Abstract: Water dams are prone to various types of damage under the coupled effects of external erosion and complex loads, particularly their underwater structural damage, which is often difficult to detect and requires timely monitoring to mitigate safety hazards. Previous deep learning-based methods for such damage detection suffer from limitations such as high computational demands and significant manual intervention, while commonly-used detecting devices tend to possess inadequate computational capabilities, leading to certain incompatibility. This paper presents a new intelligent detection model based on the YOLOv7 algorithm for multi-class underwater damage to water dams under the conditions of low computational capabilities. This model enhances the detection accuracy by integrating three intelligent modules－deformable convolution, SE attention mechanism, and MPDIoU loss function－and provides strong robustness for application in complicated underwater environments. It achieves lightweight operation through a structured pruning strategy at a ratio of 0.4, and reduces significantly computational power requirements. Analysis of engineering examples and comparison with the previous algorithms in literature shows that its floating-point computation and the number of its parameters are reduced by 48% and 61% respectively. It improves detection accuracy for exposed reinforcement bars and voids significantly by 18.7% and 11.9% respectively, and enhances the average detection accuracy for various types of damage by 8.3%, achieving the goal of accurate detection under the conditions of low computational resources.

Key words: underwater defects of dams, intelligent detection, weak computing power conditions, lightweight model, high robustness