Wind Turbine Gearbox Condition Monitoring Based on Class of Support Vector Regression Models and Residual Analysis

Area of Science:

• Engineering
• Data Science
• Renewable Energy

Background:

• Wind turbine downtime impacts reliability and operational costs.
• Intelligent condition monitoring systems are crucial for proactive maintenance.
• Supervisory Control and Data Acquisition (SCADA) data offers valuable insights.

Purpose of the Study:

• To develop a feature selection-based methodology for identifying faulty scenarios in wind turbines.
• To assess the prediction performance of machine learning regression models using SCADA data.
• To improve the accuracy and reliability of wind turbine condition monitoring.

Main Methods:

• Utilized Neighborhood Component Analysis (NCA) for feature selection.
• Applied machine learning regression models to predict gearbox oil and bearing temperatures.
• Analyzed Supervisory Control and Data Acquisition (SCADA) data from 1009 samples prior to failure.
• Performed statistical tests (Diebold-Mariano, Durbin-Watson) on Support Vector Regression (SVR) model residuals.

Main Results:

• Twin Support Vector Regression achieved 99.91% accuracy for gearbox oil temperature prediction.
• Decision Trees achieved 98.74% accuracy for bearing temperature prediction.
• NCA significantly increased the accuracy and reliability of the condition monitoring system.

Conclusions:

• Feature selection using NCA enhances the performance of machine learning models for wind turbine condition monitoring.
• The proposed methodology offers a reliable approach to reduce wind turbine downtime.
• Statistical validation confirms the robustness of the employed regression models.