Advances in data-based modelling for structural health monitoring systems

Probability and statistical applications can be found spread across various scientific and engineering disciplines. In the field of Structural Health Monitoring (SHM), promising advances have been made possible with the development of statistical models. Challenges that were once too complex to solve analytically can now be addressed with the assistance of intelligent monitoring systems, which are, fundamentally, driven by statistical pattern
recognition and machine learning algorithms. Over the years, numerous data-driven approaches have emerged, collectively aiming to make SHM a standard practice. This thesis explores the use of novel statistical models to facilitate the implementation of intelligent health-monitoring systems; namely, by focussing on nonparametric Bayesian modelling and, to a lesser extent, autoencoders, to address a few prevalent challenges currently encountered in SHM.
Within this framework, the problem of model selection becomes central, determining how well a system is represented in a statistical sense. The main body of this thesis thus delves into this issue, alongside the rationale for adopting models that are both nonparametric and Bayesian in engineering applications.
A series of case studies are presented, each highlighting unique challenges in SHM. These case studies are approached with models based on either Gaussian Processes (GPs) or Dirichlet Processes (DPs). Initially, GPs are employed to enhance localisation techniques in SHM. These methodologies demonstrate their capabilities to detect abnormal operations in a journal bearing using ultrasonic measurements. Additionally, they are used to simplify the process of localising damage sources in composite structures using Acoustic Emission (AE) data. The following component of this thesis introduces an approach to identifying AE events in time-series signals. It further incorporates a DP prior into a mixture model, designed to autonomously capture changes introduced by different sources of damage. The methodology is validated using AE data collected from a fatigue test of an Airbus A320 landing gear. While a significant portion of the SHM literature focusses on data-driven models, there is a growing interest in integrating physics into these models. Therefore, this thesis concludes with insights into the input-state-parameter estimation of journal bearings by combining GPs with the dynamics of journal response in a state-space representation. The miscellany of methodologies presented in this thesis is but a small contribution to bridging the gap between state-of-the-art machine learning techniques and their application in SHM.