Data-Driven Structural State Estimation via Multi-Fidelity Gaussian Process Models

Abstract

Guided wave-based techniques have been used extensively in Structural Health Monitoring (SHM). Models using guided waves can provide information from both time and frequency domains to make themselves accurate and robust. Probabilistic SHM models, which have the ability to account for uncertainties, are developed when decision confidence intervals are of interest. Most active-sensing guided-wave methods rely on the assumption that a large dataset can be collected, making them impractical when data collection is constrained by time or environmental factors. Meanwhile, although simulation results may lack the accuracy of real-world data, they are easier to obtain. In this context, models that integrate data from multiple sources have the potential to combine the accuracy of experimental data with the convenience of simulated data, without requiring large and potentially costly experimental datasets. The goal of this work is to introduce and assess a probabilistic multi-fidelity Gaussian process regression framework for damage state estimation via the use of both experimental and simulated guided waves. The main differences from previous works include the integration of damage-sensitive features (damage indices, DIs) extracted from both experimental and numerical sources, as well as the use of a relatively small amount of real-world data. The proposed model was validated by two test cases where multiple data sources exist. For each test case, experimental data were collected from a piezoelectric sensor network attached to an aluminum plate with various structural conditions, while simulated data were generated using either multiphysics finite element model (FEM) or physics-based signal reconstruction approaches under the same conditions.