MSDA-YOLO: a multiscale direction-aware network for surface defect detection in arc additive manufacturing

Abstract

Wire and arc additive manufacturing (WAAM) enables efficient fabrication of complex metal components, but surface defects can compromise structural integrity. Detecting such defects from point-cloud projection images remains challenging due to low contrast, scale variation, and detail loss. This paper proposes MSDA-YOLO, a multi-scale direction-aware defect detection network based on YOLOv11. The network incorporates a local multi-scale channel– spatial attention mechanism, a multi-pooling direction fusion module, and a direction-aware upsampling module. Experiments on a self-built WAAM dataset and the public NEU-DET dataset show that MSDA-YOLO achieves mAP@0.5 scores of 89.49% and 84.92%, respectively, with an inference speed of 68 FPS, demonstrating superior performance for real-time WAAM surface defect detection.