Defect Detection and Monitoring in Metal Additive Manufactured Parts through Deep Learning of Spatially Resolved Acoustic Spectroscopy Signals
Defect Detection and Monitoring in Metal Additive Manufactured Parts through Deep Learning of Spatially Resolved Acoustic Spectroscopy Signals
复制标题
DOI：
10.1520/ssms20180035
复制
发表时间：
2018-01-01
影响因子：
1
通讯作者：
Samal, Ashok
中科院分区：
文献类型：
--
作者：
Williams, Jacob;Dryburgh, Paul;Samal, Ashok
激光粉末床熔融（LPBF）是一种增材制造（AM）工艺，通过消除传统减材和成形制造工艺的许多设计和材料相关限制，有望开创制造业的新时代。然而，在安全关键应用中，由当前等级的LPBF系统生产的部件中观察到的缺陷的水平和严重性是不可容忍的。因此，有必要引入信息丰富的过程监控，以评估零件的完整性，同时与制造，以便采取适当的纠正措施，以尽量减少零件缺陷。空间分辨声光谱（SRAS）是一种独特的定位无损声学显微传感方法，已成功地用于探测LPBF零件的机械性能和评估缺陷的存在。然而，该技术是敏感的外来现象，如表面反射率，这发生在LPBF系统内，并可能遮挡识别表面断裂和表面下的缺陷。为了将SRAS技术应用于生产规模LPBF环境中的过程中监测并克服上述挑战，本研究提出使用深度学习卷积神经网络，称为用于多模态图像回归的密集连接卷积块架构（DCB-MIR），其调用该部件的SRAS导出的声速图作为输入数据，并将它们转换成类似于光学显微照片的输出。通过这种方法，我们证明，缺陷，如孔隙度和表面缺陷的钛合金和镍合金试样使用LPBF，这是不清楚地辨别在作为测量的SRAS声学地图，并掩盖了在光学图像中的文物，准确地识别。为了量化该方法的准确性，预测输出图像和目标光学图像之间的余弦相似性被用作DCB-MIR的目标函数。所获取的SRAS特征与样品的相应测量光学显微照片之间的余弦相似性通常在-0.15至0.15之间。相比之下，当光学显微图像来自DCB-MIR在这项工作中提出的光学签名进行比较，余弦相似性改善的范围为0.25至0.60。
Laser powder bed fusion (LPBF) is an additive manufacturing (AM) process that promises to herald a new age in manufacturing by removing many of the design and material-related constraints of traditional subtractive and formative manufacturing processes. However, the level and severity of defects observed in parts produced by the current class of LPBF systems will not be tolerated in safety-critical applications. Hence, there is a need to introduce information-rich process monitoring to assess part integrity simultaneously with fabrication so that opportune corrective action can be taken to minimize part defects. Spatially Resolved Acoustic Spectroscopy (SRAS) is a uniquely positioned nondestructive acoustic microscopy sensing approach that has been successfully used to probe the mechanical properties and assess the presence of defects in LPBF parts. However, the technique is sensitive to extraneous phenomena, such as surface reflectivity, which occur within the LPBF system and may occlude identification of surface breaking and subsurface defects. With a view to applying the SRAS technique for in-process monitoring in a production-scale LPBF environment and to overcome the foregoing challenge, this study proposes the use of a deep learning convolutional neural network that is termed Densely connected Convolutional Block Architecture for Multimodal Image Regression (DCB-MIR), which invokes SRAS-derived acoustic velocity maps of the part as input data and translates them to an output resembling an optical micrograph. Through this approach, we demonstrate that defects, such as porosity and surface imperfections in titanium alloy and nickel alloy specimens made using LPBF, which were not clearly discernable in the as-measured SRAS acoustic map and were obscured by artifacts in the optical image, are accurately identified. To quantify the accuracy of the approach, the cosine similarity between the predicted output images and target optical images was used as the objective function of DCB-MIR. The cosine similarity between the acquired SRAS signatures and the corresponding as-measured optical micrographs of samples typically ranged between -0.15 and 0.15. In contrast, when the optical micrograph-like images derived from DCB-MIR proposed in this work were compared with the optical signatures, the cosine similarity improved in the range of 0.25 to 0.60.