The geological condition monitoring and identification based on TBM-equipped sensors is of great significance for efficient and safe tunnel construction. Full-scale rotary cutting experiments are carried out using tunnel-boring machine disc cutters. Thrust, torque and vibration sensors are equipped on the rotary cutting machine (RCM). A stacking ensemble-learning model for real-time prediction of rock mass classification using features of mathematical statistics is proposed. Three signals, thrust, torque and a novel vibration spectrogram-based local amplification feature, are fed into the model and trained separately. The results show that the stacked ensemble-learning model has better accuracy and stability than any single model, showing a good application prospect in the rock mass classification.
As tunnel construction advances toward greater depths and lengths, full-face tunnel boring machines (TBMs) have become the preferred method for large-scale excavation. The operational efficiency of TBMs significantly impacts the progress, cost, and safety of tunnel projects. With the rapid development of machine learning and big data technologies, data-driven models based on cutterhead response signals have emerged as a key approach to improving TBM perception and decision-making. However, conventional rock-breaking indicators predominantly rely on single physical variables, limiting their ability to capture the complex dynamic interactions between the TBM and surrounding rock during excavation, thereby restricting their engineering applicability. To address this limitation, this study proposes a knowledge-driven data processing and indicator construction method to more accurately represent TBM operational states and surrounding rock properties. First, a novel excavation phase division algorithm based on time-domain and penetration-depth features is developed to accurately distinguish different tunneling stages. Subsequently, using data from the YC and YE projects, thrust- and torque-driven rock-breaking indicators are formulated, and the relationship between penetration depth and thrust/torque is optimized via power function fitting. Optimal exponents are determined through algorithmic optimization. Validation with field data confirms that the proposed indicators significantly enhance the accuracy and generalization of surrounding rock classification and control parameter prediction models.