Image Recognition of Wind Turbine Blade Defects Using Attention-Based MobileNetv1-YOLOv4 and Transfer Learning

Area of Science:

• Engineering
• Computer Science
• Renewable Energy

Background:

• Machine vision and deep learning are advancing wind turbine blade surface damage detection.
• Current methods face challenges with feature extraction, computational complexity, and data limitations.

Purpose of the Study:

• To propose an efficient and accurate image recognition method for wind turbine blade defects.
• To enhance defect detection by optimizing feature extraction and network convergence.

Main Methods:

• Developed an attention-based MobileNetv1-YOLOv4 model, replacing YOLOv4's backbone with lightweight MobileNetv1.
• Integrated SENet, ECANet, and CBAM modules for adaptive feature optimization.
• Employed a two-stage transfer learning approach to address insufficient data and improve convergence.

Main Results:

• The proposed model achieved higher detection accuracy and faster response speeds compared to state-of-the-art networks.
• Demonstrated significantly reduced computational complexity.
• A sensitivity study confirmed the impact of training dataset size on performance.

Conclusions:

• The attention-based MobileNetv1-YOLOv4 model offers a promising solution for automated wind turbine blade defect detection.
• The method effectively balances accuracy, speed, and computational efficiency for practical applications.