Anomaly Detection in Wind Turbines: Persistence-Based Alarm Confirmation for False-Alarm Mitigation and...

Area of Science:

• Wind Turbine Engineering
• Condition Monitoring Systems
• Data Analytics

Background:

• Anomaly detection models for wind turbines often use only healthy data due to limited failure data.
• These models can generate false alarms from normal operational variations, not just faults.
• Alarm system effectiveness relies on both the detection model and the decision rule for alarm activation.

Purpose of the Study:

• To evaluate the impact of persistence-based alarm confirmation on wind turbine anomaly detection.
• To analyze how different temporal persistence settings affect alarm outcomes across various detection methods.
• To provide guidance on configuring alarm systems for effective wind turbine condition monitoring.

Main Methods:

• Evaluated four anomaly detection techniques: Isolation Forest, One-Class SVM, PCA, and Autoencoder.
• Used OpenFAST simulations for controlled rotor unbalance scenarios.
• Analyzed an industrial SCADA dataset with a known main bearing fault.

Main Results:

• Temporal persistence significantly influences alarm outcomes.
• Moderate persistence reduces false positives while maintaining detection for severe faults.
• Excessive persistence leads to delayed detection and missed faults, especially for subtle issues.

Conclusions:

• Persistence-based alarm confirmation is a critical decision-level parameter.
• Optimizing persistence settings is key to balancing detection speed and accuracy.
• This approach enhances the reliability of anomaly detection in wind turbine monitoring.