Efficient Defect Detection Method for Wire and Arc Additive Manufacturing Based on Modified YOLOv8 Model

Abstract

Surface defect detection of parts manufactured by wire arc additive manufacturing (WAAM) is an important step for subsequent process improvement, optimization, and defect suppression. However, traditional methods and existing detection models suffer from high parameter counts, hardware requirements, and low accuracy. We presents a WAAM weld surface defect detection method derive from YOLOv8n, called high-efficiency new YOLO (HEN-YOLO). To address these limitations, a novel feature interaction detection head (NFIDH) is designed to enhance the feature learning and selectivity, reducing parameters and calculate losses. Subsequently, a lightweight and efficient local attention (ELA) mechanism was introduced to enhance both computational efficiency and detection accuracy of the model. Furthermore, the advanced screening feature fusion pyramid (HS-FPN) was employed to achieve cross-scale feature fusion and improve feature representation. Additionally, ConvTranspose2d deconvolution was utilized to optimize the upsampling process in the neck network, enabling the extraction of more effective and richer features. Finally, Experiments on 3440 WAAM weld surface defect dataset and the NEU-DET are maded to test the validity of HEN-YOLO. Results show that the mAP@.5(%) and mAP@.5:.95(%) of the HEN-YOLO are 2.4% and 8.3% higher than the YOLOv8n, respectively, which significantly improves the precision of weld surface defects detection; afterwards, it achieves a model parameters of 2.897 M and an 11.2% increase in FPS, surpassing the original YOLOv8n, which demonstrates that the HEN-YOLO has superior detection performance. This demonstrates that HEN-YOLO is efficient and can meet the practical detection requirements, and provides an efficient detection scheme for the weld defects in WAAM.