Abstract: Image-based prediction of rock mass ahead of the tunnel face is vulnerable to environmental disturbances, whereas the longitudinal displacement profile (LDP) reflects the intrinsic strength attributes of the rock mass. The authors propose a lightweight prediction approach using LDP as the model input. Numerical simulations are conducted to analyze and summarize LDP evolution patterns for three rock mass qualities (geological strength indices (GSIs) of 15, 30, and 50) under nine stratigraphic combinations. Based on these results, a classification mode for rock mass ahead of the tunnel face is developed using a one-dimensional convolutional neural network (1D-CNN) enhanced with spatiotemporal attention, with LDP sequences as inputs. The main findings are as follows: (1) For homogeneous strata, deformation generally increases and subsequently decreases as excavation advances; at a GSI of 15, the overall deformation is larger. For strata containing an interface, deformation increases and subsequently decreases before the interface, with postinterface deformation dependent on lithological parameters. (2) The proposed 1D-CNN with spatiotemporal attention achieves a test accuracy of 0.88 for strata classification from LDP sequences. (3) Comparative experiments with variant models show that the 1D-CNN with spatiotemporal attention attains the highest accuracy with a relatively low parameter count. By combining LDP evolution characteristics with deep learning, this lightweight method offers a new approach for predicting rock mass class ahead of the tunnel face.

Key words: tunnel, rock mass ahead of tunnel face, numerical simulation, longitudinal displacement profile, prediction model, spatiotemporal attention mechanism, convolutional neural network