Automatic PAUT crack detection and depth identification framework based on inspection robot and deep learning method

Orthotropic steel bridge decks (OSD) are widely acclaimed for their lightweight, high load-carrying capacity, and adaptability, making them a popular choice in steel structure bridges. However, the complex nature of their structure makes them susceptible to fatigue cracking, posing significant safety concerns. To address the issues above, this study employs a robot equipped with an ultrasonic phased array probe to automate the detection of internal cracks within Orthotropic Steel Decks (OSD). A Deep Convolutional Generative Adversarial Network (DCGAN) is utilized to augment the training dataset of Phased Array Ultrasonic Testing (PAUT) images. The YOLO series algorithms are applied and compared for crack localization, with YOLO v7-tiny exhibiting the highest accuracy and speed. Integrating attention mechanisms into the YOLO v7-tiny algorithm to facilliate rapid and high-precision crack detection. Analyzing the echo region with an echo intensity bar enabled the identification of crack depth, with an identification error within 5%. • Developed a robot with an ultrasonic phased array probe for automatic, unmanned OSD crack detection in darkness. • Deep Convolutional Generative Adversarial Network (DCGAN) enhances Phased Array Ultrasonic Testing (PAUT) images. • YOLOv7-tiny, enhanced with ODDC, detects OSD cracks with high speed and accuracy. • Incorporating echo intensity bar charts, the YOLOv7-tiny algorithm achieves crack depth identification with an error margin of 5%.