Deep Gaussian process regression for damping of a long-span bridge under varying environmental and operational conditions

Abstract

Modal parameters are critical for wind resistant design and vibrational serviceability assessments of long-span cable-supported bridges. In contrast to the successful research efforts into natural frequencies, there are still challenges in modeling the damping ratio due to the following aspects: (1) inherent errors in damping estimates, (2) lack of insight into the damping mechanisms, and (3) epistemic uncertainties on the effects of environmental and operational conditions (EOCs). This paper proposes a probabilistic regression model for damping using Deep Gaussian Processes (DGP) on damping estimates compiled from 2.5 years of structural health monitoring (SHM) data from a cable-stayed bridge. Input features representative of EOCs theorized to be related to damping ratios from past literature were used. Two data cleaning strategies based on statistics and knowledge-based criteria were used for enhancing the model performance. A comparative study with DGPs and different regression models were carried out to confirm the robustness of DGPs across different datasets. A knowledge-based feature engineering process examined the most significant predictor of the damping ratios. The proposed data-driven regression model can enable a probabilistic consideration of damping in structural design and vibrational serviceability assessments.

Doyun Hwang

Ph.D. Student

(Incoming) Ph.D. student at the Structures as Sensors laboratory at Stanford University.