Computer Vision-Based Structural Health Monitoring and Damage Prognostics of Composite Wind Turbine Blades

The shift toward renewable energy has driven the development of larger wind turbines with longer blades to maximize energy production. This increase in size raises repair and maintenance costs. An effective strategy to reduce these costs is the timely detection of subsurface blade damage while making inspection faster, safer, and less disruptive to turbine operation through automation and digitalization. In addition to early detection, tracking the progression of damage on each individual blade over time is crucial for informed maintenance decisions and more efficient operation. Furthermore, localizing detected damage across the entire blade through 2D captured images and projecting it into a 3D reconstruction provides valuable information about the damage location and helps infer the potential damage type, enabling targeted repairs. Finally, these capabilities highlight the importance of an effective, automated structural health monitoring (SHM) system for wind turbine blades (WTBs) that minimizes operational disruption. A typical SHM system is designed to achieve four specific targets: damage detection, classification, localization, and quantification. Among available inspection methods, computer vision (CV) uniquely has the potential to address all four targets, whereas other techniques only partially fulfill them. Accordingly, the objective of this work is to advance automated CV-based SHM of WTBs by addressing critical gaps in current inspection practices, while aiming to achieve robustness under varying real-world conditions. Specifically, this thesis aims to (1) evaluate the capability of state-of-the-art CV methods and imaging modalities to identify the most effective combinations for different inspection scenarios; (2) explore the capability of combining thermography with a self-learning deep learning model that can learn discriminative texture and morphology patterns, rather than relying only on color variation as in traditional CV models to be able to automatically detect subsurface blade damage under varying illumination conditions (3) investigate the potential of AI-based similarity retrieval models to identify individual blades by retrieving prior inspection images of the same blade as a newly taken one, based on visual similarity, thereby enabling tracking of individual blade damage progression over time without requiring retraining of the entire network as is the case with traditional classification models (4) assess the potential of dense and unconstrained AI-based 3D reconstruction models to reconstruct entire wind turbine blades from unconstrained 2D images without prior knowledge of camera calibration. Accordingly, the first part of this thesis presents a comprehensive evaluation of state-of-the-art CV-based SHM techniques used for WTBs, identifies existing challenges, and proposes potential solutions. The second part introduces an innovative method that integrates passive thermography with a deep learning-based feature learning approach to automatically detect subsurface damage under varying environmental conditions, while being robust against color-similar non-damage regions. The third part of this thesis presents a novel image similarity retrieval model based on Siamese convolutional neural networks to automatically search an existing inspection image database, taken under varying lighting conditions, and retrieve images of the same blade as a query image to identify individual blades. The fourth part of this thesis introduces an AI-based 3D reconstruction model, specifically DUSt3R, capable of predicting the geometry of visible blade surfaces under diverse conditions without any explicit knowledge of camera parameters, thereby providing accurate localization and depth information of the damaged areas.
In summary, this thesis contributes significant advancements to the CV-based SHM of WTBs through automation. The proposed approaches are designed to be robust under environmental variability and have the potential to transform the current labor-intensive and operation-interfering inspection practices in the wind energy sector.