Study on Cutter Torque Prediction based on Bayesian Optimisation Random Forests

Abstract

During shield tunnelling operations, cutterhead torque serves as a critical parameter reflecting ground conditions, equipment performance, and construction risks. Accurate torque prediction provides operators with technical guidance, enables early warning of potential anomalies, and reduces failure risks. Traditional machine learning methods have been applied in engineering prediction, but limitations remain in their predictive accuracy and parameter dependency. This paper constructs a Bayesian-optimised Random Forest (BO-RF) model based on actual engineering data to predict cutterhead torque, comparing it with conventional RF, BR, and SVR models. Unified evaluation metrics (MAE, RMSE, R²) are established to analyse the predictive performance of the four models. Results demonstrate that the RF-BO model significantly enhances prediction performance through automated hyperparameter optimisation, exhibiting superior fitting accuracy and robustness compared to other models. This study confirms that Bayesian optimisation effectively leverages the strengths of random forests, providing a feasible and efficient technical approach for high-precision cutterhead torque prediction and construction safety in shield tunnelling.