In order to reflect the stochastic and time-varying characteristics of seepage failure for earth-rock dams more truthfully, the dynamic seepage failure probability of earth-rock dams is characterized by means of geotechnical dry-wet cycle tests and permeability coefficient inversion. The fusion of RBF neural network and LHS method is used to improve the random sampling method, and the shortcomings of the traditional Monte Carlo random sampling method with heavy calculation burden are overcome. The accuracy and efficiency of the improved method are proved by comparison. The improved method is applied to the dynamic analysis of the seepage failure probability of an earth-rock dam, and the conclusion shows that the risk of seepage failure of the earth-rock dam is high and it increases year by year, indicating the need to be reinforced.