自己組織化学習による逐次ベイズ更新に基づく履歴変数の区間推定

抄録

Accurate identification of hysteretic restoring‑force parameters in civil structures usually requires displacement or strain measurements that are rarely available in field monitoring. This study proposes a sequential Bayesian updating framework that estimates confidence intervals for seven hysteresis parameters and four algorithmic hyperparameters (11 variables in total) solely from absolute acceleration records. Initial particle populations are generated by Evolutionary Computation, and the posterior distribution is progressively refined with a Sequential Monte Carlo sampler in which a kernel‑density‑estimated likelihood accommodates multi‑modal residuals. The approach eliminates manual tuning of algorithmic hyper‑parameters by learning the KDE (Kernel Density Estimation) bandwidth and PSO (Particle Swarm Optimization) coefficients simultaneously with the physical parameters. Numerical verification using a single‑degree‑of‑freedom system with a bilinear skeleton curve lacking point symmetry about the origin demonstrates that, after assimilating 100 % of the simulated earthquake data, the posterior standard deviations shrink by more than 70 % while the 95 % confidence intervals capture the true values of all variables. Throughout the three data‑assimilation stages, the effective sample size remains above 90 %, indicating negligible particle degeneracy.