Brain–computer interface (BCI) technology based on motor imagery (MI) can establish the connection between the brain and the outside world; this has gradually become an important application of human–machine hybrid intelligence enhancement, especially for medical rehabilitation treatment. Because of the nonlinear, nonstationary, and low signal-to-noise ratio (SNR) characteristics of electroencephalogram (EEG) signals, it is a great challenge to accurately classify MI-EEG signals. Toward this end, in this study, both variational mode decomposition (VMD) and a deep belief network (DBN) were applied to MI classification based on the dataset of a previous BCI competition. Firstly, the EEG signal was decomposed by VMD to obtain the narrow-band component. Then the marginal spectrum, the instantaneous energy spectrum under the characteristic frequency band, and time-frequency joint features were extracted by Hilbert transform to achieve feature fusion. Finally, the DBN was used to reduce the dimensions of high-dimensional features and recognize MI patterns. The experimental results show that the joint VMD feature extraction and DBN feature classification method avoids information omission caused by the manual optimization of the period and the frequency band of artificial determination imagery. On the other hand, the method of using VMD and DBN to automatically extract the characteristics of the optimal period and optimal frequency band can effectively improve the recognition rate of MI.