In nondestructive testing (NDT), geometrical features of a flaw embedded in the material such as its location, length, and orientation are critical factors to assess the severity of the flaw and make post-manufacturing decisions to improve the design. In this study, artificial intelligence (AI) based NDT approach was applied to the ultrasonic oscillograms obtained from virtual ultrasonic NDT to estimate geometrical features of a flaw. First, a numerical model of NDT specimen was constructed using acoustic finite element analysis (FEA) to produce the ultrasonic signals. The model was validated by comparing the simulated signals produced from the numerical model with the experimental data from actual NDT tests. Then, 750 numerical models containing flaws with different locations, lengths, and orientation angles were generated by FEA. Next, the oscillograms produced by the models were divided into 3 datasets: 525 for training, 113 for validation, and 112 for testing. Training inputs of the network were parameters extracted from ultrasonic signals by fitting them to sine functions. Lastly, to evaluate the network performance, outputs of the network including flaw’s location, length, and angle were compared with the desired values for all datasets. Deviations of the outputs from desired values were calculated by a regression analysis. Statistical analysis was also performed by measuring root mean square error (RMSE) and efficiency. RMSE in x-location, y-location, length, and angle estimations are 0.09 mm, 0.19 mm, 0.46 mm, and 0.75°, with efficiencies of 0.9229, 0.9466, 0.9140, and 0.9154, respectively for the testing dataset. Results suggest that the proposed AI-based method has the potential to interpret the oscillograms from ultrasonic NDT to estimate geometrical features of flaws embedded in the material.

在无损检测（NDT）中，嵌入材料中的缺陷的几何特征（例如其位置，长度和方向）是评估缺陷严重性并做出制造后决定以改进设计的关键因素。在这项研究中，基于人工智能（AI）的NDT方法应用于从虚拟超声NDT获得的超声波形图，以估计缺陷的几何特征。首先，使用声学有限元分析（FEA）构建NDT标本的数值模型以产生超声信号。通过将数值模型产生的模拟信号与实际NDT测试的实验数据进行比较，对模型进行了验证。然后，通过FEA生成了750个包含不同位置，长度和方向角的缺陷的数值模型。下一个，模型产生的波形图分为3个数据集：用于训练的525个，用于验证的113个和用于测试的112个。网络的训练输入是通过将超声波信号拟合为正弦函数而从超声波信号中提取的参数。最后，为了评估网络性能，将网络输出（包括缺陷的位置，长度和角度）与所有数据集的期望值进行了比较。通过回归分析计算出输出与期望值的偏差。还通过测量均方根误差（RMSE）和效率来进行统计分析。测试数据集的x位置，y位置，长度和角度估计的RMSE分别为0.09 mm，0.19 mm，0.46 mm和0.75°，效率分别为0.9229、0.9466、0.9140和0.9154。