Prediction of Weld Current Using Deep Transfer Image Networks Based on Weld Signatures for Quality Control

Abstract

The utilizing electrical resistivity tomography (ERT) in land surveying, one approach is using trailers to locate fluid or underground utilities in large-scale subsurface constructions. It is known that welding is an ideal manufacturing process to incorporate a well-assembled finished product such as a trailer system used in subsurface imaging. However, there are conditions where inconsistencies are generated in a weld metal called welding defects that results from poor welding procedure or improper welding patterns. Welding flaws can develop on both the exterior and interior of the weld metal if the current setup is too low or high. The objective of this paper is to identify and categorize weld current using integrated computer vision and deep learning. There are nine deep transfer image networks in MATLAB namely, DarkNet53, DenseNet201, EfficientNetB0, InceptionV3, MobilenetV2, NASNetLarge, ResNet18, ResBet101, and Xception, which were trained and tested for weld current detection and classification. Augmented images were pre-clustered into four current levels: 60 A, 80A, 100A, and 140 A. The test phase confirmed that ResNetl01 exhibited the highest accuracy among other trained deep learning models. The image spectral thresholding used in the extraction of image texture features explained the reason for the low accuracy in other networks. Overall, this study will have a contribution to the quality control of the welding process outcomes by introducing another phase of non-invasive determination of current used in welding the metal joints and sections of a road tomography trailer.