A knowledge-data dual-driven framework for predicting surrounding rock classification in TBM tunneling and generalization analysis

• A novel knowledge-data dual-driven framework for surrounding rock classification prediction is proposed combining the CSM model and APCA method. • A TPE-LightGBM voting model based on a time series segmentation algorithm is developed. • A strong negative linear correlation between dataset distribution discrepancies and prediction performance is identified. • Joint multi-project learning enhances prediction accuracy and generalization during the early stages of construction. Accurate and real-time prediction of surrounding rock classification (SRC) in tunnel boring machine (TBM) operations is critical for construction risk management and dynamic support design. Although machine learning has shown potential in improving SRC prediction accuracy, its generalization performance in practical engineering applications remains insufficiently explored. This study proposes a knowledge-data dual driven SRC prediction framework (KD-SRC) based on rock–machine interaction principles and evaluates its performance using historical excavation sequences. The framework incorporates physical relationships into feature engineering through the Colorado School of Mines (CSM) model and the adaptive piecewise constant approximation (APCA) method. A TPE-LightGBM voting model, which employs a tree-structured Parzen estimator to optimize the light gradient boosting machine, is used to achieve SRC prediction at the boring cycle level. Case studies on two deep-buried, long tunnel projects demonstrate that the proposed framework achieves over 89 % accuracy, with a recall exceeding 0.94 under adverse geological conditions. When dynamically updating the prediction model through excavation history, a strong negative linear correlation was observed between dataset distribution discrepancies and prediction performance, confirming that model generalization is influenced by similarities in feature distribution. Cross-project learning demonstrates that joint multi-project training improves prediction accuracy during the early stages of construction, and SHAP analysis validates the model’s credibility by aligning learned patterns with rock-machine interaction principles. Overall, this study not only improves the accuracy and efficiency of SRC prediction but also offers insights for developing more robust cross-project applications.