Flash Defect Detection System of Friction Stir Welding Process Based on Convolutional Neural Networks for AA 6061-T651

Abstract

An early detection control system for high-speed and objectivity welding defects is needed. Visual Inspection (VT) is an important method and the initial stage before a welded material will be tested at destructive testing. So far, VT has only used human vision, which takes a protracted process and is highly subjective. This paper will contribute to the VT method to control the Friction Stir Welding (FSW) process by detecting the flash defect using image processing and Convolutional Neural Network (CNN). Thus, flash defects in the FSW process can be minimised and detected as early as possible. Image processing and CNN serve as a substitute for human vision. The selection of CNN is considered suitable for detecting an image because the process is fast and detects key features without human supervision, which is carried out by a continuous learning process. 620 images from the FSW process were processed into two groups of datasets. It was processed with two types of CNN architecture, including AlexNet and VGG16. Based on the VT results by CNN, the AlexNet model showed a detection accuracy of 91.03%, while the VGG16 model showed a detection accuracy of 77.35%. From these results, CNN’s success in conducting VT on FSW process control is relatively high and can play a more significant role in checking the results of the FSW process. Therefore, the possibility of flash defects can be minimised and detected as early as possible.