Reinforcement Learning for Integrated Structural Control and Health Monitoring

Abstract

Structural systems are vulnerable to dynamic loading and need special protection while facing extreme conditions. This study proposes an integrated structural control and health monitoring (ISCHM) system to enhance the safety and performance of building structures subjected to seismic loading. The system encompasses a semiactive controller based on reinforcement learning (RL) as well as a real-time damage identification (RTDI) system for structural health monitoring. The controller uses Deep Q-Networks (DQNs) to operate a semiactive control device and suppress the dynamic vibrations. The DQN controller was integrated with the RTDI system proposed in the preliminary phase of the project. The damage information provided by the RTDI was used to train the DQN controller and optimize the control policy in different damage conditions. The performance of the ISCHM system was evaluated through the numerical example of a building structure with variable viscous dampers installed between adjacent floors. OpenAI Gym and Keras were used to create a custom environment, define the DQN agent, simulate the interaction between the agent and environment, and train the agent while exploring the environment. The smart structure equipped with the ISCHM system was subjected to earthquake loading and the performance was compared with conventional semiactive control alternatives including skyhook and Lyapunov controllers. The results show the effectiveness of the proposed ISCHM system especially in the presence of damage. The ISCHM system can enhance the episode score by up to 58%.