Data-Driven Methods for structural health monitoring of offshore wind turbines and (pedestrian) bridges

Structural Health Monitoring (SHM) is essential for ensuring the operational safety and longevity of
critical infrastructure, such as Offshore Wind Turbines (OWTs) and (pedestrian) bridges. This dissertation
explores data-driven methodologies for SHM, focusing on two primary approaches: event-driven and
natural frequency-based monitoring.
The first part addresses event-driven SHM, activated by significant load events, such as vehicle
crossings on bridges. It is demonstrated that if events can be reliably detected and the induced fatigue
damage predicted using a Machine Learning (ML) model, scenario-based estimations of remaining
structural lifetime become feasible. However, event detection is often complicated by overlapping events
and the temperature sensitivity of (optical) strain sensors. To address this, a novel strain monitoring
methodology is introduced, emphasizing temperature compensation through the use of Particle Filters
(PFs). This method is validated through case studies on a bio-composite pedestrian bridge, showcasing
its effectiveness in event-driven SHM applications.
The second part investigates natural frequency-based SHM, which continuously monitors structures by
detecting deviations in natural frequencies—an indicator of potential damage. An initial unsupervised ML
approach is introduced for novelty detection in OWTs, leveraging tracked natural frequencies. However,
due to the laborious nature of manual frequency tracking and the challenges posed by Environmental and
Operational Variability (EOV), an unsupervised clustering approach is developed to automate this process.
To further address these challenges, a novel methodology is proposed that uses the same ML model for
EOV normalization and mode tracking, in combination with Uncertainty Quantification (UQ) techniques.
This integrated approach enables smart tracking of natural frequencies and improves the reliability of
novelty detection. The methodology is tested on both real-world and synthetic OWT monitoring data,
leading to enhanced accuracy in damage detection and overall structural health assessment. Additionally,
a hybrid damage quantification approach is introduced, combining physical modeling with data-driven
techniques to assess detected damage in OWT foundations.
This dissertation advances the integration of data-driven methods in SHM, particularly emphasizing
the use of PFs for temperature compensation in strain measurements and the combination of ML with UQ
for automated mode tracking in OWTs. It establishes a foundation for future research aimed at optimizing
SHM strategies for both event-driven and continuous SHM.