Direct metal laser solidification (DMLS) has been receiving increasing research interest in the additive manufacturing (AM) industry due to its outstanding performance in producing parts with ultra-high precision and variable geometries. However, the lack of appropriate in-situ disturbance detection techniques specialized for DMLS makes real-time quality control extremely difficult. In this work, a first-principles-based finite element method (FEM) model focusing on the heat transfer behavior of the DMLS process is developed. The model-produced thermal data is processed to reproduce the outputs of two industrially used sensors: the optical tomography (OT) sensor and the melt pool monitoring (MPM) sensor. In-situ sensor data analytics is then performed through the use of a convolutional neural network (CNN) to automatically detect disturbances from each sensor data during the build process in real-time. Transfer learning is applied to the pre-trained AlexNet backbone to efficiently construct the disturbance detection CNNs for each sensor, and overall accuracy of 73.4% and 82.8% is observed for the melt pool and the OT CNN, respectively. To further increase the accuracy, a cross-validation scheme is proposed to combine the results from both CNNs based on their confidence levels. The final results demonstrate that the testing accuracy increases to 90.4% when cross-validation between both sensor data analytics is applied. The developed machine-learning-based algorithms provides a methodology to perform real-time process data analytics and thereby reducing defect formation and increasing production throughput.

直接金属激光固化（DMLS）在增材制造（AM）行业中受到越来越多的研究兴趣，这是由于其在生产具有超高精度和可变几何形状的零件方面的出色表现。但是，缺乏专用于DMLS的适当的原位干扰检测技术，这使得实时质量控制极为困难。在这项工作中，建立了以第一原理为基础的有限元方法（FEM）模型，该模型专注于DMLS过程的传热行为。处理模型产生的热数据，以重现两个工业上使用的传感器的输出：光学层析成像（OT）传感器和熔池监控（MPM）传感器。然后，通过使用卷积神经网络（CNN）进行原位传感器数据分析，以在构建过程中实时自动检测来自每个传感器数据的干扰。将转移学习应用于经过预先训练的AlexNet主干，以有效地构建每个传感器的干扰检测CNN，并且熔池和OT CNN的总体准确度分别为73.4％和82.8％。为了进一步提高准确性，提出了一种交叉验证方案，以基于两个CNN的置信度来组合它们的结果。最终结果表明，在两个传感器数据分析之间进行交叉验证时，测试准确性提高到90.4％。