2025A-IS-2 Development of a Real-Time Defect Detection and Classification Technique for DC-GMAW Using CNN

抄録

Welding automation is increasingly adopted in industrial manufacturing, yet its efficiency is limited due to the manual defect inspection during the welding process. This study presents a real-time defect detection and classification method for DC-GMAW using Convolutional Neural Networks (CNNs). The model leverages both time-series data - current and voltage signals - and welding bead images. To utilize 1D time-series data in the CNN architecture, Reccurence Plot (RP) transformation is applied to convert the signals into 2D image-line feature maps. These maps are then combined with bead images to create three-channel inputs for CNN training. Experimental data were collected during actual welding, and the model was trained using synchronized signal and image datasets.

The proposed approach achieved over 96% classification accuracy, demonstrating its potential for intelligence, real-time welding quality monitoring. This technique significantly reduces the need for manual inspection, contributing to improved reliability and efficiency in automated welding systems.