Yin Chao Wu

Graduation Semester and Year

2023

Language

English

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Civil Engineering

Department

Civil Engineering

First Advisor

Suyun Ham

Abstract

The dissertation explores into structural damage identification and prediction through the application of machine learning. Firstly, it enhances impact echo delamination mapping by incorporating two signal image parameters—signal duration and starting time after zero-crossing. The research also explores the efficiency of machine learning concerning image resolution and the number of images. This approach refines the evaluation of bridge damage in field tests. Two crucial signal parameters are investigated: signal duration and signal starting time after the zero-crossing point. These parameters exert a direct impact on the frequency domain, a crucial metric in impact echo analysis. This metric discerns signals reflected from areas with delamination and those from non-delaminated areas by assessing the accumulated low-frequency energy within the range of 1kHz to 5kHz. Considering the complexity of signals collected during field tests, artificial neural network (ANN) and convolution neural network (CNN) are employed to identify delamination signals using these two parameters. The study reveals that a signal duration of 1 millisecond with a starting time at 0.1 millisecond yields the highest accuracy in damage identification. A comparative analysis between delamination maps before and after applying deep learning (DL) suggested signal parameters demonstrates the updated results' enhanced accuracy in distinguishing delamination areas correctly. Secondly, the study differentiates between homogeneous and inhomogeneous mediums with random aggregate sizes and distributions, utilizing wave scattering models and wave response variations. The internal crack geometry is identified through wave response variation with machine learning, and the finite element model design for random aggregate is detailed. Analyzing cracks at different depths, the wave response variation (WRV) study observes lower impulse frequencies for shallower cracks, computed from forward and incident waves, and higher frequencies for deeper cracks. Inhomogeneous medium (IHM) complexity increases with random aggregate size and distribution, where larger aggregates significantly impact the WRV pattern, causing more energy attenuation in the forward wave than smaller aggregates. Machine learning (ML) techniques accurately predict cracks and elucidate the black box process, emphasizing the impact of aggregate information on WRV. The study outlines the framework of internal damage in IHM using ML technology. Lastly, the dissertation employs bridge weight-in-motion signals to predict varying levels of structural damage. A finite element model using the kinetic contact method is developed and verified with bridge-vehicle motion theory. The study focusses on the development of moving load FE model. It introduces a comprehensive assessment of 1) a unique finite element (FE) simulation approach, which leverages the kinematic contact enforcement (KCE) method, verified with the vehicle-bridge interaction (VBI) theory, 2) laboratory tests and field test were conducted and performed to calibrate and verify the FE model, and 3) ML techniques to identify and automatically predict structural damages from the structural response. The KCE method is a new approach to simulating vehicle motion in a BWIM model, which is used to carry out actual structural responses to motion. By providing contact conditions between elements, including contact type, material properties, and element speed, the KCE method enables realistic simulation of both vehicle motion and structural response. The responses generated by the FE simulation are further analyzed using a feature selection method that ranks the importance of various ML models. Specifically, the prediction model includes a decision tree, a support vector machine (SVM), backpropagation (BP), and XGBoost. The results show that XGBoost with its assembly decision tree provides the most reliable outcomes. Three distinct studies leverage machine learning algorithms for tasks such as signal identification, damage identification, and the prediction of structural damages.

Keywords

Damage detection, NDT, Machine leanring, Signal processing, Impact-echo

Disciplines

Civil and Environmental Engineering | Civil Engineering | Engineering

Comments

Degree granted by The University of Texas at Arlington

Download

Available for download on Sunday, February 01, 2026

Share

COinS