The observation series of concrete dam deformation is composed of the trend component, the periodic component and the stochastic component.In this paper, several stages are separated from the whole runtime according to the state of the dam.The trend component at each stage is extracted with the wavelet decomposition method, and modeled by the auto-regression method respectively. Then the characteristic polynomials of the auto-regression models are built. When solving these characteristic polynomials, the characteristic roots can be got. At last, to make a decision whether the dam is safe or not according to the trend of the minimum moduli of the roots. The application example shows that the conclusions of the method are consistent with the fact.