Damage identification and level evolution probabilistic risk assessment of corroded reinforced concrete beams using multi-domain dynamic and static data with deep learning

Abstract

The dynamic response of post-fire corroded concrete structures under time-varying loads continues to evolve, resulting in the progressive development of damage levels. However, existing standards often fail to fully account for this entire process. To investigate the evolution of damage levels in structures, this study proposes a framework that integrates finite element models (FEMs) with deep learning for damage identification and probabilistic risk assessment. The method first utilizes multi-domain experimental data, including dynamic and static responses, to correct the physical parameters of the initial FEM via residual network. Based on the updated high-precision model, the static reduction coefficients can be effectively predicted with an error controlled within 5%. Furthermore, by combining the static reduction coefficients with multi-domain dynamic information, the multi-granularity spatial-temporal attention learning evaluation branch is employed to derive the damage contribution rates and the comprehensive index, multi-variable feature index (MVFI). Using the MVFI, the probability density fitting analysis based on R² is conducted to determine the optimal distribution models for different damage levels. Finally, by integrating the prior-based minimum expected loss criterion with posterior probabilities, a comprehensive assessment of damage level evolution is achieved. Compared with conventional static-load tests that provide only local, static information, the proposed multi-domain probabilistic estimation method can effectively evaluate the damage evolution throughout the service life of the structure and offers the potential for local deployment.