Ensuring safety in tunnel construction necessitates a rapid and objective evaluation of exposed tunnel faces for proactive rock mass risk assessment. The diversity of rock types and discontinuities contributes to structural complexity and local variability, introducing subjectivity and uncertainty in traditional evaluations based on the experience of field engineers. In this study, a convolutional neural network (CNN) is employed for regression to enhance the objectivity and consistency of rock mass classification via rock mass rating (RMR). The model interprets the entire tunnel face image holistically by identifying and learning from complex patterns, rather than detecting specific geological features like fractured or discontinuity zones. The network model, comprising feature extraction and regression blocks, utilizes the EfficientNet family for computational efficiency and accuracy. Data augmentation and ensemble learning address prediction variability due to image quality, optimizing the model for accurate RMR predictions and providing error ranges. The RMR predictions using excavation site images closely follow the field-evaluated evolution. A server communication-based mobile application is developed for real-time RMR evaluation, enhancing its practical field applicability. In geology, tunnel, and geotechnical engineering, where decisions rely on extensive experience, our approach demonstrates that deep learning can enhance decision-making by analyzing large accumulated datasets to predict optimal outcomes.
Read full abstract