Screw Conveyor Speed Prediction for EPB-TBM Excavations Using Hybrid Deep Learning Models

• Novel hybrid deep learning framework combining SAINT, TabM, and KAN for TBM prediction. • BBOA-TabM achieved highest accuracy (R²=0.898 train, 0.855 test) among 9 models. • Dependent features introduced bias despite high R²; independent features ensure robustness. • Four empirical equations enable practical screw speed prediction (R² up to 0.70). Accurate prediction of screw conveyor speed in Earth Pressure Balance Tunnel Boring Machines (EPB-TBMs) is critical for maintaining operational stability and efficiency. This study proposes a hybrid deep learning framework integrating three architectures: Self-Attention and Intersample Attention Transformer, ensemble multilayer perceptron, and Kolmogorov-Arnold Networks, optimized through four different metaheuristic algorithms: Brown Bear Optimization, Fick’s Law Algorithm, Giant Trevally Optimizer, and Success History Intelligent Optimizer. Data from 930 excavation rings of Tabriz Metro Line 2 was utilized to evaluate the models in two modes: independent features and a dependent parameter called Screw Working Pressure. The results revealed higher performance of the Brown Bear Optimization-optimized ensemble multilayer perceptron model with superior training and testing correlations and lower prediction errors. The self-attention-driven transformer model followed closely, while the Kolmogorov-Arnold Networks trailed with a small margin largely due to limitations in modeling nonlinear tabular data. Metaheuristic experiments showed that Brown Bear Optimization demonstrated comparable or superior performance compared to other algorithms in each case. Moreover, four empirical models have been introduced in this work, differentiating the input variables used, where acceptable correlations were achieved with all features included. This highlights the need for rigorous independence of features to allow for generalization of the model. Overall, hybrid frameworks show promise for real-time prediction, though risks of overfitting and geological variability are present. Hybrid architectures and multi-site validation should be a focus for further work to increase robustness.