Abstract With the continuous development of additive manufacturing technique, the issue on built quality has caught increasing attentions. To improve the quality of built parts, the process monitoring and control has been emphasized as a promising solution. Despite a large number of studies on the development of sensors and instrumentations, the investigation on statistical analysis, modelling and automatic anomalies detection is still at an infant stage. To advance the related research, the intelligent classification methods, support vector machines (SVM) and convolutional neural network (CNN), were proposed for quality level identification in this work. A vision system with high speed camera was used for process images acquisition. The features of different objects including melt pool, plume and spatter were extracted based on the AM process understanding. The corresponding feature vectors were used as the input for the SVM classification. The results indicated the information from different objects is sensitive to different types of quality anomalies. Moreover, the combination of features from these three objects can significantly improve the classification accuracy to 90.1%. Additionally, the comparison between SVM and CNN was also conducted, the high accuracy of 92.7% for the CNN model demonstrated that it is a promising method for quality level identification by using the vision system.

中文翻译：

---

粉末床熔融增材制造过程监测中熔池、羽流和飞溅信息的提取和评估

摘要 随着增材制造技术的不断发展，建筑质量问题越来越受到关注。为了提高构建部件的质量，过程监控已被强调为一种有前途的解决方案。尽管对传感器和仪器的发展进行了大量研究，但对统计分析、建模和自动异常检测的研究仍处于起步阶段。为了推进相关研究，本文提出了智能分类方法、支持向量机 (SVM) 和卷积神经网络 (CNN) 用于质量级别识别。带有高速相机的视觉系统用于过程图像采集。不同物体的特征，包括熔池、羽流和飞溅是基于对 AM 过程的理解来提取的。相应的特征向量被用作 SVM 分类的输入。结果表明，来自不同对象的信息对不同类型的质量异常敏感。此外，这三个对象的特征组合可以将分类准确率显着提高到 90.1%。此外，还进行了 SVM 和 CNN 之间的比较，CNN 模型的 92.7% 的高准确率表明它是一种使用视觉系统进行质量级别识别的有前途的方法。这三个对象的特征组合可以将分类准确率显着提高到 90.1%。此外，还进行了 SVM 和 CNN 之间的比较，CNN 模型的 92.7% 的高准确率表明它是一种使用视觉系统进行质量级别识别的有前途的方法。这三个对象的特征组合可以将分类准确率显着提高到 90.1%。此外，还进行了 SVM 和 CNN 之间的比较，CNN 模型的 92.7% 的高准确率表明它是一种很有前途的使用视觉系统进行质量级别识别的方法。