Damage Detection Approach for Bridges under Temperature Effects using Gaussian Process Regression Trained with Hybrid Data

Abstract

The success of detecting damage robustly relies on the availability of long periods of past data covering multiple weather scenarios and on the information contained in the data used during the learning process. Thus, the innovation of this paper is to apply a hybrid data set to train a Gaussian process regression, assuming a practically plausible range of environmental conditions. The proposed model presents a satisfactory performance to detect damage when structural changes caused by damage are blurred with changes caused by temperature. Rather than relying exclusively on experimental data, this strategy use finite-element models to generate complementary data when the structure is undamaged under a broad spectrum of temperature variations that are not measured. Once the stochastic interpolation is defined, the damage detection model is tested using experimental data considering different damage levels and temperature conditions. Induced settlements of a bridge pier are used as realistic damage scenarios. The Z24 prestressed concrete highway bridge in Switzerland is used to demonstrate the applicability of the proposed strategy.