In-situ defect detection in wire arc additive manufacturing of mild steel using deep learning

Abstract

Wire arc additive manufacturing (WAAM) has emerged as a promising technique for large-scale metal fabrication due to its high deposition rate and cost efficiency; however, its widespread adoption is hindered by the occurrence of process-induced defects such as porosity, spattering, and burn-through. This study presents an in-situ defect detection framework based on acoustic signal analysis and deep learning. Acoustic emissions generated during the welding process are captured using a low-cost electret microphone and transformed into time–frequency spectrograms via short-time Fourier transform (STFT). These spectrograms are utilized as inputs to a ResNet18 convolutional neural network with transfer learning to perform multi-class classification of welding conditions. The proposed framework achieves an overall accuracy of 93.67% across four classes, demonstrating that acoustic signals contain discriminative features associated with weld quality and defect formation mechanisms. Compared to conventional monitoring approaches that rely on complex sensing systems, the proposed method offers a simple, non-intrusive, and cost-effective solution. The findings highlight the potential of acoustic-based monitoring for real-time WAAM process control and its integration into intelligent and digital manufacturing systems.