Research On Portable Motion Imagination EEG Signal Decoder Based On Convolutional Neural Networks And Microcontrollers

Brain-computer interface technology detects brain electrical signals and decodes corresponding movement instructions to control assistive devices for the human body.Deep learning technology has been widely used for decoding motor imaginary(MI)electroencephalogram(EEG)signals,with high classification accuracy.However,due to the large computational requirements,deep learning decoders for EEG signals usually need to be implemented on bulky and heavy computing devices,which is not conducive to practical applications.At the same time,maintaining the accuracy of brain electrical signal classification is essential when making brain-computer interfaces lightweight.Currently,the application of deep learning technology in portable brain-computer interface(BCI)devices has not been widely explored.This study focuses on portable brain-computer interface technology and proposes and implements a high-precision MI-EEG signal decoder based on convolutional neural networks that can be deployed on microcontrollers.The main research work of this article is as follows:(1)MI-EEG signals are mainly concentrated in the low-frequency range.In this paper,the GigaDB MI dataset is filtered through an 8-30 Hz bandpass filter,downsampled,subjected to independent component analysis(ICA),noise removal,and finally normalized to improve the neural network training speed and accuracy.(2)To address the problem of many parameters and complex network structure in deep learning models,this paper proposes a compact convolutional neural network(CompactConvNet)for MI-EEG signal classification.After preprocessing the data,the motion-related channel EEG signals are used as input to train and test the CompactConvNet model,including channels FC3-FC4,C1-C2,CP1-CP2,and C3-C4.Finally,by comparing with ShallowConvNet,DeepConvNet,and EEG-Inception networks,the average classification accuracy of the proposed model for MI-EEG signals is higher than other models.(3)To address the challenge of applying deep learning algorithms in portable BCIs,this paper implements an offline portable BCI device for independently decoding left/right handopening and closing motor imagination.The deep learning model obtained in task(2)is deployed on the ARM microcontroller STM32F746 and uses the GigaDB MI dataset for MIEEG signal classification recognition to verify the proposed deep learning decoder.When using EEG signals from four channels(FC3,FC4,C1,C2)as input,the overall average accuracy reaches 93.63%.By quantizing the neural network model parameters,the average classification time when using EEG signals from two channels(FC3,FC4)as input is 1.12 s,significantly improving the classification speed of neural networks deployed on portable devices.This portable deep learning decoder has low computational costs,is easy to use in practical applications,and can be used to develop AI wearable BCI devices.