A Simplified Treed Gaussian Process Approach to the Modelling of Bridge Data for Structural Health Monitoring

Abstract

Data-based modelling in the presence of uncertainties is a key challenge for Structure Health Monitoring (SHM) practitioners. In the scenario of applying damage detection methods using time series analysis, many of the approaches are based on Statistical Process Control (SPC). The inherent limitation within a SPC-based method is the restriction to stationary data. However, given the inevitable presence of uncertainties from the environment, stationary data acquisition is not guaranteed. To explore the potential relations behind the data while accommodating the issue of non-stationary behaviour, a Treed Gaussian Process (TGP) can be brought into action. TGP models have been shown to be successful in tackling non-stationary regression problems effectively, and have been adopted and applied across numerous research fields including SHM. However, a key limitation that has been found is the computational complexity of these models. This paper introduces a simplified TGP model that is distinct from previous formulations in being based on a different and less complex statistical model. The simplified version presented here reduces the complexity by removing and replacing parameters from previous formulations, as well as sampling and evaluating using a different criterion. The objective for the algorithm is to maintain accuracy and effectiveness while reducing the computational cost of implementation. The simplified algorithm is demonstrated here for two case studies. First, a benchmark comparison is made using a well-known motorcycle accident dataset. Secondly, the algorithms will be compared using experimental data from the well-known Z24 bridge SHM dataset.