随着制造业智能化需求的不断提高，基于视觉的机械臂抓取系统的重要性日益凸显。视觉机械臂抓取系统通过视觉输入对场景中的物体进行感知和识别，能够替代人类完成复杂繁琐的重复性工作，提升流水线的生产效率。然而，工业零件通常由几何约束形成的零件模板和具体的参数值实例化得到，这造成工业零件具有类间差异大、类内实例多等特点，这些特点给工业场景下的视觉感知带来了不小的挑战。因此，本硕士论文提出了基于图增强的参数化零件点云理解方法，并初步应用到基于3D视觉的工业零件抓取系统中。本文的工作与贡献如下： 首先，本文提出了基于图增强的零件分析网络（GEPAN），该网络能够针对复杂多变的参数化零件进行特征的提取和理解，实现工业零件的种类识别和参数估计。通过学习同一模板下的部分实例，该网络能够挖掘到零件中隐含的几何关系与拓扑信息，并泛化到同模板下未见过的零件实例。同时，为了缓解工业零件训练数据的紧缺，本文构造了全点云数据集和视角点云数据集，用于评估网络学习能力和泛化能力，以支持相关任务的训练和测试。在全点云数据集上，GEPAN在参数估计准确率的学习能力和泛化能力上比最优方法分别提升了1.4%和1.8%，在视角点云数据集上分别提升了2.3%和4.2%。 其次，本文针对点云深度学习网络决策的不透明问题，开展了点云网络的可解释性研究。本文分别提出基于特征的方法和基于梯度的方法，揭示了点云学习网络在推理过程中所关注的区域，以及输入数据对输出数据的影响大小，从不同角度探索和讨论了点云学习网络的决策偏向性，解释了网络的工作原理。通过相应的可解释方法可视化实验，GEPAN展现出优于其它流行方法的推理合理性和决策可靠性，在工业落地时更为安全。 最后，本文将GEPAN初步应用到了基于3D视觉的工业零件抓取系统中。为了配合点云理解网络进行机械臂抓取控制，本文提出了基于强化学习的机械臂抓取网络。同时，本文搭建了实验室环境下的真实视觉机械臂抓取平台，并模拟了工业零件的分拣和装配任务。应用效果表明，本文的算法和平台能够很好地进行工业零件的感知与抓取，并完成多样的下游任务，拥有良好的落地潜力和应用前景。

With the increasing demand for intelligent manufacturing, the importance of vision-based robotic grasping systems is becoming more prominent. Vision-based robotic grasping systems perceive and recognize objects in the scene through visual input, enabling them to replace humans in completing complex and repetitive tasks, thereby enhancing production efficiency on the assembly line. However, industrial parts are typically formed into part templates through geometric constraints and instantiated into specific parts based on specific parameter values, leading to large differences between classes and numerous instances within classes. The diverse parametric parts present a significant challenge for visual perception in industrial settings. Therefore, this master‘s thesis proposes a graph-enhanced parametric part point cloud understanding method and applies it to a 3D vision-based industrial part grasping system. The work and contributions of this paper are as follows: Firstly, this paper introduces a graph-enhanced part analysis network (GEPAN) that can extract and understand features for complex and variable parametric parts, enabling the recognition of industrial part types and parameter estimation. By learning from partial instances under the same template, this network can uncover the hidden geometric relationships and topological information within the parts and generalize to unseen part instances under the same template. To alleviate the scarcity of training data for industrial parts, this paper constructs full point cloud datasets and part point cloud datasets for evaluating the network‘s learning and generalization capabilities to support training and testing for related tasks. On the full point cloud dataset, GEPAN improves the learning and generalization capabilities of parameter estimation accuracy by 1.4% and 1.8% respectively, and on the part point cloud dataset, it improves by 2.3% and 4.2% respectively. Secondly, this paper addresses the opacity of decisions made by point cloud deep learning networks by conducting research on the interpretability of point cloud networks. It proposes feature-based and gradient-based methods to reveal the regions of interest during the inference process of point cloud learning networks and the impact of input data on output data, exploring and discussing the decision biases of point cloud learning networks from different perspectives and explaining the network‘s workings. Through visual experiments using the corresponding interpretable methods, GEPAN demonstrates superior inference rationality and decision reliability compared to other popular methods, making it safer for industrial applications. Finally, this paper applies GEPAN to a 3D vision-based industrial part grasping system. To complement the point cloud understanding network for robotic grasping control, a reinforcement learning-based robotic grasping network is proposed. Additionally, a real-world visual robotic grasping platform is set up in the laboratory environment, simulating the sorting and assembly tasks of industrial parts. The application results indicate that the algorithm and platform in this paper can effectively perceive and grasp industrial parts, complete various downstream tasks, and have great potential and prospects for practical applications.