Interpretable predictive model for shield attitude control performance based on XGboost and SHAP

The sudden decline in the attitude control performance is a common abnormal situation during shield tunneling. When the problem happens, the shield driver will have difficulty controlling the shield's attitude, which will cause the shield to deviate from its design axis and affect the quality of the tunnel. The causes behind poor control performance are usually complicated, so how to choose appropriate countermeasures is a challenging problem. Based on the above issues, this paper proposes the Interpretable Predictive Model for Shield attitude Control Performance (IPM_SCP). The model first predicts the current shield control performance through the extreme gradient boosting (XGBoost) sub-model and then uses the Shapley additive explanation sub-model to interpret the model output. The model was tested on the left-line tunnel of the Hangzhou–Shaoxing railway project in the Ke-Feng section. The results reveal that the model could effectively predict the control performance of the shield and give the most influential parameter and the direction in adjusting the parameter to improve the shield's attitude control performance when the control performance decreases. Therefore, IPM_SCP gives the correct parameter adjustment instructions when the shield’s attitude control performance declines, and eventually improves tunnel construction quality and efficiency.