A Novel Deep Class-Imbalanced Semisupervised Model for Wind Turbine Blade Icing Detection

Area of Science:

• Renewable Energy Engineering
• Artificial Intelligence in Energy
• Machine Learning for Condition Monitoring

Background:

• Wind farms face significant operational risks due to blade icing, especially at high latitudes.
• Existing icing detection methods (manual, sensor-based, model-based) have limitations in accuracy, cost, and data requirements.
• Data-driven approaches are promising but hindered by the scarcity of labeled, imbalanced datasets common in wind turbine operations.

Purpose of the Study:

• To develop a novel deep class-imbalanced semisupervised (DCISS) model for accurate wind turbine blade icing estimation.
• To overcome challenges of limited labeled data and inherent data imbalance in icing detection.
• To enhance feature extraction from raw data for improved detection performance.

Main Methods:

• Integration of class-imbalanced learning and semisupervised learning (SSL) within a prototypical network framework.
• Utilizing a prototypical network to rebalance features and measure sample similarities between labeled and unlabeled data.
• Introduction of a channel calibration attention module for enhanced feature extraction from raw sensor data.

Main Results:

• The DCISS model demonstrated significant accuracy advantages over classical anomaly detection and state-of-the-art SSL algorithms.
• Performance of the DCISS model was competitive when compared against five different class-imbalanced loss functions.
• The model's generalization capability and practical applicability were validated through an online estimation use case.

Conclusions:

• The proposed DCISS model offers a robust and accurate solution for wind turbine blade icing estimation, particularly in data-scarce and imbalanced scenarios.
• The integration of class-imbalanced learning, SSL, and attention mechanisms provides a powerful approach for condition monitoring in renewable energy systems.
• The DCISS model shows strong potential for practical implementation in real-world wind farm operations, enhancing safety and efficiency.