Decoding Of Motor Imagery EEG Signals Based On Multi-domain Feature Selection

With the development of science technology and intelligent devices,brain-computer interface(BCI)technology has gradually entered people’s vision.BCI technology based on motion imagination is one of the highly promising directions and has achieved good results in medical,entertainment,commercial and military fields.BCI technology is developing rapidly,but there is still much room for improvement in some key issues of electroencephalography(EEG)information extraction,including decoding rate,interpretability,and availability of online applications.The decoding rate is a key research direction for EEG signal extraction.In the decoding process of motion-based imagery EEG,signal analysis and conversion algorithms are the key components to decode neural signals from the brain into a machine capable of recognition and state representation.Improving the decoding performance and optimising the control performance of BCI systems by improving the algorithms for signals analysis and conversion with constant training intensity is the focus and difficulty of research in this field.The spatial weights of Common Spatial Pattern(CSP)can be effectively applied to the decoding study of EEG signals in motion imagination-based BCI systems.In the Discriminative Filter Bank CSP(DFBCSP),an optimized method for CSP,it is necessary to find the frequency bands that are relevant to the motion imagery task.To solve this problem,the selection of sub-bands was optimized by adding channel data.Meanwhile,for the classical sample covariance estimation used in the DFBCSP algorithm,which is highly influenced by outliers and highly non-robust,the Minimum Covariance Determinant(MCD)is used to obtain robust estimates of the covariance matrix.The paper proposes a two-channel MCD-DFBCSP based algorithm for feature extraction from both frequency and space domain perspectives by filtering frequency bands and reducing the effect of outliers on feature extraction,and validates the classification by a linear discriminant analysis method.The experimental results show that the proposed method in the paper can enhance the selection of effective sub-bands in the channels,reduce the influence of outliers,effectively extract frequency and space domain features,and finally achieve a classification accuracy of 83.1% on a dichotomous EEG dataset.The paper addresses the problem of EEG signal feature extraction and classification,and further applies deep learning methods to explore the decoding study of motion imagery EEG signals.The paper takes the approach of combining Bi-directional Long Short-Term Memory(Bi LSTM)and Convolutional Neural Networks(CNN).Among them,Bi LSTM can realize the computation of neuron output at multiple time nodes on the time domain through time expansion,which is suitable for describing continuous state output;CNN network has excellent spatial expansion,and can efficiently extract spatial information on the null domain through neuron and spatial convolution operations.The network model that combines the two networks can acquire feature information in both time and space domains.In addition,the decoding accuracy of the network model is further improved by adding an Attention mechanism to the network model to enhance the weight of important features.The network model was validated on a triple-classified motion imagery EEG dataset and achieved a relatively good classification accuracy of 92%,demonstrating that the model can effectively decode motion imagery EEG signals in both the time domain as well as the frequency domain perspectives.The effectiveness of model decoding was also further verified by recall,precision,F1 score and loss function.In order to further explore the features of the motion imagery EEG signal and reduce the negative impact of too many layers of the network,the paper expands on the basis of CNN and Bi LSTM,introduces Residual Network(Res Net),and proposes the CLRNet network model.Specifically,the linear subspace with the most information in the EEG signal is identified through CNN spatial convolution operation;the time dynamics is captured through Bi LSTM to obtain the time domain information;the output of the multi-layer network of Bi LSTM is connected across layers using Res Net to enhance the data processing capability of the model,and finally the extracted features are classified through Softmax function to realize the decoding of EEG signal.The model achieved an average accuracy of 89% with quadruple classification validation on nine EEG datasets.The experimental results demonstrate that the model has strong network generalization ability and has better performance for decoding complex MI-EEG signals.