Tunnel lining defects identification using TPE-CatBoost algorithm with GPR data: A model test study

Abstract

The quality of tunnel lining is critical for both construction integrity and safety. Ground Penetrating Radar is widely employed for tunnel lining inspections, but manual analysis is time-consuming and inefficient. This paper presents TPE-CatBoost, a machine learning method designed to identify typical defects in tunnel linings using GPR data. The CatBoost algorithm is used to classify defects, while the Tree-structured Parzen Estimator is employed to optimize hyperparameters. A full-scale model test with actual engineering materials was conducted, resulting in a dataset of 249,500 GPR A-scans, effectively addressing the challenge of limited data sources with known defect information. The SMOTE algorithm was utilized to mitigate the issue of imbalanced samples within the dataset. A novel feature extraction method, incorporating GPR waveform statistics and parameters, was developed and proved effective in defect identification. TPE-CatBoost achieved an accuracy of 0.92 in just 125 s, surpassing the performance of Naive Bayes, Support Vector Machines, Random Forest, and XGBoost. The method was successfully applied in the Shanggang Tunnel using an intelligent detection vehicle, significantly enhancing GPR efficiency. The results were validated against traditional data processing methods.