YOLO11-LDConv-G-based detection of defects in TFM images of small-bore nozzle fillet welds under structural echo interference

Small-bore nozzle fillet welds present a challenging ultrasonic inspection scenario because defect responses in total focusing method (TFM) images are often weak and easily masked by structural echoes and background noise. To address this issue, this study proposes an intelligent defect detection framework that combines phased-array full matrix capture (FMC), TFM imaging, and an improved YOLO11 network for complex weld inspection. A labeled dataset was established from experimental inspections of set-on and set-in nozzle specimens with prefabricated defects, covering cracks, porosity, incomplete penetration, lack of fusion, and defect-free cases. To enhance robustness against structural echo interference and improve the representation of small, morphologically diverse defects, a gated linear deformable convolution module was introduced into YOLO11, yielding the proposed YOLO11-LDConv-G model. Compared with the baseline YOLO11, the proposed model improved Recall from 0.819 to 0.845, mAP@0.5 from 0.882 to 0.931, and mAP@0.5:0.95 from 0.543 to 0.565 on the independent test set. These results suggest the preliminary feasibility of the proposed method on the current held-out test set.