Adaptive Time-frequency Analysis And Its Application On EEG Signal Processing

Brain science is hailed as the ultimate field for human exploration of themselves and the natural world.With the official launch of the ”China Brain Project”,a major scientific and technological innovation project of 2030,brain science research has entered a new stage of development in China.In brain science and Brain Computer Interface(BCI)research,EEG signal processing is the key underlying technology.However,EEG signals have the characteristics of low signal-tonoise ratio,non-linearity,non-stationarity,aliasing of useful information,and multiple channels,and the physiological information contained in EEG signals is difficult to capture.Traditional signal processing methods need to consider fixed analysis bases or strong statistical assumptions in advance.When dealing with these characteristics,problems such as mode mixing and loss of useful information will occur.In recent years,the development of adaptive time-frequency analysis theory has brought new ideas for EEG signal processing.Such methods break through the inherent defects of traditional analysis methods and effectively deal with the nonlinearity and non-stationarity of EEG signals.Based on the adaptive time-frequency analysis method,this paper focuses on the problems that arise in the application of EEG signal processing,and achieves the following results:· A normal single-channel EEG signal consists of a set of neural oscillatory rhythms in a broad frequency band,which contain physiological characteristics that reveal human brain activity.In motor imagery tasks,these physiological features can be used to classify different imagery tasks.Compared with other adaptive time-frequency analysis methods,the current adaptive time-frequency analysis method based on the narrowband signal model,variational mode decomposition(VMD),has the limitation that only cater for narrowband signals and cannot accurately separate broadband rhythms.Therefore,in this paper,combined with the short-time Fourier Transform(STFT),a time-frequency jointly-optimization method is proposed to extend VMD to the field of broadband analysis,which is called shorttime variational mode decomposition(STVMD).Then,by constructing time-frequency demodulation atoms,this paper generalizes the time-frequency jointly-optimization strategy in STVMD to all linear time-frequency transforms,and then proposes Wavelet Transform VMD(WTVMD).The time-frequency jointly-optimization method in STVMD or WTVMD inherits the advantages of VMD in signal analysis,such as high noise robustness,strong frequency resolution,accurate mode reconstruction,etc.The neural oscillatory components of the EEG signal are well separated by the time-frequency jointly-optimization method of STVMD.STVMD simultaneously provides the instantaneous frequencies,average frequencies and corresponding time-frequency coefficients of all components.These timefrequency features can be used as motor imagery features of single-channel EEG for classification of motor imagery tasks.· In the EEG motor imagery task,the common information between multiple channels must be considered.Using a single-channel method to process multi-channel EEG signals will cause problems such as loss of common information and time-frequency aliasing.Although the existing multivariate adaptive time-frequency analysis methods can overcome the above problems,the computation consumption is large when processing multi-channel data.Even using the faster multivariate empirical mode decomposition(MEMD),its computational consumption will increase significantly with the number of channels,which cannot meet the real-time requirements of motor imagery tasks.In addition,MEMD is not robust to noise,and problems such as mode mixing and incomplete dynamic information extraction still occur when processing EEG signals.Therefore,this paper proposes fast multivariate empirical mode decomposition(Fast MEMD(FMEMD))and noise-assisted FMEMD(NA-FMEMD)for time-frequency analysis of motor imagery EEG signals.In this paper,simulation experiments are designed to verify the advantages of FMEMD in terms of runnning efficiency and the stability of filter bank property.The multi-channel signal analysis of actual motor imagery shows that FMEMD and NA-FMEMD can effectively extract EEG rhythms related to specific tasks,and the running efficiency is significantly improved.Finally,a FMEMDbased EEG classification scheme is proposed by combining common spatial pattern(CSP),single-channel time-frequency features and classifiers.· In the offline analysis of abnormal multi-channel EEG signals,the EEG signals will show mutational features such as spikes.The multivariate adaptive time-frequency analysis method has the problem of less-concentrated or lack of time-frequency information when processing such EEG spike signal.To this end,this paper proposes a bivariate local mean decomposition(Bivariate Local Mean Decomposition(BLMD))and its multivariate extension(Mul-tivariate Local Mean Decomposition(MLMD)).This paper firstly verifies the advantages of LMD in preserving mutation time-frequency features,and then introduces LMD into the field of bivariate analysis by defining bivariate local mean and bivariate local amplitude.These two definitions can also be directly generalized to multivariate filed.In this paper,we demonstrate the mode alignment property of BLMD based on the BLMD decomposition theoretical model.BLMD(MLMD)inherits the advantages of LMD,can accurately extract the common oscillation mode,can effectively retain the time-frequency mutation information of EEG spike signals,and can also provide more concentrated time-frequency representations of other nonlinear and non-stationary signals such as float displacement signals.