Machine learning–based tunnel face evaluation using observation log descriptions and text mining

Abstract

In the construction industry, the decline in skilled engineers has made the transfer of experience-based tunnel face evaluation techniques a critical challenge. This study focuses on textual data contained in tunnel face observation logs prepared by experienced engineers and investigates its effectiveness. Using data from the S Tunnel and the U Tunnel, an artificial neural network (ANN) model was developed in which the observation log data were transformed into embedding representations using BERT, and tunnel face evaluation for each item were predicted. As a reference for verifying the effectiveness of the textual information, a comparison was conducted with a conventional ANN model that used only measurement while drilling (MWD) data as input. The results showed that the model incorporating textual data improved both accuracy and F1 scores across all evaluation items, demonstrating that linguistic information reflecting the comprehensive judgment of experienced engineers—difficult to capture using numerical or image data alone is effective for ANN learning in tunnel face evaluation.