| Motor imagery(MI)is a classic brain-computer interface(BCI)paradigm.Currently,BCIbased rehabilitation mainly focuses on the identification of motor imagery.MI-based BCI systems can help restore motor function in patients with disabilities,especially in neuromuscular disorders,and are considered a promising cognitive tool for the rehabilitation of motor disorders.However,EEG signals are characterized by nonlinearity,instability,and weak signal-to-noise ratio,which makes decoding EEG signals a challenging task.The current motor imagery brain-computer interface technology suffers from low classification accuracy and long training time in practical use,which limits its application in daily life.To solve the above problems,this paper designs two classification models based on convolutional neural networks.The first classification model is proposed to improve the classification accuracy,and the second classification model is proposed to further improve the classification accuracy and shorten the training time.First,this paper proposes an input method based on wavelet transform,which converts multichannel EEG signals into two-dimensional time-frequency images using wavelet transform to obtain its comprehensive information,including time-frequency features and relative positions of electrodes.This method combines the time-frequency images of C3 and C4 channels to extract the features of motor imagery EEG signals.The multi-scale analysis of wavelet transform avoids the problem of window size selection in the Short-time Fourier Transform(STFT)method.Then,a Multi-width convolutional network(MWCNN)is built as a classifier.Using the dataset III of BCI Competition II to test,the experimental results show that the accuracy of the algorithm is as high as 96.43%.Compared with other algorithms in the open dataset BCI Competition IV dataset 2b,the algorithm can improve the classification accuracy of motor imagery and has better classification performance.Secondly,although the classification accuracy of the algorithm combining wavelet transform and neural network is good,the training time still needs to be shortened.Manifold learning algorithms have been successfully applied to data mining,pattern recognition,and other fields since they were proposed,but few manifold algorithms have been effectively applied to BCI systems,especially motor imagery.Given the limited motor imagery dataset,neural networks with complex inputs are prone to overfitting.However,the feature extraction ability of the simple neural network is limited.Therefore,this paper combines the local linearity embedding of the unsupervised learning manifold algorithm with the one-dimensional Simplified Convolutional Neural Network(1DSCNN).First,the preprocessed data is dimensionally reduced using Locally linear embedding(LLE),and then input into 1DSCNN for classification.Using the dataset III of BCI competition II to test,the experimental results show that the accuracy of the algorithm is as high as 98.21%.In the open dataset BCI Competition IV dataset 2b,compared with other algorithms,the algorithm can improve the recognition rate of motor imagery by 6-13% on average,with better classification performance and shorter training time. |
