Structural damage detection using dictionary learning

Update Item Information
Title Structural damage detection using dictionary learning
Publication Type thesis
School or College College of Engineering
Department Mechanical Engineering
Author Melville, Joseph
Date 2017
Description Nondestructive evaluation (NDE) is a means of assessing the reliability and integrity of a structural component and provides such information as the presence, location, extent, and type of damage in the component. Structural health monitoring (SHM) is a subfield of NDE, and focuses on a continuous monitoring of a structure while in use. SHM has been applied to structures such as bridges, buildings, pipelines, and airplanes with the goal of detecting the presence of damage as a means of determining whether a structure is in need of maintenance. SHM can be posed as a modeling problem, where an accurate model allows for a more reliable prediction of structural behavior. More reliable predictions make it easier to determine if something is out of the ordinary with the structure. Structural models can be designed using analytical or empirical approaches. Most SHM applications use purely analytical models based on finite element analysis and fundamental wave propagation equations to construct behavioral predictions. Purely empirical models exist, but are less common. These often utilize pattern recognition algorithms to recognize features that indicate damage. This thesis uses a method related to the k-means algorithm known as dictionary learning to train a wave propagation model from full wavefield data. These data are gathered from thin metal plates that exhibit complex wavefields dominated by multipath interference. We evaluate our model for its ability to detect damage in structures on which the model was not trained. These structures are similar to the training structure, but variable in material type and thickness. This evaluation will demonstrate how well learned dictionaries can both detect damage in a complex wavefield with multipath interference, and how well the learned model generalizes to structures with slight variations in properties. The damage detection and generalization results achieved using this empirical model are compared to similar results using both an analytical model and a support vector machine model.
Type Text
Publisher University of Utah
Subject Electrical engineering; Mechanical engineering; Computer science
Dissertation Name Master of Science
Language eng
Rights Management © Joseph Melville
Format application/pdf
Format Medium application/pdf
ARK ark:/87278/s6rn7mg8
Setname ir_etd
ID 1400330
Reference URL https://collections.lib.utah.edu/ark:/87278/s6rn7mg8
Back to Search Results