Identifying evolving leading edge erosion by tracking clusters of lift coefficients

Imad Abdallah, Gregory Duthé, Sarah Barber, Eleni ChatziJournal of Physics Conference Series 2022 · DOI

Abstract

This work proposes an approach to identify Leading Edge Erosion (LEE) of a wind turbine blade by tracking evolving and emerging clusters of lift coefficients $C_L$ time-series signals under uncertain inflow conditions. Most diagnostic techniques today rely on direct visual inspection, image processing, and statistical analysis, e.g. data mining or regression on SCADA output signals. We claim that probabilistic multivariate spatio-temporal techniques could play an eminent role in the diagnostics of LEE specifically leveraging $C_L$ time-series signals form multiple sections along the span of the blade. The proposed method extracts clusters’ features based on Variational Bayesian Gaussian Mixture Models (VBGMM) and tracks their spatial and temporal changes, as well as interpret the evolution of the clusters through prior physics-based assumptions. The parameters of the VBGMM are the mean, the eigenvalues and eigenvectors of the covariance matrix, and the angle of orientation of the eigenvectors. We show that the distribution of the $C_L$ data may not show statistically separable clusters, however, the parameters of the VBGMM clusters fitted to the $C_L$ data, allows to discriminate moving clusters primarily due to varying inflow and operating conditions, versus emerging clusters primarily due to evolving severity of the blade LEE.

Introduction: The Challenge of Leading Edge Erosion

Wind turbine blades are constantly exposed to harsh environmental elements like rain, sand, and hail, leading to Leading Edge Erosion (LEE). Often described as a "silent killer" of turbine efficiency, LEE begins with minor surface pitting but eventually degrades the blade's aerodynamic profile. This degradation results in significant power production losses and can eventually compromise structural integrity.

Schematic of wind turbine blade sections showing Leading Edge Erosion (LEE) and relative velocity vectors.

Detecting LEE in its early stages is notoriously difficult. Current industry standards rely on expensive manual inspections via drones or climbers. While SCADA (Supervisory Control and Data Acquisition) data is available, it is often too "noisy" to provide clear diagnostic signals. Turbulent wind conditions create performance fluctuations that are hard to distinguish from permanent structural damage. This research aims to isolate the true signal of erosion from environmental noise.

Methodology: Spatio-Temporal Clustering

The proposed approach moves beyond single-point analysis by comparing lift coefficients ($C_L$) across multiple sections of the blade span.

Variational Bayesian Gaussian Mixture Models (VBGMM)

The researchers employed VBGMM to analyze the data. Unlike standard clustering methods that require a pre-defined number of clusters, VBGMM automatically determines the optimal number of clusters using a Bayesian framework. This makes the model more robust and adaptable to the complex, multi-modal nature of aerodynamic data.

Feature Extraction and Physics-Informed Tracking

The core innovation lies in tracking the geometric properties of these clusters over time rather than just the raw data points. The model extracts:

  • Mean: The central tendency of the lift coefficients.
  • Eigenvalues and Eigenvectors: Describing the spread and shape of the data clusters.
  • Angle of Orientation: The direction of the cluster's main axis.

By integrating prior aerodynamic knowledge—specifically how erosion affects the lift curve slope and stall angle—the authors can interpret shifts in these parameters as physical changes in blade health.

Data Generation

The study utilized high-fidelity simulations to test the model:

  • OpenFAST: Used to model the NREL 5-MW reference wind turbine.
  • Stochastic Erosion Model: A Non-Homogeneous Compound Poisson Process (NHCPP) simulated 20 years of wear, modeling erosion as discrete "shocks" (e.g., hailstorms).
  • Turbulent Inflow: Realistic wind variables, including turbulence intensity and inflow skew, were included to ensure robustness against environmental noise.

Key Insights and Results

A primary finding of the study is the "inseparability" problem: at high wind speeds, raw $C_L$ data for clean and eroded blades often overlap because the turbine's pitch control system masks the damage.

Lift coefficient (CL) clusters at different erosion severity levels, showing overlapping data but distinct orientations.

However, while the data points overlap, the parameters of the fitted VBGMM clusters remain distinct. The angle of orientation of the cluster emerged as a highly sensitive health indicator. As LEE severity increases, the blade stalls earlier, changing the relationship between lift coefficients at different sections. This physical change manifests as a systematic rotation of the cluster's main axis, allowing for clear discrimination between erosion levels even when raw performance metrics appear stable.

The method successfully captured rapid degradation events and proved robust under highly uncertain inflow conditions, suggesting that cluster orientation can serve as a reliable predictor of blade health.

Conclusion and Future Directions

While these results are promising, several challenges remain for real-world implementation. The transition from high-fidelity simulations to field data involves managing sensor noise, drift, and calibration issues. Furthermore, this diagnostic method requires high-frequency $C_L$ measurements from multiple blade sections—hardware that is not yet standard on commercial turbines, though emerging MEMS-based pressure systems may bridge this gap.

Future work may explore more advanced machine learning architectures, such as Variational Autoencoders (VAEs) or contrastive learning, to perform clustering in latent spaces, potentially uncovering even more subtle indicators of structural degradation.

Summary updated

Abstract

Abstract This work proposes an approach to identify Leading Edge Erosion (LEE) of a wind turbine blade by tracking evolving and emerging clusters of lift coefficients C L time-series signals under uncertain inflow conditions. Most diagnostic techniques today rely on direct visual inspection, image processing, and statistical analysis, e.g. data mining or regression on SCADA output signals. We claim that probabilistic multivariate spatio-temporal techniques could play an eminent role in the diagnostics of LEE specifically leveraging C L time-series signals form multiple sections along the span of the blade. The proposed method extracts clusters’ features based on Variational Bayesian Gaussian Mixture Models (VBGMM) and tracks their spatial and temporal changes, as well as interpret the evolution of the clusters through prior physics-based assumptions. The parameters of the VBGMM are the mean, the eigenvalues and eigenvectors of the covariance matrix, and the angle of orientation of the eigenvectors. We show that the distribution of the C L data may not show statistically separable clusters, however, the parameters of the VBGMM clusters fitted to the C L data, allows to discriminate moving clusters primarily due to varying inflow and operating conditions, versus emerging clusters primarily due to evolving severity of the blade LEE.
BibTeX
@article{abdallah2022identifying,
  author = {Imad Abdallah and Gregory Duthé and Sarah Barber and Eleni Chatzi},
  title = {{Identifying evolving leading edge erosion by tracking clusters of lift coefficients}},
  journal = {Journal of Physics Conference Series},
  year = {2022},
  doi = {10.1088/1742-6596/2265/3/032089},
}
Figure