A-BiYOLOv9: An Attention-Guided YOLOv9 Model for Infrared-Based Wind Turbine Inspection

Abstract

This work examines how thermal turbulence patterns can be identified on the blades of operating wind turbines-an issue that plays a key role in preventive maintenance and overall safety assurance. Using the publicly available KI-VISIR dataset, containing annotated infrared images collected under real-world operating conditions, four object detection architectures were evaluated: YOLOv8, the baseline YOLOv9, the transformer-based RT-DETR, and an enhanced variant introduced as A-BiYOLOv9. The proposed approach extends the YOLOv9 backbone with convolutional block attention modules (CBAM) and integrates a bidirectional feature pyramid network (BiFPN) in the neck to improve feature fusion. All models were trained for thirty epochs on single-class turbulence annotations. The experiments confirm that YOLOv8 provides fast and efficient detection, YOLOv9 delivers higher accuracy and more stable convergence, and RT-DETR exhibits strong precision and consistent localization performance. A-BiYOLOv9 maintains stable and reliable accuracy even when the thermal patterns vary significantly between scenes. These results confirm that attention-augmented and feature-fusion-centric architectures improve detection sensitivity and reliability in the thermal domain. Consequently, the proposed A-BiYOLOv9 represents a promising candidate for real-time, contactless thermographic monitoring of wind turbines, with the potential to extend turbine lifespan through predictive maintenance strategies.