Real-Time Prediction of TBM Response Parameters Based on Temporal Convolutional Network
Yuan-en Pang, Zi-kai Dong, Hongwei Yu, Hao Cai, Guo-shuai Tian, Yuan Ji-dong, Yan Liu, Yu Wang, Xu Li
Abstract
Establishing an accurate predictive model for response parameters is the foundation of control parameter optimization for tunnel boring machines (TBMs). However, existing research mostly focuses on mean values during stable stages, and lacks real-time prediction throughout the entire process, failing to meet the demand for fine-tuned parameter recommendations. This paper proposes the weight matrix method for feature selection, which provides specific numerical values and rankings of each feature’s contribution. A deep learning model based on temporal convolutional network (TCN) is proposed to achieve real-time prediction of cutterhead torque (T) and total thrust (F), which is compared with the gated recurrent unit (GRU) and long short-term memory (LSTM). The proposed method was validated on the Yinchao project, and the results demonstrated that (1) the weight matrix method outperforms the Pearson coefficient method in terms of model accuracy, and (2) the TCN model performs better than GRU and LSTM. The method proposed in this paper achieves high precision in predicting T and F, and holds promise as a core algorithm for automatic control in TBM and providing crucial support for TBM’s advancement into the era of autonomous driving.