Data-driven transformer-based explainable deep learning model for real-time TBM performance prediction

Abstract

Aiming to effectively and efficiently predict tunnel boring machine (TBM) penetration rates to assist in guiding the operation management, this research proposes TransXTBMNet, a Transformer-based explainable network designed for predicting TBM penetration rates in real-time, which integrates ensemble bidirectional Long Short-Term Memory (BiLSTM) and improved Transformer to capture the complex interative relationship between operational parameters of TBM and penetration rates, enabling real-time precise predictions. Furthermore, the method encompasses various components: the data preparation process, the SHAP-based model explainability, a loss function, and performance metrics to collectively enhance robust data quality and ensure accurate predictions in dynamic excavation environments. Validation on the real TBM operation dataset demonstrates TransXTBMNet’s exceptional performance, with metrics like Mean Squared Error (MSE) = 0.2529, Root MSE (RMSE) = 0.5029, Mean Absolute Error (MAE) = 0.2991, Mean Absolute Percentage Error (MAPE) = 1.4984 %, and R2 = 0.9770. Comparative analysis shows superior results over other state-of-the-art methods such as conventional recurrent architectures (simple RNN (SRNN), GRU, LSTM) and their bidirectional variants (BiSRNN, BiGRU). SHAP analysis reveals that penetration rate at the previous time step, long with thrust groups A and F play crucial roles in predicting penetration rates. Additionally, the proposed method exhibits improved performance when considering the recorded penetration rate data, highlighting its capability to provide a better performance. Overall, the proposed method TransXTBMNet proves to be an effective and efficient solution in predicting TBM performance, providing valuable real-time guidance for optimizing TBM operations.