Exploring Synthetic Dataset Generation for Improving Computer Vision System Performance in Hydropower Turbine Facility

Hydropower turbines are critical for sustainable energy generation, but their blades are frequently exposed to damage from cavitation, pitting, and corrosion. Regular visual inspections are essential but are often hazardous, time-consuming, and constrained by the limited availability of high-quality inspection images. This study explores the use of synthetic image generation to overcome these challenges by leveraging the Stable Diffusion model with ControlNet reference-only methods. A dataset of 62 real turbine blade damage images was augmented to 836 synthetic images, enhancing data diversity for training computer vision models. Expert evaluations revealed an average accuracy of 64%–68% in distinguishing real from synthetic images, with variations in precision and recall between evaluators. These findings highlight the potential of synthetic datasets to address data scarcity, improve model robustness, and support safer and more efficient turbine inspections.