Damage Detection of Fiber Reinforced Polymer Plate Repaired Steel Structure Using Percussion and Machine Learning

Aug 4, 2026

Upload by Jiven

Recent structural failures highlight the need for reliable infrastructure rehabilitation and verification. Fiber reinforced polymer (FRP), particularly CFRP, is an effective and cost-efficient approach for strengthening or repairing steel members, yet damage detection is essential to ensure repaired structural integrity and performance. This thesis develops a percussion-based damage detection method combined with machine learning. A CFRP-bonded steel beam with known bonding defects is tapped at different locations; percussion signals are recorded and converted into discriminative features using MFCC. Supervised learning with SVM and RNN predicts healthy states with high accuracy, while unsupervised k-means and GMM clustering further distinguish defect conditions. The proposed approach avoids installing sensors or deploying data acquisition systems.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages63

Reading Time2 h 39 min

AccessFree to download

I. Introduction

1.1 Fiber reinforced polymer in civil engineering

1.2 CFRP repaired steel structure

1.3 Current defect detection method for CFRP repaired structure

1.4 Percussion based structure health monitoring

1.5 Machine learning in sound recognition

II. Experiment setup and data collection

2.1 Experiment setup

2.2 Data collection

III. Defect detection of CFRP repaired steel beam

3.1 Damage detection using SVM

3.1.1 Introduction to SVM

3.1.2 Feature transformation and SVM

3.1.3 SVM experiment result

DAMAGE DETECTION OF FIBER REINFORCED POLYMER PLATE REPAIRED STEEL STRUCTURE USING PERCUSSION AND MACHINE LEARNING by Yong Xu A thesis submitted to the Department of Mechanical Engineering, Cullen College of Engineering in partial fulfillment of the requirements for the degree of MASTER OF SCIENCES in Mechanical Engineering Chair of Committee: Gangbing Song Committee Member: Zheng Chen Committee Member: Xuemin Chen University of Houston May 2022 ACKNOWLEDGMENTS I would like to express my great appreciation to the Department of Mechanical Engineering and College of Engineering, for helping and supporting my journey at University of Houston. When I enrolled in the Machine Learning course offered by Professor Gangbing Song in the semester ofFall 2021, I thought I could gain some general concept. The results are astonishing, Professor Song has taught me the latest development of machine learning in multiple fronts, furthermore he has taught me how to use technology to tackle real-life engineering problems. This is eye opening forme, and the results that I achieved in the course project encouraged me to pursuit my master’s degree with a thesis on the subject. I would like to thank Professor Song for teaching me the technology with the best material summarized and highly interactive course structure. During my thesis work, Professor Song gave me close guidance and many suggestions when I was facing the challenges to improve machine learning results. I will forever be thankful to my advisor, Professor Song. I would like to offer my special thanks to my defense committee members Professor Zheng Chen and Professor Xuemin Chen for taking their time to read my thesis and serve on my defense committee. I would also like to thank Mr. JiAn Chen, a Ph.D. student in Professor Song’slab, for his generous help during my thesis work. Last, but not least, my extensive and warm gratitude go to my family for their continuous and unmatched help, support, and love. ABSTRACT The recent collapse of the Fern Hollow Bridge in Pittsburgh, Pennsylvania , brings attention to the structurally deficient infrastructure. The fiber reinforced polymer (FRP) has been proven to be a cost-effective, efficient, and reliable method for structure rehabilitation or reinforcement. Damage detection is an important measure to ensure the integrity and performance of such repairs. A novel method of using percussion and machine learning to detect the damage ofFRP plate repaired steel structure was developed and discussed in this work. A steel beam with bonded carbon fiber reinforced polymer (CFRP) and known bonding defects was used as a test specimen. Then, different locations with different bonding conditions on the beam were tapped to generate the percussion sound, which was recorded by an iPhone. The mel-frequency cepstral coefficient (MFCC) algorithm was employed to extract features from percussion sound. The support vector machine (SVM) and recurrent neural network (RNN) method were implemented to learn the training samples and achieved high accuracy when predicting the healthy status of new test samples. The SVM used a new way of feature transformation, which is based on mean and standard deviation of MFCC. The high accuracy of 98.5% demonstrated the new feature transformation method is effective for SVM in percussion application. Then, the unsupervised clustering algorithms, k-means and Gaussian mixture model (GMM), were implemented on the sample data. The accuracy ofk-means algorithm varies in a wide range from 51.5% to 70.4%, while the GMM clustering uses transformed MFCC and manually selected features, achieves an accuracy of 93.8%, The results of this work have demonstrated that the novel method of percussion and machine learning is reliable for damage detection of the FRP repaired steel structure. Compared with the conventional method, the proposed method does not require installation of sensors or implementing data a

FAQs

What repaired structure and damage type are used to validate the method?

A steel beam bonded with carbon fiber reinforced polymer (CFRP) is used, with known bonding defects. Different bonding locations and conditions are created for testing.

How are percussion signals processed to support machine learning?

Tapping generates percussion sounds recorded by an iPhone. The MFCC algorithm extracts features from the recorded signals for subsequent learning.

Which machine learning approaches are used and what performance is reported?

Support vector machine (SVM) and recurrent neural network (RNN) are used for prediction of healthy status, with high accuracy reported for new test samples. Unsupervised clustering using k-means and Gaussian mixture model (GMM) is also applied, with GMM achieving higher clustering accuracy than k-means.

Aug 4, 2026

Upload by Jiven