Abstract:Under the coupled effects of long-term traffic loads, temperature variations, and material degradation, deflection responses of existing bridges often exhibit significant non-stationary characteristics, posing challenges for traditional stationary extreme value models in assessing their long-term safety risks. Taking a bridge in Quanzhou as an engineering case, this study employs 60-minute high-frequency deflection monitoring data and introduces the non-stationary generalized extreme value distribution (NS-GEV) to conduct deflection extreme value analysis. Outliers in the monitoring data are removed using the interquartile range method, and continuous daily minimum sequences are constructed. The time-varying features of the sequence are identified through ADF, KPSS, and Mann-Kendall tests, leading to the development of a dual-time-varying M3-GEV model with both location and scale parameters. Model selection is optimized using AIC, BIC criteria, and likelihood ratio tests.Results indicate that deflection extremes follow a Weibull distribution with a physical upper bound. The calculated 100-year return period deflection extreme does not exceed historical maximum values, suggesting that the bridge retains overall safety margins. Span 01 exhibits the largest deformation and lowest reserve, requiring focused monitoring. These findings provide quantitative support for the full-life safety assessment of in-service bridges.