The rotation of rotor blades on small unmanned aerial vehicles (UAVs) introduces micro-Doppler (m-D) signatures to discriminate UAVs from other aircrafts or birds. Compared with the Doppler signal induced by the translation of the platform, the m-D signal, however, is rather weak due to the small radar cross section (RCS) of rotor blades. Moreover, the m-D signal always consists of multiple frequency components, most of which are induced by the vibration of the platform. This paper focuses on the extraction of the rotation signal via the empirical mode decomposition (EMD) algorithm, which would decompose an arbitrary signal into a series of intrinsic mode functions (IMFs). The constitution of IMF is investigated by using the partial differential equations (PDEs) to analyze the distribution of the extrema. For separating the rotation signal, the essence is to unify the frequency of extrema in the corresponding IMF exhibiting the rotation features. First, the fitting Doppler signal is constructed from the data with the Doppler property in the IMF. Second, a criterion is proposed to establish the masking signal, after which the rotation components are extracted from the residual m-D signal via another EMD. Simulations are carried out to demonstrate the effectiveness of the proposed method in solving the mode-mixing problem. The results derived from the measured data validate the feasibility of the proposed approach in the identification of small UAVs.