Vibration-Based Damage Detection Using Machine Learning Techniques with Application to Offshore Wind Turbines

Abstract

Although vibration-based damage detection (VBDD) has seen great advancements since the 1980s, it is still facing several challenges, such as identifying features of vibration measurements that are sensitive to damage, tackling nonlinear structures, and removing environmental and operational variation (EOV) effects. This thesis is devoted to addressing these challenges, with application to offshore wind turbines (OWTs) for which there is demand for VBDD techniques for cost reduction.

A new VBDD method is proposed for linear structures subject to Gaussian white noise loading. It characterises vibration responses with the multivariate probability density function (PDF) of the underlying stochastic process. The change of the PDF, measured by the Kullback-Leibler divergence, is defined as a damage index. From a case study of an OWT with a reduction to the foundation stiffness, it is shown that the proposed method offers significant performance improvements over an autoregressive-based method and an autocorrelation function-based method.

For structures that are nonlinear both before and after damage, the PDFs become non-Gaussian and difficult to estimate. Therefore, the density ratio estimation method, which directly estimates the difference of the PDFs rather than the individual PDFs themselves, is applied to data from the undamaged state and a potentially damaged state. The effectiveness and advantages of the proposed method are demonstrated in two case studies: an experimental nonlinear beam and an OWT with nonlinear pile-soil interaction.

To remove EOV effects, a new approach is proposed and applied to OWTs. It takes advantage of the fact that OWTs are built in groups and thus are subjected to similar EOVs at the same time. When the monitored features are viewed relative to each other, the complex EOV effects are probably cancelled.
The method is demonstrated on a wind farm with three OWTs affected by scouring damage, and the results show that the damage is detected and localised to a specific OWT.

Date of Award	2 Dec 2021
Original language	English
Awarding Institution	University of Bristol
Supervisor	John H G Macdonald (Supervisor), Paul W Harper (Supervisor) & Song Liu (Supervisor)