와이어-아크 적층제조 공정의 비드 결함 감지를 위한 기계학습 모델

Abstract

In the wire-arc additive manufacturing (WAAM) process, which creates metal layers with weld beads, it is important to detect weld bead defects and resolve them properly and timely. In this paper, we propose a machine learning approach for automatically detecting weld bead defects based on voltage signature data captured during the WAAM process. We adopt multi-layer perceptron (MLP) and convolutional neural network (CNN) as machine learning models, and consider three types of beads: normal bead, abnormal bead with balling effects, and abnormal bead with cuts. After capturing voltage signatures while building weld beads, we separated each voltage signature into 17 to 19 segments, from each of which a set of features are extracted. We then constructed training and test data with feature datasets. We built total 75 voltage signatures: 45 for normal beads, 15 for abnormal beads with balling effects, and 15 for abnormal beads with cuts. After training the MLP and CNN models using TensorFlow, we tested and compared their performance. We found that the two types of models works well even though the amount of data used is small, but the CNN models are more appropriate for real-time detection of weld bead defects.