Lightweight deep learning model for identifying tunnel lining defects based on GPR data.

Existing lightweight artificial intelligence models for interpreting tunnel lining Ground Penetrating Radar (GPR) data often suffer from inadequate accuracy and robustness owing to noise interference caused by in-tunnel infrastructure. This study introduces an optimised method, named MTGPR, for the automatic detection of voids and cavities in tunnel linings based on GPR radargrams. The proposed model offers strengthened feature extraction and fusion owing to the use of CAPW-YOLO, a hybrid model aimed at enhancing accuracy in the presence of infrastructure interference. To address the shortage of high-quality training samples, an augmented dataset was generated by refining synthetic radargrams. Ablation experiments showcased that the proposed scheme attained an accuracy of 88.5% and a precision of 84.0%. In comparison to the baseline model, the proposed method exhibited a 46.9% increase in recognition speed and a 10.2% reduction in weight parameter quantity. Consequently, the proposed method advances identification accuracy whilst preserving lightweight and high-speed. • The optimised, lightweight deep-learning MTGPR method is presented. • The method detects subway tunnel lining voids and cavities using radargrams • Ablation experiments showed that this method yielded accuracy/precision: 88.5/84.0%. • MTGPR achieves enhanced performance in highly interfering noisy environments. [ABSTRACT FROM AUTHOR]