In the classification and recognition of limb motion imagery features in Electroencephalogram, the problem of low action recognition accuracy exists when fusing different domain feature extractions. Based on the complex and different domain relationships of limb motion imagery brain electrical characteristics in multi-channel acquisition, this paper constructs an EEG-Symmetric Positive Definite Netmotion feature classification model for limb action recognition, which effectively extracts and fuses different domain features to realize the extraction of limb motion features based on brain electrical signals. The main research contents include: 1) A time domain feature extraction module is designed, which incorporates the channel-wise features weighted into the time domain features, and uses multi-scale separable convolution to fully extract the time domain features, improving the recognition accuracy of limb actions; 2) The extracted time domain features are mapped to a high-dimensional manifold that can more effectively describe their distribution, and a manifold mapping module is designed to solve the problem of ineffective fusion of time domain features and spatial domain features; 3) To make up for the lack of frequency domain information in the time domain, a frequency domain feature extraction module is designed to enrich the motion features such as gestures. By integrating the above multiple modules, the EEG-SPDNet classification model is constructed, and the effective recognition of limb motion actions is realized based on motor imagery brain electrical information. Experimental results show that on the BCI Competition IV 2a motion imagery dataset for recognizing four types of limb motion, the accuracy of action recognition based on the constructed classification model reaches 0.85, and the Kappa coefficient reaches 0.80, with high precision.