Intelligent recognition model for weld defects in pressure pipes based on fan-shaped Transformer knowledge distillation

Accurate recognition of weld defects in pressure pipelines is critical for structural integrity. Existing deep learning methods for phased array ultrasonic sector-scan data face three bottlenecks: insufficient exploitation of physical features, difficulty in distinguishing similar defects, and an excessive parameter count that hinders edge deployment. To address these bottlenecks, this paper proposes a lightweight, high-precision model based on fan-shaped Transformer knowledge distillation (SF-TransKD). First, polar coordinate convolution and a fan-shaped masked Transformer focus the network on the effective sector field of view and extract radial-angular features. Second, a lightweight cross-modal fusion network based on cross-attention enables adaptive interaction between A-scan signals and S-scan images. Finally, a hierarchical knowledge distillation framework with learnable projection layers transfers soft labels and intermediate features, achieving efficient compression. Experimental validation on a self-constructed dataset achieves 98% classification accuracy across five defect types, including slag inclusions and cracks. The student model contains only 0.82 M parameters, achieving over 80% compression. This provides technical support for intelligent non-destructive testing of pressure pipeline welds.