Project: #273 Vehicle-based Panhandle Bridge Monitoring Progress Report - Reporting Period Ending: Sept. 30, 2019 Principal Investigator: Hae Young Noh Status: Overdue Project Start Date: Sept. 1, 2018 End Date: Aug. 31, 2019 Research Type: Applied Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2018 Traffic21 Progress Report (Last Updated: Sept. 29, 2019, 9:27 a.m.) % Project Completed to Date: 100 % Grant Award Expended: 100 % Match Expended & Document: 100 USDOT Requirements Accomplishments The 2017 National Bridge Inventory of the Federal Highway Administration found that 47,619 out of 615,002 bridges in the U.S. were in poor condition across the nation. These make apparent the need for monitoring bridge conditions. Recently, indirect structural health monitoring of bridges has become popular since it is a low-cost and low-maintenance approach in which sensors on the vehicle are used to detect infrastructure changes and damage. The ultimate objective of this project is to develop a system that would provide continuous monitoring of bridges by collecting vibration data from sensors on in-service trains. Ideally, such a system would be able to detect, localize and quantify the damage of tracks soon after they begin to occur. To achieve the goal, we have conducted lab-scale experiments and provided evidence of the applicability of the indirect damage diagnosis approach through better accuracy from the indirect sensors than the direct bridge sensors in certain cases. To validate the robustness of this indirect damage diagnosis framework with a more complex and realistic system, field experiments on real-world bridges are needed. Over the six months, we conducted on-site tests on the Panhandle bridge, a steel truss bridge (including three Pennsylvania trusses) carrying two rail lines across the Monongahela River in Pittsburgh, Pennsylvania. To represent damage or changes in structure, we positioned one or two train cars along one of the tracks on the bridge as the extra load. The other track was used for running an operational train back and forth, carrying accelerometers to measure vibration signals. Four sets of experiments were conducted for representing four different damage scenarios: unloaded bridge; single train car loaded on the first truss; single train car loaded on the second truss; two train cars loaded on the second truss. We evaluated the ability of our previous framework to detect the position and magnitude of the damage (stationary trains). We obtained good classification accuracy for the indirect sensors, which is a step forward for proving the feasibility of the indirect bridge monitoring. However, to ensure adequate performance, our previous data-driven approach requires a large set of training data, which is expensive and impractical. In addition, the limited physical interpretability of these approaches constrains the application of the learned model on other bridges. We encountered this challenge in our field experiment on the Panhandle bridge. To address the challenge, we represented the dynamic vehicle-bridge interaction (VBI) system as a state-space model. We introduced a hybrid algorithm that uses non-linear least squares to pre-estimate modal and structural parameters of the VBI system and updates the parameters and uncertainty co-variance matrices by using the expectation-maximization algorithm. To eliminate unrealistic local optima and accelerate convergence, we constrained the space of the parameters by using the domain knowledge of the bridge. We applied the proposed methodology to the Panhandle bridge-vehicle system. Though the EM algorithm with domain knowledge for system identification has been studied before, to the best of our knowledge this is the first time that it is applied in the context of a VBI system and more importantly, the first time that it is evaluated on a full-scale bridge under operation. Compared to a data-driven method, our proposed approach improves the error of the bridge loading estimation by 36.3%. This grant has helped us to deploy vibration sensors on the Panhandle bridge, and we have been collecting data for validating our indirect damage diagnosis approaches in practice. We have presented the work at the International Workshop on Structural Health Monitoring. In addition, we are working on writing a manuscript for publishing this work in a journal. Impacts We have made several contributions as listed below: • We designed and conducted field experiments on a real-world truss bridge for evaluating the applicability of the indirect structural health monitoring. This advances the science of bridge monitoring. • We developed a data management system for collecting and processing dynamic responses from the bridge and the moving vehicle. • We published (open source) a dataset recording dynamic responses from two in-service light rail vehicles that run on a 42.2-km light rail network in the city of Pittsburgh, Pennsylvania (USA). A data descriptor journal that describes this dataset was submitted. This allows the scientific and engineering communities to expand our work or utilize the data to benefit their rail asset maintenance. • We modeled the relationship between the acceleration signals collected from the moving vehicle and the bridge conditions by applying dimensionality reduction and a regression model to the data. We obtain high classification accuracy for the indirect sensors, which provides evidence of the applicability of the indirect monitoring approach. • We introduced a vehicle-vibration-based indirect structural health monitoring (SHM) framework that uses acceleration signals collected from within a moving vehicle to identify global modal and structural parameters of a full-scale and in-service bridge. Other We collected dynamic responses of the Panhandle bridge and of a light rail vehicle passing over the bridge with different surrogate damage scenarios. We developed an indirect SHM approach after modeling the VBI system as a linear time-varying Gaussian state-space model. We used a hybrid algorithm that uses non-linear least squares to pre-estimate the parameters and applies the expectation-maximization algorithm to handle the incomplete-data situation. Outcomes New Partners None Issues None