Synchronous control of multiple hydraulic cylinders in aerial building machine using improved deep reinforcement learning

Abstract

Aerial building machine (ABM) represents an advanced integrated construction system designed for high-rise structures, where the jacking phase constitutes a critical safety determinant. Current research on multi-cylinder synchronous control during ABM jacking operations remains scarce. To address this gap, this study proposes a Lyapunov-constrained twin delayed deep deterministic Policy Gradient (TD3) framework integrated with hindsight experience replay (HER). A physics-based multi-cylinder interaction environment is established to facilitate agent training. Deep reinforcement learning is utilized to train the controller for adaptive synchronous control of the multi-cylinder system in ABM. Validation through a case study of a Chinese ABM project demonstrates the following outcomes: (1) Synchronization error is reduced to 0.46 mm, contrasting sharply with 30 mm observed under uncontrolled conditions. (2) Structural stress decreases by 29.28 % compared to conventional control methods. (3) The Lyapunov constraint theoretically ensures control stability of the algorithm, and the HER strategy facilitates faster convergence of the model. These results underscore the robustness and generalizability of the reinforcement learning controller in uncertain operational scenarios, highlighting its potential for hydraulic system applications and contributions to control theory advancement.