Sawicki Piotr, Dybała Bogdan
Center for Advanced Manufacturing Technologies, Wroclaw University of Technology, 50-370 Wrocław, Poland.
Global Engineering & Technology Center, Collins Aerospace, 51-317 Wrocław, Poland.
Materials (Basel). 2025 Jun 26;18(13):3026. doi: 10.3390/ma18133026.
In this work, we address the task of monitoring Powder Bed Fusion-Laser Beam processes for metal powders (PBF-LB/M). Two main contributions with practical merit are presented. First, we consider the comparison between a large deep neural network (VGG-19) and a small model consisting of, among others, four convolutional layers. Our study shows that the small model can compete favorably with the big model, which takes advantage of transfer learning techniques. Secondly, we present a filtering method using a semantic segmentation approach to preselect a region for the classification algorithm. The region is selected based on post-exposure images, and preselection can be easily adopted for any machine independently of the software used for the translation of process input files. To consider the task, a master dataset with over 260,000 samples was prepared, and a detailed process of preparing the training datasets was described. The study demonstrates that the classification time can be reduced by a factor of 4.51 while still maintaining the model's necessary performance to detect errors in a PBF-LB process.
在这项工作中，我们致力于监测金属粉末的粉末床熔融-激光束工艺（PBF-LB/M）任务。提出了两项具有实际价值的主要贡献。首先，我们考虑了大型深度神经网络（VGG-19）与一个小型模型（其中包括四个卷积层等）之间的比较。我们的研究表明，该小型模型能够与利用迁移学习技术的大型模型相媲美。其次，我们提出了一种使用语义分割方法的滤波方法，为分类算法预先选择一个区域。该区域基于曝光后图像进行选择，并且预选择可以轻松地应用于任何机器，而与用于处理输入文件转换的软件无关。为了考虑该任务，准备了一个包含超过260,000个样本的主数据集，并描述了准备训练数据集的详细过程。研究表明，分类时间可以减少4.51倍，同时仍能保持模型在PBF-LB工艺中检测错误所需的性能。
