Data-driven modeling of long temperature time-series to capture the thermal behavior of bridges for SHM purposes

Abstract

Bridges experience complex heat propagation phenomena that are governed by external thermal loads, such as solar radiation and air convection, as well as internal factors, such as thermal inertia and geometrical properties of the various components. This dynamics produces internal temperature distributions which cause changes in some measurable structural responses that often surpass those produced by any other load acting on the structure or by the insurgence or growth of damage. This article advocates the use of regression models that are capable of capturing the dynamics buried within long sequences of temperature measurements and of relating that to some measured structural response, such as strain as in the test structure used in this study. Two such models are proposed, namely the multiple linear regression (MLR) and a deep learning (DL) method based on one-dimensional causal dilated convolutional neural networks, and their ability to predict strain is evaluated in terms of the coefficient of determination R2. Simple linear regression (LR), which only uses a single temperature reading to predict the structural response, is also tested and used as a benchmark. It is shown that both MLR and the DL method largely outperform LR, with the DL method providing the best results overall, though at a higher computational cost. These findings confirm the need to consider the evolution of temperature if one wishes to setup a temperature-based data-driven strategy for the SHM of large structures such as bridges, an example of which is given and discussed towards the end of the article.