Self-piercing riveting (SPR) is one of the most important joining technologies used in light-weight vehicle body manufacturing. Traditionally, SPR joints are developed through trial-and-error tests on various rivet and die combinations. Given the hundreds of rivets and dies available in SPR joint design, huge combinations exist in the full solution space. How to optimize the joining process at a low cost with high reliability is critical both for new material, new vehicle and new production line development. Instead of relying on experience-based physical testing, data-driven approach has been believed to be a promising way for future automotive manufacturing and design. However, the lack of effective data acquisition and accurate characterization methods for joints becomes a barrier, which limits the volume and availability of joining data. In present research, an automatic identification framework of the geometric parameters on SPR cross-sections using deep learning is proposed, which integrates an innovated and complete flow from the image pre-process to postprocess. Firstly, cross-section images are transformed into material segmentation maps using deep learning. Then critical control point of a cross-section is identified accurately, and totally 9 key geometrical parameters are measured based on the locations and combinations of control points. The present research shows that SOLOv2 and Unet are the best models for SPR cross-section image segmentation. The proposed framework could provide high accurate measurement results (average error < 0.02 mm) with a very short process time (within secs), laying a solid foundation for data-driven design and optimization of joining processes within the whole design space at vehicle level.

自冲铆接 (SPR) 是轻量化车身制造中最重要的连接技术之一。传统上，SPR 接头是通过对各种铆钉和模具组合进行反复试验而开发的。鉴于 SPR 接头设计中可用的铆钉和模具有数百个，整个解决方案空间中存在巨大的组合。如何以低成本、高可靠性优化接合工艺对于新材料、新车辆和新生产线的开发至关重要。与依赖基于经验的物理测试不同，数据驱动的方法被认为是未来汽车制造和设计的一种有前途的方法。然而，缺乏有效的数据采集和关节的准确表征方法成为一个障碍，限制了连接数据的数量和可用性。在目前的研究中，提出了一种基于深度学习的 SPR 横截面几何参数的自动识别框架，该框架集成了从图像预处理到后处理的创新和完整的流程。首先，使用深度学习将横截面图像转换为材料分割图。然后准确识别断面的关键控制点，并根据控制点的位置和组合测量共9个关键几何参数。目前的研究表明，SOLOv2 和 Unet 是 SPR 横截面图像分割的最佳模型。所提出的框架可以提供高精度的测量结果（平均误差 < 0.02 mm），处理时间非常短（几秒内），