Chen Hsin-Yu, Lin Ching-Chih, Horng Ming-Huwi, Chang Lien-Kai, Hsu Jian-Han, Chang Tsung-Wei, Hung Jhih-Chen, Lee Rong-Mao, Tsai Mi-Ching
Department of Mechanical Engineering, National Cheng Kung University, Tainan 701, Taiwan.
Department of Computer Science and Information, National Pingtung University, Pingtung 900, Taiwan.
Materials (Basel). 2022 Aug 17;15(16):5662. doi: 10.3390/ma15165662.
Due to its advantages of high customization and rapid production, metal laser melting manufacturing (MAM) has been widely applied in the medical industry, manufacturing, aerospace and boutique industries in recent years. However, defects during the selective laser melting (SLM) manufacturing process can result from thermal stress or hardware failure during the selective laser melting (SLM) manufacturing process. To improve the product's quality, the use of defect detection during manufacturing is necessary. This study uses the process images recorded by powder bed fusion equipment to develop a detection method, which is based on the convolutional neural network. This uses three powder-spreading defect types: powder uneven, powder uncovered and recoater scratches. This study uses a two-stage convolutional neural network (CNN) model to finish the detection and segmentation of defects. The first stage uses the EfficientNet B7 to classify the images with/without defects, and then to locate the defects by evaluating three different instance segmentation networks in second stage. Experimental results show that the accuracy and Dice measurement of Mask-R-CNN network with ResNet 152 backbone can reach 0.9272 and 0.9438. The computational time of an image only takes approximately 0.2197 sec. The used CNN model meets the requirements of the early detected defects, regarding the SLM manufacturing process.
由于具有高度定制化和快速生产的优势，金属激光熔化制造（MAM）近年来已在医疗行业、制造业、航空航天和精品行业中得到广泛应用。然而，选择性激光熔化（SLM）制造过程中的缺陷可能源于选择性激光熔化（SLM）制造过程中的热应力或硬件故障。为了提高产品质量，在制造过程中使用缺陷检测是必要的。本研究利用粉末床熔融设备记录的过程图像开发了一种基于卷积神经网络的检测方法。该方法针对三种铺粉缺陷类型：粉末不均匀、粉末未覆盖和刮刀划痕。本研究使用两阶段卷积神经网络（CNN）模型完成缺陷的检测和分割。第一阶段使用EfficientNet B7对有无缺陷的图像进行分类，然后在第二阶段通过评估三种不同的实例分割网络来定位缺陷。实验结果表明，以ResNet 152为骨干网络的Mask-R-CNN网络的准确率和Dice度量分别可达0.9272和0.9438。一幅图像的计算时间仅约为0.2197秒。所使用的CNN模型满足SLM制造过程中早期检测缺陷的要求。
