Research On Multilinear Subspace Algorithm Based On EEG

Brain Computer Interface(BCI)technology is responsible for effectively solving the task of human-computer interaction in real-world scenarios,and it has become a research hotspot in many fields such as biomedicine,brain science,and intelligent information processing.Compared to other methods,the BCI system based on wearable EEG devices has been widely studied due to its non-invasive,high temporal resolution,acceptable spatial resolution and relatively acceptable economic cost.But how to extract effective ingredients from EEG signals has always been the direction of brain science research.Starting from the characteristics of the EEG signal,this article mainly carried out systematic research on the feature extraction algorithm,and especially focusing on the research of the space-time-frequency multidimensional unified feature extraction algorithm.This article is first inspired by discriminative spatial pattern(DSP),because DSP is a classic and effective feature extraction technology used for single EEG classification in motion tasks.However,it only uses the spatial information of the EEG signal.On the other hand,time information plays an important role in EEG sequences with high time resolution.Therefore,we propose a novel method called two-dimensional discriminative spatial pattern(2DDSP).This method can use the same criteria as the DSP to filter the spatio-temporal dimensions at the same time to extract the spatio-temporal information,instead of performing spatial filtering and time filtering separately like traditional spatio-temporal analysis ideas.The experimental results of single EEG classification show that the proposed 2DDSP method has better classification accuracy than DSP,and the dimension of the feature matrix generated by 2DDSP is smaller than that of DSP.Subsequently,we were inspired by tensor algebra.The original EEG signal is presented in matrix form,but it is also a special case of higher-order tensor form when the order is 2.On the other hand,we hope that the discriminant criteria based on DSP can be applied to other dimensions at the same time,including time domain and frequency domain.Therefore,this paper proposes multilinear discriminative spatial patterns(MDSP)to derive multiple lowdimensional discriminant subspaces related to low-frequency motion-related cortical potentials.MDSP hopes that these low-dimensional discriminant subspaces can work together to obtain better results.The experimental results of the two EEG data set for the finger movement task prove the effectiveness of the proposed MDSP method.Finally,in order to make full use of the tensor theory to explore the neural activity mechanism of the brain,we were inspired by the brain network theory.Since brain networks are commonly used to construct brain maps to explore human neurocognitive activities,we can incorporate tensor decomposition techniques into the brain network analysis process.This article first constructs a functional connection matrix related to frequency and time window through continuous wavelet transform(CWT)and dynamic functional connection(d FC)technology.Then we reorder these connection matrices to obtain a tensor object containing connection information and time-frequency information.Then we use Tucker decomposition technology to decompose and analyze the dimensions of the tensor object to extract the main components.The experimental results of EEG data based on intention understanding prove that the proposed joint tensor brain network analysis framework is effective and feasible.