在现代制造业中，薄壁框架零件因其轻量化和高强度的特性，在航空航天领域得到了广泛应用。然而，由于其结构刚度低，毛坯的内应力场（在成型过程中产生的残余应力）释放会导致显著的加工变形，影响最终产品的精度和性能。针对这一问题，本文提出了一种基于有限元仿真与深度学习相结合的内应力场预测方法，以解决传统测量技术破坏性强、成本高、数据获取不全等局限性。以铝合金7050-T7451为研究对象，通过构建薄壁结构的三维有限元模型，并利用Abaqus软件模拟粗加工过程，为内应力场的预测提供了新的解决方案。首先，对仿真数据经过归一化处理后，分别构建了规则三维矩阵数据和基于点云的图结构数据，为后续深度学习模型提供输入。为实现加工变形场到初始内应力场的非线性映射，设计了两种深度神经网络模型：一是基于三维卷积神经网络（3D CNN）的3D U-Net模型，采用编码器—解码器结构，通过多层卷积、最大池化和反卷积操作提取并恢复空间特征，利用跳跃连接融合低级和高级信息；二是基于图神经网络（GNN）的模型，利用GraphConv模块对点云中各节点间的空间关联进行信息聚合，挖掘数据内在联系。经多组实验验证，模型在反演初始内应力场方面均取得了较高的预测精度和稳定性。接下来，为验证所提出方法的有效性，本文将深度学习模型反演得到的内应力场与基于裂纹柔度法的实验测量数据进行对比，结果表明两者在应力分布趋势和数值上具有良好一致性，充分证明了基于有限元仿真与深度学习结合的方法能够非破坏性、快速地预测毛坯的初始内应力分布。此外，本文还开发了一套集数据采集、模型推理和结果可视化于一体的软件系统，实现了从实际加工变形场数据中迅速反演初始内应力场，并以三维云图形式直观展示预测结果，为工艺优化、夹具设计和加工参数调整提供了决策依据。本研究的学术价值与工程意义主要体现在：（1）提出了一种基于有限元仿真与深度学习相结合的内应力场预测方法，一定程度上弥补了传统测量技术破坏性强、成本高及数据获取不全的局限；（2）从数据生成机制与模型结构创新两个维度，揭示了加工变形场与初始内应力场的非线性映射规律；（3）开发了集成化内应力场反演软件系统，实现了从加工变形场到初始应力场的非破坏性快速预测，并以三维云图形式直观呈现预测结果。
In modern manufacturing, thin-walled frame beam parts are widely used in the aerospace industry due to their lightweight and high-strength characteristics. However, their low structural stiffness leads to significant machining deformation caused by the release of internal stress field(residual stress produced by forming processes) of the blank, adversely affecting the precision and performance of final products. To address this issue, this study proposes a internal stress field prediction method that combines finite element simulation with deep learning, overcoming the limitations of traditional measurement techniques (e.g., drilling method, contour method, and X-ray diffraction) such as destructiveness, high costs, and incomplete data acquisition. Focusing on aluminum alloy 7050-T7451, a 3D finite element model of thin-walled structures was constructed, and the rough machining process was simulated using Abaqus software, providing a novel solution for internal stress field prediction.First, after normalizing the simulation data, this study generated two types of input data for subsequent deep learning models: regular 3D matrix data and point cloud-based graph-structured data. To achieve nonlinear mapping from the machining deformation field to the initial internal stress field, two deep neural network models were designed: (1) A 3D U-Net model based on a 3D Convolutional Neural Network (3D CNN), featuring an encoder-decoder architecture that extracts and restores spatial features through multi-layer convolution, max-pooling, and deconvolution operations, with skip connections to fuse low-level and high-level information; (2) A Graph Neural Network (GNN)-based model that aggregates spatial correlations among nodes in point clouds using GraphConv modules to uncover intrinsic data relationships. Extensive experiments validated that both models achieved high prediction accuracy and stability in reconstructing initial internal stress fields.Then, to verify the method’s effectiveness, the internal stress fields derived from the deep learning models were compared with experimental measurements based on the crack compliance method. The results demonstrated strong consistency in both stress distribution trends and numerical values, confirming that the integrated finite element-deep learning approach enables non-destructive and rapid prediction of the initial internal stress distribution in blanks.Furthermore, a software system integrating data acquisition, model inference, and result visualization was developed. This system rapidly reconstructs initial internal stress fields from actual machining deformation data and visually presents predictions through 3D cloud diagrams, offering decision-making support for process optimization, fixture design, and machining parameter adjustments.The academic and engineering significance of this research lies in three key contributions:(1) Methodological Innovation: A internal stress field prediction framework combining finite element simulation and deep learning was established, partially overcoming the destructive nature, high costs, and data limitations of traditional measurement techniques.(2) Mechanistic Insights: The nonlinear mapping relationships between machining deformation fields and initial internal stress fields were elucidated from dual perspectives of data generation mechanisms and model structural innovation.(3) Practical Tool Development: An integrated software system was created to enable non-destructive, rapid prediction of initial stress fields from deformation data, visualized via 3D cloud diagrams.
