A real-time end-to-end detector for detecting surface defects on oversized rings

Area of Science:

• Mechanical Engineering
• Electrical Engineering
• Computer Vision

Background:

• Oversized rings are critical load-bearing components in wind turbines.
• Surface defects on these rings can lead to operational disruptions.
• Detecting small defects on large rings is challenging, time-consuming, and often inaccurate.

Purpose of the Study:

• To develop a lightweight and accurate model for detecting surface defects on oversized wind turbine generator rings.
• To balance model accuracy with computational efficiency for practical applications.
• To improve the speed and reliability of wind turbine inspection processes.

Main Methods:

• Proposed a Lightweight Multiscale High-Efficiency Detector (LMHD) using RepViT backbone and Bi-directional Feature Pyramid Network (BiFPN).
• Introduced Diverse View Group Shuffle Cross Stage Partial Network (DVOV-GSCSPM) for efficient multiscale feature fusion.
• Integrated Self-Calibrated Convolutions (SCConv) and Efficient Local Attention (ELA) modules to reduce complexity.
• Utilized Powerful-IoUv2 loss function for improved convergence and generalization.

Main Results:

• Achieved a detection accuracy of 87.0% on the NEU-DET dataset.
• Demonstrated a processing speed of 70.4 frames per second (FPS).
• The LMHD model offers a balance between accuracy and model size.

Conclusions:

• The proposed LMHD model effectively addresses the challenges of detecting surface defects on oversized wind turbine generator rings.
• The model's lightweight design and high efficiency make it suitable for real-world inspection tasks.
• This advancement contributes to improved maintenance strategies and operational reliability for wind turbines.