A Research Of Motor Imagery EEG Signal Recognition Based On Neural Network

Motor imagery EEG signal recognition refers to inferring human movement inten-tion through analysis and recognition of EEG signals,so as to realize direct communication between humans and machines.At present,the electroencephalogram recognition tech-nology of motor imagery has been widely used in medical fields such as mobile assisted robots,upper limb assisted robots and stroke rehabilitation,as well as entertainment fields such as brain-controlled games and virtual reality.This thesis first studies related work at home and abroad in this field,from EEG sig-nal recognition based on traditional machine learning to the latest EEG signal recognition based on Convolutional Neural Network.Two shortcomings of recognition algorithms:The first drawback is that these methods use a single convolution scale in CNN for feature extraction and classification.However,the best convolution scale often differs from sub-ject to subject,or from time to time for the same subject,it leads to limited classification accuracy.The second shortcoming is that under the condition of limited training data,the neural network is prone to overfitting,which leads to a reduction in the accuracy of EEG signal classification in motor imagery.In order to solve the problem of convolutional neural networks with a single convo-lution scale,this thesis proposes a hybrid convolutional scale CNN structure.The multi-branch structure of the hybrid convolutional scale CNN can extract the motion-related information of EEG signals in different domains(time,frequency,and space)as much as possible,which improves the classification accuracy of motor imagery EEG signals.On the other hand,in order to solve the problem of limited training data,this thesis proposes a data enhancement algorithm for EEG signals,which further improves the classification accuracy of motor imagery EEG signals.At the same time,in order to meet the real-time and miniaturization requirements,a dedicated acceleration hardware is designed for the hybrid convolutional scale CNN algorithm and implemented by FPGA.This hardware can configure the network structure through instructions,with the layer as the basic unit,each instruction Corresponding to a layer of neural network,hardware resource overhead is reduced through hardware reuse,while real-time requirements are met.Then,experiments and analysis are performed on the proposed algorithm and hard-ware.Compared with several state-of-the-art methods,the proposed algorithm achieves an average classification accuracy of 91.6% and 87.6% on the classification of the two public data sets Data sets 2a and Data sets 2b,respectively,which outperforms other ex-isting methods.At the same time,the design of the motion imagery EEG signal recog-nition hardware can work at 100 MHz,and it only takes 10.224 ms to process a frame of motion imagery EEG signal,which meets the real-time requirements and achieves a small hardware overhead.Finally,the thesis summarizes the whole paper,analyzes the shortcomings of the proposed algorithm and hardware,and gives a prospect for the future work.