Selective laser melting (SLM) is an additive manufacturing technology that has an extensively applied foreground and practical value in many fields. Despite its powerful manufacturing ability, defects are prone to occur and therefore a more reliable and repeatable manufacture process is in high demand. During the SLM process, the melt pool signature is the key to understanding the dynamic process status, with which it is possible to predict process failure and give guidance to real-time feedback control. In this paper, a novel method to capture melt pool signature using a U-Net-based convolutional neural network is described. A lightweight architecture was used to reduce the inference time, and an improved loss function with penalty maps was applied to better remove interferences. The model performance was evaluated by comparing both the processing time and accuracy with two conventional image segmentation algorithms, including the threshold segmentation method and the active contour method. Mean intersection over union (MIoU) was chosen as the segmentation metric. Unlike traditional algorithms, U-Net successfully eliminated the interferences, and reached the highest MIoU (0.9806) at a relatively low computational cost of 37 ms on average. The collected information from the melt pool area in various scenarios was analyzed, and its potential to indicate the problem of melt pool overheating was investigated.

选择性激光熔化（SLM）是一种增材制造技术，在许多领域具有广泛的应用前景和实用价值。尽管其制造能力强大，但很容易出现缺陷，因此迫切需要更可靠和可重复的制造工艺。在SLM过程中，熔池特征是了解动态过程状态的关键，通过它可以预测过程故障并指导实时反馈控制。在本文中，描述了一种使用基于 U-Net 的卷积神经网络捕获熔池特征的新方法。使用轻量级架构来减少推理时间，并应用带有惩罚图的改进损失函数以更好地去除干扰。通过比较两种传统图像分割算法的处理时间和精度来评估模型性能，包括阈值分割方法和活动轮廓方法。选择联合平均交集（MIoU）作为分割指标。与传统算法不同，U-Net 成功消除了干扰，并以平均 37 ms 的相对较低的计算成本达到了最高的 MIoU（0.9806）。分析了在各种情况下从熔池区域收集的信息，并研究了其指示熔池过热问题的潜力。与传统算法不同，U-Net 成功消除了干扰，并以平均 37 ms 的相对较低的计算成本达到了最高的 MIoU（0.9806）。分析了在各种情况下从熔池区域收集的信息，并研究了其指示熔池过热问题的潜力。与传统算法不同，U-Net 成功消除了干扰，并以平均 37 ms 的相对较低的计算成本达到了最高的 MIoU（0.9806）。分析了在各种情况下从熔池区域收集的信息，并研究了其指示熔池过热问题的潜力。