Progressive Damage Modeling in Steel Connections Using Deep Reinforcement Learning

Progressive damage in steel connections under cyclic loading poses a critical challenge to the safety and resilience of steel structures. Accurate modeling of damage evolution is essential for informed decision-making in performance-based design and structural health monitoring (SHM). Traditional finite element (FE) approaches, while detailed, are computationally intensive and lack adaptability for real-time applications. In response to this limitation, the present study explores the application of deep reinforcement learning (DRL) for progressive damage modeling in steel connections. A hybrid framework was developed by integrating high-fidelity FE simulation data with DRL algorithms, enabling sequential prediction of damage progression. Steel connection types, including moment-resisting beam-to-column joints and end-plate bolted connections, were modeled in ABAQUS under various cyclic loading protocols. State-space representations were carefully designed to capture key indicators of damage evolution, such as cumulative plastic strain, energy dissipation, stiffness degradation, and damage indices. DRL models including DQN, Double DQN, and PPO were trained and evaluated. The PPO model achieved superior performance, accurately predicting temporal damage trends and exhibiting strong generalization across unseen test cases. Additionally, the trained PPO model demonstrated computational efficiency suitable for real-time SHM integration. The findings suggest that DRL offers a promising complementary approach to conventional FE-based damage modeling. Future work will focus on expanding the dataset, exploring transfer learning, and validating the DRL framework in live monitoring scenarios. This research contributes to the advancement of intelligent structural systems and opens new pathways for adaptive and data-driven engineering solutions.

کلیدواژه ها:

Damage Modeling ، Steel Connections ، Deep reinforcement learning ، Structural health monitoring