Laser powder bed fusion (L-PBF) is a metal additive manufacturing (AM) process with great potential in producing high performance metal components. Due to lack of stability and repeatability of the building process, its wide application in industry is limited. The process monitoring and control are import to ensure product quality. The size and shape of the melt pool are continuously changing during the L-PBF process, which may lead to the generation of defects. To represent the melt pool variations more accurately, a new motion feature is extracted and a classification model is constructed to identify the melting state. Firstly, a 36-dimensional motion feature is obtained by contour unwrapping with respect to the melt pool centroid. Subsequently, a sample dataset of melt pool image including four categories of melting states is established. Finally, a Gaussian process classification (GPC) model is constructed to identify the melting state based on motion feature. To verify the performance of GPC, it is also given that the recognition results based on support vector machine (SVM) model, multilayer perceptron (MLP) and long short-term memory (LSTM) neural network. The research results show that under the advantages of automatically optimizing hyperparameters and providing probability distribution information of melting state, the GPC model can still achieve a better recognition result. The overall recognition rate reaches 87.1%, and the melting state can be better identified. A novel in-situ monitoring idea is provided for the L-PBF in this research.

激光粉末床熔合 (L-PBF) 是一种金属增材制造 (AM) 工艺，在生产高性能金属部件方面具有巨大潜力。由于构建过程缺乏稳定性和可重复性，限制了其在工业上的广泛应用。过程监控是进口的，以确保产品质量。L-PBF 过程中熔池的大小和形状不断变化，这可能导致缺陷的产生。为了更准确地表示熔池变化，提取了新的运动特征并构建了分类模型来识别熔化状态。首先，通过相对于熔池质心的轮廓展开获得 36 维运动特征。随后，建立了包含四类熔化状态的熔池图像样本数据集。最后，构建高斯过程分类（GPC）模型以基于运动特征识别熔化状态。为了验证GPC的性能，还给出了基于支持向量机（SVM）模型、多层感知器（MLP）和长短期记忆（LSTM）神经网络的识别结果。研究结果表明，在自动优化超参数和提供熔化态概率分布信息的优势下，GPC模型仍能取得较好的识别效果。整体识别率达到87.1%，可以更好的识别熔化状态。本研究为L-PBF提供了一种新颖的原位监测思路。为了验证GPC的性能，还给出了基于支持向量机（SVM）模型、多层感知器（MLP）和长短期记忆（LSTM）神经网络的识别结果。研究结果表明，在自动优化超参数和提供熔化态概率分布信息的优势下，GPC模型仍能取得较好的识别效果。整体识别率达到87.1%，可以更好的识别熔化状态。本研究为L-PBF提供了一种新颖的原位监测思路。为了验证GPC的性能，还给出了基于支持向量机（SVM）模型、多层感知器（MLP）和长短期记忆（LSTM）神经网络的识别结果。研究结果表明，在自动优化超参数和提供熔化态概率分布信息的优势下，GPC模型仍能取得较好的识别效果。整体识别率达到87.1%，可以更好的识别熔化状态。本研究为L-PBF提供了一种新颖的原位监测思路。研究结果表明，在自动优化超参数和提供熔化态概率分布信息的优势下，GPC模型仍能取得较好的识别效果。整体识别率达到87.1%，可以更好的识别熔化状态。本研究为L-PBF提供了一种新颖的原位监测思路。研究结果表明，在自动优化超参数和提供熔化态概率分布信息的优势下，GPC模型仍能取得较好的识别效果。整体识别率达到87.1%，可以更好的识别熔化状态。本研究为L-PBF提供了一种新颖的原位监测思路。