Aiming at the problem that the common spatial pattern algorithm cannot obtain sufficient spatial distribution information when processing EEG signals with low channel counts and multi-modal imaginative actions, resulting in low classification accuracy, proposing a feature extraction method in time-frequency and spatial domains based on event-related spectral perturbations. Firstly, according to the characteristics of the independent functional mapping area of the limb imagination in the motor sensory cortex region, the event-related time-frequency feature information with significant differences under specific leads was extracted and fused with the spatial feature information of specific leads. Finally, the parameter-optimized support vector machine was used to identify different classes of limb-imagined actions. Comparison of the experimental results shows that the recognition performance of fused features in multi-modal imagined actions is significantly improved over single features. It can not only obtain more comprehensive EEG feature information, but also effectively reduce the demand for multi-channel number, and its average classification accuracy reaches 93.1%.