A Data-Driven Diagnostic Framework for Wind Turbine Structures: A Holistic Approach

Simona Bogoevska, Minas Spiridonakos, Eleni N. Chatzi, Elena Dumova-Jovanoska, Rudiger Höffer

TL;DR

This paper introduces a novel "bi-component" framework for continuously monitoring the health of wind turbine structures. It combines short-term time-varying models to capture rapid dynamic changes with long-term probabilistic models to account for environmental and operational influences, enabling more robust and automated structural health monitoring.

What Problem Does This Paper Address?

Wind turbine structures are complex dynamic systems subjected to constantly changing environmental conditions (wind, temperature) and operational regimes (rotor speed, pitch angle). Existing structural health monitoring (SHM) methods, often based on operational modal analysis (OMA), struggle to accurately capture this inherent time-variability and non-stationarity. This can lead to false alarms or missed damage detection, hindering efficient operation and maintenance. The challenge is to develop an SHM framework that can holistically model the complete operational spectrum of a wind turbine, distinguishing between benign variations and actual structural degradation.

Methods

The proposed framework is a "bi-component" approach:

  • Short-Term Framework (SP-TARMA): This component uses Smoothness-Priors Time-Varying Autoregressive Moving Average (SP-TARMA) models. Unlike traditional ARMA models that assume stationarity, SP-TARMA models can adapt to time-varying parameters, effectively capturing the rapid fluctuations in the turbine's dynamics caused by aerodynamic loads and operational changes. This is achieved by introducing stochastic difference equations that govern the evolution of the AR and MA coefficients.
  • Long-Term Framework (PCE): This component addresses the influence of environmental and operational variables (EOPs) on the structural response. It employs Polynomial Chaos Expansion (PCE). PCE is a non-sampling method that constructs a functional dependence of the model output (e.g., residuals from the SP-TARMA model) on a set of independent random variables representing the EOPs. This allows for the quantification of uncertainty in the structural response due to these external factors.
  • Integration: The SP-TARMA model provides short-term performance indicators (residuals), which are then used as inputs for the PCE model. The PCE model quantifies how these indicators vary with EOPs, leading to a long-term diagnostic index.
  • Data Preprocessing: To satisfy the PCE requirement of independent input variables, Independent Component Analysis (ICA) is used to transform correlated SCADA data into independent latent variables. These are then further transformed into uniformly distributed variables.

Datasets

The framework was validated using real-world data from two operating wind turbine structures in Germany:

  • Case A: A 2MW WT in Lübbenau, Germany. Data included 10-minute acceleration datasets (sampled at 1600 Hz) over 29 days, along with 10-minute averaged SCADA data. For the long-term analysis, a dataset spanning 1388 10-minute intervals over two months was used.
  • Case B: A 0.5MW WT near Dortmund, Germany. Data included acceleration response (sampled at 100 Hz) and SCADA data over four months (May-August 2010), comprising 17,706 1-hour datasets. For the long-term analysis, a full year of 2012 data (49,956 datasets) was utilized, with temperature added as an input.

Key Insights and Ideas

  • Holistic Modeling: The bi-component approach is crucial. Simply modeling vibration response without considering environmental and operational factors leads to inaccurate diagnostics. Conversely, only looking at EOPs misses the structural dynamics.
  • Time-Varying Dynamics are Key: The SP-TARMA model's ability to capture time-varying parameters is essential for accurately representing the dynamic behavior of operating wind turbines, which standard OMA methods fail to do.
  • Quantifying Uncertainty: PCE provides a powerful way to quantify how EOPs influence structural response, allowing for more robust condition monitoring by accounting for expected variations.
  • Automated SHM Potential: The framework is designed to be automated, moving towards a diagnostic tool that can continuously track structural health over the entire lifecycle of a wind turbine.
  • Distinguishing Benign vs. Malignant: The diagnostic index developed aims to differentiate between normal operational fluctuations and potential structural issues, reducing false alarms.

Key Results

  • The SP-TARMA models successfully tracked the evolution of time-varying frequencies in operating wind turbines, outperforming stationary ARMA models.
  • The PCE model demonstrated a good ability to simulate the standard deviation of the SP-TARMA residuals, indicating its effectiveness in capturing the influence of EOPs.
  • The developed Diagnostic Index (DI) showed consistency in tracking structural behavior over extended periods (one year for Case B), with errors falling within 99.7% confidence intervals.
  • The study identified necessary training periods for different SCADA variables to capture their full range of values, highlighting the importance of sufficient historical data for robust modeling.

Limitations and Future Work

  • Parameter Tuning: While the paper demonstrates parameter tuning for the SP-TARMA models using BIC and AIC, it notes that optimal tuning might require further investigation, especially concerning the residual variance ratio.
  • Damage Detection: The current study focuses on demonstrating the framework's applicability and robustness in monitoring healthy structures under varying conditions. The explicit detection and localization of structural damage were not the primary focus but are a clear next step.
  • Model Retraining: The paper suggests that instead of alarming for potential maintenance, model retraining should be considered when new ranges of measured input data are encountered. This implies a need for adaptive retraining strategies.
  • Future Work: The authors propose expanding validation data to subsequent years, interrelating typical operating regimes with specific structural behavioral patterns, and developing an algorithm to separate benign pattern distortions from malignant alterations due to structural damage. Further work will also explore fusing these surrogate models with simulation-based design procedures for fatigue analysis.

Summary updated

BibTeX
@inproceedings{bogoevska2017data,
  author = {Simona Bogoevska and Minas Spiridonakos and Eleni N. Chatzi and Elena Dumova-Jovanoska and Rudiger Höffer},
  title = {{A Data-Driven Diagnostic Framework for Wind Turbine Structures: A Holistic Approach}},
  booktitle = {Sensors},
  year = {2017},
  url = {https://openreview.net/forum?id=Ui54N1btps},
}
Figure