The data-driven approach of machine learning enables the intelligent construction of TBM (tunnel boring machines), which is crucial for optimizing the tunneling process, improving the safety of tunneling and reducing labor costs. In order to solve the problems of excessive noise, redundant parameters and difficult effective feature extraction in TBM operation data, a data-driven machine learning method was used to mine the complex machine-soil interaction contained in the data and realize the classification and prediction of TBM surrounding rock mass. First, for the large amount of operational data generated during TBM tunneling, the KDE(kernel density estimation) method was used to extract features from typical tunneling parameter curves, and the maximum probability of the key operating parameters during stable tunneling stage of TBM is obtained. Then, based on the actual TBM operation data, an integrated learning algorithm for surrounding rock classification stacking was proposed. The algorithm is further optimized through k-fold cross-validation, and the complex relationships in the data are mined by using the two-layer learning framework of base classifier and meta-classifier. Finally, a data set of 5 868 TBM segments was used to verify the effectiveness of the proposed algorithm. The results show that the average F₁ of the four-classification problem is 0.705, and the average F₁ of the two-classification problem is 0.797, which are better than the four selected base classifiers.