Real-time perception of rock conditions based on continuously collected data to meet the requirements of continuous Tunnel Boring Machine (TBM) construction presents a critical challenge that warrants increased attention. To achieve this goal, this paper establishes real-time prediction models for fractured and weak rock mass by comparing 6 different algorithms using real-time data collected by the TBM. The models are optimized in terms of selecting metric, selecting input features, and processing imbalanced data. The results demonstrate the following points. (1) The Youden’s index and area under the ROC curve (AUC) are the most appropriate performance metrics, and the XGBoost Random Forest (XGBRF) algorithm exhibits superior prediction and generalization performance. (2) The duration of the TBM loading phase is short, usually within a few minutes after the disc cutter contacts the tunnel face. A model based on the features during the loading phase has a miss rate of 21.8%, indicating that it can meet the early warning needs of TBM construction well. As the TBM continues to operate, the inclusion of features calculated from subsequent data collection can continuously correct the results of the real-time prediction model, ultimately reducing the miss rate to 16.1%. (3) Resampling the imbalanced data set can effectively improve the prediction by the model, while the XGBRF algorithm has certain advantages in dealing with the imbalanced data issue. When the model gives an alarm, the TBM operator and on-site engineer can be reminded and take some necessary measures for avoiding potential tunnel collapse. The real-time predication model can be a useful tool to increase the safety of TBM excavation.