Robust post-earthquake structural safety evaluation from limited acceleration measurements using sparse Bayesian broad learning with dimensionally reduced damage-sensitive features

After an earthquake, the safety evaluation of buildings is crucial for functionality restoration. However, inevitable uncertainties from multiple sources (e.g., sparse instrumentation, poor data quality and modeling errors) necessitate more robust approaches to ensure reliable assessments accounting for uncertainties and minimizing the risk of misclassification (i.e., incorrect safety tagging). This paper proposes a novel framework for robust post-earthquake safety evaluation of sparsely instrumented buildings. Accelerations from only the top and ground floors are used to extract damage-sensitive features (DSFs) as the input of the sparse Bayesian broad learning (SBBL) model to probabilistically predict the peak inter-story drift ratio (PIDR), a key engineering demand parameter. The posterior mean and variance provided by the SBBL are used to quantify the prediction uncertainty. Based on the predicted PIDR distribution, three safety tags (Safe, Restricted Use and Unsafe) are assigned to the instrumented building according to the exceeding probability of the thresholds defined in engineering guidelines or codes. The methodological contributions are threefold. First, a hierarchical SBBL architecture combining robust sparse Bayesian learning algorithm ensures accurate and robust learning of nonlinear DSF-PIDR mappings. It enforces balance between prediction accuracy and parsimony to enhance generalization and robustness under uncertainty. Second, an adaptive input feature selection procedure is incorporated to automatically identify an informative and compact subset of DSFs, enhancing efficiency and robustness. Finally, unlike conventional point-estimate-based tagging, the framework incorporates the predicted posterior to support uncertainty-aware robust safety evaluations, thereby controlling the risk of unconservative assessments. Numerical and full-scale shake table test case studies validate the framework. Results confirm accurate uncertainty quantification, reliable safety evaluations, and robustness against measurement noise and modeling errors. This framework offers a practical solution for rapid post-earthquake safety evaluation of sparsely instrumented buildings.