基于微特征分析的加工表面微缺陷智能辨识方法研究

The automatic detection and high-precision identification of machined surface micro defects is the important demand to improve the surface quality, predict the fault , extend the service life and improve the safety of microelectronics and semiconductor, aerospace and other high-end manufacturing complex parts. In view of the problem that this kind of precision manufacturing industry, which because the identification accuracy of micro defects is not high enough, leading to the early warning of major quality accident is not effective enough. And this project will carry out the research on intelligent identification approach of machining surface defect based on micro-feature analysis. This research includes: (1) studying the spatiotemporal distribution of micro-feature components and the micro-feature self-extraction method, and using these to solve the problem that it is difficult to build an accurate model because of the complicated processing environment; (2) studying the mapping relationship between machined surface features, defect morphology and defect feature descriptions, as well as the micro-feature deep self-learning model, which can solve the problem that the micro-feature is too difficult to describe them accurately; (3) studying the coupling mechanism of mixed micro-feature and unsupervised decoupling method, and using these to solve the problem that the defect features are confused with the noise and difficult to be identified accurately. Through the studies in the project, the theory and method of micro-feature analysis will be established, and a new approach for the intelligent detection of machined surface micro defect which is integrated the function of "Micro Defect Self-extraction, Pseudo Defect Self-taught Learning, Automatic Classification of the True-false Defect" will also be developed. Accordingly, this study will solve the crucial problem that restricting the improvement of product quality in key industries such as microelectronics and semiconductors, aerospace and so on.

加工表面微缺陷的自动检测和高精度辨识，是微电子和半导体、航天航空复杂零件等高端制造领域提高表面质量、预测故障、延长服役寿命和提高安全性的重要需求。针对这类精密制造行业因微小缺陷的辨识准确率不高，导致重大质量事故预警不足的问题，本项目开展基于微特征分析的加工表面缺陷智能辨识方法研究，包括：(1) 研究微缺陷成份频空分布规律及微缺陷无模板自提取方法，解决加工环境复杂难以精确建模的问题；(2) 研究微特征对缺陷形态辨识的作用机理及其自学习描述方法，解决微特征难以准确描述的问题；(3) 研究混合微特征耦合机理及无监督解耦方法，解决微缺陷特征与噪声混淆，难以准确辨识的问题。通过本项目研究，建立微特征分析理论和方法，发展集成“微缺陷自提取-伪缺陷自学习-真伪缺陷自动分类”功能的加工表面微缺陷智能检测新方法，解决制约微电子和半导体、航天航空等重点行业产品质量提升的关键问题。

缺陷检测是微电子和半导体、航天航空复杂零件等工业制造领域提高表面质量、预测故障、延长服役寿命和提高安全性的重要需求，微小缺陷对比度低，绝对尺寸小，相对尺度大且隐藏于复杂背景纹理，导致辨识准确率不高，致使重大质量事故预警不足。现有检测方法面临难以精确建模、缺陷样本稀缺、训练时间长、实时性低、算力代价高的瓶颈问题。上述难题的核心是微特征的深度描述问题，本项目深入研究微特征分析理论和方法，为产品质量的退化预测、微缺陷检测和工艺条件的演化分析提供一种新的方法。 .研究进展包括：.一，针对缺陷低对比度问题，提出了小波域的自适应双伽马矫正算法，从局部和整体提升光照和对比度，多项增强后的评价指标超过现有的先进方案。.二，针对现有语义学习的问题，提出了一种基于相对位置关系的自监督学习算法来提取零件的语义特征，设计亲和力因子增强异常和正常特征的差异。公开数据集上，检测和分割结果分别在公榜上排名第四和第六。.三，针对大尺度缺陷检测的难题，提出预训练分层特征融合算法，设计多元高斯函数对正常特征进行建模，通过马氏距离计算正常与异常特征的差异，进而融合成缺陷定位图。大分辨率图像的平均训练时间<40s，平均检测时间<1s。 .四，发明了一种基于特征自愈的缺陷检测方法，将异常特征敲除，用正常特征替换，将缺陷样本修复成正常样本，再与原样本求差。公开数据集上，纹理类产品取得SOTA检测效果。.五，针对纹理类缺陷检测的实时性需求，提出了一种轻量化重构网络，网络大小<1MB，对1K˟1K的图像，处理时间仅2.84ms。.六、研究激光加工的微孔特征对湍流特性的影响机理，量化在一定的温压差条件下，微穿孔特征包括尺寸、总交换面积和材料厚度对气体渗透率的影响。数值计算结果与实验结果相差4%以内。已开发出在线检测装置。.七、对上微特征分析方法开展实验研究，建立了3种典型产品表面缺陷样本库。.成果与人才培养包括：.发表论文8篇，其中第一标注SCI论文5篇，EI论文3篇；参加国际会议3次，授权国家专利6项和软件著作权1项。培养毕业硕士研究生7人；指导优秀本科生4人。.成果应用：.微特征分析方法和技术在多家企业和国家重大项目上得到有效的应用。