Degree

Doctor of Philosophy (PhD)

Department

Civil and Environmental Engineering

Document Type

Dissertation

Abstract

This dissertation develops and validates an integrated framework for predicting fatigue life in wind turbine blades, combining physics-based simulations, full-scale aeroelastic testing, and machine learning surrogates. The goal is to clarify fatigue mechanisms under turbulent inflows and provide decision-quality tools for durability assessment. The first part establishes a turbulence-resolved OpenFAST-ANSYS framework for fatigue analysis of an NREL-5MW blade. By preserving the aero-structural chain from inflow turbulence to stress analysis, the study isolates the influence of ply-level mechanics on durability. Material comparison revealed that Kevlar 49 provides orders-of-magnitude longer fatigue lives in endurance-controlled regimes, T300 carbon/epoxy offers intermediate performance, and the hybrid P2B laminate fails almost immediately due to weak transverse and shear strengths. These mechanistic insights explain the abrupt collapse of P2B and highlight stress-range mitigation as essential at high wind speeds, where all materials converge to short lifetimes. The second part extends validation through full-scale experiments in LSU’s WISE Lab. Non-contact modal tests and accelerometer-based response measurements under turbulent inflows were used to validate a calibrated LES-WALE CFD model. When coupled with ANSYS, this experiment-anchored framework identified the blade root as the controlling hotspot, with life consumption dominated by high-wind turbulence. The staged Tier 1 (coupon S-N) and Tier 2 (aeroelastic response) validations provide the highest feasible fidelity short of full-scale fatigue-to-failure testing. The final part explores machine learning surrogates. Random Forest (RF) and Multi-Layer Perceptron (MLP) models were trained on Tier 1-2 validated datasets. RF outperformed MLP across materials and wind speeds, achieving lower errors and showing resilience under limited training data. Its ensemble structure effectively captured abrupt behaviors such as P2B’s strength-controlled failures and Kevlar 49 exponential life regime. MLP captured nonlinear fatigue patterns but required larger, richer datasets to remain stable. Beyond predictive accuracy, ML models achieved unprecedented computational gains: whereas ANSYS simulations required 22 hours to 2.5 days per case, RF completed predictions in ~30 minutes and MLP in ~110 minutes, offering speedups of two to three orders of magnitude. These results establish ML not as a replacement, but as a transformative complement to physics-based methods, enabling real-time monitoring, accelerated design iterations, and large-scale parametric studies. Together, this work establishes an experimentally anchored, turbulence-resolved, and computationally scalable framework for fatigue assessment, offering a foundation for more durable and cost-effective wind turbine blade designs.

Date

11-3-2025

Committee Chair

Aly, Aly

Download

Available for download on Monday, November 01, 2032

Share

COinS