Abstract: Rock mass classification has a crucial impact on the performance of full-face tunnel boring machines (TBM).Due to high testing cost,rock mass data are sparse along tunnel alignments.Leveraging the extensive monitoring data collected during TBM excavation processes,data-driven models have been introduced to address rock classification challenges.Most existing models rely on sparse labeled data and empirical interpolations,often leading to training and evaluation based on labels inconsistent with actual rock conditions.A semi-supervised learning-based approach for rock mass classification is proposed,utilizing limited labeled data and abundant unlabeled tunneling parameters.The model is first pre-trained using the limited labeled data and then iteratively self-trained by gradually assigning pseudo-labels to the unlabeled data.Four evaluation metrics (accuracy,precision,recall,and F1 score) are used to assess the model performance on the labeled data,and the torque penetration index (TPI) is employed to verify model classification accuracy on the unlabeled data.The proposed method is illustrated using the Yinchuan Jiliao project in northeastern Inner Mongolia.The results show that under the specified dataset and testing conditions,the proposed method outperforms the supervised and unsupervised models in terms of accuracy and TPI.Using tunneling data from stable segments improves model accuracy.The model also outputs classification confidence,enhancing reliability.