This study proposed deep learning-based ultrasonic nondestructive testing for porosity evaluation of additively manufactured components. First, porosity mechanisms according to additive manufacturing (AM) processing conditions were studied using traditional scanning acoustic microscopy and optical microscopy. Second, correlations between ultrasonic properties and porosity content were analyzed. The correlation results showed that the increased porosity content resulted in a decreased ultrasonic velocity and increased ultrasonic attenuation coefficient. Third, various levels of porosities were evaluated using a deep learning model based on a fully connected deep neural network that was trained on raw ultrasonic signals measured in the AM samples. After training, the testing performance of the trained model was evaluated. Additionally, the generalization performance of the pre-trained model was assessed using newly fabricated AM samples that were not used for training. The results showed that the porosity content evaluated by the pre-trained model matched well with that measured via traditional scanning acoustic microscopy, thus demonstrating the feasibility of deep learning-based ultrasonic nondestructive testing for porosity evaluation of additively manufactured components.

这项研究提出了一种基于深度学习的超声无损检测方法，用于评估增材制造组件的孔隙率。首先，使用传统的扫描声学显微镜和光学显微镜研究了根据增材制造（AM）工艺条件的孔隙率机理。其次，分析了超声性能与孔隙率之间的关系。相关结果表明，孔隙率的增加导致超声速度的降低和超声衰减系数的提高。第三，使用基于完全连接的深度神经网络的深度学习模型评估各种孔隙率，该模型基于AM样品中测量的原始超声信号进行训练。训练后，评估训练后模型的测试性能。此外，使用未用于训练的新制造的AM样本评估了预训练模型的泛化性能。结果表明，通过预训练模型评估的孔隙度与通过传统扫描声学显微镜测得的孔隙度非常吻合，从而证明了基于深度学习的超声无损检测在增材制造部件孔隙度评估中的可行性。