A1302
Title: Deriving Gaussian processes for physics-informed structural health monitoring
Authors: Matthew Jones - University of Sheffield (United Kingdom) [presenting]
Daniel Pitchforth - University of Sheffield (United Kingdom)
Elizabeth Cross - University of Sheffield (United Kingdom)
Abstract: In engineering, it is vital that high-value assets such as bridges and aircraft are managed, operated and maintained in both a cost-effective and safe manner. Structural health monitoring has established itself as a promising solution to intelligent asset management, where measurement data obtained from sensors attached to the structure are used to monitor its condition in real time. Machine learning tools such as Gaussian processes are then used to interpret the data and subsequently identify the health state of the structure. One limitation of the use of machine learning tools in structural health monitoring is that training data are required to represent all of the environmental and operational conditions that the structure will experience, in addition to all damage states of interest. In many practical monitoring scenarios, collecting a data set that is extensive is infeasible and limits the effectiveness of data-driven monitoring strategies. The aim is to propose fusing physical knowledge inside Gaussian processes to mitigate against limited training data. By fusing known physics into a data-driven learner, modelling practitioners can combine the expressive power of machine learners with known mechanistic laws, offering enhanced predictive performance where training data is lacking, in addition to improving model generalization. The method will be demonstrated in a real-life case study.
