Data visualization strategy for predicting rock mass classification with TBM key rock-fragmentation parameters

Abstract

This study proposes a data visualization strategy for predicting rock mass classification, which only adopts 5 tunnelling parameters related to rock fragmentation as input. For implementation, first, the database of TBM tunnelling parameters was established after data cleaning, in which 4,172 boring cycles from the Yinchao project were acquired. Subsequently, five tunnelling parameters of each cycle were plotted automatically into 9 single images in a defined manner. AIF (Assembled Image Fusion) image was composed of 9 single images. Finally, a convolutional neural network model (ResNet) was employed to recognize AIF fused image for rock mass classification. The results show that: (1) The AIF fused image outperforms single images, with an F₁ score of 0.83 compared to 0.77 for the best single image (image of cutterhead torque T vs cycle duration time t). (2) In comparison, LightGBM and RF models, using mean values of tunneling parameters, achieved F₁ scores of 0.73 and 0.71, respectively, which are inferior to the AIF fused image approach. (3) The proposed visualization strategy suggests that image plotting should consider the statistical distribution of the dataset for axis scaling, while color choice has minimal impact on recognition. This strategy offers an effective framework for prediction of rock mass classification.