Abstract: To quantitatively evaluate the safety performance of engineering structures in probability, proposed is a nonlinear probability model identification method based on variational Bayesian (VB) theory, being used to identify structural nonlinear probability model subjected to seismic loads. Afterwards, the probability density evolution theory is further employed for reliability assessment of structures under dynamic loads. Different from the classical Bayesian theory, which usually adopts time-consuming random sampling methods to approximate the posterior distribution of model parameters, in this study, the Gaussian mixture model (GMM) is conducted to approximate the posterior probabilistic density function (PDF) of nonlinear parameters. To achieve the variational solution of nonlinear parameters, the instantaneous amplitude of the acceleration responses of structures subjected to external excitations are considered as measured data, and the stochastic gradient descent algorithm is further used to optimize the parameters of the GMM. The optimal parameter values of the predefined GMM are obtained by maximizing the Evidence lower bound (ELBO). Based upon the identified nonlinear probabilistic model of the structures and on first-passage failure criterion, the reliability of structures subjected to earthquake excitations can be evaluated by taking the safety threshold as constraint conditions. To validate the feasibility and effectiveness of the method proposed, a three-storey steel frame structure under earthquake excitations is conducted as a numerical simulation, and the reliability of the method is further verified by a scaled precast segmental column shake table structure. The numerical and experimental results indicate that the method proposed can effectively be used for nonlinear probabilistic model identification of the structures, and the identified results can further be conducted for structural reliability evaluation under strong external loads.