A method for EEG signal recognition based on PLSR and extended FBCCA

Abstract

The present invention discloses an EEG signal recognition method based on PLSR and extended FBCCA. The method comprises the following steps: regressing a sinusoidal reference template signal onto a multi-channel sampled EEG signal to obtain an EEG estimation signal; performing correlation analysis on the decomposed sub-band components of the multi-channel sampled EEG signal and the estimation signal to obtain a set of extended correlation coefficients of the sub-band components; and obtaining a set of extended correlation coefficients of the EEG signal for different targets from a set of extended correlation coefficients of different sub-band components of the EEG signal for the estimation signal. The fundamental frequency of the reference template signal corresponding to the maximum value is the target frequency of the EEG signal to be identified. The accuracy and applicability of steady-state visual evoked potential signal recognition are improved by extracting the spatial distribution relationship in the EEG signal. At the same time, the influence of multiple groups of typical variables is taken into account by expanding the main characteristic components, thereby improving the accuracy of EEG signal classification and recognition in the steady-state visual evoked potential paradigm.

Description

Technical Field

[0001] The present invention belongs to the technical field of EEG signal feature extraction and classification recognition of brain-computer interface, and specifically relates to an EEG signal recognition method based on PLSR and extended FBCCA. Background Art

[0002] A brain-computer interface (BCI) is a software and hardware communication system that enables humans to interact with the external environment without the involvement of peripheral nerves and muscles through control signals generated by brain activity. Based on the steady-state visual evoked potential (SSVEP) paradigm, BCIs are important tools in neuroengineering and clinical neuroscience, and are of great significance in rehabilitation medicine. They have been gradually extended to fields such as intelligent interaction and virtual reality, significantly improving the efficiency and convenience of human life.

[0003] EEG signals are non-stationary, nonlinear signals. Their strong individual variability significantly limits the accuracy of signal recognition for brain-computer interfaces (BCIs) and their widespread adoption in practical applications. Research on BCIs focuses on algorithms for analyzing and processing EEG signals. These algorithms aim to improve the accuracy of systems through signal feature extraction and classification, and enhance the performance of SSVEP-based BCI devices such as brain-controlled robotic arms and EEG typewriters.

[0004] Filter Bank Canonical Correlation Analysis (FBCCA) is a model-based method for SSVEP (Sensory Spectrum Video Encephalogram) (SSVEP) EEG signal recognition. It measures the canonical correlation coefficients between the EEG signal's subband components and a reference template signal. Currently, FBCCA can only extract the canonical correlation coefficient of a single feature, and the constructed standard reference template signal cannot reflect the distribution relationship between the sampled EEG signal and the original. Therefore, it is important to expand and improve the FBCCA algorithm to enhance the accuracy and applicability of signal recognition. Summary of the Invention

[0005] In view of this, the present invention provides an EEG signal recognition method based on PLSR and extended FBCCA, which can improve the limitations of the FBCCA method and enhance the classification and recognition accuracy and applicability of SSVEP signals.

[0006] The present invention provides an EEG signal recognition method based on PLSR and extended FBCCA, comprising the following steps:

[0007] For the multi-channel SSVEP EEG signal to be identified, N sinusoidal reference template signals are established according to the number of target stimulus sources N. The fundamental frequency of each template signal corresponds to the frequency of the target stimulus source. PLSR is used to regress the sinusoidal reference template signal onto the multi-channel sampled EEG signal to obtain N groups of EEG estimation signals.

[0008] Decompose the multi-channel sampled EEG signal through a filter bank to generate k EEG signal sub-band components;

[0009] For each sub-band component, the weighted sum of the first n canonical correlation coefficients of each estimated signal is obtained by applying an extended canonical correlation analysis algorithm to the sub-band component, which is used as the extended correlation coefficient of the sub-band component to the single estimated signal.

[0010] The extended correlation coefficients of the k sub-band components decomposed from the EEG signal to a single estimated signal are weighted and summed as the overall correlation coefficient of the EEG signal to the single estimated signal. For N target stimulus sources, a total of N overall correlation coefficients can be obtained. The fundamental frequency of the reference template signal corresponding to the largest one is the target frequency of the multi-channel SSVEP EEG signal to be identified.

[0011] Furthermore, the method of using PLSR to regress the multi-channel sampled EEG signal to obtain the EEG estimation signal is:

[0012] Build a PLSR model for the sinusoidal reference template signal:

[0013] Y f =UC T +E

[0014] X=VD T +G=UR T +F

[0015] Among them, U and V are the extracted latent component matrices, E and G are the residual matrices of U and V respectively, and F is the residual matrix with U as the dependent variable; the partial least squares regression coefficient matrix B is obtained by iterative calculation M times f , M satisfies M<min{c,2h}; using matrix B f The estimated EEG signal is X f :X f =Y f B f , Y f is the sinusoidal reference template signal.

[0016] Furthermore, for each subband component, obtaining a weighted sum of the first n canonical correlation coefficients of each estimated signal through an extended canonical correlation analysis algorithm as the extended correlation coefficient of the subband component includes:

[0017] For the multi-channel SSVEP EEG signal X to be identified b ∈R c×s and the EEG estimation signal X f ∈R 2h×s , calculate X b The sample covariance C with itself XX 、Xf The sample covariance with itself is C YY 、X b With X f The sample covariance of XY 、X f With X b The sample covariance of YX , calculate the canonical correlation matrix A:

[0018]

[0019] Obtain the c eigenvalues ​​of the typical correlation matrix A and arrange them in descending order as [λ1,λ2,…,λ c ], from which n eigenvalues ​​are selected: ρ in For the i n Typical correlation coefficients under heavy features; different weights are set for typical correlation coefficients under different heavy features: Among them, k3 and k4 are adjustable weight parameters, so the obtained X b Relative to X f The expanded correlation coefficient is:

[0020] Furthermore, the extended correlation coefficients of the k sub-band components decomposed from the EEG signal to the single estimated signal are weighted summed as the overall correlation coefficient of the EEG signal to the single estimated signal, and the calculation method is:

[0021]

[0022] Among them, r(X,X f ) is the overall correlation coefficient, w FB (i k ) is the subband weight.

[0023] Beneficial effects:

[0024] The present invention constructs a sinusoidal reference template signal for the EEG signal to be identified according to the frequency of the target stimulus source, and regresses the sinusoidal reference template signal to the multi-channel sampled EEG signal to obtain an EEG estimation signal, and then performs correlation analysis on the decomposed sub-band components of the multi-channel sampled EEG signal and the estimation signal to obtain a set of extended correlation coefficients of the sub-band components, and then obtains a set of extended correlation coefficients and a set of extended correlation coefficients of the sampled EEG signal from the set of extended correlation coefficients of the sub-band components, wherein the fundamental frequency of the reference template signal corresponding to the maximum value is the target frequency of the EEG signal to be identified, and the accuracy and applicability of signal recognition are improved by extracting the spatial distribution relationship in the EEG signal. At the same time, since the influence of multiple groups of typical variables is taken into account in the form of expansion of the main characteristic components, the accuracy of EEG signal splitting and recognition can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of a method for classifying and recognizing EEG signals based on PLSR and extended FBCCA provided by the present invention.

[0026] Figure 2 This is a flow chart of the filter bank algorithm in the EEG signal classification and recognition method based on PLSR and extended FBCCA provided by the present invention.

[0027] Figure 3 This is an algorithm flow chart of the extended FBCCA method in the EEG signal classification and recognition method based on PLSR and extended FBCCA provided by the present invention.

[0028] Figure 4 This is a comparison chart of offline experimental results of an EEG signal classification and recognition method based on PLSR and extended FBCCA provided by the present invention on the San Diego Square Joint Frequency-Phase Modulation SSVEP Dataset open source dataset.

[0029] Figure 5 This is a comparison chart of offline experimental results of an EEG signal classification and recognition method based on PLSR and extended FBCCA provided by the present invention on the Tsinghua Benchmark SSVEP Database open source dataset. DETAILED DESCRIPTION

[0030] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0031] PLSR is a regression algorithm based on partial least squares. By establishing a regression model of the latent variables of the independent variable with respect to the latent variables of the dependent variable, the two sets of latent variables are designed to retain the characteristic information of the independent and dependent variables to the greatest extent possible, while also maximizing their covariance. FBCCA is a model-based SSVEP EEG signal feature extraction and classification algorithm that extracts features from EEG sampling signals and classifies and identifies the target stimulus source based on the extracted features. The significance of feature expansion of the FBCCA algorithm lies in the fact that the introduction of the influence of multiple groups of typical variables can improve the accuracy of EEG signal splitting and recognition. The significance of combining the two is that they can improve the accuracy and applicability of EEG signal classification and recognition based on individual differences and spatiotemporal differences in EEG signals. PLSR performs regression estimation of the template signal, while the extended FBCCA extracts features of the multi-channel EEG sampling signal pair estimation signal and performs classification tasks.

[0032] The present invention provides an EEG signal classification and recognition method based on PLSR and extended FBCCA, the core idea of ​​which is: first, for the multi-channel SSVEP EEG signal to be identified, a sinusoidal reference template signal is constructed according to the target stimulus source frequency, and PLSR is used to regress it to the multi-channel sampled EEG signal to obtain an EEG estimation signal; then the multi-channel sampled EEG signal is decomposed into k EEG signal sub-band components through a filter bank; then, each sub-band component and each estimated signal are subjected to an extended canonical correlation analysis algorithm to obtain a weighted sum of the first n canonical correlation coefficients, thereby obtaining a set of extended correlation coefficients of the sub-band components; finally, the extended correlation coefficients of the k sub-band components are weighted and summed to obtain a set of extended correlation coefficients sum of the sampled EEG signal, wherein the fundamental frequency of the reference template signal corresponding to the largest item is the target frequency of the EEG signal to be identified.

[0033] The present invention provides a method for classifying and recognizing EEG signals based on PLSR and extended FBCCA. The process is as follows: Figure 1 As shown, the following steps are included:

[0034] Step 1: Based on the scale of the brain-computer interface and the multi-channel sampled EEG signal, a sinusoidal reference template signal is established and regression estimation is performed using PLSR, as described below:

[0035]

[0036] Among them, s represents the number of multi-channel EEG signal sampling data points, f s represents the sampling rate of the multi-channel EEG signal, f represents the fundamental frequency of the reference template signal, h represents the number of harmonics of the sinusoidal reference template signal, and the sinusoidal reference template signal thus established satisfies Y f ∈R 2h×s , there are N groups in total, where N is the number of stimulation sources and targets of the brain-computer interface system; a PLSR model is established for each sinusoidal reference template signal, which is defined as:

[0037]

[0038] Among them, U and V are the extracted latent component matrices, E and G are the residual matrices of U and V respectively. After using U to represent the dependent variable, the residual matrix is ​​correspondingly represented by F. The iterative calculation is performed with the number of iterations being M, which needs to satisfy M<min{c,2h}. For the i-th m The iteration is defined as i m principal components and Then we have:

[0039]

[0040] Among them, the coefficient and The calculation is based on the partial least squares regression formula as follows:

[0041]

[0042] It can be written in matrix form as:

[0043]

[0044] Where B = PRT is the regression coefficient matrix; after executing M iterations, the partial least squares regression coefficient matrix B that can be used for estimation is obtained f , the EEG estimation signal can be defined as:

[0045] X f =Y f B f (6)

[0046] By estimating N groups of reference template signals, N groups of EEG estimation signals can be obtained for classification and recognition.

[0047] Step 2: Use FB analysis to perform sub-band decomposition of multi-channel EEG signals. The process is as follows: Figure 2 As shown, it is described as follows: The frequency domain characteristic distribution of the EEG signal is measured with the overall upper limit cutoff frequency f of the filter group H and the lower cutoff frequency f L Then, k filters are set according to the preset parameters. The upper and lower cutoff frequencies of each filter are the same, both are f H ; The lower cutoff frequency is different, the i k The lower cutoff frequency of the filter is:

[0048]

[0049] Thus, a set of k filters can be obtained. The multi-channel sampled EEG signal can be passed through the set of filters to obtain k EEG signal sub-band components X b , set different component weights for different sub-band components, the i-th k The weight of the subband components is:

[0050]

[0051] Where k1 and k2 are adjustable weight parameters.

[0052] Step 3: For each EEG signal subband, use the extended FBCCA algorithm to calculate the extended correlation coefficient between it and each estimated signal. The process is as follows: Figure 3 As shown, the description is as follows:

[0053] For multi-channel sampling EEG signal X b ∈R c×s and the EEG estimation signal Xf ∈R 2h×s , calculate X b The sample covariance with itself is denoted as C XX 、X f The sample covariance with itself is denoted as C YY , X b With X f The sample covariance of XY and C YX , calculate the canonical correlation matrix A, specifically:

[0054]

[0055] Obtain the eigenvalues ​​of the typical correlation matrix A and arrange them in descending order as [λ1,λ2,…,λ c ], set the number of features n to be selected to satisfy n<c, then:

[0056]

[0057] in is the i n Typical correlation coefficients under heavy features; different weights are set for typical correlation coefficients under different heavy features, specifically:

[0058]

[0059] Where k3 and k4 are adjustable weight parameters, the EEG signal subband component X obtained by the extended canonical correlation analysis algorithm is b Relative to the estimated signal X f The expanded correlation coefficient is:

[0060]

[0061] Step 4: Calculate the overall correlation coefficient of the EEG signal to the estimated signal, as described below:

[0062]

[0063] Among them, r(X,X f ) is the EEG signal with respect to the estimated signal X f The overall correlation coefficient of is the EEG signal i k subband components about the estimated signal X f The expanded correlation coefficient, w FB (i k ) is the subband weight, from which a set of N overall correlation coefficients can be obtained, where the largest term is the recognition result of the EEG signal to be recognized, which is defined as:

[0064] fresult =f i |maxr(X,X fi )i=1,2,...N (14)

[0065] Among them, f i represents the frequency of the i-th target stimulus source, Represents the EEG signal with respect to the i-th estimated signal The overall correlation coefficient.

[0066] To verify the effectiveness of the method proposed in this invention, offline experiments were conducted on the open source datasets San Diego Square Joint Frequency-Phase Modulation SSVEP Dataset from the University of California, San Diego, and the open source dataset Tsinghua Benchmark SSVEP Database from Tsinghua University. By using the open source datasets for offline experiments, the multi-channel EEG data in the datasets replaced the collected signals in the online device. The effectiveness of the method can be verified by comparing its performance with the benchmark data.

[0067] The San Diego Square Joint Frequency-Phase Modulation SSVEP Dataset (hereinafter referred to as the San Diego dataset) is a public dataset commonly used in the industry. Its stimulus source contains 12 visual stimulus targets. During the experiment, it was displayed on a 27-inch LCD monitor at a refresh rate of 60Hz. The experimental data of 10 different subjects were recorded. The experiment sampled 8 channels of SSVEP data segments within a 4s flickering period at a sampling rate of 2048Hz, and the signal was downsampled to 256Hz. The performance of the algorithm under different data lengths was tested on this dataset with a time step of 0.25s, and the accuracy was compared with the FBCCA method commonly used in the industry. The parameters N=12, c=8, and f=1 were set. s =256, k=3, h=2, n=3.

[0068] Figure 4 The signal recognition accuracy of the method provided by the present invention and the signal recognition accuracy of the FBCCA method at different sampling times on the San Diego dataset are shown. The accuracy here is the average value of the data of 10 subjects in the entire dataset. The recognition accuracy of the method provided by the present invention is higher than that of the FBCCA method at all sampling times.

[0069] Tsinghua Benchmark SSVEP Database (hereinafter referred to as Tsinghua dataset) is another public dataset commonly used in the industry. Its stimulus source contains 40 visual stimulus templates. During the experiment, it was displayed on a 23.6-inch LCD monitor at a refresh rate of 60Hz. The experimental data of 35 different subjects were recorded. The experiment collected 64-channel EEG data over a full 6s process at a sampling rate of 1000Hz and downsampled to 250Hz. The performance of the algorithm under different data lengths was tested on this dataset with a time step of 0.25s. The accuracy was compared with the FBCCA method commonly used in the industry. The parameters were set as N=40, c=9, and f=1. s =250, k=5, h=5, n=3.

[0070] Figure 5 The signal recognition accuracy of the method provided by the present invention and the signal recognition accuracy of the FBCCA method at different sampling times on the Tsinghua dataset are shown. The accuracy here is the average value of the data of 35 subjects in the entire dataset. The recognition accuracy of the method provided by the present invention is higher than that of the FBCCA method at all sampling times.

[0071] The above experiments verify the effectiveness of the method provided by the present invention, which can improve the recognition accuracy and applicability of the SSVEP paradigm brain-computer interface. Specifically, for multi-channel sampled EEG signals of different subjects and different spatiotemporal scenarios, the spatial distribution relationship in the EEG signal can be extracted, and the estimated situation of the reference template signal under the current spatial distribution can be reflected in the estimated signal, thereby improving the accuracy and applicability of signal recognition. When the number of EEG signal acquisition leads is large, the SSVEP EEG signal recognition method based on PLSR and extended FBCCA takes into account the influence of multiple groups of typical variables in the form of main characteristic component expansion, which can improve the accuracy of EEG signal splitting recognition.

[0072] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for EEG signal recognition based on PLSR and extended FBCCA, characterized in that: The following steps are involved: For the multi-channel SSVEP EEG signal to be identified, N sinusoidal reference template signals are established according to the number of target stimulus sources N. The fundamental frequency of each template signal corresponds to the frequency of the target stimulus source. PLSR is used to regress the sinusoidal reference template signal onto the multi-channel sampled EEG signal to obtain N groups of EEG estimation signals. Decompose the multi-channel sampled EEG signal through a filter bank to generate k EEG signal sub-band components; For each sub-band component, the weighted sum of the first n canonical correlation coefficients of each estimated signal is obtained by applying an extended canonical correlation analysis algorithm to the sub-band component, which is used as the extended correlation coefficient of the sub-band component to the single estimated signal. The extended correlation coefficients of the k sub-band components decomposed from the EEG signal to the single estimated signal are weighted and summed as the overall correlation coefficient of the EEG signal to the single estimated signal. For N target stimulus sources, a total of N overall correlation coefficients can be obtained. The fundamental frequency of the reference template signal corresponding to the largest one is the target frequency of the multi-channel SSVEP EEG signal to be identified. For each sub-band component, obtaining a weighted sum of the first n canonical correlation coefficients of each estimated signal through an extended canonical correlation analysis algorithm as the extended correlation coefficient of the sub-band component includes: For the multi-channel SSVEP EEG signal X to be identified b ∈R c×s and the EEG estimation signal X f ∈R 2h×s , calculate X b The sample covariance C with itself XX 、X f The sample covariance with itself is C YY 、X b With X f The sample covariance of XY 、X f With X b The sample covariance of YX , calculate the canonical correlation matrix A: Obtain the c eigenvalues ​​of the typical correlation matrix A and arrange them in descending order as [λ1,λ2,…,λ c ], from which n eigenvalues ​​are selected: For the i n Typical correlation coefficients under heavy features; different weights are set for typical correlation coefficients under different heavy features: Among them, k3 and k4 are adjustable weight parameters, so the obtained X b Relative to X f The expanded correlation coefficient is:

2. The method for recognizing an electroencephalogram signal according to claim 1, wherein: The method of using PLSR to regress the multi-channel sampled EEG signal to obtain the EEG estimation signal is: Build a PLSR model for the sinusoidal reference template signal: Y f =UC T +E X=VD T +G=UR T +F Among them, U and V are the extracted latent component matrices, E and G are the residual matrices of U and V respectively, and F is the residual matrix with U as the dependent variable; the partial least squares regression coefficient matrix B is obtained by iterative calculation M times f , M satisfies M<min{c,2h}; using matrix B f The estimated EEG signal is X f :X f =Y f B f , Y f is the sinusoidal reference template signal.

3. The method for recognizing an electroencephalogram signal according to claim 1, wherein: The weighted sum of the extended correlation coefficients of the k sub-band components decomposed from the EEG signal to the single estimated signal is used as the overall correlation coefficient of the EEG signal to the single estimated signal. The calculation method is: Among them, r(X,X f ) is the overall correlation coefficient, w FB (i k ) is the subband weight.

Images