Real-time hard-rock tunnel prediction model for rock mass classification using CatBoost integrated with Sequential Model-Based Optimization: Fid-Bau Portal

Real-time hard-rock tunnel prediction model for rock mass classification using CatBoost integrated with Sequential Model-Based Optimization

Bo, Yin / Liu, Quansheng / Huang, Xing / Pan, Yucong

Abstract

Graphical abstract Display Omitted

Highlights • A new hybrid SMBO-CatBoost algorithm that integrates CatBoost with the Sequential Model-Based Optimization is developed. • The present model predicts with more accuracy than the optimized XGBoost, SVM, RF, KNN, LR, and AdaBoost. • Unnecessary input variables are removed to improve the models’ field applicability. • A new method for more accurately calculating the rock mass boreability is proposed based on the big data.

Abstract In-time perception of changing geological conditions is crucial for safe and efficient TBM tunneling. Precisely detecting or predicting the rock mass qualities ahead of the tunnel face can forewarn the geological disasters (e.g., burst or squeezing behaviors of surrounding rock mass). A novel hybridization model based on CatBooost and Sequential Model-Based Optimization (SMBO) is proposed in this study. Firstly, a database incorporating 4464 samples acquired from the Songhua River Water Diversion Project is established using the capping method. Owing to SMBO’s different surrogate types (GP, RF, and GBRT) and performance validation, the comparisons of SMBO-CatBoost’s three types and other six hybridized models (SMBO-XGBoost, SMBO-AdaBoost, SMBO-RF, SMBO-SVM, SMBO-KNN, and SMBO-LR) are successively carried out. As a result, in terms of the optimization speed, performance, and sensitivity to poor geological conditions, SMBO(RF)-CatBoost is the most suitable model for rock mass class prediction; furthermore, it achieves the best performance $\bar{ACC}$ = 0.9207 and $\bar{F 1}$ = 0.9178 among the seven hybridized models. Next, the scientific feature selection methods (i.e., filter, embedded) are used to reduce the model’s complexity (i.e., feature dimensions) step by step to increase the model’s on-site practicality. The determined ten influential features still can keep the model’s $\bar{ACC}$ and $\bar{F 1}$ greater than 0.85, and only respectively declines 5.4% and 5.6% in contrast to the original performance. Subsequently, in order to explore the importance of the first-hand features and the second-hand features (i.e., composite features), a new method for more accurately calculating the rock mass boreability indices (regarded as the second-hand features) is proposed based on the big data at a relatively high sampling frequency of 1 Hz, this newly-proposed method could make these indices more of significance under the complex geological conditions. With the SHAP technique, the modified torque penetration index (TPI’) is more valuable than other second-hand and some first-hand features.