Multi-Branch Deep Fusion Network-Based Automatic Detection of Weld Defects Using Non-Destructive Ultrasonic Test
Kyeeun Kim, Keo Sik Kim, Hyoung-Jun Park
Abstract
This study introduces a deep learning engine designed for the non-destructive automatic detection of defects within weld beads. A 1D waveform ultrasound signal was collected using an A-scan pulser receiver to gather defect signals from inside the weld bead. We established 5,108 training datasets and 500 test datasets for five pass/fail labels in this study. We developed a multi-branch deep fusion network (MBDFN) model that independently trains 1D-CNN for local pattern learning within a sequence and 2D-CNN for spatial feature extraction and then combines them in an ensemble method, achieving a classification accuracy of 92.2%. The resulting deep learning engine has potential applications in automatic welding robots or welding inspection systems, allowing for rapid determination of internal defects without compromising the integrity of the finished product.