Wind Turbine Fault Detection Through Autoencoder-Based Neural Network and FMSA.

Area of Science:

• Renewable Energy Engineering
• Artificial Intelligence in Industrial Applications
• Condition Monitoring Systems

Background:

• Wind power is crucial for clean energy, but turbine reliability is key.
• Existing fault detection methods can be limited in complex systems.
• Predictive maintenance is essential for optimizing wind farm operations.

Purpose of the Study:

• To develop a novel hybrid fault detection approach for wind turbines.
• To integrate expert knowledge (Failure Mode and Symptoms Analysis) with data-driven models (autoencoders).
• To enhance anomaly detection, feature selection, and fault localization for improved reliability.

Main Methods:

• Utilized Failure Mode and Symptoms Analysis (FMSA) for failure mode identification.
• Developed autoencoder neural networks trained on healthy SCADA data.
• Implemented an anomaly detection strategy using reconstruction error and a persistence-based rule.
• Employed a fault-specific modeling strategy for customized turbine and failure mode analysis.

Main Results:

• Achieved 99% classification accuracy on simulated data.
• Successfully detected anomalies up to 60 days before reported failures in real-world data.
• Identified degradations in key components like the transformer, gearbox, generator, and hydraulic group.
• FMSA integration improved feature selection and fault localization.

Conclusions:

• The hybrid approach significantly enhances fault detection accuracy and early warning capabilities.
• The system improves the interpretability and precision of wind turbine condition monitoring.
• This methodology offers a robust solution for predictive maintenance in wind energy systems.