Improved Intelligent Condition Monitoring with Diagnostic Indicator Selection

Area of Science:

• Engineering
• Data Science
• Artificial Intelligence

Background:

• Traditional predictive maintenance (PdM) systems for wind turbines often use excessive vibration-derived features, potentially degrading performance.
• Generic feature selection methods may misinterpret or lose critical fault information in PdM applications.

Purpose of the Study:

• To develop a structured feature selection method for wind turbine blade defect detection and classification.
• To improve the efficiency and interpretability of PdM systems by reducing feature dimensionality.

Main Methods:

• A novel feature selection approach combining correlation analysis and visual evaluation was proposed.
• This method focuses on preserving domain-specific information, unlike generic dimensionality reduction techniques.
• The reduced feature set was used to train both machine learning (ML) and deep learning (DL) models.

Main Results:

• The proposed method reduced the number of condition indicators by approximately 3.27 times.
• Machine learning models trained on the reduced dataset matched or surpassed complex deep learning models in performance.
• Computational time for prediction was significantly reduced, up to 50% for complex models, without compromising ML model accuracy.

Conclusions:

• A structured, domain-relevant feature selection method enhances PdM systems by streamlining data input and reducing computational demands.
• This approach enables simpler ML models to achieve high accuracy, offering interpretable and cost-effective solutions for industrial applications.
• While DL model metrics were not similarly improved, the study underscores the value of optimized feature sets for efficient PdM.