An intelligent recognition method for weak ultrasonic images of internal defects in composite material based on the improved YOLOv11 model

Phased Array Ultrasonic Testing (PAUT) is a critical technique for the in-service inspection of large composite structures like wind-turbine blades. However, the interpretation of ultrasonic signals remains heavily dependent on operator expertise, limiting inspection efficiency and accuracy. Moreover, large-area scanning presents challenges for consistent identification, particularly for weak-echo defects where signal interpretation becomes ambiguous. To overcome these limitations, this study proposes LWA-YOLO, an intelligent PAUT framework that integrates a simplified deep-learning detector with high-efficiency inference. The framework is derived from YOLOv11 by: (1) inserting a dual-level CA block for multi-scale feature discrimination; (2) replacing the regression loss with Wise-IoU for faster convergence; (3) pruning redundant detection heads, reconstructing the feature-fusion neck and optimising the backbone architecture. Guided by the defect characteristics of an authentic wind turbine blade cut-out, eleven glass-epoxy samples containing delamination, folds and sundries were manufactured using RTM to simulate realistic conditions. A dataset of 1516 B-scan images was constructed and annotated by professional NDT staff. Experimental results demonstrate that LWA-YOLO achieves accuracy rates of 98.9%, 98.0%, and 96.3% for the three defect types, respectively. With an inference speed of 158.7 FPS, the proposed method balances high precision with the efficiency required for intelligent industrial inspection.