As the global trend toward carbon neutrality accelerates the scale-up and offshore installation of wind power facilities, ensuring the maintenance and safety of blades—a core component—has emerged as an imperative task. However, the actual operati...
As the global trend toward carbon neutrality accelerates the scale-up and offshore installation of wind power facilities, ensuring the maintenance and safety of blades—a core component—has emerged as an imperative task. However, the actual operating environment of wind farms involves mixed background noise, resulting in a low Signal-to-Noise Ratio. Furthermore, the scarcity of defect data for training deep learning models has limited the direct application of existing diagnostic technologies to the field.
To overcome these environmental constraints and enhance diagnostic reliability, this thesis proposes a hybrid defect diagnosis framework that hierarchically combines unsupervised learning-based signal separation and supervised learning-based Deep Neural Networks. The core approach of this study is materialized as a data-driven pipeline that fuses noise reduction in field data with precise pattern learning from experimental data.
In the first phase of the proposed system, a two-stage signal processing process linking Non-negative Matrix Factorization and K-means clustering was applied to resolve the high-noise issue in operational data. This effectively separated background noise from valid signals based on statistical characteristics without relying on complex physical modeling, thereby significantly improving the input quality for the diagnostic model. In the second phase, a defect diagnosis model based on DNN was constructed utilizing wind tunnel experimental data, securing robust detection performance even in imbalanced data environments.
In conclusion, this study establishes a practical defect detection methodology applicable to wind power sites with limited data and high environmental variables by employing a dual-track strategy that complementarily utilizes noisy field data and precise experimental data. This is expected to contribute to building an intelligent maintenance system optimized for the domestic wind power environment.