Overloading phenomena during tunnel boring machine (TBM) operations in adverse geological conditions can lead to equipment damage or jamming disasters, severely impeding construction progress. To address this challenge, this paper proposes a TBM load prediction method based on causal explainable artificial intelligence (AI). Specifically, the independent component analysis–linear non-Gaussian acyclic model (ICA-LiNGAM) algorithm is utilized to analyze causal relationships among tunneling parameters and select optimal input features. Subsequently, a Transformer-based time-series prediction model, incorporating causal strength relationships, is constructed to achieve real-time prediction of two critical load parameters: cutterhead torque and thrust force. Furthermore, the SHapley Additive exPlanations (SHAP) method is employed to elucidate the internal decision-making mechanism of the model. The accuracy and generalizability of the proposed method were validated using excavation data from two distinct TBM tunneling projects. Comparative experiments were also conducted to analyze the impact of the AI model architecture and causal discovery on prediction performance. The results indicate that feature selection and the integration of the causal adjacency matrix as an input feature significantly enhance both prediction accuracy and computational efficiency. The proposed CX-Transformer achieved coefficients of determination (R2) of 0.877 and 0.954 for cutterhead torque and thrust force, respectively, demonstrating robust generalizability across projects with differing geological conditions. Feature importance analysis revealed that parameters such as gripper pressure, thrust pressure, and thrust cylinder rod side pressure significantly influence thrust force predictions, while thrust force, thrust speed, and belt speed are primary determinants for cutterhead torque. The proposed method exhibits strong robustness in multi-step prediction tasks and provides a theoretical reference for TBM load prediction in practical engineering, offering vital support for safe on-site construction operations.Highlights<UnorderedList Mark="Bullet"> <ItemContent> Screening the input features of the model based on causal discovery and incorporating causal strength relationships, a Transformer-based time-series prediction model (CX-Transformer) was constructed to enable real-time prediction of two load parameters, namely cutterhead torque and thrust force. </ItemContent> <ItemContent> The model performance is evaluated using MSE, MAE, and R2, and the SHAP method is employed to analyze the contribution of each input feature and time step to the prediction results, enabling both global and local model interpretation. </ItemContent> <ItemContent> The proposed CX-Transformer modal achieved R2 values of 0.877 and 0.954 for cutterhead torque and thrust force, respectively. Compared to the causal-free Transformer model, CX-Transformer demonstrated superior prediction performance and improved computational efficiency by 11.72%. </ItemContent> <ItemContent> The CX-Transformer model demonstrates strong generalization performance. Validation using excavation data from a water supply project in Xinjiang shows that the prediction model performs well, with R2 values of 0.953 and 0.948 for cutterhead torque and thrust force, respectively. </ItemContent> </UnorderedList>

TBM Load Time-Series Prediction Based on Causally Interpretable Artificial Intelligence

期刊论文
发表日期：

2026年2月9日

查看全文

Zhenliang Zhou,
Man Wang,
Yimei Yue,
Yanrui Liu,
Haokai Li

摘要

Overloading phenomena during tunnel boring machine (TBM) operations in adverse geological conditions can lead to equipment damage or jamming disasters, severely impeding construction progress. To address this challenge, this paper proposes a TBM load prediction method based on causal explainable artificial intelligence (AI). Specifically, the independent component analysis–linear non-Gaussian acyclic model (ICA-LiNGAM) algorithm is utilized to analyze causal relationships among tunneling parameters and select optimal input features. Subsequently, a Transformer-based time-series prediction model, incorporating causal strength relationships, is constructed to achieve real-time prediction of two critical load parameters: cutterhead torque and thrust force. Furthermore, the SHapley Additive exPlanations (SHAP) method is employed to elucidate the internal decision-making mechanism of the model. The accuracy and generalizability of the proposed method were validated using excavation data from two distinct TBM tunneling projects. Comparative experiments were also conducted to analyze the impact of the AI model architecture and causal discovery on prediction performance. The results indicate that feature selection and the integration of the causal adjacency matrix as an input feature significantly enhance both prediction accuracy and computational efficiency. The proposed CX-Transformer achieved coefficients of determination (R²) of 0.877 and 0.954 for cutterhead torque and thrust force, respectively, demonstrating robust generalizability across projects with differing geological conditions. Feature importance analysis revealed that parameters such as gripper pressure, thrust pressure, and thrust cylinder rod side pressure significantly influence thrust force predictions, while thrust force, thrust speed, and belt speed are primary determinants for cutterhead torque. The proposed method exhibits strong robustness in multi-step prediction tasks and provides a theoretical reference for TBM load prediction in practical engineering, offering vital support for safe on-site construction operations.

Highlights

Screening the input features of the model based on causal discovery and incorporating causal strength relationships, a Transformer-based time-series prediction model (CX-Transformer) was constructed to enable real-time prediction of two load parameters, namely cutterhead torque and thrust force.

The model performance is evaluated using MSE, MAE, and R², and the SHAP method is employed to analyze the contribution of each input feature and time step to the prediction results, enabling both global and local model interpretation.

The proposed CX-Transformer modal achieved R2 values of 0.877 and 0.954 for cutterhead torque and thrust force, respectively. Compared to the causal-free Transformer model, CX-Transformer demonstrated superior prediction performance and improved computational efficiency by 11.72%.

The CX-Transformer model demonstrates strong generalization performance. Validation using excavation data from a water supply project in Xinjiang shows that the prediction model performs well, with R2 values of 0.953 and 0.948 for cutterhead torque and thrust force, respectively.