AI and knowledge driven computation of rock mass characteristic parameters across engineering projects Authors: Zewang Zheng, Yunpei Zhang, Quan Xu, Kezhong Wang, Changchun Ren, Hao Cai Published: 2026 ABSTRACT 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. Introduction Since the beginning of China’s 14th Five-Year Plan, the development and utilization of underground space have continued to accelerate, and the scale of tunnel construction has expanded steadily. By the end of 2024, 424 new railway tunnels had been opened to traffic, with a total length of 738 km; 2521 railway tunnels were under construction and 5634 were in planning, with a total length of approximately 13,742 km1. TBMs (tunnel boring machines), owing to their advantages of safety, high efficiency and low labor intensity, have been widely adopted in tunnel construction2.In recent years, benefiting from the rapid development of computer and artificial intelligence (AI) technologies and the accumulation of TBM operational data3, research on smart TBM technologies has gained increasing momentum. Among these, machine learning (ML) learns data-driven mappings between inputs and targets to enable recognition and prediction, and has therefore been widely adopted. In addition, the digital twin (DT) paradigm has emerged as another major research direction. DT constructs an “physical entity-virtual counterpart” mapping by integrating multi-source information, thereby enabling real-time analysis and decision support. For example, Latif et al.4 proposed a hybrid DT approach for TBM performance visualization, improving information representation and management efficiency during construction. Sharafat et al.5 further explored DT-driven optimization of underground construction schemes to support decision-making on design parameters. DT is oriented toward process management and decision support and thus exhibits strong development potential. However, continuous updating of twin information in practical tunneling projects typically relies on high-density sensing, reliable communication, and substantial computing and system-integration capabilities, which increases the difficulty of field deployment and long-term maintenance. By contrast, ML emphasizes learning “input–output” relationships from historical or streaming data. Common algorithms include random forest (RF), support vector machine (SVM), gradient boosting/boosting methods, and deep learning models for time-series data(e.g., LSTM/GRU). Owing to its relatively lightweight implementation and ease of rapid deployment, ML is selected in this study as the primary approach to develop and validate practical models for key tasks under modest on-site requirements. Based on these methodological paradigms, current TBM intelligence research mainly focuses on two categories of tasks: (1) prediction/optimization of tunneling parameters and (2) identification of surrounding rock quality. Parameter optimization often involves multi-objective trade-offs among efficiency, energy consumption, cutter wear, vibration, and safety6,7. Similar attention has also been paid in shield tunnelling, where multivariable parameter prediction and combination models have been investigated to support performance assessment and construction control8. Meanwhile, studies on excavation-induced deformation analysis, nonlinear response prediction of adjacent pipelines, and jacking-force analysis in pipe-jacking construction indicate that underground construction performance is influenced not only by operational parameters, but also by excavation effects and tunnel–structure interaction9,10. Together, these studies provide a broader theoretical and application background for both parameter-driven performance assessment and tunnel–structure interaction analysis in underground engineering. However, the objective functions, constraints, and definitions of “optimality” vary substantially across projects, resulting in relatively inconsistent research boundaries. In contrast, surrounding rock quality identification/classification aims to characterize objective geological conditions; the task definition is clearer and more generalizable, and it also constitutes a key input prioritized by on-site operations. Therefore, this work investigates ML-based prediction of surrounding rock quality. Although ML is advantageous in handling complex nonlinear data, its prediction quality and accuracy strongly depend on the rational selection of input features11. TBM equipment typically acquires on the order of hundreds of data channels. In existing research, this is often handled through feature engineering—that is, extracting a limited number of parameters that exhibit strong mapping relationships with the output variables to serve as model inputs. This approach reduces the dimensionality of the input vector while improving computational efficiency, and is widely recognized as one of the most important parts of ML12–14. In addition, ML prediction performance depends heavily on the scale and quality of the sample data. Data quality directly affects model training, such that feature engineering can be regarded as setting the upper bound of achievable accuracy, whereas data quality sets its lower bound. Raw TBM construction data contain a large amount of noisy data caused by operator behavior, equipment faults and other factors. Standardized data cleansing is therefore the basis for subsequent analysis15,16 and is essential for studying rock–machine interaction and extracting reliable characteristic parameters. There are currently two main strategies for computing and extracting features: (1) data-driven approaches, in which intelligent algorithms automatically select features based on statistical patterns; and (2) knowledge-driven approaches, in which features are selected based on physical and mechanical understanding and professional expertise17. Because TBM rock breakage is similar to in-situ indentation–shear testing, knowledge-driven approaches based on geotechnical expertise and TBM operating principles are generally considered more interpretable and are thus more widely used than purely data-driven methods18–24.In addition, data-driven approaches often suffer from limited interpretability, and variations in data collection types and volumes across projects make cross-project generalization challenging. Most studies agree that cutterhead thrust F, cutterhead torque T, advance rate v and rotation speed n are the core parameters that reflect TBM mechanical status and rock-breaking performance11,19,25–29. To further reduce dimensionality, many researchers17,30 have linearly combined and transformed these basic parameters to derive rock mass boreability indexes, such as the field penetration index (FPI) and torque penetration index (TPI), which are then used as model inputs. Extensive research on FPI and TPI has demonstrated their feasibility for predicting rock mass quality17,18,30–32. In practice, to make effective use of such characteristic parameters, they must be computed and processed in real time. In most current studies, FPI and TPI are directly computed from their definitions, which results in multiple values per boring cycle and considerable fluctuation, making it difficult to link them directly to rock mass quality. To address this, Chen et al.17 proposed a regression-based method for computing characteristic parameters, using the fitting coefficients to represent FPI and TPI for each cycle. This method is suitable for TBM tunnelling analysis under big-data conditions. However, its applicability is limited by the quality of curve fitting. Under certain geological conditions, such as fractured soft rock zones or ultra-hard rock segments, the relationships among parameters may become unclear and regularity is weak, leading to low-quality characteristic parameters that poorly reflect rock mass conditions. Therefore, a key challenge is how to adapt to varying geological conditions, scientifically extract effective rock-breaking data, and achieve fast and accurate computation of FPI and TPI across different projects to support real-time evaluation of rock mass quality ahead of the TBM face. Against this background, this study aims at rock mass quality identification during TBM advance. By coupling TBM disc cutter energy conversion with rock fracture characteristics, we identify high-efficiency rock-breaking segments, compute characteristic parameters across multiple projects and perform real-time rock mass quality prediction. The method is validated on TBM projects with different diameters and geological conditions, demonstrating its applicability and supporting the integration of big data and AI algorithms in TBM engineering for safer and more efficient construction. The main innovations of this study can be summarized as follows. Limitations of existing FPI/TPI calculation methods in engineering applications. Limitations of existing FPI/TPI calculation methods in engineering applications. Limitations of existing FPI/TPI calculation methods in engineering applications. Existing FPI/TPI-based methods are effective for rock mass quality prediction, but their applicability in modern TBM data environments remains limited because they often rely on direct calculation or project-specific regression, leading to unstable characteristic parameters, sensitivity to geological variability and data quality, and poor cross-project robustness. (2)New capabilities of the proposed method for different engineering projects. (2)New capabilities of the proposed method for different engineering projects. New capabilities of the proposed method for different engineering projects. Unlike conventional methods that rely on large prior datasets and fixed penetration thresholds, this study identifies high-efficiency rock-breaking stages using local maxima of the FPI–t curve during tunnelling, enabling real-time characteristic-parameter computation with good timeliness. Validation across TBM projects with different diameters and geological conditions further demonstrates the generality of the proposed method. Engineering background and database Project background This study uses data from three TBM projects with different diameters and overburden depths: the Yinsong Water Diversion Project (YS), the Yinchao Water Diversion Project (YC), and the Huanbei Water Diversion Project (HB). Plan views are shown in Fig. 1, and the main TBM design and technical parameters are given in Fig. 2; Table 1. Table 1Main technical parameters of the TBMs used in the three projects.ParameterYS projectYC projectHB projectCutterhead diameter/m7.935.26.53Maximum overburden/m260250600Number of disc cutters563441Number of data channels199400391Maximum stroke/m1.81.81.9Rated torque/(kN m)841033404322Rated thrust/kN23,26011,34021,991Disentangling torque/(kN m)12,61549006483Maximum advance rate/(mm min−1)120120120Maximum rotation speed/(r min−1)7.611.458.4Number of boring cycles13,8846419253 Table 1Main technical parameters of the TBMs used in the three projects.ParameterYS projectYC projectHB projectCutterhead diameter/m7.935.26.53Maximum overburden/m260250600Number of disc cutters563441Number of data channels199400391Maximum stroke/m1.81.81.9Rated torque/(kN m)841033404322Rated thrust/kN23,26011,34021,991Disentangling torque/(kN m)12,61549006483Maximum advance rate/(mm min−1)120120120Maximum rotation speed/(r min−1)7.611.458.4Number of boring cycles13,8846419253 Main technical parameters of the TBMs used in the three projects. The YS project employs an open-type TBM with a diameter of 7.93 m to construct a 17.45 km tunnel through surrounding rock classes II–V. Tunnelling started in May 2015 and was completed in January 2018. The YC project diverts water from the Chuoer River to the Xiliao River through a circular non-pressurized tunnel with a diameter of 5.2 m and a TBM-driven length of 18.8 km; tunnelling started in September 2020 and ended in December 2023. The HB project uses a TBM with a diameter of 6.53 m to excavate a 10.09 km tunnel, mainly through class III surrounding rock, followed by classes II, IV, and V. Tunnelling started on 17 October 2024 and was still ongoing at the time of this study. Fig. 1Schematic diagrams of TBM tunnelling in the three projects. (a)YS project, (b) YC project, (c) HB project. The maps were compiled by the authors based on non-confidential engineering geological information from the corresponding projects. Geological information related to the YS project has been previously reported in Ref.24. Schematic diagrams of TBM tunnelling in the three projects. (a)YS project, (b) YC project, (c) HB project. The maps were compiled by the authors based on non-confidential engineering geological information from the corresponding projects. Geological information related to the YS project has been previously reported in Ref.24. Fig. 2Schematic diagram of TBM design. Schematic diagram of TBM design. Database construction Effective rock-breaking data During TBM operation, hundreds of channels are recorded every second. Because of equipment faults, adverse geology, and other factors, downtime occupies a large proportion of the data33, and even working cycles contain many invalid records16. For example, on 18 September 2020, the YC project recorded 85,197 data rows but only nine valid boring cycles, with an effective working time of 17,196 s (20.2% of the total), as shown in Fig. 3. Rigorous preprocessing is therefore necessary. Fig. 3Tunnelling record of the YC project on 18 September 2020. Tunnelling record of the YC project on 18 September 2020. The first preprocessing step is to identify the TBM operating state from cutterhead thrust, torque, advance rate, and rotation speed22,29,34–39, and divide the data into downtime and tunnelling stages. Abnormal conditions and invalid data are then removed using the rules summarized in Table 2. These methods mainly rely on parameter reasonableness and trends, but do not explicitly reflect the actual rock-breaking state. Further improvement in data quality is therefore still possible. Table 2Identification and handling of invalid TBM data.Abnormal operation conditionCharacteristicsIdentification methodProcessing methodExcessively short boring distanceShort cycle lengthLength < 0.3 mDiscard the boring cycleShort stoppageTemporary standstillMain tunnelling parameters drop to zero within a short periodDelete the corresponding data rowsAdvance rate out of rangeAbnormally high advance rateAdvance rate v > 120 mm·min−1Set v = 120 mm min−1; delete out-of-range rows; remove outliers using the 3σ criterionData transmission anomaly15Communication interruption or failureMain tunnelling parameters remain unchanged for 10 sDelete the corresponding data rows Table 2Identification and handling of invalid TBM data.Abnormal operation conditionCharacteristicsIdentification methodProcessing methodExcessively short boring distanceShort cycle lengthLength < 0.3 mDiscard the boring cycleShort stoppageTemporary standstillMain tunnelling parameters drop to zero within a short periodDelete the corresponding data rowsAdvance rate out of rangeAbnormally high advance rateAdvance rate v > 120 mm·min−1Set v = 120 mm min−1; delete out-of-range rows; remove outliers using the 3σ criterionData transmission anomaly15Communication interruption or failureMain tunnelling parameters remain unchanged for 10 sDelete the corresponding data rows Identification and handling of invalid TBM data. TBM rock breaking is analogous to a large-scale in situ torsional shear test17. Rock fragmentation is mainly produced by indentation from thrust and cutting by torque40,41. Focusing on a single disc cutter, this study evaluates whether cutter work is effectively converted into rock-breaking energy, as illustrated in Fig. 4. If either thrust work or torque work at a given time step is non-positive, the cutter cannot provide effective rock-breaking energy; the corresponding records are therefore treated as non-rock-breaking and removed. This criterion is necessary but not sufficient: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{W}>0\:$$\end{document}does not necessarily imply efficient rock breaking. Low-efficiency cutting is retained and labeled as valid but low-efficiency data for subsequent modelling. In Fig. 4, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{F}\:$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{T\:}$$\end{document}denote the work done per second by the single-cutter thrust \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{n}=F/N$$\end{document}and torque \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{s}=T/N$$\end{document}, respectively (J), where Fis cutterhead thrust (kN), Tis cutterhead torque (kN m), and Nis the number of cutters. Here, tis time (s), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Delta\:}x={x}_{t}-{x}_{t-1}$$\end{document}is the advance increment (mm), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Delta\:}\theta\:\:$$\end{document}is the rotation angle increment (rad), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t-1}$$\end{document}are the rotation speeds (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{r}\cdot\mathrm{m}\mathrm{i}\mathrm{n}}^{-1}$$\end{document}). Fig. 4Flowchart of rock-breaking data filtering. Flowchart of rock-breaking data filtering. To verify applicability, conventional cleaning was first performed on YC, YS, and HB, followed by the proposed disc-cutter-energy filtering. This yielded 6005, 12,150, and 228 effective rock-breaking cycles for YC, YS, and HB, respectively. The changes in data volume and parameter statistics are summarized in Fig. 5. Fig. 5Comparison of parameter statistics before and after rock-breaking data filtering. (a) Reduction in data volume, (b) Changes in mean parameter values, (c) Reduction in coefficients of variation. Comparison of parameter statistics before and after rock-breaking data filtering. (a) Reduction in data volume, (b) Changes in mean parameter values, (c) Reduction in coefficients of variation. As shown in Fig. 5, the proposed filtering method performs well in all three projects. After non-rock-breaking data are removed, the average number of data points per cycle decreases markedly, whereas the parameter means and coefficients of variation change only slightly. Thus, essential information is preserved while redundant or invalid records are reduced. To further assess the robustness of using zero as the work threshold, the YC project was taken as an example. Thresholds of 0, 0.1, and 0.2 were applied to the effective rock-breaking cycle samples, and the numbers of samples in each surrounding rock class were compared, as shown in Table 3. Only slight changes were observed, indicating that small perturbations around zero do not materially affect class distributions; this further confirms that zero is both physically justified and robust, and it is therefore retained as the final work threshold. Table 3Statistics of sample numbers for each surrounding rock category in effective rock-breaking cycle segments under different thresholds (YC project).ThresholdSamplesClass-IIClass-IIIClass-IVClass-V0.0106833879456050.1105933829456020.210483382945585 Table 3Statistics of sample numbers for each surrounding rock category in effective rock-breaking cycle segments under different thresholds (YC project).ThresholdSamplesClass-IIClass-IIIClass-IVClass-V0.0106833879456050.1105933829456020.210483382945585 Statistics of sample numbers for each surrounding rock category in effective rock-breaking cycle segments under different thresholds (YC project). High-efficiency rock-breaking stage Within a complete boring cycle, the TBM typically passes through idle advance, initial contact with the face, stable high-efficiency rock breaking, and shutdown18. Based on the evolution of tunnelling parameters, a cycle can be divided into idling, ascending, stable, and descending stages16. Taking the YC project at chainage 59,957.4 m as an example, the segmentation based on raw data is shown in Fig. 6a. Some studies further subdivide the ascending stage. Wang et al.34 divided it into two parts: an initial stage of rapid thrust and torque growth, corresponding to initial tool-rock contact and rock breakage, and a second stage beginning when the applied load exceeds the critical rock-breaking force \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}}_{\boldsymbol{b}}$$\end{document}, after which high-efficiency rock breaking starts. In practice, because direct in situ determination of rock parameters is difficult, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\boldsymbol{F}}_{\boldsymbol{b}}$$\end{document}is usually approximated by a penetration-per-revolution threshold32,34. In this study, non-rock-breaking data are first removed using the cutter-energy criterion in “Effective rock-breaking data”, and the high-efficiency rock-breaking stage is then identified. The end of the ascending stage is determined using histogram statistics of tunnelling parameters40(We construct an equal-width histogram (B bins) of the smoothed tunnelling parameter (W-point centered moving average). The modal bin is identified, and its upper edge is used as a threshold. The stable-start is the first sample whose raw value exceeds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{T}\mathrm{-}\mathrm{m}$$\end{document}, subject to jitter and trend gates based on relative change-rate thresholds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta_{\mathrm{j}}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta_{\mathrm{t}}$$\end{document}. For multiple parameters, candidate start indices are fused by IQR-based outlier removal and the median index). Figure 6b shows the resulting segmentation for the same cycle after rock-breaking filtering. Fig. 6Comparison of tunnelling data before and after rock-breaking data filtering. (a) Original boring cycle and tunnelling stages, (b) Effective rock-breaking cycle and rock-breaking stages. Comparison of tunnelling data before and after rock-breaking data filtering. (a) Original boring cycle and tunnelling stages, (b) Effective rock-breaking cycle and rock-breaking stages. Computation of rock mass characteristic parameters based on rock-breaking features Rock mass boreability indexes Rock mass quality is a key basis for identifying adverse geology and optimizing TBM performance11,42. In practice, thrust and torque are often used directly to infer rock conditions. However, analysis of the YC project shows that raw tunnelling parameters do not correlate clearly with rock mass classes overall (Fig. 7a). To overcome this limitation, boreability indexes have been proposed. FPI and TPI characterize rock-mass behavior by integrating the rock-machine interaction process17,30 and are defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{FPI}}=\frac{{{F_n}}}{p}=\frac{{F/N}}{{Pr/RPM}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{TPI}}=\frac{{{T_s}}}{p}=\frac{{T/N}}{{Pr/RPM}}$$\end{document} where Fn is the average single-cutter thrust (kN); F is the total cutterhead thrust (kN); N is the number of cutters; Pr is the penetration rate (mm min− 1); RPM is the cutterhead rotation speed (r min− 1); and T and Ts are the total and single-cutter torques, respectively (kN m). Using these definitions, the mean FPI and TPI of each effective rock-breaking cycle in the YC project were calculated, and their distributions across rock mass classes are shown in Fig. 7b. Compared with the raw tunnelling parameters, FPI and TPI exhibit clearer and generally positive correlations with rock mass quality. This is further supported by the inter-class difference analysis in Table 4, which shows that the mean inter-class variation increases from 19.35% for the raw parameters to 28.38% for FPI/TPI, indicating that the boreability indexes are more informative for distinguishing rock mass quality. Table 4Inter-class difference test.Input parametersParameterVariation ranges between rock mass classesVariation (Class II → III)Variation (Class III → IV)Variation (Class IV → V)MeanRaw tunnelling parametersCutterhead torque5.21%8.87%30.21%14.77%Cutterhead thrust13.79%27.65%46.07%29.17%Advance rate3.42%23.89%37.21%21.51%Cutterhead rotational speed15.80%0.19%19.84%11.95%Average9.56%15.15%33.34%19.35%Rock-breakability indicesFPI15.29%32.78%39.51%29.19%TPI1.73%34.18%46.77%27.56%Average8.51%33.48%43.14%28.38% Table 4Inter-class difference test.Input parametersParameterVariation ranges between rock mass classesVariation (Class II → III)Variation (Class III → IV)Variation (Class IV → V)MeanRaw tunnelling parametersCutterhead torque5.21%8.87%30.21%14.77%Cutterhead thrust13.79%27.65%46.07%29.17%Advance rate3.42%23.89%37.21%21.51%Cutterhead rotational speed15.80%0.19%19.84%11.95%Average9.56%15.15%33.34%19.35%Rock-breakability indicesFPI15.29%32.78%39.51%29.19%TPI1.73%34.18%46.77%27.56%Average8.51%33.48%43.14%28.38% Inter-class difference test. Fig. 7Distribution of tunnelling parameters and rock mass boreability indexes across rock mass classes in the YC project. (a) Tunnelling parameters, (b) Boreability indexes. Distribution of tunnelling parameters and rock mass boreability indexes across rock mass classes in the YC project. (a) Tunnelling parameters, (b) Boreability indexes. Fitting-based parameter calculation method Direct calculation of characteristic parameters (FPI/TPI) is straightforward but often results in strong fluctuations32. To address this issue, Chen Zuyu’s team proposed a regression-based method in which characteristic parameters are computed by fitting tunnelling data34. In this framework, the coefficients of the thrust–penetration ( F– p) relationship, a and b, are used to characterize FPI: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:a\:$$\end{document}denotes the thrust increment per unit penetration during the stable rock-breaking stage, reflecting the integrated boreability of the surrounding rock under sustained cutting, whereas b, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\approx\:b\:$$\end{document}as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\to\:0$$\end{document}, represents the initial breakage threshold associated with tool–rock engagement and crack initiation. The fitted torque–penetration ( T– p) relationship is used to characterize TPI. In this way, each boring cycle is uniquely represented by a single parameter set (a, b, TPI). The effectiveness of this method depends strongly on fitting quality. To illustrate this, 2299 effective rock-breaking cycles from the YC project (chainages 66,059–55,721 m and 25,916–25,525 m) were analyzed. The characteristic parameters were computed by curve fitting, and the coefficients of variation (CVs) of a, b, and TPI were compared between cycles with goodness-of-fit (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{2}$$\end{document}) lower than 0.6 and those with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{2}$$\end{document}higher than 0.6. The results are shown in Fig. 8. As \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{2}\:$$\end{document}increases, the CVs of a, b, and TPI decrease across all rock mass classes. This indicates that higher fitting quality leads to lower within-class variability and, therefore, higher-quality characteristic parameters. In other words, the quality of curve fitting directly determines the quality of the derived characteristic parameters. Accurate selection of fitting data is therefore critical for effective regression-based parameter calculation. Fig. 8Coefficients of variation of a, b and TPI for different levels of goodness-of-fit. (a) Distribution of the CV of a across different surrounding rock categories. (b) Distribution of the CV of b across different surrounding rock categories. (c) Distribution of the CV of TPI across different surrounding rock categories. Coefficients of variation of a, b and TPI for different levels of goodness-of-fit. (a) Distribution of the CV of a across different surrounding rock categories. (b) Distribution of the CV of b across different surrounding rock categories. (c) Distribution of the CV of TPI across different surrounding rock categories. As discussed in “High-efficiency rock-breaking stage”, a key issue is how to determine the boundary of the high-efficiency rock-breaking stage. Existing studies usually rely on project-specific statistics and adopt a fixed penetration threshold (e.g., p = 2 mm·rev− 1) to approximate the rock-breaking force \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{b}$$\end{document}and identify the boundary of the ascending stage. However, because geological conditions vary across projects, this fixed penetration-based criterion is often unsuitable for cycles with strongly fluctuating penetration under special conditions. For example, in the YC project at chainage 65,823.2 m, penetration ranges from 2 to 3.5 mm rev− 1 and contains obvious outliers. As shown in Fig. 9, the F– p and T– p relationships extracted by the penetration-threshold method exhibit poor goodness-of-fit. In addition, at the initial stage of construction, the accumulated data may be insufficient to determine a reliable project-specific threshold. Fig. 9Correlation of parameters in the high-efficiency rock-breaking stage extracted using the penetration-threshold method. Correlation of parameters in the high-efficiency rock-breaking stage extracted using the penetration-threshold method. Knowledge-driven feature parameter extraction method The stress–strain curve of rock failure (Fig. 10) shows that stress first compacts the rock and increases its load-bearing capacity; once the ultimate strength is exceeded, the strength and elastic modulus decrease and macro-failure occurs. Similar progressive failure characteristics have also been reported in studies on deep hard rock fracture and rock mass kinematic evolution43,44. Recent studies further suggest that this macroscopic response is controlled by heterogeneous microstructure evolution, fracture architecture, and stress-dependent nonlinear deformation behavior45–47. During TBM cutting with disc cutters, the rock–cutter interaction follows a similar progressive process. At the initial penetration stage, pores and microcracks are compacted, so the cutter thrust generally rises with penetration. As penetration continues, damage accumulates and cracks propagate beneath the cutter, and the thrust reaches its peak. Once unstable macro-fracture and rock chipping occur, the local load-bearing capacity decreases, leading to a reduction in thrust. Therefore, the thrust evolution described by the model reflects the progressive damage evolution of the rock mass, rather than merely a phenomenological variation in cutter load. This trend has been confirmed by the disc-cutter indentation tests conducted by Wang48 (Fig. 11) and by the TBM rock-breaking model tests reported by Jiang et al.49. Accordingly, thrust evolution can be used to distinguish different rock-breaking stages: in the ascending phase, FPI increases with penetration, whereas its decrease indicates the onset of unstable failure. The turning point of the FPI–time (FPI–t) curve therefore marks the transition from stable damage development to unstable macro-fracture propagation, indicating that FPI evolution provides a model-consistent mechanical basis for identifying the onset of the high-efficiency rock-breaking stage, rather than serving as a purely empirical criterion. Fig. 10Stress–strain curve of rock failure process48. Stress–strain curve of rock failure process48. Fig. 11Thrust load–penetration distance curve obtained from disc cutter rock-breaking test48. Thrust load–penetration distance curve obtained from disc cutter rock-breaking test48. On the basis of the above analysis of thrust–rock fragmentation characteristics, the local maximum of the FPI–t curve is adopted as the onset of the high-efficiency rock-breaking stage in the ascending phase. The implementation includes signal smoothing and local peak detection. The signal is first segmented using a 15-point window (window_size = 15). A candidate onset is identified when two consecutive segments satisfy the following pattern: the mean of the first 7 points is greater than the global mean, whereas the mean of the last 7 points is lower than the global mean. Then, over the full sequence, a 3-point moving average is iteratively updated with a step size of 3, and the index corresponding to the maximum 3-point mean is selected as the reference point. To evaluate robustness, Monte Carlo simulations were conducted under different sampling rates (1/2/5/10 Hz) and noise levels (Gaussian noise plus random spikes). The results show that the proposed detection strategy remains stable over the tested range: Under a ± 5 s tolerance criterion, the miss rate is zero for all tested combinations, and the success rate remains nearly 100%, decreasing slightly to 0.995 only when fs = 1 Hz and σ = 1.0. Although the localization error increases monotonically with noise level, it remains within approximately 0.3–1.5 s. These results confirm that the local-maximum method is robust to practical variations in sampling frequency and noise level. Taking the YC project at chainage 65,823.2 m as an example, Fig. 12 shows the FPI– t curve and the selected data segment to the right of the local maximum. The corresponding T– p and F– p relationships are shown in Fig. 13. Compared with Fig. 9, the anomalous penetration data are effectively removed, and the goodness-of-fit of the F– p and T– p regressions improves from 0.3876 to 0.7180 to 0.8433 and 0.9866, respectively. Fig. 12Schematic illustration of selecting the high-efficiency rock-breaking stage based on the FPI–t curve. Schematic illustration of selecting the high-efficiency rock-breaking stage based on the FPI–t curve. Fig. 13Relationships between cutterhead torque and penetration and between cutterhead thrust and penetration for the selected high-efficiency rock-breaking data. Relationships between cutterhead torque and penetration and between cutterhead thrust and penetration for the selected high-efficiency rock-breaking data. Among the 6005 boring cycles in the YC project, 1400 cycles (23.31%) exhibit strong penetration fluctuations. For these cycles, the high-efficiency rock-breaking stage was extracted using both the penetration-threshold method and the proposed FPI–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{t}$$\end{document} local-maximum method. The resulting characteristic parameters were then statistically compared, and the results are summarized in Table 5. Table 5Comparison of different methods for extracting the high-efficiency rock-breaking stage in problematic cycles.Criterion for extracting high-efficiency rock-breaking stageCVGoodness-of-fitabTPIFPITPIPenetration > 2 mm·rev⁻¹0.5680.3530.3880.3550.602Local maximum of FPI–t curve0.5310.2450.3350.6050.753Change (%)− 6.39%− 30.51%− 13.73%+ 70.47%+ 25.17%Average change− 16.88%+ 47.82% Table 5Comparison of different methods for extracting the high-efficiency rock-breaking stage in problematic cycles.Criterion for extracting high-efficiency rock-breaking stageCVGoodness-of-fitabTPIFPITPIPenetration > 2 mm·rev⁻¹0.5680.3530.3880.3550.602Local maximum of FPI–t curve0.5310.2450.3350.6050.753Change (%)− 6.39%− 30.51%− 13.73%+ 70.47%+ 25.17%Average change− 16.88%+ 47.82% Comparison of different methods for extracting the high-efficiency rock-breaking stage in problematic cycles. Table 5 indicates that, compared with the penetration-threshold method, the local-maximum method based on the FPI–t curve reduced the average coefficient of variation of the characteristic parameters by 16.88% and improved the average goodness of fit by 47.82%. This suggests that, in abnormal cycles with significant penetration fluctuations, the proposed method can extract more stable characteristic parameters with better fitting performance. From an engineering perspective, these improvements mean that the extracted parameters more accurately reflect the actual rock-breaking state, especially in sections with poor surrounding rock conditions or large parameter fluctuations. As a result, the method can reduce fitting errors and feature distortion, lower the risk of misclassification in real-time surrounding rock classification, and support earlier and more accurate identification of poor surrounding rock zones for tunneling parameter adjustment and safety decision-making. To further assess the consistency of this improvement, the per-sample improvement distribution for the 1400 cycles was summarized and a paired Wilcoxon signed-rank test was performed (Table 6). The median absolute improvement is 0.2409 (IQR: 0.2068), and the median relative improvement is 62.05% (IQR 120.41%). The Wilcoxon test gives \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p=8.67\times\:{10}^{-209}$$\end{document}, confirming that the improvement in goodness-of-fit is systematic rather than a random fluctuation. Table 6Distributional statistics of per-sample improvements and paired Wilcoxon test.ItemsMetric definitionMedianIQRSample size(n)Distribution (absolute improvement)diff = new − old0.24090.20681400Distribution (relative improvement)(new − old)/|old| × 100%62.05%120.41%1400Significance testPaired Wilcoxon signed-rankp = 8.67 × 10⁻209–1400 Table 6Distributional statistics of per-sample improvements and paired Wilcoxon test.ItemsMetric definitionMedianIQRSample size(n)Distribution (absolute improvement)diff = new − old0.24090.20681400Distribution (relative improvement)(new − old)/|old| × 100%62.05%120.41%1400Significance testPaired Wilcoxon signed-rankp = 8.67 × 10⁻209–1400 Distributional statistics of per-sample improvements and paired Wilcoxon test. In summary, the proposed knowledge-driven real-time feature-parameter extraction method, derived from the analysis of disc-cutter thrust–rock fragmentation characteristics, consists of three steps: removing non-rock-breaking data based on effective cutter-work energy conversion, identifying the high-efficiency rock-breaking stage from the local maximum of the FPI–t curve, and fitting the F–p and T–p relationships to obtain the characteristic parameters a, b, and TPI. AI-based rock mass quality prediction with cross-project transferability Model development and validation for the YC project Comparative evaluation of AI-driven machine learning models In this subsection, multiple AI-driven machine learning models are developed and systematically compared to evaluate predictive performance and identify the most suitable model configuration. To evaluate the proposed knowledge-driven characteristic-parameter computation method, the 6,005 effective rock-breaking cycles of the YC project were used to calculate the input features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{a}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{b}$$\end{document}, FPI goodness-of-fit, TPI, and TPI goodness-of-fit. Five machine-learning algorithms were considered for rock mass quality prediction: XGBoost, CatBoost, Random Forest (RF), LightGBM, and Logistic Regression (LR). Accuracy (ACC) and F1-score were used as evaluation metrics, and the results are shown in Fig. 14. CatBoost achieves the best overall accuracy and the clearest advantage in F1-score. It was therefore selected as the base model for this study. Fig. 14Comparison of rock mass quality prediction performance for different ML models. Comparison of rock mass quality prediction performance for different ML models. Furthermore, to substantiate the knowledge-driven nature of the proposed framework and improve interpretability, SHAP (Shapley Additive Explanations) was used to quantify the contributions of the input features to the CatBoost predictions. As shown in Fig. 15, TPI has the highest global importance (Mean(|SHAP|) = 1.8154), followed by b(0.4939), a(0.4099), TPI_R2 (0.3139), and FPI_R2 (0.2540). This indicates that the model mainly relies on the knowledge-based information encoded by TPI, while aand bprovide complementary contributions. In contrast, the relatively low contribution of the R2-type goodness-of-fit indicators reduces the risk that the model relies on fitting quality or other potential spurious correlates. Fig. 15Global feature importance of the CatBoost model based on Mean(|SHAP|). Global feature importance of the CatBoost model based on Mean(|SHAP|). Model construction, optimization and testing CatBoost is a gradient-boosting algorithm that uses ordered target statistics and ordered boosting to reduce target leakage and gradient bias. Its symmetric tree structure improves stability and generalization in practice50–52.The CatBoost model contains multiple hyperparameters. Among them, iterations and learning_rate control the learning rate and convergence process; depth and l2_leaf_reg control model complexity and prevent overfitting; and random_strength and bagging_temperature adjust model randomness to enhance generalization. To maximize the predictive power of CatBoost, the input feature vectors are first filtered based on the conclusions of “Fitting-based parameter calculation method”: the goodness-of-fit of the FPI regression is used as a criterion for screening characteristic parameters. The parameters computed in “Comparative evaluation of AI-driven machine learning models” are grouped according to FPI goodness-of-fit, and four prediction models are built for different ranges of goodness-of-fit. Each model was evaluated using stratified five-fold cross-validation. Hyperparameters were optimized by Bayesian optimization with Optuna TPE (TPESampler(seed = 42)) to maximize the five-fold macro-F1. The search space was: iterations [300, 2000], learning_rate [0.01, 0.3] (log), depth [4, 10], l2_leaf_reg [1, 20] (log), random_strength [0, 20], and bagging_temperature [0, 1], with bootstrap_type = Bayesian. Thirty trials were run for each dataset with fixed seeds (42), single-thread execution (thread_count = 1), and early stopping (od_wait = 100, use_best_model=True). The average metric across the five-folds serves as the objective function to be maximized, and the hyperparameter combination that achieves the best validation performance is selected as the optimal setting. The corresponding sample numbers and prediction results are shown in Fig. 16. As the goodness-of-fit increases, both accuracy and F1-score generally improve, while the number of available cycles decreases. The best performance is obtained when the FPI goodness-of-fit exceeds 0.6. When it exceeds 0.8, the number of cycles drops sharply and prediction accuracy declines because of insufficient training data. Therefore, 0.6 is selected as the boundary between high- and low-quality feature sets. The optimal hyperparameters of the YC CatBoost model are listed in Table 7, and the final model achieves an accuracy of 85.60%. Five-fold cross-validation further indicates stable performance, with an ACC of 0.8560 ± 0.0083 and an F1-score of 0.7751 ± 0.0147. Detailed results are summarized in Table 8. Table 7Optimal hyperparameter combination of the best CatBoost model for the YC project.HyperparameterDescriptionOptimal valueIterationsNumber of boosting iterations786Learning_rateLearning rate0.2950DepthMaximum tree depth7l2_leaf_regL2 regularization coefficient2.4733Random_strengthRandomness of feature importance3.7809Bagging_temperatureSampling temperature0.2179 Table 7Optimal hyperparameter combination of the best CatBoost model for the YC project.HyperparameterDescriptionOptimal valueIterationsNumber of boosting iterations786Learning_rateLearning rate0.2950DepthMaximum tree depth7l2_leaf_regL2 regularization coefficient2.4733Random_strengthRandomness of feature importance3.7809Bagging_temperatureSampling temperature0.2179 Optimal hyperparameter combination of the best CatBoost model for the YC project. Table 8Summary of five-fold cross-validation results for the YC, YS, and HB projects.ProjectMetricFold 1Fold 2Fold 3Fold 4Fold 5MeanSD95% CIYCACC0.86670.84500.86020.85710.85110.85600.00830.8457–0.8664F1-score0.77270.76750.77250.80040.76240.77510.01470.7568–0.7934YSACC0.86690.86920.86230.86340.86210.86480.00310.8609–0.8687F1-score0.84640.85510.83940.84750.84220.84610.00600.8387–0.8535HBACC0.84070.81480.84310.84070.88890.84560.02680.8123–0.8789F1-score0.81200.87670.87230.83810.89100.85800.03220.8180–0.8980 Table 8Summary of five-fold cross-validation results for the YC, YS, and HB projects.ProjectMetricFold 1Fold 2Fold 3Fold 4Fold 5MeanSD95% CIYCACC0.86670.84500.86020.85710.85110.85600.00830.8457–0.8664F1-score0.77270.76750.77250.80040.76240.77510.01470.7568–0.7934YSACC0.86690.86920.86230.86340.86210.86480.00310.8609–0.8687F1-score0.84640.85510.83940.84750.84220.84610.00600.8387–0.8535HBACC0.84070.81480.84310.84070.88890.84560.02680.8123–0.8789F1-score0.81200.87670.87230.83810.89100.85800.03220.8180–0.8980 Summary of five-fold cross-validation results for the YC, YS, and HB projects. Fig. 16Rock mass quality prediction performance for cycle groups with different FPI goodness-of-fit levels. Rock mass quality prediction performance for cycle groups with different FPI goodness-of-fit levels. Model application for scheme validation and comparison To further validate the proposed characteristic parameter computation method, three calculation schemes are compared using the YC project data and the optimized CatBoost model from Section. Method 0: No data preprocessing; high-efficiency rock-breaking stage identified by the local maximum of the FPI–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{t}$$\end{document} curve; characteristic parameters computed by regression. Method I: Conventional preprocessing; high-efficiency stage identified by the penetration-threshold method(p = 2 mm·rev−1); characteristic parameters computed by regression. Method II: Conventional preprocessing; high-efficiency stage identified by the local maximum of the FPI–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{t}$$\end{document} curve; characteristic parameters computed by regression. Method III (proposed): Rock-breaking data preprocessing to obtain effective rock-breaking cycles; high-efficiency stage identified by the local maximum of the FPI–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{t}$$\end{document} curve; characteristic parameters computed by regression. After filtering using the FPI goodness-of-fit threshold of 0.6, 858, 833, 2020 and 1919 groups of characteristic parameters are obtained by Methods I, II and III, respectively. A total of 170 common cycles are randomly selected as the test set (stratified random sampling by rock mass class), and the remaining cycles are used for training. Overall accuracy (ACC) and F1-score are used as evaluation metrics. The results are summarized in Tables 9 and 10. Table 9Comparison of different characteristic parameter computation methods in the YC project.MethodMethod for calculating input feature parametersNumber of high-quality feature setsPrediction of surrounding rock qualityData preprocessing resultCriterion for starting high-efficiency rock-breaking stageTraining set/segmentsTest set/segmentsAccuracy ACC/%F1-scoreMethod 0Original cyclesLocal maximum of FPI–t curve85868817079.170.8016Method IValid boring cyclesPenetration > 2 mm·rev−183366380.590.7699Method IILocal maximum of FPI–t curve2020185082.940.8134Method IIIEffective rock-breaking cyclesLocal maximum of FPI–t curve1919174984.120.8244 Table 9Comparison of different characteristic parameter computation methods in the YC project.MethodMethod for calculating input feature parametersNumber of high-quality feature setsPrediction of surrounding rock qualityData preprocessing resultCriterion for starting high-efficiency rock-breaking stageTraining set/segmentsTest set/segmentsAccuracy ACC/%F1-scoreMethod 0Original cyclesLocal maximum of FPI–t curve85868817079.170.8016Method IValid boring cyclesPenetration > 2 mm·rev−183366380.590.7699Method IILocal maximum of FPI–t curve2020185082.940.8134Method IIIEffective rock-breaking cyclesLocal maximum of FPI–t curve1919174984.120.8244 Comparison of different characteristic parameter computation methods in the YC project. Table 10Detailed predictive performance metrics for the YC project.MethodRock mass classIIIIIⅣⅤPrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecallMethod 00.28570.38100.86360.77030.87760.89581.00001.0000Method I0.20000.04760.76670.93240.84000.84001.00001.0000Method II0.45450.23810.80720.90540.86270.88001.00001.0000Method III0.33330.19050.81610.95950.93480.86001.00001.0000 Table 10Detailed predictive performance metrics for the YC project.MethodRock mass classIIIIIⅣⅤPrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecallMethod 00.28570.38100.86360.77030.87760.89581.00001.0000Method I0.20000.04760.76670.93240.84000.84001.00001.0000Method II0.45450.23810.80720.90540.86270.88001.00001.0000Method III0.33330.19050.81610.95950.93480.86001.00001.0000 Detailed predictive performance metrics for the YC project. Compared with Method I, Methods II and III increase the number of high-quality feature sets by 142.50% and 130.37%, respectively. Method III gives the best prediction performance, improving ACC by 3.53% points and F1-score by 0.0545 over Method I. Compared with Method 0, Method III improves ACC by 4.95% points and F1-score by 0.0228. These results confirm that the proposed rock-breaking preprocessing and FPI–\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{t}$$\end{document} local-maximum method better capture the rock-machine interaction and improve rock mass quality prediction. Engineering portability of the proposed method YS project To investigate transferability, the same modelling strategy was applied to the YS project. The optimal CatBoost model and its performance are listed in Table 11, with an ACC of 86.48%. Five-fold cross-validation indicates high stability, with an ACC of 0.8648 ± 0.0031 and a F1-score of 0.8461 ± 0.0060. Detailed results are summarized in Table 8. Table 11Parameter settings and performance of the optimal CatBoost model for the YS project.CategoryParameterValueHyperparametersIterations1247Learning_rate0.1878Depth7l2_leaf_reg12.2816Random_strength0.0810Bagging_temperature0.4017Evaluation metricsF1-score0.8461ACC0.8648 Table 11Parameter settings and performance of the optimal CatBoost model for the YS project.CategoryParameterValueHyperparametersIterations1247Learning_rate0.1878Depth7l2_leaf_reg12.2816Random_strength0.0810Bagging_temperature0.4017Evaluation metricsF1-score0.8461ACC0.8648 Parameter settings and performance of the optimal CatBoost model for the YS project. The same three methods (I–III) used for the YC project are then applied to the YS project to evaluate improvements in rock mass quality prediction. The results are summarized in Tables 12 and 13 . Table 12Comparison of rock mass quality prediction performance for different methods in the YS project.MethodNumber of high-quality feature setsPrediction of surrounding rock qualityTraining set/segmentsTest set/segmentsACC/%F1-scoreMethod I2950255139982.71%0.8129Method II3952355383.46%0.8175Method III4319392084.96%0.8346 Table 12Comparison of rock mass quality prediction performance for different methods in the YS project.MethodNumber of high-quality feature setsPrediction of surrounding rock qualityTraining set/segmentsTest set/segmentsACC/%F1-scoreMethod I2950255139982.71%0.8129Method II3952355383.46%0.8175Method III4319392084.96%0.8346 Comparison of rock mass quality prediction performance for different methods in the YS project. Table 13Detailed predictive performance metrics for the YS project.MethodRock mass classIIIIIⅣⅤPrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecallMethod I1.00000.66670.87060.92120.75000.77270.50000.1000Method II1.00000.50000.85500.95440.79670.74240.50000.1000Method III1.00000.50000.87170.95850.80310.77270.75000.1500 Table 13Detailed predictive performance metrics for the YS project.MethodRock mass classIIIIIⅣⅤPrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecallMethod I1.00000.66670.87060.92120.75000.77270.50000.1000Method II1.00000.50000.85500.95440.79670.74240.50000.1000Method III1.00000.50000.87170.95850.80310.77270.75000.1500 Detailed predictive performance metrics for the YS project. Methods II and III increase the number of high-quality feature sets by 33.97% and 46.41%, respectively, compared with Method I. The corresponding ACC values are 82.71%, 83.46% and 84.96%. These results indicate that the proposed rock-breaking data preprocessing and FPI–t local-maximum method are also applicable to the YS project, mitigating the influence of geological conditions and tunnel diameter and exhibiting good robustness across projects. HB project For the HB project, an intelligent TBM tunnelling system has been developed for TBM 9 in the Genzi branch tunnel, incorporating the characteristic parameter computation method (Method III) proposed in this study. During on-site deployment, a data acquisition unit was installed in the TBM electrical cabinet and connected to the Programmable Logic Controller(PLC) of TBM. The acquisition unit was then configured and commissioned through wiring inspection, signal verification, and synchronization tests. In addition, an industrial computer and an operator display were mounted in the control cabin to visualize real-time data and prediction results. Photographs of the data acquisition unit, the installation process, and the operator display interface are shown in Fig. 17. The system computes characteristic parameters in real time and predicts rock mass quality during tunnelling, as illustrated in Fig. 18. Fig. 17On-site deployment of the intelligent TBM tunnelling system. (b) Installation process, (c) Operator display interface. On-site deployment of the intelligent TBM tunnelling system. (b) Installation process, (c) Operator display interface. Fig. 18Real-time rock mass quality prediction in the TBM intelligent tunnelling system. Real-time rock mass quality prediction in the TBM intelligent tunnelling system. Using on-site monitoring data collected from 23 November to 24 December 2024, a total of 228 effective cycles were used for model training and prediction. Five-fold cross-validation was performed, and the results are summarized in Table 8. For the HB project, the model achieved a mean accuracy (ACC) of 0.8456 ± 0.0268 (95% CI 0.8123–0.8789) and a mean F1-score of 0.8580 ± 0.0322 (95% CI 0.8180–0.8980). Compared with YC and YS, HB exhibits relatively larger variability across folds, whereas YS is the most stable and YC lies in between. Overall, the variation across all three projects remains limited, indicating reasonably stable model performance. For reference, the confusion matrix of one representative fold (Fold 5) is shown in Fig. 19, where the overall accuracy reaches 88.89% and the F1-score is 0.8910. Fig. 19Confusion matrix of the rock mass quality prediction model for the HB project test set. Confusion matrix of the rock mass quality prediction model for the HB project test set. Beyond random five-fold cross-validation, we also conducted a time-based evaluation. Specifically, the most recent continuous period (e.g., 16–24 December 2024) was held out as the test set, while the remaining historical data were used for training and hyperparameter tuning; the trained model was then directly applied to the held-out period without retraining. This design aims to assess robustness under potential changes in geology and operating conditions. The resulting confusion matrix is shown in Fig. 20. Under this temporal hold-out scheme, the model achieved an accuracy (ACC) of 0.8182 and an F1-score of 0.8228. Relative to the random-split setting, ACC decreased by 0.0274 and the F1-score decreased by 0.0352, indicating only a modest degradation. Overall, the model remains practically usable and engineering-stable under temporal distribution shift. Fig. 20Confusion matrix for the time-based test set. Confusion matrix for the time-based test set. The prediction results for a verifiable subset of boring cycles are further evaluated, as summarized in Table 14. Ground-truth labels for this subset were obtained from the established on-site geological rechecking and exploration record system routinely used in tunnelling projects. In Table 14, the “Observed geological conditions” are determined according to an operational criterion based on the widely adopted rock mass classification–support correlation framework in engineering practice. The in situ surrounding rock conditions at chainages 11,938.59, 12,247.27, and 12,328.00 are presented in Fig. 21. Overall, the model predictions agree well with the corresponding engineering interpretations on this subset, demonstrating reliable field follow-up performance and highlighting the practical applicability of the proposed method. Table 14Verification of prediction results for a subset of boring cycles in the test set.Chainage (m)ParametersRock mass quality and support measuresabTPIPredicted classObserved geological conditionsSupport method11,938.5980.08129.417.93IIIntactNo support11,959.4452.84161.767.18IIIntactRock-bolt support12,175.7638.48195.667.45IIIntactShotcrete support12,313.8537.51159.565.91IIIModerately intactLocal steel ribs12,247.2724.54175.895.86IIIModerately intactRebar mesh12,103.9722.77221.464.45IIILocally brokenLocal steel ribs and rebar mesh12,328.0011.18119.622.78IVCavity collapseSystematic steel ribs and rebar rows Table 14Verification of prediction results for a subset of boring cycles in the test set.Chainage (m)ParametersRock mass quality and support measuresabTPIPredicted classObserved geological conditionsSupport method11,938.5980.08129.417.93IIIntactNo support11,959.4452.84161.767.18IIIntactRock-bolt support12,175.7638.48195.667.45IIIntactShotcrete support12,313.8537.51159.565.91IIIModerately intactLocal steel ribs12,247.2724.54175.895.86IIIModerately intactRebar mesh12,103.9722.77221.464.45IIILocally brokenLocal steel ribs and rebar mesh12,328.0011.18119.622.78IVCavity collapseSystematic steel ribs and rebar rows Verification of prediction results for a subset of boring cycles in the test set. Fig. 21Field-observed surrounding rock conditions at different boring cycles. (b) Chainage 12,247.27 m, (c) Chainage 12,328.00 m. Field-observed surrounding rock conditions at different boring cycles. (b) Chainage 12,247.27 m, (c) Chainage 12,328.00 m. Conclusions This study targets cross-project rock mass quality prediction. By filtering TBM raw data based on the effective energy conversion of disc cutter work and extracting high-efficiency rock-breaking stages considering rock fracture characteristics, a knowledge-driven real-time computation method for rock mass characteristic parameters applicable across different projects is developed. An AI-driven rock mass quality prediction model based on CatBoost is then constructed and validated using multi-project datasets. The main conclusions are as follows. Rock-breaking data filtering based on disc cutter energy. Rock-breaking data filtering based on disc cutter energy. Rock-breaking data filtering based on disc cutter energy. A rock-breaking data filtering method is proposed based on effective energy conversion of TBM disc cutter work. Compared with conventional data-driven cleaning methods, the proposed method achieves better performance on three real projects. The resulting CatBoost models attain prediction accuracies of 85.60% and 86.48% for the YC and YS projects, respectively, demonstrating good applicability across projects. (2)FPI–t local-maximum method for extracting high-efficiency rock-breaking stages. (2)FPI–t local-maximum method for extracting high-efficiency rock-breaking stages. FPI–t local-maximum method for extracting high-efficiency rock-breaking stages. A real-time method is developed to identify the onset of the high-efficiency rock-breaking stage from the local maximum of the FPI– t curve. In YC and YS, this method increases the number of high-quality characteristic-parameter sets by 142.50% and 33.97%, respectively, compared with the penetration-threshold method, and improves prediction performance. (3)Knowledge-driven cross-project computation of rock mass characteristic parameters. (3)Knowledge-driven cross-project computation of rock mass characteristic parameters. Knowledge-driven cross-project computation of rock mass characteristic parameters. Building on rock-breaking data preprocessing and high-efficiency rock-breaking stage extraction, this study proposes a knowledge-driven near-real-time method for computing rock-mass characteristic parameters. The results demonstrate good performance and engineering reproducibility in the YC and YS projects, as well as practical utility in the ongoing HB project under comparable open-type TBM conditions. Rather than claiming universal transferability, the present evidence supports promising deployment potential for similar open-type TBM projects with diameters of approximately 5–8 m. Generalization to substantially different conditions, such as shield TBM operation, ultra-deep tunnels ( >1000 m), or severe class imbalance, still requires further verification.