Abstract: In view of the shortcoming of traditional empirical mode decomposition (EMD), the method is improved by using signal frequency modulation and optimizing mode selection criterion. The relationship of decomposition ability versus selection criterion, iteration number, and frequency modulation was studied, and a detailed procedure of frequency modulation is given. A closely-spaced mode identification method was developed using window functions, signal frequency modulation, and improved selection criterion, and it was validated via simulations of vibration signals observed in-situ. The results show that the improved method can effectively identify the parameters of closely-spaced modes, thus laying a basis for EMD application to power plant analysis and identification of its vibration parameters.