Hongwei Huang

Machine Learning for Safety Risk Assesment on Rock Tunnel Driving

The excavation process of conventional tunneling method encounters complex uncertainties in geological conditions, posing an exceptional challenge to establishing a universally applicable rock mass classification system. Consequently, methodologies urgently necessitate comprehensive, objective, and accurate acquisition of geological information, despite the inherent abundance of information in excavation fronts. The unresolved persistence of relying on subjective expertise of on-site engineers for decisions regarding rock mass classification and stability assessment is noteworthy. There is an urgent need to effectively gather multisource datasets during excavation and objectively evaluate the rock mass quality and safety stability of tunnel excavation. In this context, this lecture proposes a set of excavation front rock mass intelligent classification based on computer vision and a stability assessment mechanism based on the upper bound limit analysis method, aiming to achieve intelligent management of safety risks associated with complex excavation fronts. 

Specifically, a framework has been established for extracting key information of rock mass working faces based on convolutional neural networks by collecting visual data from hundreds of working faces from over twenty tunnels. This framework encompasses the rock structure classification model based on Inception-ResNet-V2, the quantitative segmentation model for weak interlayers and seepage based on improved DeepLab V3+, and the statistical model for joint fractures based on the self-developed FraSegNet algorithm. Subsequently, 3 dimensional point cloud models were reconstructed using Structure from Motion (SfM) technology. Correspondingly, algorithms were developed to automate the extraction of 3D discontinuous faces and trace lines, thus establishing a multi-source heterogeneous rock mass working face dataset based on computer vision extraction algorithms. By correlating the acquired information with drilling data, mechanical and geometric data of the surrounding rock, the main types of indicators were determined: (1) primary parameters of discontinuous rock faces; (2) basic parameters of surrounding rock; and (3) on-site parameters. 

Based on this, research efforts were directed towards establishing an intelligent rock mass classification system based on computer vision and data-driven approaches. Machine learning models were utilized to predict rock mass grades. To enhance model performance, a Tree-structured Parzen Estimator based Bayesian optimization method was employed to select optimal model hyperparameters, followed by a robustness analysis of the model. The study indicates that the proposed hybrid ensemble learning models perform well in rock mass quality assessment. Subsequently, considering the structural characteristics of rock masses, the impact of groundwater, and the influence of artificial factors such as construction excavation schemes and parameters, a theoretical model for stability analysis of working faces was constructed within the framework of limit analysis. By integrating multiple regression algorithms and response surface theory, a rapid calculation formula for the safety factor of tunnel working faces was developed. By incorporating multifaceted information, a predictive model for on-site construction schemes was constructed. 

The superiority of the proposed methodology in practical engineering rock mass classification was validated by comparing the application of existing classification standards in the Qingdao Jiaozhou Bay Subsea under construction tunnel and the Yunnan Mountain tunnel. Combining considerations of complex geological environments and construction parameters enabled effective evaluation of stability and on-site validation, thereby providing an important foundation for risk management of subsequent excavation fronts.