Algorithm for detecting surface defects in wind turbines based on a lightweight YOLO model.

Area of Science:

• Renewable Energy Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Early detection of wind turbine surface defects is crucial for optimizing energy generation efficiency and minimizing maintenance costs.
• Traditional defect detection methods face challenges in lightweight deployment and accuracy.
• Developing robust and efficient automated systems is essential for the wind power industry.

Purpose of the Study:

• To propose a high-precision object detection model for identifying surface defects on wind turbines.
• To address the limitations of existing methods regarding deployment and accuracy.
• To enhance the efficiency and reduce the operational costs of wind power generation through improved defect diagnosis.

Main Methods:

• A novel PC-EMA block integrating partial convolution (PConv) and efficient multiscale attention (EMA) was developed to replace the YOLOv8 backbone's bottleneck layer.
• GSConv was incorporated during feature fusion to preserve channel information and balance model complexity and accuracy.
• A PConv head was designed for reduced model complexity, and WIoUv3 was used as the regression loss function.

Main Results:

• The proposed model achieved a 5.07% improvement in average accuracy compared to the original YOLOv8.
• Model size was compressed by 32.5%, facilitating lightweight deployment.
• On a Jetson Nano, a 11 frames/s increase in processing speed was observed after TensorRT acceleration.

Conclusions:

• The enhanced YOLOv8 model with PC-EMA blocks offers a significant improvement in accuracy and efficiency for wind turbine surface defect detection.
• The model's reduced complexity and size make it suitable for lightweight deployment in real-world applications.
• This approach contributes to more reliable and cost-effective wind power generation through advanced automated inspection.