Physics-data driven multi-objective optimization for parallel control of TBM attitude

Abstract

To more accurately control the attitude of the tunnel boring machine (TBM), this study proposes a physics-data driven multi-objective optimization (MOO) method. The proposed method combines the dynamics theory of the shield propulsion hydraulic system with deep neural networks (DNN) to generate a physics-informed deep learning (PIDL) model that is capable of accurately estimating oil cylinder strokes. Furthermore, a simulation model integrating the PIDL and the non-dominated sorting genetic algorithm III (NSGA-III) is established to perform optimization of shield attitude deviation. A field test of synchronous excavation and segment assembly TBM (S-TBM) is used as a case study to confirm the proposed method’s reliability. The results indicate that: (1) The developed PIDL model accurately predicts oil cylinder strokes under different geological conditions with R² values of 0.99. (2) For all strata, the proposed shield attitude control framework achieves an average overall improvement rate of 19.57% while considering regulation time, overshoot, and accumulative error simultaneously. (3) The proposed PIDL stands out with an advantage of 0.40 higher R² mean value than that of existing methods. (4) Compared to other popular MOO algorithms, the NSGA-III employed in this study generates Pareto fronts with the highest hypervolume mean value of 7.25, demonstrating better convergence and diversity. The novelty of this study lies in proposing an optimization framework with the integration of PIDL, NSGA-III, and virtual model to realize effective control of shield attitude.