中国科学院自动化研究所机构知识库

增材制造又称3D打印、快速成型，它以灵活、方便、高度数字化的特点，在多个领域展现出了巨大的潜力。如今，愈加复杂的应用场景对增材制造的效率和零件精度提出了更高的要求。为了实现增材制造在社会范围内的大规模应用，利用人工智能算法对逆向工程获取的缺失点云进行补全，使3D数字模型成为人人都可以轻易获取的资源是很重要的一个方面。此外，受3D打印机系统误差、工艺参数设置欠佳和材料收缩等多种因素的影响，制造出的零件很容易产生变形，直接导致其在一些领域如航空航天中的可用性降低。本文围绕3D模型形状补全、增材制造变形预测和补偿两方面开展研究。首先，通过优化深度神经网络的结构来提高3D形状补全任务的精度，以获取高质量的3D模型。进一步地，对获取的3D模型进行前馈补偿以抵消制造过程中产生的变形，提高增材制造零件精度。本文的研究内容和主要贡献总结如下：

 (1) 本文提出了一种结合体素和隐函数表示的3D形状补全方法。由于遮挡、光反射以及表面材料的透明性等因素，利用3D传感器获取的点云时常会存在孔洞等缺陷，难以重建为增材制造所需的三角网格。基于体素的3D形状补全方法旨在将点云转换为体素表示，利用3D卷积神经网络实现特征提取和形状生成。然而，这种方法受内存和效率的限制难以生成高分辨率3D形状。针对这一问题，本文结合体素和隐函数两种表示方式提升预测的3D形状的几何精度。首先，利用由3D卷积模块组成的编解码器网络预测粗糙但完整的3D形状；然后，在粗糙预测结果的引导下对3D形状轮廓处的点进行采样；最后，提出点预测网络为每个采样点分配一个占用率以实现对3D形状隐式曲面的预测。在通用形状数据集ShapeNet和机械形状数据集CADNET上的实验结果表明，所提方法实现在仅增加少量参数的情况下，使神经网络输出任意分辨率和更高精度的3D形状，在交并比指标上较前沿方法分别取得2.7%和3.4%的提升。   

(2) 本文提出了一种基于注意力增强上下文感知神经网络的点云补全方法。与体素相比，点云更加接近传感器的直接输出形式。点云补全网络直接处理点云数据，无需对其进行体素化预处理，因此大大提升了时间效率。另外，点云只保存3D形状表面的信息，从根本上减小了内存占用。然而，现有点云补全方法独立地预测每个点的位置，忽略了上下文信息，并且，这些方法通常根据从输入不完整点云中提取的全局特征向量来预测完整的3D形状，导致了部分形状细节的丢失。鉴于此，本文提出了一种点云补全解码器实现对上下文信息的利用和探索，并设计了一种注意力增强的跳跃连接结构，通过学习点云不同部分结构的相关性实现高精度点云补全。在通用形状数据集PCN、室外场景数据集KITTI和机械形状数据集CADNET上的实验结果表明，所提方法在倒角距离指标上较前沿方法获得了有竞争力的性能，且更好地实现了精度和复杂度的平衡。

(3) 本文设计了基于神经网络的变形预测和变形补偿框架，并在此基础上提出了一种多尺度特征融合的增材制造零件变形预测与补偿方法。现有基于神经网络的变形补偿方法执行点级别补偿量预测，难以有效理解3D形状，从而不具备对不同3D形状的泛化性。鉴于此，本文构建了变形预测网络和变形补偿网络，作为变形预测和补偿框架的核心组件，实现对增材制造过程中的变形函数及其反函数的近似。进一步地，本文将增材制造中的变形预测和补偿任务建模为逐点偏差预测任务，以3D模型表面点云和法向作为输入，利用多尺度特征融合深度神经网络获取每个点的局部和全局特征，并将所获取特征映射为逐点偏差。本文收集了真实场景牙冠变形数据集，并在所收集数据集上开展计算实验。实验结果表明，变形预测网络的拟合度可以达到47.9%，变形补偿网络将增材制造树脂牙冠的误差由0.0816 mm降为0.0402 mm。

(4) 本文提出了一种基于级联跨模态生成对抗网络的无监督3D牙冠模型生成方法，并开展了虚实数据融合的增材制造零件变形预测和补偿实验。在增材制造变形预测和补偿任务中，训练集样本数量有限导致神经网络泛化性能不强，利用生成对抗网络生成的虚拟数据扩充数据集是提升神经网络性能的重要手段。现有基于生成对抗网络的点云生成结果中含有较大的噪声和伪影，难以重建出平滑的表面。鉴于此，本文设计了一个两阶段的无监督牙冠模型生成框架。在第一阶段，采用位移图引导的生成对抗网络生成具有丰富形状多样性的粗糙牙冠模型。在第二阶段，使用基于图像的生成对抗网络将细粒度细节添加到粗糙网格中。实验结果表明，所提方法能够生成表面平滑、语义合理的3D牙冠形状，并且在弗雷歇DGCNN距离FDD、基于推土机距离的覆盖率COV-EMD、基于推土机距离的最小匹配距离MMD-EMD和基于倒角距离的最小匹配距离MMD-CD四种指标下均优于前沿方法。本文利用生成的虚拟牙冠扩充牙冠变形数据集，并展示了虚拟数据和真实数据在变形趋势上的一致性。实验结果表明，生成的虚拟数据将变形预测和补偿网络的拟合度分别提升了3.8%和2.1%。

综上所述，本文的研究工作均是在加工制造之前对3D数字模型进行优化，设计特定的神经网络提升3D形状补全、增材制造变形预测和补偿任务的性能。本文所设计的3D形状补全算法能够在一定程度上替代费时费力的传统手工建模的方法，提高3D模型获取的效率和质量。在加工制造之前对零件变形进行预测和补偿，可以提高增材制造零件的精度和可用性，降低产品不良率，从而促进增材制造技术在实际应用中的推广和发展。本文从精确化数字模型和高质量3D打印两个方面提升增材制造系统的智能化水平，不仅能够促进增材制造在社会范围内的普及，还能助力增材制造在高端制造领域的进一步应用。
 

Additive manufacturing (AM), also known as 3D printing or rapid prototyping, has shown great potential in many fields due to its flexibility, convenience, and high digitization. Nowadays, increasingly complex application scenarios place higher demands on the efficiency and part accuracy of AM. To achieve the widespread application of AM in society, it is important to use artificial intelligence algorithms to complete defective point clouds obtained by reverse engineering, so that 3D digital models become a resource that everyone can easily access. In addition, due to factors such as 3D printer system errors, poorly set process parameters, and materials shrinkage, AM parts are susceptible to deformation, making it difficult to apply them directly to fields such as aerospace. This dissertation is organized from two aspects: 3D shape completion, and deformation prediction and compensation in AM. First, deep neural networks are designed to improve the accuracy of 3D shape completion for obtaining high-quality 3D models. Second, feedforward compensation is applied to the acquired 3D models to offset the deformation during the AM process and thus improve the accuracy of AM parts. The contents and contributions of this dissertation are summarized as follows.

(1) This dissertation proposes a 3D shape completion method that combines voxel and implicit function representations. Due to factors such as occlusion, light reflection, and the transparency of materials, point clouds obtained by 3D sensors often contain defects such as holes, making it difficult to reconstruct them into triangular meshes. Voxel-based 3D shape completion methods convert point clouds into voxel grids and use 3D convolutional neural networks to achieve feature learning and 3D shape generation. However, limited by memory and efficiency, these methods are difficult to yield high-resolution 3D shapes. To address this problem, this dissertation combines both voxel and implicit function representations to improve the quality of predicted complete 3D shapes.
First, an encoder-decoder network consisting of 3D convolution layers is used to predict a coarse yet complete 3D shape. Then, points at the surface of the 3D shape are sampled with the guidance of the coarse prediction. Finally, a point prediction network is proposed to assign an occupancy to each sample point for predicting the implicit surface of the 3D shape. Experimental results on the generic shape dataset ShapeNet and the mechanical shape dataset CADNET show that the proposed method can output 3D shapes under arbitrary resolution and high accuracy with only a small increase in the number of parameters. It achieves 2.7% and 3.4% improvements in terms of the intersection over union metric compared to state-of-the-art methods, respectively.

(2) This dissertation proposes an attention-enhanced context-aware neural network for point cloud completion. Compared to voxels, point clouds are closer to the direct output from the 3D sensors. The point cloud completion network directly processes point clouds without the need for voxelization preprocessing, which greatly improves time efficiency. In addition, point clouds only store information about the surfaces of 3D shapes, radically reducing memory usage. However, previous approaches predict each point independently and ignore contextual information, and they usually predict a complete 3D shape based on a global feature vector extracted from an incomplete input, resulting in the loss of fine-grained details. This dissertation handles the point clouds directly and proposes a novel decoder to realize the exploitation and exploration of contextual information. Besides, elaborated attention-enhanced skip connections are also devised, which can learn the correlations across different regions in the 3D shape for high-accuracy point cloud completion. Experimental results on the generic shape dataset PCN, the outdoor scene dataset KITTI, and the mechanical shape dataset CADNET show that the proposed method achieves competitive performance over state-of-the-art methods in terms of the Chamfer distance metric as well as a better balance of accuracy and complexity.

(3) This dissertation introduces neural network-based deformation prediction and compensation frameworks in AM, and based on this, it presents deformation prediction and compensation methods that combine multi-scale features. Existing neural network-based deformation compensation methods predict point-level compensation value, making the trained neural network difficult to effectively understand and generalize to different 3D shapes. In this dissertation, a deformation prediction network and a deformation compensation network are constructed as the core components of the deformation prediction and compensation framework, aiming at approximating the deformation function and its inverse function in the AM process, respectively.
Furthermore, this dissertation formulates the deformation prediction and compensation tasks in AM as a point-wise deviation prediction task.
It takes the point cloud and normals of a 3D model as input, uses a deep neural network to learn local and global features for each point, and maps the obtained features to point-wise deviations. This work collects a real-world dental crown deformation dataset and conducts computational experiments on it. The experimental results show that the proposed deformation prediction network can achieve the goodness of fit of 47.9%, and the deformation compensation network can reduce the deformation errors of the printed dental crowns from 0.0816 mm to 0.0402 mm.

(4) This dissertation proposes an unsupervised 3D dental crown model generation method based on a cascaded cross-modality generative adversarial network (GAN), and conducts experiments of AM deformation prediction and compensation in a mixed dataset that combines generated and real data. In AM deformation prediction and compensation tasks, the limited dataset size leads to weak generalization performance of neural networks, and augmenting the dataset with generated data by GAN is a solution. Current GAN-based point cloud generation results contain large noise and artifacts, and it is hard to reconstruct a smooth surface from them. This dissertation presents a tailored two-stage framework for dental crown model generation.  In the first stage, a displacement map-guided GAN is used to generate coarse meshes with diverse shapes. In the second stage, fine-grained details are added to the coarse meshes using an image-based GAN. Experimental results show that the proposed method can generate smooth and semantically meaningful 3D crown shapes, outperforming state-of-the-art methods in four metrics: Fréchet Distance based on Deep Graph Convolutional Neural Network (FDD), Coverage based on Earth Mover's Distance (COV-EMD), Minimum Matching Distance based on Earth Mover's Distance (MMD-EMD), and Minimum Matching Distance based on Chamfer Distance (MMD-CD). This study augments the dental crown deformation dataset with the generated dental crown models, and demonstrates the consistency of the generated and real data in terms of deformation. The virtual data improved the goodness of fit of the AM deformation prediction and compensation network by 3.8% and 2.1%, respectively.

In summary, this dissertation focuses on optimizing the shapes of 3D digital models before an AM process. It designs specific neural networks to improve the performance of the 3D shape completion task, and the AM deformation prediction and compensation tasks. The proposed 3D shape completion approaches can to some extent replace the time-consuming and labor-intensive traditional manual modeling methods, and improve the efficiency and quality of 3D model acquisition. Predicting and compensating for the deformation before the AM process contributes to improving the accuracy and usability of the AM parts, reducing the defect rates of products, and facilitating the promotion and development of AM technology in practical applications. The dissertation aims to improve the intelligence level of the AM system in terms of two aspects: acquiring accurate 3D digital models and performing high-quality 3D printing, which not only promotes the popularization of AM in society but also contributes to its application in the high-end manufacturing field.