Incorporating smart computer vision and in-drilling information into rock quality evaluation via incomplete data-driven Bayesian networks

Tunnelling is a challenging task due to a lack of full understanding of the surrounding rock quality. This study proposes a solution driven by a refined computer vision (CV) method, complemented by rock mass drilling tests and Bayesian networks, to address this issue through a multi-source heterogeneous data approach. Initially, improvements are made to the popular Swin Transformer to improve the recognition and segmentation of intricate rock features. Notably, refined smart CV, owing to its U-shaped architecture and smart window self-attention computation, exhibits segmentation performance superior to that of conventional CV methods such as Swin Transformer, Deeplab V3+, and UNet. Building upon the segmentation outcomes of the refined CV, a parameter set comprising apparent rock parameters is established. Then, two datasets encompassing rock internal drilling parameters and mechanics, as well as design parameters, are curated. The combination of the aforementioned parameter sets is referred to as the rock quality comprehensive evaluation dataset. However, analysis reveals data incompleteness issues within these datasets. To mitigate this problem, a novel tree-augmented Bayesian network is designed, and a prediction accuracy of 91% is realized, surpassing popular decision trees, ensemble learning, and deep learning methods. Furthermore, evaluation services are provided in mountain and submarine tunnels, suggesting that drilling parameters significantly enhance the evaluation performance. Moreover, employing two sensitivity analysis metrics underscores the prominent influence of rotating pressure and drilling speed parameters. This study endeavor presents diverse solutions for achieving precise and expeditious predictions of rock quality through various parameter sets, tailored to cater to diverse requirements of tunnels.