Real-time prediction of TBM penetration rates using a transformer-based ensemble deep learning model

Abstract

Targeted to address the challenge of accurately predicting Tunnel Boring Machine (TBM) penetration rates in real-time, this paper explores how to develop a deep learning method that effectively and efficiently predicts penetration rates. A deep learning method termed a transformer-based ensemble bi-directional Long Short-Term Memory network (TransBiLSTMNet) is developed, comprising several modules, namely, the data processing, a backbone ensemble model, an improved transformer, loss function, and evaluation metrics. Validated on an actual TBM operation database, the developed method attains excellent performance with Mean Squared Error (MSE) of 0.1372, Mean Absolute Error (MAE) of 0.2099, Root MSE (RMSE) of 0.3704, Mean Absolute Percentage Error (MAPE) of 0.7091 %, and $R^2$ of 0.9961. Furthermore, the ablation experiments and comparative results illustrate the superior predictive accuracy. Accordingly, the TransBiLSTMNet provides a robust solution for real-time TBM operation management. Future research could focus on refining the model and exploring its application to other predictive scenarios.