Anomaly Detection of Wind Turbine Driveline Based on Sequence Decomposition Interactive Network

Area of Science:

• Engineering
• Data Science
• Artificial Intelligence

Background:

• Wind turbines (WTs) generate crucial data via Supervisory Control and Data Acquisition (SCADA) systems.
• Accurate health state identification is vital for efficient wind energy production and maintenance.
• Existing anomaly detection methods may not fully leverage spatio-temporal information in SCADA data.

Purpose of the Study:

• To propose a novel anomaly detection method, the decomposed sequence interactive network (DSI-Net), for identifying the health state of wind turbines.
• To effectively utilize comprehensive spatio-temporal information from SCADA data for enhanced anomaly detection.
• To improve the accuracy and reliability of wind turbine health monitoring.

Main Methods:

• A DSI-Net model is trained on preprocessed SCADA data from healthy wind turbines.
• The DSI-Net extracts underlying spatio-temporal features through an interactive learning process, decomposing data into trend and seasonality subsequences.
• Online data is processed to calculate residuals, which are then analyzed using Exponential Weighted Moving Average (EWMA) on residuals and Root Mean Square Error (RMSE) for anomaly identification.

Main Results:

• The DSI-Net model successfully identifies underlying spatio-temporal features from SCADA data.
• Residuals and RMSE, processed by EWMA, effectively flag anomalies in wind turbine operations.
• Control experiments demonstrate that DSI-Net outperforms existing anomaly detection models in accuracy and effectiveness.

Conclusions:

• DSI-Net provides an effective and accurate method for wind turbine anomaly detection using SCADA data.
• The model's ability to capture spatio-temporal features enhances the identification of wind turbine health states.
• This approach offers a significant advancement in condition monitoring for wind energy systems.