Detection of Defects in Welded Butt Joints Based on Laser Scanning: Neural Networks Approach

The work is devoted to the detecting of defects in welded butt joints based on the data from the computer vision system (iRVision 3DL) installed on a Fanuc welding robot. The main approach to solving the problem is deep learning of artificial neural networks. The images of the welded joint were obtained as a result of laser scanning. The obtained images of scanned surfaces were represented in a vector form of boundary points' values. The neural network is based on one-dimensional convolutional layers, which are good for searching patterns in the data. The task of the research was to obtain a neural network model with the maximum potential value of the metric because the missing defects are unacceptable in manufacturing and always have a bad influence on the properties of welded steel structures. The results indicate the applicability of this approach to the problem of defect detection in welded joints.