CASIA OpenIR  > 毕业生  > 博士学位论文
多机器人舱段装配系统建模与智能测控方法研究
刘兆阳
2024-05-18
Pages210
Subtype博士
Abstract

当前海洋、航天、航空工业中的大型设备,一般由多段舱段构成,主要采用模块化生产,总装车间装配集成的方式制造。随着国家加快由制造大国向制造强国转型的进程,大国重器智能制造依赖的大型舱段装配方法也成为了产业升级的重点对象。同时,随着舱段制造中小批量、多样化的柔性生产需求日益增加,研究在非结构化环境下基于机器人系统的自主智能装配,具有重要的理论意义和应用价值。

实现非结构化环境下的舱段的自主智能装配,当前的理论技术在装配机构、测量方法、控制方法等方面存在诸多挑战:装配机构难以承载不同重量、形状的大型舱段,负载扩展能力有限;在非接触、无靶标的舱段测量场景下,深度学习训练数据采集成本高,位姿估计算法精度不足,无法兼顾大型舱段的大范围快速定位与小范围高精测量;舱段柔性力控算法高度依赖建模,对不同舱段机械接口的泛化性不足。

为此,本文旨在针对上述挑战,提供理论与技术支撑,以实现舱段自主智能装配。研究按装配机器人系统设计与建模、自动测量、智能控制三个方面展开,其主要成果如下:

1. 装配冗余并联机器人执行系统建模方法。针对装配任务研制了基于 4PPPS 型冗余并联机器人的装配执行系统,配合搭载多种视觉传感器的机械臂装配测量系统,共同构成多机器人协调的舱段装配系统。针对所设计的冗余并联机器人,提出一种基于旋量理论的快速运动学、动力学建模方法。基于所建立的机器人运动模型,进一步提出了一种基于微小形变分配的冗余并联机器人负载分配优化算法。同时,分析了 4PPPS 型并联机构的操作灵活性,提出其全姿态空间的快速求解算法。通过一系列仿真实验验证了所提出算法以及所建立模型的有效性及优越性。
    
2. 全局到局部的高精度舱段位姿测量方法。基于本文所构建的主动测量系统,提出一种由全局到局部的舱段相对位姿主动测量流程。通过主动选取测量视点,在大视野范围使用图像数据搜索舱段,定位舱段局部结构,再计算定位结构小范围测量视点,提高局部定位结构点云采集质量,从而提升相对位姿计算精度。针对在测量中使用的舱段图像检测、分割算法,提出一种基于数字孪生合成数据的深度网络训练方法,实现了无需现场数据的图像处理。所提出的网络训练方法,在舱段真实数据上分割精度达到 89.7%,超越真实数据所训练的相同网络。针对流程使用的点云配准算法,提出邻居点卷积算子,以及以其为核心、使用非均匀采样、渐进式重叠区域估计的配准算法框架,实现了低重合度舱段点云的精确、高速、鲁棒配准。所提出的点云配准算法,在公开数据集上多项指标达到最优或次优,在舱段键-槽数据上定位误差低至1.01°/0.94 mm 及 0.74°/1.95 mm。最终,结合孪生系统,实现了所提出的全局到局部测量视点规划流程,实现了无需预设装配机器人工位的自动舱段相对位姿测量,舱段相对位姿测量精度可达 1.82°/3.37 mm。通过视觉引导方法装配 5 mm 间隙舱段,证明所提出测量流程及配套算法的有效性及准确性。
    
3. 融合传统控制与强化学习的智能柔性控制方法。针对超出视觉定位系统精度的装配任务,本文提出一种结合力位混合控制与软演员评论家框架的多阶段舱段装配强化学习柔性控制算法,智能地生成轴孔搜索阶段的搜索轨迹,以及轴孔插入阶段的底层控制器参数,实现舱段的柔性装配。算法通过奖励函数的改进,提升了柔性控制器对轴孔机械接口形状的泛化能力;通过仿真环境分布式训练算法,并使用课程学习方法规划算法目标,实现了强化学习算法的快速稳定收敛。最终,通过在真实系统中部署算法,装配 0.5 mm 间隙舱段,证明了所设计方法在高精度舱段装配任务上的可行性及有效性。

Other Abstract

In contemporary marine, aerospace, and aviation industries, large equipment typically comprises multiple segment modules, predominantly manufactured through modular production and assembled in final assembly workshops. With the nationwide acceleration from a manufacturing powerhouse to a leading manufacturing nation, intelligent manufacturing of large segment assembly methods has become a focal point for industrial upgrading. Additionally, the increasing demand for flexible production in small batches and diversified manufacturing within segment construction underscores the significance and value of researching autonomous intelligent assembly based on robotic systems in unstructured environments.

Current theoretical and technological challenges in achieving autonomous intelligent assembly of cabin sections in unstructured environments include: assembly mechanisms struggling to handle large cabins of varying weights and shapes, with limited load expansion capability; high costs of deep learning training data collection and insufficient accuracy of pose estimation algorithms in non-contact, targetless section measurement scenarios; inability to balance broad-range rapid positioning with small-range high-precision measurement; and a high dependency of flexible force control algorithms on modeling, lacking generalizability across different mechanical interfaces of cabins.

To address these challenges, this paper aims to provide valuable theoretical research and technical support for the autonomous intelligent assembly of cabin sections. It delves into assembly robot system design and modeling, automatic assembly measurement, and intelligent assembly control. The main achievements can be summarized as follows:

1. A modeling method for redundant parallel robotic execution systems was developed for assembly tasks, employing a 4PPPS type redundant parallel robot and an assembly measurement system equipped with various vision sensors, constituting a coordinated multi-robot assembly system. For the designed redundant parallel robot, a rapid kinematic and dynamic modeling method based on screw theory was proposed. Based on the established robot motion model, an optimization algorithm for load distribution of redundant parallel robots based on minor deformation distribution was further proposed. The operational flexibility of the 4PPPS type parallel mechanism was analyzed, introducing a rapid solution algorithm for its total orientation workspace. The effectiveness and superiority of the proposed algorithms and models were validated through a series of simulation experiments.
    
2. A high-precision global-to-local section pose measurement method was introduced, combining with the constructed active measurement system. This method proposes an active global-to-local measurement process of the section's relative pose, by actively selecting measurement viewpoints to search for the section in a wide field of view, locate local structures of the section, then calculate the small-range measurement viewpoints to improve the quality of local structure point cloud, thereby enhancing the calculation accuracy of the relative poses. For the cabin image detection and segmentation algorithms used in measurement, a deep network training method based on synthetic data from a digital twin was proposed, achieving image processing without the need for on-site data. The proposed network training method achieves a segmentation accuracy of 89.7\% on real data from the cabin segment, surpassing the same network trained on real data. For the point cloud registration algorithm used in the process, a neighbor point convolution operator was proposed, along with a registration framework using non-uniform sampling and progressive overlap estimation based on it, achieving precise, fast, and robust registration of low-overlap cabin point clouds. The proposed point cloud registration algorithm achieves optimal or sub-optimal results on multiple metrics on public datasets, and on the key-slot data of the cabin segment the positioning errors are reduced to 1.01°/0.94 mm and 0.74°/1.95 mm. Finally, combined with the twin system, the proposed global-to-local measurement viewpoint planning process is implemented, achieving automatic relative pose measurement of the cabin segment without pre-setting assembly robot positions, with a relative pose measurement accuracy of up to 1.82°/3.37 mm. Assembling a 5 mm gap cabin segment through visual guidance proves the effectiveness and accuracy of the proposed measurement process and supporting algorithms.
    
3. An intelligent compliant control method combining traditional control and reinforcement learning was proposed for assembly tasks exceeding the precision of visual positioning systems. This method integrates hybrid force-position control with soft actor-critic framework in a multi-stage section assembly reinforcement learning compliant control algorithm, intelligently generating search trajectories for the peg-hole searching phase and bottom-level controller parameters for the peg-hole insertion phase, achieving compliant assembly of the cabin sections. The algorithm improves the generalization capability of the compliant controller for the mechanical interfaces of the cabin sections through refinement of reward functions. By employing distributed training algorithms in simulation environments and using curriculum learning methods to plan learning objectives, rapid and stable convergence of the reinforcement learning algorithm was achieved. Finally, by deploying the algorithm in a real system and assembling a 0.5 mm gap segment, the feasibility and effectiveness of the designed method for high-precision segment assembly tasks are demonstrated.

Keyword轴孔装配 冗余并联机器人建模 工业数字孪生 点云配准 强化学习
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/56501
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
刘兆阳. 多机器人舱段装配系统建模与智能测控方法研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
Thesis.pdf(126199KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[刘兆阳]'s Articles
Baidu academic
Similar articles in Baidu academic
[刘兆阳]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[刘兆阳]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.