@article{Wang2026, 
author = {Linlin Wang and Xuemin Chen and Gangbing Song},
title = {A comparative study of machine learning methods for detecting debonding in carbon-fiber-plate reinforced concrete beam using percussion induced sound},
year = {2026},
journal = {Lifeline Emergency and Safety},
volume = {1},
number = {1},
pages = {9660013},
keywords = {Carbon-fiber-plate reinforced concrete beam, structural health inspection and monitoring, percussion-based approach, debonding detection, shallow supervised learning, deep supervised learning},
url = {https://www.sciopen.com/article/10.26599/LLES.2025.9660013},
doi = {10.26599/LLES.2025.9660013},
abstract = {Carbon fiber reinforced polymer (CFRP) plates are widely used to strengthen deteriorated concrete beams by bonding them to cracked surfaces. However, poor construction quality, aging, and deterioration can result in debonding at the carbon-fiber-plate repaired concrete beam (CFP-RCB) interface, compromising structural integrity and increasing failure risks. To address this, the study introduces a nondestructive, cost-effective structural health inspection and monitoring (SHIM) method for detecting debonding in CFP-RCB. This method utilizes frequency analysis of percussion-induced sound from tapping the CFP-RCB surface, combined with machine learning (ML) models, offering significant cost savings compared to traditional sensor-based systems and scalability for large-scale applications. Various ML models, including K-Nearest Neighbors (KNN), Decision Tree (DT), Naive Bayes (NB), Logistic Regression (LGR), Support Vector Machine (SVM), and Recurrent Neural Network (RNN), are evaluated. Results demonstrate that the deep learning model (RNN) outperforms shallow learning models (KNN, DT, NB, LGR, SVM) by approximately 10%, using Mel-frequency cepstral coefficients (MFCC) for feature extraction, reducing human error compared to manual Power Spectral Density (PSD) methods. Among shallow learning models, KNN, DT, and SVM exhibit higher accuracy than NB and LGR, though DT shows instability. The accuracy achieved on the unseen test dataset is generally lower than the accuracy achieved on the training and validation datasets across all models. This ML-driven approach enables rapid and reliable damage detection, enhancing preventive maintenance to ensure safer and longer-lasting infrastructure.}
}