A Novel Approach to Fatigue: Grey-box Modelling for Probabilistic Damage Assessment

There is a growing interest in monitoring the loads that engineering structures withstand in order to better predict the fatigue damage accrual that has been accumulated. This is done with two goals: improving the safety of structures and enabling a greater useful life to be achieved. However, when structures operate in harsh environments, such as aircraft or offshore wind turbines, maintaining sensing networks for measuring stress and strain at crucial locations is difficult. As a result, virtual loads monitoring, or inferential sensing, in which machine learning methods are used to predict the stress at critical locations is becoming increasingly popular.

In the first part of this thesis, Gaussian process (GP) regression is used to develop a probabilistic approach for fatigue. The choice of a GP for a virtual sensor is not uncommon, however, developing a probabilistic view of fatigue by propagating the model uncertainty throughout the fatigue assessment procedure is novel. By doing this, a more robust assessment of the damage state of the structure is achieved. Furthermore, a discussion is facilitated around the causes and consequences of uncertainty in data-driven models with respect to fatigue assessment.

Treating fatigue analysis probabilistically is considered to be one way of reducing the conservatism that is common as a result of many uncertainties in the assessment procedure. In this work, it is the uncertainty from a loading perspective that is considered, but this thesis will also discuss how this could fit in with other existing probabilistic methods.

Following this, grey-box modelling - the introduction of our knowledge of physics into data-driven models – is considered. This is a new area of research currently attracting significant interest and different methods of inputting this physical knowledge into the model are presented. By considering different loading scenarios (both dynamic and quasi-static), it is shown that domain-specific knowledge can both improve the accuracy, and also reduce the uncertainty, of model predictions, with the impact on probabilistic fatigue prediction being significant.