In general, a high-fidelity finite element (FE) model consists of the massive number of elements and each element has various physical parameters such as mass density and Young’s modulus. Therefore, a high-fidelity FE model is in an extremely high-dimensional parameter space, which can pose extra difficulties in understanding the relationships between model outputs and the status of a model. The objective of this research is to automatically find physically-meaningful and efficient groups of model parameters in an effort to reduce the dimension of parameter space into a more tangible format. We developed a machine-learning-based parameterization method, which can determine efficient groups of model parameters based on the sensitivities of model parameters. The developed method has applied to sensitivity-based and Bayesian model updating practices to make these inverse problems more affordable and tractable by efficiently reducing the parameter space of a sophisticated full-scale model.
Civil structures are influenced by various operational and environmental conditions, such as traffic, wind, humidity, and most importantly, temperature. It is fundamentally important to distinguish temperature-related changes in structural responses from those related to structural damages since changes in model outputs caused by environmental effects can mask the effect of structural damage. This research work aims to investigate the effects of temperature distribution on natural frequencies with a full-scale finite element model. Temperature profiles of the model are hypothetically simulated and used to generate a large number of simulation datasets for training data-driven models. Developed data-driven models lead to an in-depth understanding of the effect of temperature distributions on modal properties of a structure.
The objective of this research is to develop an innovative road surface monitoring system that can provide continuous city-wide information for the pavement distress management system. This proposed road surface monitoring system relies on various sensor data collected from connected vehicles and state-of-the-art data analytics at a backend server. The suggested monitoring system includes an embedded data logging algorithm to maximize the efficiency of data collection. It also utilizes a machine learning algorithm to classify collected sensor data into different types of pavement distress. More importantly, it leverages a data mining technique to integrate sensor data from the various connected vehicle based on their proximity and directionality to maximize the accuracy and reliability of outcomes of the suggested monitoring system.
A massive amount of fleet trajectory data has become available to cities and communities due to the development of sensors and wireless technology. It gives great opportunities for cities and communities to utilize these newly-available data formats for their decision-making and planning processes. Online trajectory clustering algorithm, which necessitates fewer computation resources to process big trajectory data, was developed to pinpoint the hotspots of hard accelerations and braking. A map-matching algorithm, which can fuse another data source with GPS data to increase the computational efficiency and the accuracy of a result, was developed to clean up the measurement noise and to analyzes massive trajectory data on the basis of map data.