Phased array ultrasonic and deep learning based internal defect detection in underwater concrete bridge structures

Underwater concrete structures, which are critical load-bearing components in water-crossing bridges, are often prone to internal defects that undermine their long-term integrity. Traditional inspection methods, which mainly focus on surface defects, fall short in detecting internal structural issues, creating a significant gap in maintenance and safety assessments. This study proposes a phased array ultrasonic testing (PAUT) and deep learning based internal defect detection method in underwater concrete bridges. First, a new underwater PAUT device is developed, incorporating non-local means filtering to reduce noise and enhance ultrasonic imaging quality. Second, a small-scale dataset of ultrasonic images depicting internal defects in underwater concrete is created, and StyleGAN2-ADA was employed for effective data augmentation, addressing the challenge of limited data availability. To further improve defect detection, an enhanced SAM-based method is introduced, utilizing both serial and parallel depth-wise adapter fine-tuning strategies, which significantly boost the model's adaptability to complex underwater ultrasonic imaging tasks. Additionally, a pyramid-module-based self-generating prompt encoder is developed to optimize feature extraction and reduce manual intervention. Experimental results on both laboratory specimens and real-world underwater bridge piers demonstrate the superior performance of this method. The detection model achieves an Intersection over Union (IoU) score of 0.772—an improvement of 4.49% over the baseline SAM model—and outperforms other approaches in precision, recall, and F1-score. This research presents a robust, efficient, and scalable solution for internal defect detection in underwater concrete bridge structures, with strong potential for real-world applications in structural health monitoring. • Developing an underwater PAUT device, integrating NLM to enhance imaging quality and ensure stable detection in turbid underwater conditions. • Creating an underwater concrete internal defect dataset with StyleGAN2-ADA-based data augmentation to overcome data scarcity. • Proposing an enhanced SAM model for ultrasonic defect segmentation, using a Depth-wise Adapter and pyramid-based prompt encoder to improve accuracy and robustness. • Real-world validation on underwater bridge piers, demonstrating the method’s reliability in detecting internal defects and potential for structural health monitoring.