A Two-Stage Classification Method for Improved Fault Detection in Wind Turbines Based on SCADA Data

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Wind turbine operation relies on effective fault detection.
• Supervised methods struggle with imbalanced SCADA data (normal vs. fault samples).

Purpose of the Study:

• To develop an improved fault detection method for wind turbines using SCADA data.
• To overcome limitations of traditional supervised methods in handling imbalanced datasets.

Main Methods:

• A two-stage approach integrating unsupervised and supervised learning.
• Stage 1: Unsupervised One-Class Support Vector Machine (OCSVM) with anomaly scores to identify deviations.
• Stage 2: Supervised Convolutional Neural Network (CNN) applied to flagged abnormal data for fault identification.

Main Results:

• The two-stage method significantly enhances fault detection performance.
• Demonstrated improvements in accuracy and reduction in missed fault rates compared to purely supervised methods.
• Effective in discriminating between normal and abnormal operational conditions.

Conclusions:

• The proposed two-stage fault detection strategy is highly effective for wind turbine SCADA data.
• This hybrid approach offers a robust solution for reliable wind turbine operation.
• Addresses the challenge of imbalanced datasets in wind turbine fault diagnostics.