Intelligent adaptive beam string structures: A deep reinforcement learning framework for active control

To address the limitations of passive load-bearing in traditional beam string structures (BSS) when facing complex variable loads, this study proposes an Intelligent Adaptive Beam String Structure (IABSS). Functioning intrinsically as an "Autonomous Policy-Evolving Adaptive Structural System," this framework seamlessly couples deep reinforcement learning (DRL) with nonlinear finite-element analysis, endowing the structure with an "intelligent brain" for autonomous decision-making. To avoid unsafe trial-and-error exploration on the physical structure, the "trial-and-error evolution" of the control policy is strictly confined within the offline virtual training phase. During formal service evaluation, the IABSS relies exclusively on frozen terminal networks for rapid, deterministic forward inference. Using an integrated ANSYS-Gym framework, the research follows a progressive path: a single-strut DQN case first validates the feasibility of data-driven discrete actuation. Subsequently, under a rigorously matched three-strut evaluation protocol, two DRL-based IABSS controller formulations are evaluated: a DQN-based controller and a TD3-based controller. Across formal evaluations comprising unseen static loads, fixed load histories, and out-of-distribution (OOD) scenarios, the continuous TD3 formulation demonstrated generalization. Relying on a reduced macroscopic observation state, the TD3 policy achieved a peak displacement reduction of 89.21 ± 1.47%. It outperformed the discrete DQN baseline while maintaining zero violations of engineering physical constraints, including stress, cable force, actuator stroke, and Euler buckling. These results validate the efficacy of continuous multi-actuator regulation via a DRL-based control framework, providing an algorithmic basis for intelligent adaptive structural control.