A Physics-Guided Bayesian Neural Network for Sensor Fault Detection in Wind Turbines
Predictive maintenance is essential for ensuring the reliability and efficiency of wind energy systems. Traditional deep learning models for sensor fault detection rely solely on data-driven patterns, often lacking interpretability and robustness. This article proposes a Physics-Guided Bayesian Neur...
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Open Journal of the Computer Society |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11027711/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Predictive maintenance is essential for ensuring the reliability and efficiency of wind energy systems. Traditional deep learning models for sensor fault detection rely solely on data-driven patterns, often lacking interpretability and robustness. This article proposes a Physics-Guided Bayesian Neural Network (PINN-BNN) model that integrates physics-informed learning with Bayesian inference to improve fault detection in wind turbines. The proposed approach enforces domain-specific constraints to ensure physically consistent predictions while quantifying uncertainty for risk-aware decision-making. The model is evaluated using a real-world wind turbine sensor dataset, achieving an accuracy of 97.6%, a recall of 91.8%, and an AUC-ROC of 0.987. The SHapley Additive exPlanations (SHAP) analysis reveals that gearbox temperature, blade vibration, and generator torque are the most critical features influencing failure predictions. Bayesian uncertainty estimation further improves interpretability by assigning confidence levels to each prediction. A comparative study with ten baseline models, including Long Short-Term Memory (LSTM), Transformer-based models, and traditional machine learning classifiers, demonstrates that the PINN-BNN model outperforms existing approaches while maintaining computational efficiency with a training time of 39.8 minutes and an inference time of 1.7 ms per sample. The integration of physics-informed learning ensures that the model generalizes well to varying environmental conditions, reducing false negatives and minimizing unexpected system failures. The proposed methodology presents a step toward interpretable and reliable predictive maintenance in wind energy systems. |
|---|---|
| ISSN: | 2644-1268 |