影分布机理和特性还不明确，导致边缘提取困难检测精度不高。本项目针对该问题，采用仿真和实验相结合策略进行以下研究：基于多能工业CT扫描机理，分析多材质部件投影图像特性，建立投影图像精确分离数学模型，提取结合面伪影；考虑结合面特征，分析伪影分布规律，建立伪影校正方法；融合单材质体数据，基于伪影特性，结合概率统计理论，建立高精度边缘提取计算方法。通过以上研究，可望形成面向多材质部件结构工业CT高精度测量方法，项目的研究成果对于我国复杂零部件精密检测及工业CT高精度测量技术发展具有重要意义。

工业CT高精度检测技术是当前无损检测与精密测量交叉领域的研究新热点，可实现对多材质部件结构精密检测，本项目针对多材质部件工业CT高精度检测难题，采取理论建模、分析、数值仿真与实验相结合的方法，建立了X射线仿真计算方法，分析了稀疏角和有限角重建图像伪影分布特点及投影值波动对图像质量的影响，为伪影机理分析和量化校正奠定了基础；针对CT图像伪影分布复杂，结合等值面方法与概率统计方法，提取基准边缘网格面片提取，以网格面片为基本单元，提出网格面片顶点位置微移动策略，建立了边缘提取高精度计算方法，在一定程度上降低了伪影的不利影响，以提取的边缘点云为输入，建立圆柱面几何特征参数自动计算方法，实现了CT软件操作的简化，同时实验结果显示所提方法计算精度高于行业常用软件VG结果；以射线检测DR图像为输入，基于深度学习U-Net网络模型进行缺陷精确识别，引入在线数据增强算法，提高模型的鲁棒性，设计混合损失函数以缓解类不平衡问题，采用AdamW优化器使模型加速收敛和检测精度提升，改进U-Net模型内部，优化激活函数以提高模型泛化性，完善模型Encoder阶段以提高模型对缺陷特征的提取能力，解决了工业零部件射线检测缺陷自动检测问题，为下一步CT图像3D缺陷智能识别的实现建立了基础。上述研究成果为多材质部件工业CT高精度测量提供了一定的方法和技术支撑，对零部件内部结构精密检测及工业CT高精度测量技术发展具有重要理论意义。