A data-driven structural health monitoring framework for damage identification of composite beams under fire-induced extreme events

Accurate post-fire assessment requires a direct connection between measured structural response, finite element model fidelity, and quantitative damage indicators. This study develops a data-driven structural health monitoring framework that integrates Bi-LSTM: bidirectional long short-term memory (Bi-LSTM) regression with finite element model updating and Bi-LSTM–based damage identification for composite beams. Acceleration time histories are used to infer five continuous updating parameters, E , D s , K 1 , K 2 , and K 3 from a database of 48,000 simulated samples. The updated finite element model is then used together with fire exposure time and reduction factors for post-fire damage assessment. The maximum parameter error in the numerical model updating tests is 7.4%, and the maximum fire exposure time error in the independent numerical damage cases is 5.8%. The six fire-test specimens are excluded from neural network training and hyperparameter selection. Across these specimens, the absolute relative errors are 9.91 ± 3.05% for fire exposure time, 5.05 ± 1.23% for the stiffness reduction factor β t , and 3.24 ± 1.26% for the bearing capacity reduction factor K t (mean ± sample standard deviation). The updated finite element model also reproduces the measured temperature, deflection, and fundamental frequency attenuation trends. The results support a simulation-trained sequence modeling framework with fixed preprocessing and experimental finite element verification for post-fire assessment of composite beams.