To address the issues of weak generalization capability and difficult model iteration in rock mass class identification models for tunnel boring machine (TBM) tunneling, this study proposes an incremental data acquisition method for identifying tunnel rock mass classes. Based on the XE-III section of the second phase of the northern Xinjiang water supply project (YEGS project), a vibration monitoring system was deployed on the open-type TBM to collect vibration signals from the cutterhead and gripper, as well as TBM tunneling parameters. A database was constructed by integrating these data with the tunnel rock mass classes. Rock mass class identification models for the tunnel face and the gripper position were established using the XGBoost algorithm. The prediction results of the two models were matched according to the tunneling step, and the consistent results were automatically labeled as the rock mass classes for subsequent chainage, thereby achieving incremental data acquisition and iterative model updating. By introducing a manual review mechanism and relabeling the identification results of minority-class samples, the generalization capability and iterative performance of the model were further improved. This incremental data acquisition method can drive continuous model updating and enhance identification accuracy, thereby providing strong support for TBM intelligent tunneling.

Image analysis of rock chips is an important approach for TBM intelligent tunneling, and it can be used for real-time evaluation of rock-breaking efficiency and optimization of tunneling parameters. In order to investigate the relationships between TBM tunneling parameters and rock chip characteristics, a multi-step field tunneling test was carried out in an intact granite formation at a TBM construction section of Qingdao Metro Line 6. A rock chip image acquisition system was installed on the TBM to real-time collect rock chip images. Then, the relationship between the tunneling parameters and the size and shape characteristics of rock chips obtained through image processing was analyzed. The results showed that the median particle size d₅₀, maximum particle size d_max and roughness index CI all present a positive correlation with thrust, torque and penetration, while a good negative correlation with FPI and SE. When the thrust exceeds a critical value for rock-breaking, the increase in thrust will significantly increase d₅₀ and CI, while d_max increases significantly only when the thrust exceeds a larger value. Besides, as the thrust increases, the shape of the large rock chips gradually becomes flattened and the aspect ratio increases. However, after the thrust reaches a certain level, the aspect ratio will remain stable. The results can provide a basis for real-time optimization of TBM tunneling parameters based on rock chip information.