UC San Diego

Recently, advances in sensing techniques, internet technologies, and wireless communications are increasingly facilitating and allowing practical deployment of large and dense sensor networks for structural health monitoring. Thus, it is vital to develop efficient techniques to process and analyze the massive amount of sensor data in order to extract essential information on the monitored structures. The efforts in this dissertation are mainly dedicated to this direction, including studies on structural damage identification and traffic pattern recognition. In these studies, traditional analysis tools for structural engineering (e.g., finite element (FE) based simulation) are utilized, and the potential of machine learning techniques is extensively explored. Using different strategies, three structural damage identification approaches (referred to as approach A, B, and C, respectively) are developed in this dissertation. Both approaches A and B adopt decentralized analysis frameworks, which define substructures in the monitored system according to the sensor spatial distribution. Within approach A, each substructure is represented by a dynamic model, and the system properties are periodically identified based on sensor measurements for monitoring purposes. Herein, approach A is applied to seismic downhole data to monitor changes in soil layer properties. As for approach B, a neural network is developed for each substructure to predict the dynamic response at a selected sensor location from measurements of neighboring sensors. Thus, the dynamic characteristics of the substructure are represented by the network, and changes in the statistical distribution of network prediction error are evaluated and utilized as a damage indicator. Three applications of approach B are presented, two based on experimental data and one of bridge pier damage identification based on simulation data. Approach C adopts a statistical pattern recognition paradigm. Within this framework, a range of damage patterns of interest is provided by numerical simulation, the Principal Components Analysis (PCA) technique is employed for feature extraction, and a neural network is developed for damage pattern identification. Herein, approach C is also applied to the bridge pier damage identification problem. Finally, the combination of neural networks and PCA is also employed to develop a strain-based vehicle classification approach, based on a unique strain-video dataset