Spot Welding Defect Detection Using Synthetic Image Dataset on Convolutional Neural Networks

Abstract

Deep learning techniques using convolutional neural networks (ConvNet) was used to detect welding defects. This technique requires a large number of input images in order to train the network. This is not practical as the defective workpieces are undesirable and often rare. This study proposes the use of synthetic images which imitate the defective spot welding characteristic as the input dataset. By replicating Heat Effected Zone Ring(HAZ), Fusion Zone Ring (FZ)and Melt Ring in different size, color, and shape, both abnormal and typical spot welding images can be generated using image processing program written in Python. These images were then used to train two different levels of classification ConvNet. The results showed that by using two thousand artificial images, the ConvNet can classify the defective spot welding at the accuracy above 98%. Finally a test set of real defective spot welding images were carried out. The outcome also yielded the similar performance.