Abstract: A hydraulic turbine unit is a complicated nonlinear dynamic system and its vibration signals are usually non-stationary and nonlinear. Empirical mode decomposition (EMD) is a new method for analysis of time domain signals; it has good adaptability and is more suitable for non-stationary signals, but it has serious problems with end effect, mode-mixing stack, and other adverse effects. The ensemble EMD (EEMD) reduces mode-mixing stack to a certain extent, but it improves little on new mode mixing and suffers from spectrum losing and higher computational cost. In this study, we have developed a complex data-based EMD (CEMD) and applied it to fault diagnosis of a turbine water guide bearing, adding white noise as the imaginary part to construct complex signals. This method adjusts the extremum points of a signal by projecting the white noise in all the directions, and it can reduce mode-mixing stack problem through eliminating the effect of noise projection in calculation of the envelope barycenter. Its application to the signals measured at a hydropower station verifies its validity and reliability.