Abstract
Oversized rings in wind turbines are regarded as crucial components because they often serve as the main load-bearing and connector structures. Surface defects on these rings can disrupt the normal operation of the entire unit. Detecting surface defects on oversized rings in wind turbine generators (WTGs) is highly challenging due to the huge ring size and small target defects, which will cause the detection process to be very time-consuming and difficult to achieve the expected accuracy. To address this challenge, we propose a new lightweight multiscale high-efficiency detector (LMHD) that balances accuracy and model size. The framework utilizes RepViT as the detection backbone and incorporates a bi-directional feature pyramid network (BiFPN) in the neck to achieve bi-directional feature transfer and aggregation. Additionally, it includes a new lightweight, efficient, multi-scale cross-stage partition module called the Diverse View Group Shuffle Cross Stage Partial Network (DVOV-GSCSPM), which employs a rational architecture and multiscale information fusion to ensure that the overall model is lightweight while maintaining a rich gradient flow. Self-Calibrated Convolutions (SCConv) and Efficient Local Attention (ELA) modules are introduced into the neck network to reduce computational complexity and the number of parameters while ensuring model accuracy. Ultimately, we incorporate the Powerful-IoUv2 loss function to enhance the rate of model convergence and generalization capabilities. The model is experimentally validated on the public dataset NEU-DET, achieving a detection accuracy of 87.0% with 70.4 frames per second (FPS).