Automated Classification of Tunnel Face Apparent Structures
Jiayao Chen, Hongwei Huang, Mingliang Zhou · 2025
Automated rock mass classification remains challenging in tunnel construction due to geological complexity and limited field applicability of conventional methods. This study develops a deep learning framework for automated identification of tunnel face apparent structures using image datasets from the Yunnan Mengzi–Pingbian Expressway. Focusing on five BQ-classified structures—granular (GS), mosaic (MS), block (BS), layered (LS), and fragmentation (FS)—the Inception-ResNet-v2 architecture integrates inception modules and residual connections to enhance feature extraction under field constraints. Comparative analysis with ResNet-50, ResNet-101, and Inception-v4 demonstrates superior performance (92.7% F1-score) in classifying low-quality images with illumination variations (50–800 lux) and occlusions. The model reduces false positives by 18%–24% in layered/fragmented structures through multi-scale feature fusion, validated by cross-tunnel testing ( R² = 0.89). Optimized batch normalization and adaptive pooling layers improve robustness against sample scarcity, achieving 31% higher recognition accuracy than conventional neural networks (CNNs). Results establish a machine vision-driven workflow for real-time structural mapping, enabling precise deformation prediction while addressing limitations of laboratory-based and simulation approaches. This framework integrates machine vision with geotechnical engineering, providing a scalable solution for automated stability assessment in complex tunneling environments.