Developing a DSS for Enhancing Weldment Defect Detection, Classification, and Remediation Using HDR Images and Adaptive MDCBNet Neural Network

Abstract

This study presents a Decision Support System (DSS) designed for Non-Destructive Online Evaluation in welding. Based on the Multi-Scale Dense Cross Block Network (MDCBNet), it is able to detect, classify, and recommend remedial actions to prevent surface defects in welding. The performance of the network architecture is enhanced with synthetic defect samples generated through image augmentation techniques. By employing gradient attribution and t-SNE plot methods, we gained insights into the network’s predictions and comprehensively analyzed decision-making process. Comparative evaluations against pre-trained deep learning techniques revealed that our proposed model exhibits significant improvements, ranging from 2 to 10% across various performance metrics. Extensive comparisons with state-of-the-art methods underscored the effectiveness of our approach in detecting and classifying weld defects. Notably, our network, initially trained on Gas Tungsten Arc Welding images, demonstrated remarkable adaptability and versatility by effectively classifying images from Gas Metal Arc Welding processes. These findings emphasize the potential of the MDCBNet-based DSS to enhance welding practices, thereby contributing to producing high-quality weldments. The successful implementation of our DSS recommendations further supports its capacity to optimize the welding process and facilitate improved weld quality.