Aiming at the problem that a single sensor signal can only reflect the local characteristics of a structure and is susceptible to interference, a data level fusion method based on correlation variance contribution rate is proposed. Data fusion can be realized through allocating the fusion coefficients by calculating the variance contribution rate of the signals from multi-sensors. The data fusion results of the observed prototype signal of Jinping I high arch dam and the measured signal of a self-made RCC dam hydroelastic model show that the proposed method has good noise resistance, and it can mine high-frequency weak information and dense frequencies in vibration signals. Compared with the method based on correlation function, this method make the fused signal can fully reflect the overall vibration characteristics of the dam body, and it is suitable for modal parameter identification of the vibration signals from high arch dams and roller compacted concrete dams.