基于显微视觉与微力信息的微装配技术研究

CASIA OpenIR > 毕业生 > 博士学位论文

	基于显微视觉与微力信息的微装配技术研究
	刘松
	2017-05-24
学位类型	工程博士
中文摘要	本论文主要针对工业中薄壁环形微零件以及细长轴型零件装配过程中的相关问题进行了研究，包括如何快速、有效、可靠的获取多路显微视觉系统之间的相对视角信息，如何充分利用多路显微视觉系统提取的图像信息准确估计微零件在三维空间下六个自由度上的相对位姿误差，如何对操作手以及运动平台进行精密运动控制以实现微零件在三维笛卡尔空间下的位姿对准，以及如何合理利用装配插入过程中微零件之间相互接触的力信息在不损失最终装配精度的前提下实现微零件的高效快速装配。论文的主要研究工作以及贡献如下：首先，基于搭建的微装配平台，设计开发了三维空间下薄壁环形零件自动装配流程以及相应的控制策略。在流程上，装配过程包括对准阶段以及插入阶段，而对准阶阶段被解耦为角度对准阶段以及位置对准阶段。此外，提出了基于图像雅可比矩阵的多路显微视觉信息融合方法来估计微零件在三维空间的相对位姿误差，然后基于PI控制器实现微零件的六自由度对准。最后，提出了一种基于微零件之间的接触力信息的微零件插入逻辑控制算法。最终的实验结果表明，采用所提出的装配方法，能够在一分钟内实现薄壁环形零件的装配，最终的装配成功率为95%。第二，设计了一种用于在多孔微零件图像上高效快速提取微孔特征的算法。该算法主要解决了现有算法不能很好满足在大尺寸多孔图像上快速准确定位小孔位置的要求。所提出的检测算法较累计随机霍夫变换（PPHT）在实时性以及检测准确性等关键指标上都有大幅提高。本算法在合理利用大尺寸多孔图像自身特征的基础上，将小孔特征提取分为如下四步：第一步，对原始图像下采样，获得的图像尺寸较原始图像大幅缩小；第二步，在下采样后的图像上快速定位孔内参考点，用于粗略定位目标孔在原始图像上位置；第三步，对孔内参考点做进一步分析，剔除假的孔内参考点以避免第四步不必要的计算，同时对剩下的参考点分组，判断其是否在位置上属于同一个小孔区域；第四步，基于前面提取的孔内参考点在原始图像上扫描小孔边缘点，然后基于RANSAC算法拟合获取小孔的图像几何信息。第三，提出了针对细长轴型微零件装配的相对位姿估计方法。首先，细长轴型零件的轴向尺寸远大于径向尺寸，从而不能采用在薄壁环形零件中的基于垂直路视觉提取的截面图像特征定位零件的中心位置；然后，由于本文细长轴型零件具有奇数棱特征，在水平路视觉中提取的其边缘线中线特征与零件轴线在水平路视觉中的投影不重合，使得无法仅仅通过水平路视觉定位微零件。为解决这些技术问题，提出了一种充分利用多路视觉系统之间相对视角关系的位姿估计方法。实验结果表明，本方法能有效实现细长轴型零件在六个自由度上的精确对准，并最终顺利将轴形零件引导至孔形零件中。最后，针对轴形微零件插入装配过程中效率低的问题，提出了一种基于随机状态转移过程的插入建模与控制方法。该插入控制方法较现有插入控制方法在效率上提高了四倍。与现有插入控制方法相比，本方法有两个内在的优势确保其具有较高的装配效率。首先，通过将针对水平力调整的预准备动作融入到插入动作中，避免了因为将水平力调整与插入进给分离造成的效率低下问题；然后，本方法针对当前装配状态采用一种概率方法最大限度的估计进给深度，而并非采用的固定进给深度。具体来说，微零件的插入装配过程被建模为随机状态转移过程，而状态转移的不确定性由高斯分布所描述。用于描述状态转移过程的状态转移方程建立在大量分析历史装配数据以及零件的通用机械性能的基础上。此外，算法引入并精心设计了评价函数用于评估状态转移函数的性能，评价函数的评估结果则决定了算法在插入过程中采用保守亦或激进的插入策略。最后，针对当前状态所采取的动作通过N步迭代估计的算法决定。在薄壁环形零件上的装配实验结果表明：基于本方法的插入过程能够在大约30个控制周期内实现其插入装配，相较与现有装配方法的120个控制周期，本方法具有明显的效率优势。最后，论文对文中所提出的研究成果进行了总结，并对未来的研究工作进行了分析与展望。
英文摘要	This thesis focuses on investigating the common key techniques in the field of microassembly to contribute to the development and practicality of microassembly in social industrial economy. Under the support of National Natural Science Foundations of China, we studied the technical problems in the assembly of long cylindrical components and thin annular components, including how to fully utilize the pose information extracted by multi microscopic vision to estimate the relative pose differences between components in 3-Dimensional (3-D) space, how to effectively get components aligned in the Cartesian space in 6 degree of freedoms (DOFs), how to both effectively and rapidly insert one component into the other while the assembly precision is still guaranteed. The main work and contributions are listed as follows: Fristly, we developed an integrated assembly procedure and the corresponding control strategy for the automatic precision assembly of thin annular components in 3-D space. We proposed the aligning approach for the first time in precision assembly, which decouples the alignment of millimeter level size components into the orientation alignment and the position alignment. We also developed a robust method to fuse the pose information extracted from three orthogonally mounted microscopic cameras in order to reliably get the components aligned. The insertion of the components is based on a force-guided controller, which can reliably realize the insertion task. Experiment results demonstrate that the proposed assembly approach and the correspondingly established assembly system can finish one assembly task within one minute, with success rate of 38/40. Second, we developed a fast and accurate circle detection algorithm for prous components. The algorithm is dedicatedly developed for circle localization in large size images for industrial applications. The algorithm outperforms the traditional circle detection method, i.e., Progressive Probablistic Hough Transform in both accuracy and real-time performance. There are four necessary steps for the algorithm to work, such as the down sampling of the original image, the fast localization of candidate points inside the circles, the analysis of the candidate points, and circle edge points detection and fitting. Thirdly, we also developed the relative pose estimation method for the assembly of long cylindrical components, which is different from the method for thin annular components. Firstly, the axial length is far larger than the diameter for long cylindrical components. Therefore, it’s not pratical to use the vertical microscopic camera for component positioning as we do for thin annular components. Secondly, since the projection of the central axis of the component in the horizontal cameras is not the middle line of the visial side edges, the component cannot be positioned with just horizontal cameras. To solve the problems, the proposed relative pose estimation method fully utilizes the relative view relations among different cameras. Experiment results demonstrate that the proposed relative pose estimation method can reliably realize the alignment of long cylindrical components and finally get the components assembled. Finally, we conducted research aiming to improve the efficiency of the traditional force-guided insertion control method for precision assembly of cylindrical components. Our proposed approach improved the efficiency dramatically for at least 4 times better. It has two inherent advantages compared with existing insertion methods in the domain of precision assembly. Firstly, it does not need preparatory actions dedicated for horizontal forces before insertion actions, which is realized by integrating the preparations into the insertion step. Secondly, instead of inserting with fixed incremental depth, it estimates the insertion depth for every action to be as large as possible with a probabilistic approach. Specifically, the insertion of components is modeled as a stochastic state transition process with the uncertainty described by Gaussian distribution. The state transition function is well defined based on analysis of the historical assembly data and the universal mechanical properties of the components. An assessment function is also elaborately designed to evaluate the performance of the state transition function in order to stimulate the control strategy to behave progressively or conservatively. Finally, the action to be taken for the current state is decided by iterative calculations in N steps estimation manner. Experiments on the insertion of thin annular components demonstrate that: with the proposed insertion control method, the insertion process can be finished within 30 steps, which is a great improvement compared with the more than 120 steps of traditional insertion control methods. Finally, the results of the research are summarized, and the future wok is analyzed and prospected.
关键词	微装配+显微视觉+位姿检测+图像处理+视觉伺服控制+力伺服控制
学科领域	先进机器人系统与控制
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/14710
专题	毕业生_博士学位论文
作者单位	中国科学院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	刘松. 基于显微视觉与微力信息的微装配技术研究[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
基于显微视觉与微力信息的微装配技术研究.（7010KB）	学位论文		限制开放	CC BY-NC-SA