AI and knowledge driven computation of rock mass characteristic parameters across engineering projects.

With the continuous development of underground construction, represented by transportation tunnels, municipal utility tunnels and hydraulic tunnels, accurately perceiving the rock mass quality ahead of the tunnel boring machine (TBM) face has become critical for ensuring construction quality and improving construction efficiency. Over the past 5 years, predicting rock mass quality using machine learning and other artificial intelligence (AI) algorithms has gradually become a research hotspot. However, the raw TBM tunnelling data are massive and noisy, and key challenges remain, including how to reasonably select effective input features and how to cope with the limited data available from newly built projects. Although the academic community has proposed indices such as the TPI and FPI to quantify rock mass boreability and to incorporate them as input features, their application in practical engineering remains challenging. Owing to the complexity of tunnelling operating conditions and the variability in data quality, existing approaches exhibit notable limitations. In particular, physics-based "definition-based" methods typically require the processing of large volumes of high-frequency tunnelling data, and their computed results are highly sensitive to data fluctuations, resulting in poor stability. In contrast, "fitting-based" methods constructed on regression relationships, while showing reasonable effectiveness under single-project conditions, rely heavily on prior assumptions (e.g., penetration-based hypotheses of rock-breaking force) and are strongly influenced by fitting performance. As a result, these methods struggle to adapt to varying geological and engineering conditions, leading to limited generalizability and robustness in multi-project scenarios, and thus remain insufficient to support large-scale engineering applications. To address these issues, this study proposes an AI- and knowledge-driven method for computing rock mass characteristic parameters. Firstly, a rock-breaking data filtering method is developed based on the effective conversion of disc cutter energy. Engineering data from three TBM projects with different diameters and geological conditions are utilized, including the Yinchuo Water Diversion Project (YC), the Yinsong Water Diversion Project (YS), and the Huanbei Water Diversion Project (HB). Based on these datasets, this paper proposes an innovative methodology for identifying high-efficiency rock-breaking stages. Then, knowledge-driven rock mass characteristic parameters (a, b, and the torque penetration index, TPI) are computed as input features. Lastly, an AI-driven strategy is then adopted to determine the prediction model, whereby CatBoost is selected to develop the rock mass class predictor and is tested across different projects. The results indicate that the proposed characteristic parameters are strongly correlated with rock mass quality. The prediction accuracies in the YC, YS and HB projects reach 85.60%, 86.48% and 88.89%, respectively, outperforming conventional methods overall. The proposed method provides new technical support for cross-project data utilization and real-time prediction of rock mass quality, and has important implications for construction safety and efficiency in newly built tunnel projects.