Abstract: Fatigue S-N curves are the most fundamental material input to design the lightweight structures and assess the service safety of high-speed railway vehicles. How to reduce the dispersion of fatigue life data and to improve the prediction precision of fatigue performance are frequently the principal attention on long life service of railway vehicle structures. Classical sample polymerization principle (SPP) can effectively solve the small size sample, while the accuracy of fatigue life is still of primary concern. An improved SPP (iSPP) to build the fatigue probabilistic S-N (P-S-N) curve is therefore proposed to pursue an optimized search parameter related with the standard deviation (SD), thus obtaining the optimal life SD under different stress levels. Results show the slope and intercept of newly-built P-S-N curves from (X-x-x-x) type life data have the relative error less than 3%, and predicted fatigue life is approximately 5% of the group method (TGM) estimation in contrast with the TGM. Moreover, the predicted fatigue life is only 0.1% of traditional fitted method for (x-x-x-x) type life data. For the welded joints with defect induced dispersion, iSPP can achieve more conservative and reliable life predictions.It shows the iSPP and optimal parameter search method can not only ensure the accuracy of small sample but also acquire a more conservative P-S-N curve into engineering applications.