ABFormer: Acoustic Boundary-aware Transformer for Spar Adhesive-layer Segmentation in Wind Turbine Blade Ultrasonic...

The adhesive bonding quality between the spar cap and shear web is critical to the structural reliability of wind turbine blades. Phased array ultrasonic testing (PAUT) can non-destructively acquire volumetric echo data from adhesive bonding regions. Pixel-level segmentation of PAUT-derived ultrasonic D-scan images provides a direct way to identify defect-related echo regions. However, these D-scans exhibit weak defect boundaries and strong speckle noise interference. To address these challenges, this paper proposes a new method, ABFormer. In the encoder, the boundary-guided deformable routing attention (BDRA) module leverages acoustic boundary priors to guide deformable sampling and window-level top-k routing, enabling sparse cross-region interactions. The acoustic-aware local block dual attention (A-LBDA) module adopts a dual-branch structure to extract channel and acoustic structural features within local windows, enhancing fine-grained local feature representation. In the decoder, an adaptive-frequency dynamic decoder (AFDD) is proposed to preserve effective high-frequency information through frequency selection while reducing random speckle noise interference. Experiments on a self-constructed D-scan dataset from real manufacturing scenarios show that the proposed method outperforms mainstream segmentation methods, especially in detecting defective regions.