Infrared-based real-time leakage detection in earth–rock dams using an enhanced YOLO algorithm integrated with a biomimetic quadruped robot

Dam leakage poses a critical challenge to the stability and safety of earth–rock dams. However, traditional manual inspection methods are inefficient, limited in scope, and can often fail in complex environments due to subjectivity, time consumption, and human errors. To address these limitations, this study integrates for the first time an optimized You Only Look Once Leakage Enhanced (YOLO-LE) algorithm with infrared thermography and a biomimetic quadruped robot to enable real-time, autonomous leakage detection. The proposed YOLO-LE framework incorporates squeeze-and-excitation attention mechanisms, adaptive spatial feature fusion, and transfer learning to improve detection precision, robustness against environmental interference (e.g., humidity and vegetation), and recognition of small-scale leakage points. The proposed framework is verified by indoor and field experiments using an FLIR A50 infrared camera mounted on a quadruped robot, selected for its terrain adaptability. A custom dataset of 1000 thermal images, augmented with rotation and noise injection, is used for training and evaluation. The results show that the proposed YOLO-LE framework achieves an average precision (AP) of 0.873 and an F1-score of 0.865, outperforming the YOLOv5, single-shot multi-box detector, and faster region-convolutional network models in AP by 7.2 %, 6.9 %, and 1.4 %, respectively. In addition, comparative experiments under different environmental disturbances demonstrate the superior resilience of the proposed framework, with an inference speed of 51 frames per second, ensuring real-time monitoring capability. Finally, the results validate the feasibility of combining deep learning with robotic systems for dam safety, providing a scalable, accurate, and automated solution for monitoring critical infrastructure.