【作者】 王谦；

【导师】 徐征； 李经民；

【作者基本信息】 大连理工大学 ， 智能制造技术， 2025， 硕士

【摘要】 微小遮光组件是某传感装置中的关键部件,其装配精度直接影响信号传输的准确性和效率。然而,该组件包含硅基薄片等精密易损零件,传统人工装配效率低下且精度有限。同时在微米尺度下,传统视觉系统受限于高倍率成像的景深极浅、视觉模块重复定位精度较低,零件稍有偏移就可能导致图像失焦、特征难以提取。因此,实现微小遮光组件清晰成像、高精度定位的自动装配具有重要意义。本文针对以上挑战,研究并实现了一套视觉引导的微小遮光组件自动装配方法。在硬件方面,搭建了由远心显微视觉、精密运动平台、吸附与夹持末端等模块组成的装配实验平台。各模块紧密协同:视觉模块提供清晰成像,精密平台实现微米级运动定位,吸附夹持机构确保零件无损拾取与安装。同时,通过在线标定与误差补偿校正视觉轴线偏差及各模块安装误差,提高系统定位精度并保证运行稳定。在算法方面,提出了基于远心成像的自动对焦方法,引入多源模糊度评价机制:对采集图像进行Sobel梯度、傅里叶频谱、拉普拉斯锐度等多特征提取融合生成模糊度评价图,再利用预训练的ResNet50深度回归网络估计离焦量,结合斐波那契搜索迭代快速收敛至最佳焦面,实现精确自动对焦。针对微小零件特征提取难的问题,提出了“YOLOv5目标初筛-几何特征精定位”的组合识别策略:先利用YOLOv5神经网络快速检测获取零件粗位置,再基于边缘轮廓进行亚像素级精确定位,并设计鲁棒的局部特征拟合算法以识别零件细小边缘;在初筛-精准定位方法基础上,面向零件短边中点的提取引入了基于辅助线启发的特征点定位方法,解决短边中点提取漂移问题。基于上述软硬件方案,开发了MVC架构的控制系统,制定了包含上料、识别、点胶、插装、压合、检测等环节的视觉引导自动装配流程和错误监控机制,实现了微小遮光组件装配的全流程自动化控制。实验结果表明,该系统能够稳定、高精度地完成微小遮光组件的装配。远心视觉自动对焦后图像清晰度和识别率显著提升,聚焦重复精度控制在±90μm以内,自动聚焦成功率超过95%;视觉检测准确率由85%提升至98%以上;装配成功率由约75%提高到接近90%。本研究提出的视觉引导控制方法有效解决了微小遮光组件装配中的对焦难、定位难和精密对位难题,实现了软硬件一体化的微装配系统。研究成果大幅提高了装配效率和成功率,证明了该套视觉引导控制方法的有效性。更多还原

【Abstract】 The micro-shading component is a critical part of a sensing device,whose assembly accuracy directly affects the signal transmission accuracy and efficiency.However,this component contains precision and fragile parts such as silicon-based thin sheets,and traditional manual assembly is inefficient and accuracy-limited.Additionally,at the micrometer scale,traditional vision systems face challenges due to the extremely shallow depth of field under high-magnification imaging and the low repeatability of visual module positioning.Slight part misalignments can lead to image defocusing and difficulty in feature extraction.Therefore,achieving clear imaging and high-precision positioning for automated assembly of micro-shading components is of significant importance.This paper addresses these challenges by proposing and implementing a vision-guided automated assembly method for micro-shading components.On the hardware side,an experimental assembly platform was developed,integrating modules such as telecentric microscopic vision,precision motion stages,and suction-gripping end-effectors.These modules collaborate closely:the vision module provides clear imaging,the precision stages achieve micrometer-level motion control,and the suction-gripping mechanisms ensure damage-free part pickup and installation.Furthermore,online calibration and error compensation techniques were applied to correct visual axis deviations and module installation errors,enhancing system positioning accuracy and operational stability.On the algorithmic front,a telecentric imaging-based autofocus method was proposed,incorporating a multi-source blur evaluation mechanism.Features such as Sobel gradients,Fourier spectra,and Laplacian sharpness were extracted and fused from captured images to generate a blur evaluation map.A pre-trained ResNet50 deep regression network was used to estimate defocus amounts,which were then combined with Fibonacci search iterations to rapidly converge to the optimal focal plane,achieving precise autofocus.To address the difficulty of feature extraction for tiny parts,a hybrid recognition strategy of"YOLOv5preliminary screening+geometric feature refinement"was designed.First,the YOLOv5 neural network detects coarse part positions,followed by sub-pixel-level precise localization based on edge contours.A robust local feature fitting algorithm was developed to identify fine edges.Additionally,an auxiliary line-guided feature point localization method was introduced to resolve midpoint drift on short edges.Building on the above hardware and software solutions,an MVC architecture-based control system was developed.A vision-guided automated assembly workflow was established,encompassing steps such as material feeding,recognition,dispensing,insertion,pressing,and inspection,along with error monitoring mechanisms,enabling full-process automated control of micro-shading component assembly.Experimental results demonstrate that the system achieves stable and high-precision assembly of micro-shading components.After telecentric vision-based autofocus,image clarity and recognition rates significantly improved,with focusing repeatability controlled within±90μm and autofocus success rates exceeding 95%.Visual detection accuracy increased from85%to over 98%,and assembly success rates improved from approximately 75%to nearly90%.The proposed vision-guided control methodology effectively addresses challenges such as defocusing,positioning difficulties,and precise alignment in micro-shading component assembly,realizing a hardware-software integrated micro-assembly system.The research outcomes substantially enhance assembly efficiency and success rates,validating the effectiveness of the vision-guided control approach.更多还原