Weld Geometry Prediction Based on Binocular Vision and Deep Learning

Abstract

To improve the level of welding automation in the industry, there are increasing requirements for highly intelligent and accurate inspections of the welding process in real time. This paper proposed a new method for predicting weld dimensions based on binocular imaging information and a deep learning system. The binocular imaging information was acquired by binocular vision equipment and an image processing algorithm. A convolutional neural network structure was developed by adding a fully connected block and loss function judgment. A new calculating procedure was proposed to extract and link the information of the processed weld pool image and the weld parameters effectively. With the help of 7394 training samples, the results of 1849 testing samples showed that the overall accuracy of the proposed model was higher than 93% for the prediction of weld dimensions, which could meet the requirements in practical applications.