Detection of Defects on Single-Bead Welding by Machine Learning Methods

Abstract

This study classified the defects that were encountered on single-bead welding, which was made by MIG/MAG welding machines in Hursan Press company. 2250 images were taken on weldings that were made by five different welders. This study classified defects into three classes as good welding, porosity, and discontinuities. The images in the dataset, which have three classes, were classified by two stages. In the first stage, the texture features of the welding area were extracted. In the second stage, Artificial Neural Networks (ANN) method classified the extracted features. Sensitivity, specificity, accuracy, precision, and F-score metrics were used to measure the classification performance. 1500 images were used to train the system and training and validation performances were obtained as 94.03% and 94.19%, respectively. 750 images were used to test the performance of the proposed method and the test performance was obtained as 94.31%. The proposed method detected a defect on single-bead welding in 0.98 seconds.