CN112364812B - aVEPs electroencephalogram identification method based on TRCA-WPTD - Google Patents
aVEPs electroencephalogram identification method based on TRCA-WPTD Download PDFInfo
- Publication number
- CN112364812B CN112364812B CN202011343524.2A CN202011343524A CN112364812B CN 112364812 B CN112364812 B CN 112364812B CN 202011343524 A CN202011343524 A CN 202011343524A CN 112364812 B CN112364812 B CN 112364812B
- Authority
- CN
- China
- Prior art keywords
- matrix
- data
- aveps
- electroencephalogram
- wptd
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
- G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Signal Processing (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention discloses a TRCA-WPTD-based aVEPs electroencephalogram identification method, which comprises the following stages: a training stage: filtering the training data respectively through WPTD algorithms with different threshold setting modes; respectively calculating SNR and RMSE parameters of the filtered training data, and selecting the data with the highest SNR value and the lowest RMSE value for data combing; calculating a generalized right eigenvector matrix W; reducing the dimension of the data to obtain a template matrix Temp; and (3) a testing stage: and (4) screening the test data, and calculating a correlation coefficient matrix r with the generalized right eigenvector matrix W and the template matrix Temp. And a decoding stage: and decoding the correlation coefficient matrix r obtained in the test stage, generating a predictive coding matrix, comparing the predictive coding matrix with the actual coding matrix, and outputting an identification result. According to the invention, the WPTD algorithm is used for filtering, the signal-to-noise ratio of the aVEPs electroencephalogram is improved, the root mean square error is reduced, the TRCA algorithm is used for carrying out template matching for identification, and the character identification accuracy and identification speed based on the aVEPs electroencephalogram are improved.
Description
Technical Field
The invention relates to the field of electroencephalogram identification based on Event-Related Potential (ERP), in particular to an aVEPs electroencephalogram identification method based on TRCA-WPTD, which is a method for identifying asymmetric Visual Evoked Potentials (aVEPs) based on Task-Related Component Analysis (TRCA) in combination with Wavelet Packet Threshold Denoising (WTD) algorithm.
Background
With the continuous progress of computer-based biological communication technology, brain-computer interface technology has been developed as an important component thereof and is gradually applied clinically. In the research of brain-computer interface technology, the development of ERP-based electroencephalogram identification technology is stable and the application is wide, especially the ERP electroencephalogram based on P300 potential is most popular, such as: a character spelling system based on the P300 brain electricity, a target detection system based on the P300 combined with Rapid Serial Visual Presentation (RSVP), and the like. At present, a matrix flicker paradigm and an RSVP paradigm are commonly used for inducing P300 electroencephalograms, but both of the two paradigms require a testee to watch a flicker screen, and the generated stimulation appears in the central visual field of the testee, so that visual fatigue is easily caused, so that a new method for inducing the aVEPs electroencephalograms through coding by code division multiple access and space division multiple access by utilizing the space-to-side dominance characteristic of the human brain to the stimulation response is provided for character recognition. The existing method for processing the electroencephalograms of aVEPs adopts traditional typical Correlation Analysis (CCA) to calculate a Correlation coefficient and Linear Discriminant Analysis (LDA) to output a classification result, the average Recognition Rate (RR) of the electroencephalograms of aVEPs is not high, the Information Transfer Rate (ITR) of the electroencephalograms is slow, and the RR and the ITR need to be further improved when the aVEPs are actually applied to life.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide an aVEPs electroencephalogram identification method based on TRCA-WPTD, wherein the filtering is carried out through a WPTD algorithm, the Signal-to-Noise Ratio (SNR) of the aVEPs electroencephalogram is improved, the Root Mean Square Error (RMSE) is reduced, the identification is carried out through template matching through the TRCA algorithm, and the character identification accuracy and identification speed based on the aVEPs electroencephalogram are improved.
In order to achieve the purpose, the technical solution of the invention is as follows:
a TRCA-WPTD-based aVEPs electroencephalogram identification method comprises the following 3 stages:
(1) Training phase
Importing preprocessed aVEPs electroencephalogram data TrainData1 for training, and respectively filtering by using 8 WPTD algorithms with different threshold setting modes; generating 8 kinds of filtered TrainData2, calculating SNR and RMSE parameters for each TrainData2, selecting the TrainData2 with the highest SNR value and the lowest RMSE value for data combing, and generating TrainData3; calculating a generalized right eigenvector matrix W through a TRCA algorithm; and reducing the dimension of the TrainData3 based on the matrix W to obtain a template matrix Temp.
(2) Testing phase
Importing preprocessed aVEPs electroencephalogram data TestData1 for testing, and respectively filtering by using 8 WPTD algorithms with different threshold setting modes to generate 8 filtered TestData2; calculating SNR and RMSE parameters for each TestData2, selecting the TestData2 with the highest SNR value and the lowest RMSE value to screen data, and generating TestData3; calculating a correlation coefficient matrix r by the TestData3 and the generalized right eigenvector matrix W and the template matrix Temp generated in the training stage;
(3) Decoding stage
And decoding the correlation coefficient matrix r to generate a predictive coding matrix, comparing the predictive coding matrix with an actual coding matrix set in an experimental paradigm, and outputting an identification result when the predictive coding matrix is completely matched with the actual coding matrix.
Preferably, the WPTD algorithm of 8 different threshold setting manners in the stage (1) is specifically as follows:
decomposing aVEPs electroencephalogram data into various scales by wavelet packet decomposition, and generating 8 denoising methods by adopting a combination of a threshold value and a threshold value function, namely: the method comprises the steps of respectively removing wavelet coefficients belonging to noise in each scale by using each method, retaining and enhancing the wavelet coefficients belonging to aVEPs electroencephalogram data, and finally reconstructing the processed wavelet coefficients by using inverse wavelet transform to obtain noise-reduced aVEPs electroencephalogram data.
Preferably, the data combing process in the stage (1) is as follows: starting from the first character data, finding out the sampling point positions with all label categories of '0', and backward intercepting the set data segment length by taking the positions as initial positions, and sequentially storing all the data segments in an array named as Lefttemp according to the appearance order of the labels; finding out all data with the label type of 1 and storing the data in an array named as Righttemp; repeating the circulation until the data of all the characters are completely read; and merging the Data in the array of the Lefttemp and the array of the Righttemp, and sequentially putting the merged Data into an array named Data _ train for storage.
Preferably, the WPTD algorithm of 8 different threshold setting manners in the stage (1) is specifically as follows:
the WPTD algorithm process is that wavelet packet decomposition is utilized to decompose aVEPs electroencephalogram data into various scales, a threshold value is set, then wavelet coefficients belonging to noise in each scale are removed, the wavelet coefficients belonging to the aVEPs electroencephalogram data are reserved and enhanced, and finally the wavelet coefficients after processing are reconstructed by wavelet inverse transformation to obtain noise-reduced aVEPs electroencephalogram.
Matrix d for aVEPs electroencephalogram data TrainData1 designed for training l Is represented by d l There are two dimensions: the method comprises the following steps of (1) channel number and sampling point number, and also label category information which does not participate in filtering calculation, wherein the specific data processing flow is as follows:
(1-1) wavelet packet decomposition TrainData1
To d is paired l And (3) according to the selected wavelet base expansion, decomposing the nth frequency band of the j +1 layer into the 2 nth high-pass filtered frequency band and the 2n +1 th low-pass filtered frequency band of the j layer during decomposition, wherein the iterative calculation formula of the decomposition is as follows:
in the formula (I), the compound is shown in the specification,
is d l At level j the 2 nth high pass filtered sub-band, device for combining or screening>
Is d l 2n +1 low-pass filtered sub-frequencies at level jBelt, or>
Is d l The k-th component, h, in the n frequency bands of the j + 1-th layer k-2l Decomposing the high-pass filter coefficients for the wavelet packets of the corresponding component, g k-2l The low-pass filter coefficients are decomposed for the wavelet packets of the corresponding component.
(1-2) selection of threshold function
The method is divided into the following 2 types:
(1-2-1) hard threshold denoising method
Let d l Obtaining n frequency bands after wavelet decomposition, namely n filter coefficients k (including high-frequency coefficients and low-frequency coefficients), starting from item 1, when the absolute value of the filter coefficient k is smaller than a given threshold a, making the filter coefficient k be 0, otherwise, making the filter coefficient k be constant, and finally, the calculation result is represented by y until each item is calculated, and the mathematical expression of each calculation is as follows:
(1-2-2) Soft threshold denoising method
Let d l Obtaining n frequency bands after wavelet decomposition, namely n filter coefficients k (including high-frequency coefficients and low-frequency coefficients), starting from item 1, when the absolute value of the filter coefficient k is smaller than a given threshold a, making the filter coefficient k be 0, otherwise, making the filter coefficient k subtract the threshold a, and finally, the calculation result is represented by y until each item is calculated, and the mathematical expression of each calculation is as follows:
(1-3) rule for selecting threshold a
The classification is as follows 4:
(1-3-1) general type
Let d l Obtaining n filter coefficients after wavelet packet decomposition, d l When the mean square deviation of (a) is recorded as sigma, the mathematical table of the threshold value aThe expression is as follows:
(1-3-2) unbiased form
Let d l After wavelet packet decomposition, n filter coefficients are obtained, and the squared values of all coefficients form a vector w, which is expressed as: w = [ w = 1 ，w 2 ，···，w n ]And w is 1 ≤w 2 ≤···≤w n Let a Risk vector Risk whose sub-element r i The expression of (c) is:
with the minimum value r in the Risk vector Risk element j As a risk value, j ∈ [1, n ]]Finding out the corresponding sub-element w in the w j Then the mathematical expression for threshold a is:
(1-3-3) Mixed type
Let d l Obtaining n filter coefficients after wavelet packet decomposition, d l The mean square error of (c) is represented as sigma, s is the sum of the squares of all coefficients, and then
In combination with w described in (1-3-2) j Then the mathematical expression for threshold a is:
(1-3-4) very large and very small
Let d l Obtained after wavelet packet decompositionn filter coefficients, the mathematical expression of the threshold a is:
generating 8 denoising methods through the combination of the threshold values and the threshold function in the steps (1-2) and (1-3), namely: the method comprises a general hard threshold denoising method, a general soft threshold denoising method, an unbiased hard threshold denoising method, an unbiased soft threshold denoising method, a mixed hard threshold denoising method, a mixed soft threshold denoising method, a very small hard threshold denoising method and a very small soft threshold denoising method.
(1-4) wavelet packet reconstruction TrainData2
After the 8 wavelet packet threshold denoising methods are adopted for calculation, 8 decomposed d are respectively subjected to l And reconstructing, wherein the reconstruction calculation formula is as follows:
in the formula (I), the compound is shown in the specification,
for the data of the n-th frequency band of the j +1 th layer reconstructed after the wavelet packet filtering, the method and the device>
Is d l The kth component in the 2n high-pass filtered subbands of the jth layer after wavelet packet filtering, is->
Is d l The kth component, h, in 2n +1 low-pass filtered sub-band at the jth layer after wavelet packet filtering k-2l Reconstructing the high-pass filter coefficients, g, for the wavelet packets of the corresponding components k-2l Low-pass filter coefficients are reconstructed for the wavelet packets of the corresponding component.
Preferably, the calculation expression of SNR and RMSE parameters in the stage (1) is as follows:
in the formula, d l (i) Representing the original aVEPs electroencephalogram data components,
representing the reconstructed aVEPs electroencephalogram data components after denoising, wherein N is the signal length.
Preferably, the data combing process in the stage (1) is as follows:
denoised reconstructed aVEPs electroencephalogram data
There are two dimensions: the channel number and the sampling point number simultaneously contain label category information, the positions of all the sampling points with the label category of 0 are found from the first character data, the set data segment length is intercepted backwards by taking the position as the initial position, and all the data segments are sequentially stored in an array named as Lefttemp according to the label appearance sequence;
in the same way, the positions of the sampling points with the label category of 1 are found, the set data segment length is intercepted backwards by taking the positions as initial positions, and all the data segments are sequentially stored in an array named Righttemp according to the label appearance order;
repeating the circulation until the data of all the characters are completely read; merging the Data in the array of the Lefttemp and the array of the Righttemp, and sequentially putting the merged Data into an array named as Data _ train for storage, wherein the Data _ train array has four dimensions: the number of channels, the number of sampling points, the total number of labels and the number of label categories.
Preferably, the TRCA algorithm in the stage (1) is specifically as follows:
selecting a type of label number from the Data _ train arrayAccording to matrix X, X has three dimensions: the number of channels, the number of sampling points and the total number of labels are set as X = [) 1 ，X 2 ，···X n ]N is the total number of the labels in the single-type label data;
for the ith element X in X i And the jth element X j The matrix S is calculated as follows:
and (3) performing transposition operation on the matrix S to obtain a matrix S', and further solving a matrix T, wherein the calculation formula is as follows:
T＝S+S′
combining the matrix X on the dimensionality of the total number of the labels to obtain a matrix Y only containing two dimensionalities of the channel number and the sampling point number, and setting Y to expand according to columns to be expressed as: y = [ Y) 1 ，Y 2 ，···Y m ]M is the column number of the matrix Y, namely the number of sampling points, and each element of the matrix Z is calculated according to the following formula:
wherein Z is i Represents the i-th element in the matrix Z, so Z = [ Z = 1 ，Z 2 ，···Z m ](ii) a And performing transposition operation on Z to obtain Z', and further solving a matrix Q, wherein the calculation formula is as follows:
Q＝Z+Z′
and calculating the generalized eigenvalue diagonal matrix D and the corresponding generalized right eigenvector matrix V for the solved T and Q, wherein the calculation formula is as follows:
T×V＝Q×V×D
extracting eigenvalue from diagonal of matrix D, sorting obtained vectors in descending mode, sorting elements of generalized right eigenvector matrix V in same descending mode and recording the sorted elements as matrix W i ；
Repeating the steps until all the category label data are calculatedObtaining the generalized right eigenvector matrix W, W = [ W ] 1 ，W 2 ，···W t ]，W i Is the ith element of W, and t is the tag class number.
Preferably, the dimension reduction algorithm in the stage (1) is specifically as follows:
for sub-elements X in the same type label data matrix X i Said W obtained by TRCA algorithm i Is transferred W i ' with said X i Multiplying to obtain a single-class label data matrix E after dimension reduction k Sub-element E of ki Namely:
′
E ki ＝W i ×X i
repeating the steps until all the data of all the labels in the same class are calculated, and obtaining a dimension-reduced single-class label data matrix E k Can be represented as E k ＝[E k1 ，E k2 ，···E kn ]，E ki Is E k N is the total number of labels in the single-class label data; then the all label category data matrix E can be represented as E = [ E = [ ] 1 ，E 2 ，···E t ]，E k Is the kth sub-element in E, t is the label category number;
and averaging the obtained matrix E on the dimension of the total number of the labels to obtain the template matrix Temp.
Preferably, the WPTD algorithm with 8 different threshold setting manners in the stage (2) is the same as the WPTD algorithm with 8 different threshold setting manners in the step (1), and filtering derived aVEPs electroencephalogram data TestData for testing needs to be performed through the WPTD algorithm.
Preferably, the SNR and RMSE parameters in the stage (2) are calculated in the same manner as the SNR and RMSE parameters in the step (1), and the corresponding SNR and RMSE parameters are calculated for the 8 filtered TestData.
Preferably, the data screening process in the stage (2) is as follows:
because the test Data hides the label category information, the positions of the electroencephalogram sampling points when all characters appear are marked by using the same label, the positions of the electroencephalogram sampling points when the label appears are found out firstly, the set Data segment lengths are intercepted backwards by taking the positions as initial positions, all the Data segments are sequentially stored in an array named as Data _ test according to the appearance sequence of the label,
preferably, the algorithm for calculating the correlation coefficient matrix r in the stage (2) is specifically as follows:
selecting 1 tag Data matrix M from the array of Data _ test, M having three dimensions: the number of channels, the number of sampling points and the total number of labels are set as M = [ M = 1 ，M 2 ，···M n ]，M i Is the ith sub-element in M, n is the total number of labels in the single-class label data, M i Transpose M of i ' with said W i Is transferred W i ' multiplication, generating matrix A ii The calculation formula is as follows:
A ii ＝M′ i ×W′ i
the result of the above formula calculation for all the data in M is recorded as matrix A, A ii Is the ith sub-element in A, calculate A ii With a sub-element Temp of said template matrix Temp ii To obtain the sub-elements R of the correlation coefficient matrix R i The calculation formula is as follows:
wherein, the first and the second end of the pipe are connected with each other,
for an average of A in the column dimension>
Is the average of Temp over the column dimension.
Repeating the steps until all the label data are calculated, and enabling the correlation coefficient matrix R = [ R ] 1 ，R 2 ，···R t ]，R i Is the ith sub-element in R, and t is the label category number.
And rearranging and combining the correlation coefficient matrix R on the dimensionality of the trial time, and averaging on the dimensionality of the trial time to obtain the correlation coefficient matrix R.
Preferably, the decoding calculation in the stage (3) is specifically as follows:
reading the element values in the correlation coefficient matrix r in sequence, reading 2 adjacent elements each time, recording the element values as 10 if the first element value is larger than the second element value, otherwise recording the element values as 01, and sequentially putting the recorded results into the predictive coding matrix until all the values are read.
The actual encoding matrix in stage (3) is composed of a plurality of vectors with only "0" and "1" elements, each representing a character, which are specified in an experimental paradigm.
The invention has obvious and prominent substantive features and remarkable advantages:
1. the invention is based on the mixed use of TRCA algorithm and WPTD algorithm, filters the aVEPs electroencephalogram, improves SNR and reduces RMSE, and finally realizes the improvement of the overall recognition rate and the recognition speed of character data based on the aVEPs electroencephalogram by a template matching method;
2. according to the method, a dynamic WPTD algorithm is adopted, filtering is carried out on each group of aVEPs electroencephalograms under different threshold setting methods, the SNR and the RMSE after filtering are calculated, and the filtered electroencephalograms with the largest SNR and the smallest RMSE are selected for template matching, so that the overall recognition rate of the aVEPs electroencephalograms is improved;
3. the invention adopts TRCA algorithm to match the template, compared with the traditional CCA algorithm, the ITR can be effectively improved, and the invention is beneficial to the further practical application of aVEPs electroencephalogram;
4. the invention is beneficial to promoting the development of aVEPs electroencephalogram, perfects the brain-computer interface system based on ERP electroencephalogram, can be used for the auxiliary output of physically disabled people, can also be used in the fields of industrial control and automatic driving, and has good market application prospect and social and economic benefits.
Drawings
FIG. 1 is a general algorithmic flow chart of the present invention.
FIG. 2 is a flow chart of the training phase algorithm of the present invention.
FIG. 3 is a flow chart of the test phase algorithm of the present invention.
Fig. 4 is a flow chart of the decoding phase algorithm of the present invention.
Fig. 5 is a flow chart of the WPTD algorithm of 8 different threshold setting modes according to the present invention.
FIG. 6 is a flow chart of the data combing algorithm of the present invention.
Fig. 7 is a flow chart of the TRCA algorithm of the present invention.
FIG. 8 is a flow chart of the dimension reduction algorithm of the present invention.
FIG. 9 is a flow chart of a data screening algorithm of the present invention.
Fig. 10 is a flow chart of the algorithm for calculating the correlation coefficient matrix r according to the invention.
Fig. 11 is a flow chart of the decoding algorithm of the present invention.
Detailed Description
The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings:
the first embodiment is as follows:
a TRCA-WPTD-based aVEPs electroencephalogram identification method comprises the following 3 stages:
the acquisition of aVEPs electroencephalogram data uses a Neuroscan synomps 2 electroencephalogram recording system. When in electroencephalogram recording, 0.1-100Hz band-pass filtering is adopted, 50Hz notch is superposed, and the sampling frequency is 200Hz.
Referring to fig. 1, a TRCA-WPTD based aVEPs electroencephalogram identification method comprises the following 3 stages:
(1) Training phase
Referring to fig. 2, preprocessed aVEPs electroencephalogram data TrainData1 for training are imported, and filtering is performed by using 8 WPTD algorithms with different threshold setting modes; generating 8 kinds of filtered TrainData2, calculating SNR and RMSE parameters for each TrainData2, selecting the TrainData2 with the highest SNR value and the lowest RMSE value for data combing, and generating TrainData3; calculating a generalized right eigenvector matrix W by a TRCA algorithm; and (5) performing dimension reduction on the TrainData3 based on the matrix W to obtain a template matrix Temp.
(2) Testing phase
Referring to fig. 3, preprocessed aVEPs electroencephalogram data TestData1 for testing are imported, and filtering is performed through a WPTD algorithm with 8 different threshold setting modes to generate 8 filtered TestData2; calculating SNR and RMSE parameters for each TestData2, selecting the TestData2 with the highest SNR value and the lowest RMSE value to perform data screening, and generating TestData3; calculating a correlation coefficient matrix r by the TestData3 and the generalized right eigenvector matrix W and the template matrix Temp generated in the training stage;
(3) Decoding stage
Referring to fig. 4, the correlation coefficient matrix r is decoded and calculated to generate a predictive coding matrix, and the predictive coding matrix is compared with an actual coding matrix element set in an experimental paradigm, and an identification result can be output only when the predictive coding matrix element is completely matched with the actual coding matrix element set in the experimental paradigm. According to the embodiment, the WPTD algorithm is used for filtering, the signal-to-noise ratio of the aVEPs electroencephalogram is improved, the root mean square error is reduced, the TRCA algorithm is used for carrying out template matching for identification, and the character identification accuracy and identification speed based on the aVEPs electroencephalogram are improved.
Example two:
the present embodiment is substantially the same as the first embodiment, and the features are as follows:
referring to fig. 5, the WPTD algorithm of 8 different threshold setting modes in the phase (1) is specifically as follows:
the WPTD algorithm process is that wavelet packet decomposition is utilized to decompose aVEPs electroencephalogram data into various scales, a threshold value is set, then wavelet coefficients belonging to noise in each scale are removed, the wavelet coefficients belonging to the aVEPs electroencephalogram data are reserved and enhanced, and finally the wavelet coefficients after processing are reconstructed by wavelet inverse transformation to obtain noise-reduced aVEPs electroencephalogram.
Matrix d for aVEPs electroencephalogram data TrainData1 for training l Is represented by d l There are two dimensions: the method comprises the following steps of (1) channel number and sampling point number, and also label category information which does not participate in filtering calculation, wherein the specific data processing flow is as follows:
(1-1) wavelet packet decomposition TrainData1
To d l Decomposing the nth frequency band of j +1 layer into the height of j layer according to the selected wavelet base expansionThe 2n th of pass filtering and the 2n +1 th of low pass filtering, the iterative calculation formula of the decomposition is:
in the formula (I), the compound is shown in the specification,
is d l At level j the 2 nth high pass filtered sub-band, device for combining or screening>
Is d l 2n +1 low pass filtered sub-bands at level j>
Is d l The k-th component, h, in the n frequency bands of the j + 1-th layer k-2l Decomposing the high-pass filter coefficients, g, for the wavelet packets of the corresponding component k-2l The low-pass filter coefficients are decomposed for the wavelet packets of the corresponding component.
(1-2) selection of threshold function
The method is divided into the following 2 types:
(1-2-1) hard threshold denoising method
Let d l Obtaining n frequency bands after wavelet decomposition, namely n filter coefficients k (including high-frequency coefficients and low-frequency coefficients), starting from item 1, when the absolute value of the filter coefficient k is smaller than a given threshold a, making the filter coefficient k be 0, otherwise, making the filter coefficient k be constant, and finally, the calculation result is represented by y until each item is calculated, and the mathematical expression of each calculation is as follows:
(1-2-2) Soft threshold denoising method
Let d l After wavelet decomposition, n frequency bands are obtained, i.e. there are n filter coefficients k (including high frequency coefficient and low frequency coefficient) starting from item 1 when the filter coefficientWhen the absolute value of the number k is smaller than a given threshold value a, the number k is set to be 0, otherwise, the threshold value a is subtracted from the number k, the final calculation result is represented by y, and until each item is calculated, the mathematical expression of each calculation is as follows:
(1-3) rule for selecting threshold a
The classification is as follows 4:
(1-3-1) general type
Let d l Obtaining n filter coefficients after wavelet packet decomposition, d l The mean square error of (c) is denoted as σ, the mathematical expression of the threshold value a is:
(1-3-2) unbiased form
Let d l After wavelet packet decomposition, n filter coefficients are obtained, and the squared values of all coefficients form a vector w, which is expressed as: w = [ w = 1 ，w 2 ，···，w n ]And w is a 1 ≤w 2 ≤···≤w n Let a Risk vector Risk whose sub-element r i The expression of (a) is:
with the minimum value r in the Risk vector Risk element j As a risk value, j ∈ [1,n ]]Finding out the corresponding sub-element w in the w j Then the mathematical expression for threshold a is:
(1-3-3) Mixed type
Let d l Obtaining n filters after wavelet packet decompositionCoefficient of device, d l The mean square error of (c) is represented as sigma, s is the sum of the squares of all coefficients, and then
In combination with w described in (1-3-2) j Then the mathematical expression for threshold a is:
(1-3-4) very large and very small
Let d l After wavelet packet decomposition, n filter coefficients are obtained, and then the mathematical expression of the threshold value a is:
generating 8 denoising methods through the combination of the threshold values and the threshold function in the steps (1-2) and (1-3), namely: the method comprises a general hard threshold denoising method, a general soft threshold denoising method, an unbiased hard threshold denoising method, an unbiased soft threshold denoising method, a mixed hard threshold denoising method, a mixed soft threshold denoising method, a very small hard threshold denoising method and a very small soft threshold denoising method.
(1-4) wavelet packet reconstruction TrainData2
After the 8 wavelet packet threshold denoising methods are adopted for calculation, 8 decomposed d are respectively subjected to l And reconstructing, wherein a calculation formula of reconstruction is as follows:
in the formula (I), the compound is shown in the specification,
for the data of the n-th frequency band of the j +1 th layer reconstructed after the wavelet packet filtering, the method and the device>
Is d l The kth component in the 2n high-pass filtered subbands of the jth layer after wavelet packet filtering, is->
Is d l The kth component, h, in 2n +1 low-pass filtered subbands at jth layer after wavelet packet filtering k-2l Reconstructing the high-pass filter coefficients, g, for the wavelet packets of the corresponding components k-2l Low-pass filter coefficients are reconstructed for the wavelet packets of the corresponding components. />
Preferably, the calculation expression of the SNR and RMSE parameters in the stage (1) is as follows:
in the formula (d) l (i) Representing the original aVEPs electroencephalogram data components,
representing the reconstructed aVEPs electroencephalogram data component after denoising, wherein N is the signal length.
Referring to fig. 6, the data combing process in stage (1) is as follows:
denoised reconstructed aVEPs electroencephalogram data
There are two dimensions: the channel number and the sampling point number simultaneously contain label category information, the positions of all the sampling points with label categories of 0 are found from the first character data, the set data segment length is backward intercepted by taking the position as the initial position, and all the data segments are sequentially stored in an array named as Lefttemp according to the label appearance order;
in the same way, the positions of the sampling points with the label category of 1 are found, the set data segment length is intercepted backwards by taking the positions as initial positions, and all the data segments are sequentially stored in an array named Righttemp according to the label appearance order;
repeating the circulation until the data of all the characters are completely read; merging the Data in the array of the Lefttemp and the array of the Righttemp, and sequentially putting the merged Data into an array named as Data _ train for storage, wherein the Data _ train array has four dimensions: the number of channels, the number of sampling points, the total number of labels and the number of label categories.
Referring to fig. 7, the TRCA algorithm in stage (1) is specifically as follows:
selecting a type of tag Data matrix X from the array of Data _ train, X having three dimensions: setting X = [ X ] for channel number, sampling point number and total number of labels 1 ，X 2 ，···X n ]N is the total number of the labels in the single-type label data;
for the ith element X in X i And the jth element X j The matrix S is calculated as follows:
and performing transposition operation on the matrix S to obtain a matrix S', and further solving a matrix T, wherein the calculation formula is as follows:
T＝S+S′
combining the matrix X on the dimension of the total number of the labels to obtain a matrix Y only containing two dimensions of the number of channels and the number of sampling points, and setting Y to expand according to columns to be expressed as: y = [ Y = 1 ，Y 2 ，···Y m ]M is the column number of the matrix Y, namely the number of sampling points, and each element of the matrix Z is calculated according to the following formula:
wherein Z is i Represents the i-th element in the matrix Z, so Z = [ Z = 1 ，Z 2 ，···Z m ](ii) a And performing transposition operation on Z to obtain Z', and further solving a matrix Q, wherein the calculation formula is as follows:
Q＝Z+Z′
and calculating the generalized eigenvalue diagonal matrix D and the corresponding generalized right eigenvector matrix V for the solved T and Q by the following calculation formula:
T×V＝Q×V×D
extracting eigenvalue from diagonal of matrix D, sorting obtained vectors in descending mode, sorting elements of generalized right eigenvector matrix V in same descending mode and recording the sorted elements as matrix W i ；
Repeating the steps until all the category label data are calculated, and obtaining the generalized right eigenvector matrix W, W = [ W ] 1 ，W 2 ，···W t ]，W i Is the ith element of W, and t is the tag class number.
Referring to fig. 8, the dimension reduction algorithm in the stage (1) is specifically as follows:
for sub-elements X in the same type label data matrix X i Said W obtained by TRCA algorithm i Is transferred W i ' with said X i Multiplying to obtain a single-class label data matrix E after dimension reduction k Sub-element E of ki Namely:
E ki ＝W′ i ×X i
repeating the steps until all the data of all the labels in the same class are calculated, and obtaining the reduced dimension single-class label data matrix E k Can be represented as E k ＝[E k1 ，E k2 ，···E kn ]，E ki Is E k N is the total number of labels in the single-class label data; then the all label category data matrix E can be represented as E = [ E = [ ] 1 ，E 2 ，···E t ]，E k Is the kth sub-element in E, t is the label category number;
and averaging the obtained matrix E on the dimension of the total number of the labels to obtain the template matrix Temp.
Referring to fig. 5, the WPTD algorithms with 8 different threshold setting modes in the stage (2) are the same as the WPTD algorithms with 8 different threshold setting modes in the step (1), and the imported aVEPs electroencephalogram data TestData for testing needs to be filtered through the WPTD algorithms.
Preferably, the SNR and RMSE parameters in the stage (2) are calculated in the same manner as the SNR and RMSE parameters in the step (1), and the corresponding SNR and RMSE parameters are calculated for the 8 filtered TestData.
Referring to fig. 9, the data screening process in stage (2) is as follows:
because the test Data hides the label type information, the same type label '1' is used for marking the sampling point positions of the electroencephalogram when all characters appear, the positions of the electroencephalogram sampling points when the label appears are found out firstly, the set Data segment length is intercepted backwards by taking the positions as the initial positions, all the Data segments are sequentially stored in an array named as Data _ test according to the appearance sequence of the label,
referring to fig. 10, the algorithm for calculating the correlation coefficient matrix r in the stage (2) is specifically as follows:
selecting 1 tag Data matrix M from the array of Data _ test, M having three dimensions: setting M = [ M ] for channel number, sampling point number and total number of labels 1 ，M 2 ，···M n ]，M i Is the ith sub-element in M, n is the total number of labels in the single-class label data, M i Is transposed M i ' with said W i Is transferred W i ' multiplication, generating matrix A ii The calculation formula is as follows:
A ii ＝M′ i ×W′ i
the result of the calculation of all the data corresponding to the total number of the labels in the single-type label data according to the formula is recorded as a matrix A, A ii Is a sub-element in A, calculate A ii With a sub-element Temp of said template matrix Temp ii To obtain the sub-elements R of the correlation coefficient matrix R i The calculation formula is as follows:
wherein the content of the first and second substances,
is an average of A over the column dimension, <' >>
Is the average of Temp over the column dimension.
Repeating the steps until all the label data are calculated, and enabling the correlation coefficient matrix R = [ R ] 1 ，R 2 ，···R t ]，R i Is the ith sub-element in R, and t is the label category number.
And rearranging and combining the correlation coefficient matrix R on the dimensionality of the trial time, and averaging on the dimensionality of the trial time to obtain the correlation coefficient matrix R.
Referring to fig. 11, the decoding calculation in the stage (3) is specifically as follows:
and sequentially reading the element values in the correlation coefficient matrix r, reading 2 adjacent elements each time, recording the result as 10 if the first element value is greater than the second element value, otherwise recording the result as 01, and sequentially putting the recorded result into the predictive coding matrix until all the values are read.
The actual coding matrix in stage (3) is composed of a plurality of vectors containing only "0" and "1" elements, each vector representing a character, which are specified in an experimental paradigm.
In conclusion, the invention discloses an aVEPs electroencephalogram identification method based on TRCA-WPTD. The method comprises the following steps: firstly, importing preprocessed aVEPs electroencephalogram data TrainData1 for training, and respectively filtering by using 8 WPTD algorithms with different threshold setting modes; generating 8 kinds of filtered TrainData2, calculating SNR and RMSE parameters for each TrainData2, selecting the TrainData2 with the highest SNR value and the lowest RMSE value for data combing, and generating TrainData3; calculating a generalized right eigenvector matrix W through a TRCA algorithm; performing dimension reduction on the TrainData3 based on the matrix W to obtain a template matrix Temp; secondly, importing preprocessed aVEPs electroencephalogram data TestData1 for testing, and respectively filtering through 8 WPTD algorithms with different threshold setting modes to generate 8 filtered TestData2; calculating SNR and RMSE parameters for each TestData2, selecting the TestData2 with the highest SNR value and the lowest RMSE value to perform data screening, and generating TestData3; the TestData3 and the generalized right eigenvector matrix W and the template matrix Temp generated in the training stage calculate a correlation coefficient matrix r; and finally, decoding the correlation coefficient matrix r to generate a predictive coding matrix, comparing the predictive coding matrix with an actual coding matrix set in an experimental paradigm, and outputting an identification result when the predictive coding matrix is completely matched with the actual coding matrix. According to the invention, the WPTD algorithm is used for filtering, the signal-to-noise ratio of the aVEPs electroencephalogram is improved, the root mean square error is reduced, the TRCA algorithm is used for carrying out template matching for identification, and the character identification accuracy and identification speed based on the aVEPs electroencephalogram are improved.
The above detailed description merely describes preferred embodiments of the present invention and does not limit the scope of the invention. Without departing from the design concept and spirit of the present invention, various changes, substitutions and improvements of the technical solutions of the present invention made by the text description and the drawings provided by the ordinary skilled in the art shall fall into the protection scope of the present invention. The scope of protection of the invention is determined by the claims.
Claims (7)
1. An electroencephalogram identification method for asymmetric visual evoked potentials aVEPs based on task related component analysis TRCA combined with wavelet packet threshold denoising WPTD is characterized by comprising the following steps: the method comprises the following 3 stages:
(1) A training stage:
respectively filtering the training data by using 8 WPTD algorithms with different threshold setting modes; respectively calculating SNR and RMSE parameters of the filtered training data, and selecting the data with the highest SNR value and the lowest RMSE value for data combing; calculating a generalized right eigenvector matrix W through a TRCA algorithm; reducing the dimensions of the data to obtain a template matrix Temp;
(2) And (3) a testing stage:
processing the test data by the same filtering and data selecting method as the training stage, then screening the data, and calculating a correlation coefficient matrix r with the generalized right eigenvector matrix W and the template matrix Temp;
(3) And a decoding stage:
decoding the correlation coefficient matrix r obtained in the test stage, generating a predictive coding matrix, comparing the predictive coding matrix with an actual coding matrix, and outputting an identification result;
the WPTD algorithms of 8 different threshold setting modes in the stages (1) and (2) are specifically as follows:
decomposing aVEPs electroencephalogram data into various scales by wavelet packet decomposition, and generating 8 denoising methods by adopting a combination of a threshold value and a threshold value function, namely: the method comprises the steps of respectively removing wavelet coefficients belonging to noise in each scale by each method, retaining and enhancing the wavelet coefficients belonging to aVEPs electroencephalogram data, and finally reconstructing the processed wavelet coefficients by utilizing wavelet inverse transformation to obtain noise-reduced aVEPs electroencephalogram.
2. The electroencephalogram identification method for asymmetric visual evoked potentials (aVEPs) based on Task Related Component Analysis (TRCA) and Wavelet Packet Threshold Denoising (WPTD) according to claim 1, which is characterized in that: the data carding process in the stage (1) comprises the following steps: starting from the first character data, finding out the positions of sampling points with all label types of 0, taking the positions as initial positions, intercepting the set data segment length backwards, and sequentially storing all the data segments in an array named as Lefttemp according to the appearance order of the labels; finding out all data with the label type of 1 and storing the data in an array named as Righttemp; repeating the circulation until the data of all the characters are completely read; and merging the Data in the Lefttemp array and the Righttemp array, and sequentially putting the merged Data into an array named Data _ train for storage.
3. The electroencephalogram identification method for asymmetric visual evoked potentials (aVEPs) based on Task Related Component Analysis (TRCA) combined with Wavelet Packet Threshold Denoising (WPTD) according to claim 2, characterized in that: the TRCA algorithm in the stage (1) is specifically:
selecting a label Data matrix X from the Data _ train array, and setting X = [) 1 ，X 2 ，···X n ]N is the total number of the labels in the single-type label data; for the ith element X in X i And the jth element X j The matrix S is calculated as follows:
adding the matrix S and the transpose of the matrix S, and further solving a matrix T;
combining the matrix X on the dimension of the total number of the labels to obtain a matrix Y, and expanding Y according to columns to be expressed as: y = [ Y) 1 ，Y 2 ，···Y m ]M is the column number of the matrix Y, namely the number of sampling points, and each element of the matrix Z is calculated according to the following formula:
wherein, Z i Represents the i-th element in the matrix Z, so Z = [ Z = 1 ，Z 2 ，···Z m ](ii) a m is the number of columns of the matrix Z, namely the number of sampling points, and the matrix Z are transposed and added to further solve Q; calculating a generalized eigenvalue diagonal matrix D and a corresponding generalized right eigenvector matrix V for the solved T and Q, extracting eigenvalues from the diagonal of the matrix D, sorting the obtained vectors in a descending manner, sorting the elements of the generalized right eigenvector matrix V in the same descending manner, and recording the sorted elements as a matrix W i (ii) a Repeating the steps until all the category label data are calculated, and obtaining the generalized right eigenvector matrix W, W = [ W ] 1 ，W 2 ，···W t ]，W i Is the ith element of W, and t is the tag class number.
4. The electroencephalogram identification method for asymmetric visual evoked potentials (aVEPs) based on Task Related Component Analysis (TRCA) and Wavelet Packet Threshold Denoising (WPTD), according to claim 3, is characterized in that: the dimension reduction algorithm in the stage (1) is specifically as follows: for sub-elements X in the same type label data matrix X i With said W i Is transposed with respect to the X i Multiplying to obtain a single-class label data matrix E after dimension reduction k Sub-elements of
Repeating the steps until all the data of all the labels in the same class are calculated, and obtaining the reduced dimension single-class label data matrix E k Can be expressed as->
Is E k The ith sub-element in the list, n is the total number of labels in the single type of label data; then the all label category data matrix E can be represented as E = [ E = [ ] 1 ，E 2 ，···E t ]，E k Is the kth sub-element in E, and t is the label category number; and averaging the obtained matrix E on the dimension of the total number of the labels to obtain the template matrix Temp.
5. The electroencephalogram identification method for asymmetric visual evoked potentials (aVEPs) based on Task Related Component Analysis (TRCA) combined with Wavelet Packet Threshold Denoising (WPTD) according to claim 1, characterized by comprising the following steps: the data screening process in the stage (2) is as follows: the method comprises the steps of firstly finding out the position of an electroencephalogram sampling point when a label type appears, then taking the position as an initial position, intercepting the set Data segment length backwards, and sequentially storing all the Data segments in an array named as Data _ test according to the appearance sequence of labels.
6. The electroencephalogram identification method for asymmetric visual evoked potentials (aVEPs) based on Task Related Component Analysis (TRCA) and Wavelet Packet Threshold Denoising (WPTD), according to claim 3, is characterized in that: the algorithm for calculating the correlation coefficient matrix r in the stage (2) is specifically as follows: selecting 1 tag Data matrix M from Data _ test array, and setting M = [ M = [ [ M ] 1 ，M 2 ，···M n ]，M i Is the ith sub-element in M, n is the total number of labels in the single-class label data, M i Is transposed with respect to the W i Is multiplied by the transpose of (c), generating matrix a ii Calculating A ii With a sub-element Temp of said template matrix Temp ii To obtain the sub-elements R of the correlation coefficient matrix R i (ii) a Repeating the steps until all the label data are calculated, and the correlation coefficient matrix R = [ R ] = 1 ，R 2 ，···R t ]，R i Is the ith sub-element in R, and t is the label category number; and rearranging and combining the correlation coefficient matrix R on the dimensionality of the trial time, and averaging on the dimensionality of the trial time to obtain the correlation coefficient matrix R.
7. The electroencephalogram identification method for asymmetric visual evoked potentials (aVEPs) based on Task Related Component Analysis (TRCA) and Wavelet Packet Threshold Denoising (WPTD) according to claim 1, which is characterized in that: the decoding calculation in the stage (3) is specifically: and sequentially reading the element values in the correlation coefficient matrix r, reading 2 adjacent elements each time, recording the element value as 10 if the first element value is greater than the second element value, otherwise recording the element value as 01, and sequentially putting the recorded result into the predictive code until all the element values are read.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011343524.2A CN112364812B (en) | 2020-11-26 | 2020-11-26 | aVEPs electroencephalogram identification method based on TRCA-WPTD |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011343524.2A CN112364812B (en) | 2020-11-26 | 2020-11-26 | aVEPs electroencephalogram identification method based on TRCA-WPTD |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112364812A CN112364812A (en) | 2021-02-12 |
| CN112364812B true CN112364812B (en) | 2023-04-14 |
Family
ID=74533598
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011343524.2A Active CN112364812B (en) | 2020-11-26 | 2020-11-26 | aVEPs electroencephalogram identification method based on TRCA-WPTD |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112364812B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104771163A (en) * | 2015-01-30 | 2015-07-15 | 杭州电子科技大学 | Electroencephalogram feature extraction method based on CSP and R-CSP algorithms |
| CN109271887A (en) * | 2018-08-29 | 2019-01-25 | 天津大学 | A kind of composite space filtering and template matching method for the identification of brain power mode |
| CN109584900A (en) * | 2018-11-15 | 2019-04-05 | 昆明理工大学 | A kind of blind source separation algorithm of signals and associated noises |
| CN110163128A (en) * | 2019-05-08 | 2019-08-23 | 南京邮电大学 | The Method of EEG signals classification of improved EMD algorithm combination wavelet package transforms and CSP algorithm |
| CN111797674A (en) * | 2020-04-10 | 2020-10-20 | 成都信息工程大学 | MI electroencephalogram signal identification method based on feature fusion and particle swarm optimization algorithm |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8315970B2 (en) * | 2008-10-20 | 2012-11-20 | Neurochip Corporation | Method and rhythm extractor for detecting and isolating rhythmic signal features from an input signal using the wavelet packet transform |
-
2020
- 2020-11-26 CN CN202011343524.2A patent/CN112364812B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104771163A (en) * | 2015-01-30 | 2015-07-15 | 杭州电子科技大学 | Electroencephalogram feature extraction method based on CSP and R-CSP algorithms |
| CN109271887A (en) * | 2018-08-29 | 2019-01-25 | 天津大学 | A kind of composite space filtering and template matching method for the identification of brain power mode |
| CN109584900A (en) * | 2018-11-15 | 2019-04-05 | 昆明理工大学 | A kind of blind source separation algorithm of signals and associated noises |
| CN110163128A (en) * | 2019-05-08 | 2019-08-23 | 南京邮电大学 | The Method of EEG signals classification of improved EMD algorithm combination wavelet package transforms and CSP algorithm |
| CN111797674A (en) * | 2020-04-10 | 2020-10-20 | 成都信息工程大学 | MI electroencephalogram signal identification method based on feature fusion and particle swarm optimization algorithm |
Non-Patent Citations (4)
| Title |
|---|
| Subject-based feature extraction by using fisher WPD-CSP in brain–computer interfaces;Banghua Yang 等;《computer methods and programs in biomedicine》;20161231;第1-8页 * |
| WASICA: An effective wavelet-shrinkage based ICA model for brain fMRI data analysis;Nizhuan Wang 等;《Journal of Neuroscience Methods》;20150316;第1-22页 * |
| 基于脑机接口的康复训练系统;杨帮华 等;《系统仿真学报》;20190228;第1-7页 * |
| 整合贝叶斯动态停止策略对 SSVEP-BCIs的性能提升研究;江京 等;《仪器仪表学报》;20180531;第1-8页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112364812A (en) | 2021-02-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Li et al. | A channel-projection mixed-scale convolutional neural network for motor imagery EEG decoding | |
| CN108829257B (en) | Feature extraction method of motor imagery electroencephalogram signal based on DTCTWT and IL-MVU | |
| CN109890043B (en) | Wireless signal noise reduction method based on generative countermeasure network | |
| Upadhyay et al. | A comparative study of feature ranking techniques for epileptic seizure detection using wavelet transform | |
| Kidron et al. | Cross-modal localization via sparsity | |
| CN114533086B (en) | Motor imagery brain electrolysis code method based on airspace characteristic time-frequency transformation | |
| CN111310570A (en) | Electroencephalogram signal emotion recognition method and system based on VMD and WPD | |
| CN106943140A (en) | A kind of Mental imagery EEG feature extraction method based on RandomSelect RCSP | |
| Ylipaavalniemi et al. | Analyzing consistency of independent components: An fMRI illustration | |
| CN110059564B (en) | Feature extraction method based on power spectral density and cross-correlation entropy spectral density fusion | |
| CN115414051A (en) | Emotion classification and recognition method of electroencephalogram signal self-adaptive window | |
| CN112465069A (en) | Electroencephalogram emotion classification method based on multi-scale convolution kernel CNN | |
| Wang et al. | Denoising auto-encoding priors in undecimated wavelet domain for MR image reconstruction | |
| Mishra et al. | Noise removal in EEG signals using SWT–ICA combinational approach | |
| CN110141211B (en) | Steady-state visual evoked potential classification method based on empirical mode decomposition | |
| CN108985301A (en) | A kind of hyperspectral image classification method of the sub- dictionary learning of multiple features class | |
| CN114403897A (en) | Human body fatigue detection method and system based on electroencephalogram signals | |
| Ku et al. | Earthquake event classification using multitasking deep learning | |
| Lee et al. | Analysis of attention deficit hyperactivity disorder in EEG using wavelet transform and self organizing maps | |
| CN112364812B (en) | aVEPs electroencephalogram identification method based on TRCA-WPTD | |
| CN116150670A (en) | Task independent brain pattern recognition method based on feature decorrelation decoupling | |
| CN114757236A (en) | Electroencephalogram signal denoising optimization method and system based on TQWT and SVMD | |
| CN113571074A (en) | Voice enhancement method and device based on multi-band structure time domain audio separation network | |
| CN113876339A (en) | Method for constructing sleep state electroencephalogram characteristic signal feature set | |
| Chavan et al. | Improved EEG signal processing with wavelet based multiscale PCA algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |