Feature fusion method for rock mass classification prediction and interpretable analysis based on TBM operating and cutter wear data

The precise identification of rock mass classification is essential for assessing stability and optimizing tunnel support design. Despite being primarily conducted using tunnel boring machine (TBM) operating data, this paper focuses on developing a bimodal feature fusion framework for predicting the rock mass class, utilizing TBM operating and cutter wear data. This framework includes field data collection, data pre-processing and feature engineering, feature fusion, and a machine learning perception module. After data collection and processing, operating features were obtained through linear fitting of thrust-penetration and torque-penetration data, and wear features were obtained by calculating wear rates from radial wear values of 34-disc cutters. Then, a spatiotemporal feature ordering and KNN interpolation (STFO-KNN) method is proposed to integrate operating and cutter wear features. The following step establishes intelligent models using the random forest and convolutional neural networks to evaluate performance according to 5537 TBM tunnelling cycles from the 9.77 km Chaor to Xiliao River tunnel. Results indicate that surrounding rock classification performance improves significantly using the coupling modal fusion method. Binary classification achieved an accuracy of 0.97, and four classifications for support design attained an accuracy of 0.94. Finally, a feature ablation test, physical explanation, and modal contribution analysis were performed, and the results aligned with human empirical knowledge. This study not only contributes to the safe and high-efficiency TBM construction but also provides insights into feature contributions for optimizing monitoring settings.