Han Rui, Wang Chenwei, Wang Yuzhong, Zhang Yihui, Guo Wenhua, Zi Yanyang, Zhao Jiyuan
State Key Laboratory for Manufacturing Systems Engineering, Xi'an Jiaotong University, Xi'an, 710049, China.
School of Automation, Beijing Information Science and Technology University, Beijing, 100192, China.
Sci Rep. 2025 Apr 7;15(1):11899. doi: 10.1038/s41598-025-96406-8.
Additive Manufacturing (AM) technology has gained widespread application across various industries due to its capability to directly produce products from computer-aided design models. Among AM techniques, the Electron Beam Selective Melting (EBSM) process has attracted significant attention, particularly in aerospace and automotive industries, owing to its high precision, speed, and excellent material properties. However, various defects, especially internal defects that inevitably arise during the manufacturing process, significantly limit the performance of EBSM parts. In this study, X-ray computed tomography (CT) was utilized to scan EBSM parts, and cross-sectional images were employed to train several state-of-the-art modern object detection models for evaluating their effectiveness in detecting internal defects. Sparse R-CNN demonstrated the best overall performance, while the YOLO series excelled in specific metrics. To further capitalize on the strengths of different detection models, a model ensemble approach, Selective Box Fusion (SBF) is proposed. This approach employs voting and weighted fusion of detection boxes to mitigate errors inherent in individual models. Experimental results show that the SBF ensemble method effectively integrates the advantages of multiple detection models, leading to improvements across various evaluation metrics compared to individual models and other ensemble methods.
增材制造（AM）技术因其能够直接从计算机辅助设计模型生产产品而在各个行业中得到广泛应用。在增材制造技术中，电子束选区熔化（EBSM）工艺因其高精度、高速度和优异的材料性能而备受关注，尤其在航空航天和汽车行业。然而，各种缺陷，特别是在制造过程中不可避免出现的内部缺陷，严重限制了EBSM零件的性能。在本研究中，利用X射线计算机断层扫描（CT）对EBSM零件进行扫描，并使用横截面图像训练几种先进的现代目标检测模型，以评估它们在检测内部缺陷方面的有效性。稀疏R-CNN表现出最佳的整体性能，而YOLO系列在特定指标上表现出色。为了进一步利用不同检测模型的优势，提出了一种模型集成方法，即选择性框融合（SBF）。该方法采用检测框的投票和加权融合来减轻单个模型中固有的误差。实验结果表明，SBF集成方法有效地整合了多个检测模型的优势，与单个模型和其他集成方法相比，在各种评估指标上都有改进。
