Towards autonomous and optimal excavation of shield machine: a deep reinforcement learning-based approach

Autonomous excavation operation is a major trend in the development of a new generation of intelligent tunnel boring machines (TBMs). However, existing technologies are limited to supervised machine learning and static optimization, which cannot outperform human operation and deal with ever changing geological conditions and the long-term performance measure. The aim of this study is to resolve the problem of dynamic optimization of the shield excavation performance, as well as to achieve autonomous optimal excavation. In this study, a novel autonomous optimal excavation approach that integrates deep reinforcement learning and optimal control is proposed for shield machines. Based on a first-principles analysis of the machine-ground interaction dynamics of the excavation process, a deep neural network model is developed using construction field data consisting of 1.1 million samples. The multi-system coupling mechanism is revealed by establishing an overall system model. Based on the overall system analysis, the autonomous optimal excavation problem is decomposed into a multi-objective dynamic optimization problem and an optimal control problem. Subsequently, a dimensionless multi-objective comprehensive excavation performance measure is proposed. A deep reinforcement learning method is used to solve for the optimal action sequence trajectory, and optimal closed-loop feedback controllers are designed to achieve accurate execution. The performance of the proposed approach is compared to that of human operation by using the construction field data. The simulation results show that the proposed approach not only has the potential to replace human operation but also can significantly improve the comprehensive excavation performance.