Detectability of Structural Damage to a Lattice Tower using Eigenfrequencies and Gaussian Process Regression

One main challenge in structural health monitoring is distinguishing between the effects of actual system changes and changing environmental conditions (EC) on the monitored parameters. For this, data normalisation can be performed. Through the utilisation of Gaussian process (GP) regression, it is possible to map the influences of the ECs on the monitored parameters and simultaneously obtain confidence intervals of the predictions. The assumption of standard GPs is homoscedasticity, meaning a constant noise variance of the parameters. However, there is still a lack of research regarding the correctness of this assumption for different types of structures and damage sensitive features. Furthermore, the potential of normalisation techniques that enable the consideration of EC-dependent uncertainties should be analysed. In this real-world case study, damage detectability is investigated using GPs and eigenfrequencies. In contrast to standard homoscedastic GPs, heteroscedastic GPs can account for input-dependent uncertainties. Both methods are compared using real data from a lattice tower. Firstly, it is shown that the homoscedastic assumption does not hold for the eigenfrequencies of the lattice tower. Consequently, the heteroscedastic GPs enable more reliable damage detection than homoscedastic GPs when using a probabilistic novelty metric. Also, in the heteroscedastic case, the additional information on the uncertainty of the prediction increases the interpretability of the results.