Based on the different micro-Doppler modulation of three motion states of rotor target, i.e., hovering, rising and falling, a convolutional neural network (CNN) is utilized for motion states classification of rotor target in this paper. Firstly, to obtain the time-frequency spectrograms of target, the short-time Fourier transform (STFT) is applied to target echo signal after pulse compression, theoretical analysis and simulation experiments show that the maximum value of micro-Doppler frequency varies with different motion states. Secondly, we partition the echo data under three different radios of training data, i.e., 20%, 33% and 50%. Finally, the spectrogram data are fed into the proposed CNN architecture, and the cross validation is utilized to investigate the robustness of the proposed method. Experimental results show that with the continuous training iteration of the network, the data fitting ability and classification accuracy increases gradually and reaches 98.23% on average with the training data radio is 50%.