A motor imagery electroencephalogram signal classification method

Abstract

The application provides a motor imagery electroencephalogram signal classification method, steps of which comprise: preprocessing original electroencephalogram signal data, inputting the preprocessed electroencephalogram signal data into an improved residual network model Resnet-18 for training, finally mapping the feature vector output by Resnet-18 to a category space and performing classification; the residual structure of the improved residual network model Resnet-18 is constructed by an improved deformable convolution, and the final processing layer of Resnet-18 adopts an adaptive spatial attention mechanism; the improved deformable convolution does not add an offset on the center sampling point of a convolution kernel, and adds an offset on the surrounding sampling points of the convolution kernel; the adaptive spatial attention mechanism automatically adjusts the weight of a feature based on the spatial distribution of an electroencephalogram signal feature map. The method can extract more distinctive features, enhance the generalization ability of the model, and improve the accuracy and performance of motor imagery electroencephalogram signal classification.

Images