Beyond statistical fitting: improving reliability analysis through physically consistent bridge traffic load model

The estimation of lifetime bridge traffic loading is important for ensuring safety.This is commonly achieved by fitting extreme value distributions, where the distribution type dictates the upper tail behavior: unbounded (Type I), heavy-tailed (Type II), or upper-bounded (Type III).The correct modeling of the upper tail is crucial, as this is where loading more likely exceeds resistance, leading to collapse.However, many studies overlook the influence of the underlying physical processes on the upper tail behavior.This research investigates how the physical bounds of traffic loading affect the resulting extreme distribution type and its applications to reliability analysis.First, we examine the effects of naturally upper-bounded and truncated systems on the resulting extreme value distribution.The resulting block maxima probability density converges to the upper-bounded extreme value distribution, even when the original distribution is heavy-tailed.Second, we propose Bayesian modeling to integrate knowledge about traffic, such as axle configurations and mass limits, into the statistical model.Hamiltonian Monte Carlo No-U-Turn Sampler (HMC NUTS) is used to estimate the upper bound of traffic.Finally, we compare the physically informed (upper-bounded) model with the uninformed (unbounded) model in bridge reliability analysis.We show that the physically informed model reduces the probability of failure over the common range of the parameter space.These findings highlight the necessity of selecting appropriate extreme value distribution types based on physical processes to enhance safety and reliability analysis.