Lightweight You Only Look Once-based automatic defect detection in wire arc additive manufacturing

Wire arc additive manufacturing (WAAM) technology has achieved significant advancements in fabricating complex metal components, yet defects such as surface porosity, lack of fusion, and slag inclusion continue to compromise quality and efficiency. To address challenges in detecting welding slag and micro-pores, we introduce an enhanced You Only Look Once (YOLO)v8n-Attention-Refined Feature Module (ARFM) architecture integrating receptive field attention convolution, an attention-based feature pyramid network, and the Focaler-minimum point distance intersection over union metric. This integration markedly improves the precision and resilience of defect recognition. Moreover, we assessed the defect detection capabilities of YOLOv3t, YOLOv5n, YOLOv6n, YOLOv8n, YOLOv9t, and YOLOv10n in WAAM and developed a composite defect dataset incorporating aluminum, titanium, and stainless steel specimens. The experiments reveal that YOLOv8n achieved the highest overall detection effectiveness (mean average precision [mAP]@0.5 = 0.921). YOLOv10n exhibited marginal superiority in slag inclusion detection (mAP@0.5 = 0.886) with a peak throughput of 54.35 frames per second (FPS), while YOLOv5n converged the fastest within 500 epochs. Following the introduction of ARFM, mAP@0.5 increased to 0.944 (+2.4%), and slag inclusion detection reached 0.929 (+7.3%). However, the frame rate declined to 38 FPS, which suffices for basic real-time monitoring but remains inadequate for high real-time scenarios. This study provides an important reference for deploying YOLO-series models for real-time monitoring of WAAM and closed-loop control systems.