Abstract: In the preparation of Ceramic reinforced metal matrix composites (CRMMC) by Laser-directed energy deposition (L-DED),the degree of segregation and dissolution of the ceramic particles is determined by the melt pool thermal history.It can lead to unstable forming quality.A method for monitoring the quality of the CRMMC deposition layer based on the melt pool thermal history is proposed.To realize real-time monitoring of the CPU hardware with poor parallel computing capability,the lightweight fully depth-separable convolutional neural network (FD-Net) with a single-path structure is constructed.Nine different laser energies were input to prepare different states of CRMMC.The corresponding infrared images of the melt pool were synchronously captured by an infrared thermal camera as the dataset for training and testing the FD-Net model.Performance comparisons were conducted between FD-Net and other state-of-the-art lightweight CNN models.The results indicate that FD-Net can achieve an inference time of 7.9 ms/frame on the Inter-CPU,which is significantly lower than other CNN models.It is proved that FD-Net can realize real-time monitoring of CRMMC quality status on industrial microcomputers.

Key words: melt pool thermal history, convolutional neural network, ceramic reinforced metal matrix composites, laser directed energy deposition, infrared image

摘要： 针对激光定向能量沉积（L-DED）制备陶瓷增强金属基复合材料（Ceramic reinforced metal matrix composites,CRMMC）过程中,成形质量不稳定的问题,提出一种基于熔池热历史的CRMMC质量监测方法。为实现在CPU硬件上的实时监测,构建了单路结构的轻量级全深度可分离卷积神经网络模型（Fully depth-separable convolutional neural network,FD-Net）。输入9个不同激光能量制备不同状态的CRMMC成形质量,使用红外热像仪同步采集熔池红外图像作为数据集训练和测试FD-Net,并与当前先进的轻量级CNN模型进行性能对比。结果表明：FD-Net在Inter-CPU上以7.90毫秒/帧的推理时间实现了高精度监测,显著低于其他CNN模型,证明所提方法可在工业微型计算机上实现CRMMC质量状态的实时监测。

关键词: 熔池热历史, 卷积神经网络, 陶瓷增强金属基复材, 激光定向能量沉积, 红外图像