Wind Turbine Blade Defect Recognition Method Based on Large-Vision-Model Transfer Learning

Xin Li, Jinghe Tian, Xinfu Pang, Lida Shen, Haibo Li, Zedong ZhengSensors 2025 · DOI

TL;DR

This paper introduces a novel three-stage framework for automated wind turbine blade defect detection. By combining YOLOv5 for precise blade localization, the powerful DINOv2 large vision model for robust feature extraction, and a Stochastic Configuration Network (SCN) for classification, the method achieves high accuracy and real-time performance, addressing limitations of previous approaches.

What Problem Does This Paper Address?

Wind turbine blades are critical components that are constantly exposed to harsh environmental conditions. Defects like cracks, chipping, and oil stains can compromise their structural integrity and lead to costly failures. Traditional manual inspection methods are dangerous, time-consuming, and prone to errors. Existing automated methods often struggle with complex backgrounds, poor generalization to varying conditions, and insufficient real-time capabilities, hindering efficient and reliable maintenance of large wind farms.

Methods

The proposed solution is an end-to-end, three-stage pipeline:

  1. Blade Localization (YOLOv5): The first step uses the YOLOv5 object detection model to accurately identify and crop the wind turbine blade from complex background images captured by drones. This is crucial for isolating the blade and reducing interference from irrelevant visual information. To enhance robustness against variations in lighting and angles, techniques like Mosaic enhancement, flipping, and random scaling are employed.
  2. Feature Extraction (DINOv2): After localization and cropping, the blade images are fed into the DINOv2 large vision model. DINOv2, a self-supervised learning model, excels at extracting rich, generalizable visual representations from images, even in challenging visual environments. This approach moves beyond traditional methods that rely on manually designed features, leveraging DINOv2's ability to learn powerful, discriminative features automatically. The ViT-S/14 variant of DINOv2 is used, producing a 384-dimensional feature vector.
  3. Defect Classification (Stochastic Configuration Network - SCN): The extracted feature vectors are then passed to a Stochastic Configuration Network (SCN) for defect classification. SCNs are adaptive, incremental learning models that can dynamically adjust their network structure. This allows them to efficiently learn classification rules from the high-dimensional feature data and achieve high accuracy with a flexible architecture.

Datasets

The paper utilizes a public online dataset for blade area positioning, consisting of 4590 high-resolution drone images of wind turbine blades. This dataset was divided into training (80%), validation (10%), and testing (10%) sets. For defect classification, a dataset of 2989 images was used, categorized into "Damage," "Paint Shedding," and "Dirt Buildup."

Key Insights and Ideas

  • Leveraging Large Vision Models for Domain-Specific Tasks: The paper demonstrates the effectiveness of using pre-trained large vision models like DINOv2, originally trained on massive general image datasets, for a specialized industrial application like wind turbine defect detection. This highlights the power of transfer learning and foundation models in adapting to new domains.
  • The Synergy of Localization, Feature Extraction, and Classification: The three-stage approach is critical. YOLOv5's precise localization acts as a vital pre-processing step, ensuring DINOv2 receives clean blade images, thereby improving the quality of extracted features. The SCN then effectively utilizes these robust features for accurate classification.
  • SCN's Adaptive Nature for Robust Classification: The SCN's ability to dynamically adjust its hidden nodes based on error tolerance makes it well-suited for complex, real-world data where defect characteristics can vary. This adaptive learning contributes to the high classification accuracy.
  • Addressing Background Interference: A significant challenge in visual defect detection is distinguishing defects from complex backgrounds. The YOLOv5 localization step effectively mitigates this, allowing subsequent models to focus solely on the blade itself.

Key Results

  • High Classification Accuracy: The proposed method achieved an impressive classification accuracy of 97.8%.
  • Real-time Performance: The entire end-to-end inference time was approximately 19.65 ms per image, which translates to over 50 frames per second (FPS), meeting real-time requirements for inspection tasks.
  • Blade Localization Performance: YOLOv5 demonstrated excellent performance in blade area positioning, achieving an AP:50 of 99.4% with an inference speed of 9.0 ms per image.
  • Ablation Study: Removing the YOLOv5 positioning module resulted in a drop in classification accuracy from 98.7% to 94.4%, underscoring the importance of the localization stage in suppressing background interference.

Limitations and Future Work

  • Lack of Quantitative Defect Analysis: The current method identifies defects but does not provide information about their shape or extent. This limits its ability to support finer-grained maintenance strategies that require quantitative defect analysis.
  • Future Work: The authors suggest integrating segmentation networks to extract precise defect locations and sizes. This would enable severity-level estimation and enhance the system's practical value for condition-based maintenance.

Summary updated

Abstract

Timely and accurate detection of wind turbine blade surface defects is crucial for ensuring operational safety and improving maintenance efficiency with respect to large-scale wind farms. However, existing methods often suffer from poor generalization, background interference, and inadequate real-time performance. To overcome these limitations, we developed an end-to-end defect recognition framework, structured as a three-stage process: blade localization using YOLOv5, robust feature extraction via the large vision model DINOv2, and defect classification using a Stochastic Configuration Network (SCN). Unlike conventional CNN-based approaches, the use of DINOv2 significantly improves the capability for representation under complex textures. The experimental results reveal that the proposed method achieved a classification accuracy of 97.8% and an average inference time of 19.65 ms per image, satisfying real-time requirements. Compared to traditional methods, this framework provides a more scalable, accurate, and efficient solution for the intelligent inspection and maintenance of wind turbine blades.
BibTeX
@inproceedings{li2025wind,
  author = {Xin Li and Jinghe Tian and Xinfu Pang and Lida Shen and Haibo Li and Zedong Zheng},
  title = {{Wind Turbine Blade Defect Recognition Method Based on Large-Vision-Model Transfer Learning}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25144414},
}
Figure