Bridge Structural Health Monitoring Using Mobile Sensor Networks

Abstract

Vehicles moving along a bridge will dynamically interact with the bridge's vibrations. An acceleration sensor on the vehicle can record these vibrations’ signals resulting from different sources of interaction, namely, the bridge vibrations, road roughness, vehicle properties, and speed. Specific algorithms must be implemented to extract the bridge properties from such a contaminated signal. This thesis develops and integrates a framework that extracts the bridge's vibrational properties via multiple vehicle sensor readings. In addition, two new techniques for source separation are introduced. The developed framework is divided into two stages. The first part deals with source separation by removing the vehicle dynamics and road roughness effects from the original signal recorded by the sensor. Two approaches, from the literature, can be used to remove the vehicle dynamics, namely, Frequency Response Function (FRF) and Ensemble Empirical Modal Decomposition (EEMD). Moreover, a new hybrid algorithm (FRF+EEMD) is introduced and tested. Next, two methods are tested to remove the road roughness profile; Second-Order Blind Identification (SOBI) algorithm, from the literature, and Signal Subtraction Algorithm (SSA). SSA is a newly proposed technique where the difference between two vehicles’ responses is used to get the pure bridge frequencies. Then these frequencies are allowed to pass by filtering out the remaining frequencies from the deconvoluted vehicle response to get the pure bridge response. Source separation is carried out for multiple vehicles, then a sparse observation matrix is generated that has bridge vibration readings in both space and time coordinates. In the second stage of the framework, the sparse matrix is completed using the ALS algorithm. New signal processing techniques are introduced to compute the initial guess of the frequencies and prepare the data for structured optimization analysis. The new hybrid technique (FRF+EEMD) and SSA have proven to be comparable or superior to other algorithms found in the literature. Moreover, the introduced signal processing techniques were able to distinguish between vertical and torsional modes of vibration and automatically detect initial guess values close to the actual values of fundamental frequencies.