Exploring Environmental and Operational Variations in SHM Data Using Heteroscedastic Gaussian Processes

Abstract

The higher levels of Structural Health Monitoring (SHM)—localisation, classification, severity assessment—are only accessible using supervised learning in the data-based approach. Unfortunately, one does not often have data from damaged structures; this forces a dependence on unsupervised learning i.e. novelty detection. This means that detection is sensitive to benign environmental and operational variations (EOVs) in or around the structure. In this paper a two-stage procedure is presented: identify EOVs in training data using a nonlinear manifold approach and remove EOVs by utilising the interesting tool of heteroscedastic Gaussian processes (GPs). In Classical GPs models, the data noise is assumed to have constant variance throughout the input space. This assumption is a drawback most of the time, and a more robust Bayesian regression tool where GP inference is tractable is needed. In this work a combination of data projection and a non-standard heteroscedastic GP is presented as means of visualising and exploring SHM data.