Tunnel face rock mass class identification based on multi-domain feature extraction and selection of TBM cutterhead vibration signals

Abstract

The rock mass class identification of the tunnel face is a key problem for TBM operating parameters optimization and subsequent tunnel support measures selection. This study presents a rock mass class identification method by monitoring and classifying TBM cutterhead vibration signals. Firstly, vibration signals were collected by a set of cutterhead vibration monitoring system installed on the TBM cutterhead during TBM tunnelling. The corresponding rock mass classification were conducted along the excavated tunnel field investigation. Secondly, time statistics and waveform, power spectrum frequency, nonlinear and time-frequency domain were extracted from the TBM cutterhead vibration signal. 18 features were selected by Boruta-SHAP feature selection method as important feature set. Based on the result analysis of different machine learning models, the XGBoost model was the best model used to identify the rock mass class. Its accuracy was up to 98.79 % on the test set. Finally, the feature sensitivity analysis by SHAP interpretation showed that energy entropy, Imf6e and kurtosis were the most sensitive features for different rock mass classes.