ABSTRACT
Cavities behind tunnel lining lead to detrimental effects on structural stability. Cavity detection is normally performed using non-destructive devices such as ground penetrating radar (GPR), though this is a time-consuming process requiring extensive experience and sophisticated filtering methods. In this regard, we propose a deep learning (DL) based cavity information (i.e., location, depth, size) prediction model. The model extensively employs A-scan data to extract important cavity-related features. For that purpose, and to take advantage of available time series GPR data, the 1D deep convolutional neural network (1D DCNN) framework along with a conditional regression model and a Long Short Term Memory (LSTM) sub-module were adopted. Field experiments on tunnel lining with and without cavities were performed, and the results were compared with the proposed DL model. It was found that cavity location could be predicted to a maximum accuracy of 94-99% according to curing time and antenna frequency.
