Quantitative Rock Tunnel Face Risk Assessment via Multi-Source Data and Stacked Deep Learning Models

Quantitative tunnel face risk assessment is the characteristic challenge of rock tunnel excavation projects. This study establishes a multi-source database and proposes a stacked deep learning method for the quantitative tunnel face risk assessment. Contact and non-contact methods were used to collect multisource data of the tunnel (e.g., face images, site geological information, and rock mass properties). Thirteen multi-source variables describing the rock tunnel faces are considered inputs, and the rock tunnel face deformations computed by the numerical simulation are the target outputs. A staking model architecture is proposed to combine a range of well-performing models, which can make accurate predictions of tunnel face rock mass quality. The Tree-structured Parzen Estimator (TPE) algorithm is applied to determine the optimized model hyper-parameters automatically. The experimental results of a tunnel project in China suggest that the proposed stacking deep learning model performs well at assessing rock tunnel face stability.