LSTM-Autoencoder Based Anomaly Detection Using Vibration Data of Wind Turbines.

Area of Science:

• Renewable Energy Engineering
• Machine Learning for Predictive Maintenance
• Signal Processing for Fault Detection

Background:

• Wind energy is crucial for addressing energy depletion, but wind turbine breakdowns pose significant challenges.
• Effective predictive maintenance is essential to ensure the reliability and longevity of wind power infrastructure.
• Current methods for detecting wind turbine failures require improvement in speed and accuracy.

Purpose of the Study:

• To develop an unsupervised learning method for accurate outlier detection in wind generator vibration signals.
• To enhance the early identification of wind turbine failures, thereby minimizing downtime and maintenance costs.
• To contribute a novel approach for ensuring the stability and efficiency of wind energy systems.

Main Methods:

• Vibration data analysis using wavelet packet conversion to identify critical frequency bands.
• Application of high-pass filters to accentuate differences between normal and abnormal operational data.
• Dimensionality reduction via principal component analysis (PCA) for enhanced data preprocessing.
• Training a long short-term memory (LSTM) autoencoder on normal vibration data for outlier detection.

Main Results:

• Achieved a 97% outlier detection performance, indicating high accuracy in identifying abnormal conditions.
• Successfully identified specific frequency bands indicative of wind generator anomalies.
• Demonstrated the effectiveness of the proposed data preprocessing and LSTM autoencoder approach.

Conclusions:

• The proposed unsupervised learning method effectively detects wind generator failures using vibration signals.
• This approach offers a reliable and efficient solution for predictive maintenance in wind energy systems.
• The findings support the use of advanced machine learning techniques to overcome challenges in renewable energy infrastructure.