Abstract: To analyze the reliability of complex mechanical structures more efficiently， an innovative adaptive analysis method is proposed based on the Kriging model and the relative error estimation of failure probability. The range of the number of actual failure samples in the region where the signs of samples are uncertain is derived using the normal distribution that approximates to the Poisson binomial distribution. The lower limit of the range is redefined by introducing a scale factor to ensure that the number of actual failure samples can accurately fall within the range. Then， a more accurate estimation of relative error of failure probability is provided. The adaptive design of experiments is implemented by the learning function U. Two examples are employed to verify the accuracy， generality and efficiency of the presented relative error estimation of failure probability and the adaptive analysis method. The results indicate that the as-introduced approach can not only accurately estimate the relative error of failure probability， but also significantly decrease the calls to performance function.

Key words: objective, interval particle, intervalstructural reliability analysis, failure probability, relative error, Kriging model, adaptive design of experiments