A deep mutual learning-based framework for wind turbine blade defect detection in multimodal phased array ultrasonic...

The internal adhesive layer of the wind turbine blade's spar is designed to transfer loads during operation, the quality of which directly affects the structural reliability of blades. Defects in the adhesive layer are characterized by phased array ultrasonic testing (PAUT) during the manufacturing phase. However, defect detection in multimodal PAUT data is mainly performed by inspectors, which is highly subjective and time-consuming. Therefore, this article proposes an orthogonal view alignment network (OVANet) for adhesive defect detection based on deep mutual learning. By guiding two parallel detection branches for B-scan and C-scan to mutually learn aligned target distributions at both the decision and feature levels, the proposed method can enhance detection performance across modalities while enabling decoupled inference. First, an orthogonal projection Intersection-over-Union metric is designed to mutually supervise cross-modality alignment at the decision level. Second, to enhance feature-level interaction, the attention-guided multi-scale discriminator is introduced, thereby forming adversarial mutual learning between backbones. In addition, an open-source annotation tool is developed to construct paired multimodal PAUT datasets collected from real industrial scenarios. Finally, the experimental results in both planar and volumetric demonstrate that the proposed method outperforms mainstream unimodal models.