SSVEP-BCI-oriented supervision domain adaptive cross-subject migration method

Abstract

The invention provides a supervised domain adaptive cross-subject migration method oriented to SSVEP-BCI, and solves the problems that in the prior art, a brain-computer interface is poor in usability, and the migration effect is not ideal when the difference between a source domain and a target domain is too large. The implementation steps are as follows: generating a phase-locked time-shifting data set; calculating a time filter; generating an extended training set; calculating an intra-domain invariant spatial filter and a template signal; calculating an inter-domain invariant spatial filter and a template signal; calculating a domain adaptation conversion matrix; and calculating a combined correlation coefficient identification target. According to the method, the effect of multiple collection can be achieved only through a small amount of collected data, the time structure information of the SSVEP signal can be captured, the defect that visual fatigue is caused by a large amount of calibration work when a new user uses a brain-computer interface is overcome, the usability of the brain-computer interface is improved, and the method can be applied to a spelling system.

Description

Technical Field

[0001] The present invention belongs to the field of signal processing technology, and further relates to a supervised domain adaptation cross-subject migration method for a steady-state visual evoked potential brain-computer interface (SSVEP-BCI) in the field of brain-computer interface technology. The present invention can be used for a spelling system to identify a target during input. Background Art

[0002] Brain-computer interface (BCI) technology can analyze the electrical signals generated by brain activity and realize the interaction between the human brain and computers or other external devices. Non-invasive brain-computer interfaces are non-invasive and convenient, and have been widely used in reality. Among them, the brain-computer interface SSVEP-BCI based on steady-state visual evoked potentials has the characteristics of high stability, rich instruction sets and simple acquisition, which enables SSVEP-BCI to obtain a higher information transmission rate ITR (Information Translation Rate) and a better user experience. SSVEP-BCI is mainly used in spelling systems. Steady-state visual evoked potentials SSVEP (Steady-state Visual Evoked Potentials) decoding technology is the core link of SSVEP-BCI. Traditional decoding technology mainly uses reference signals or the subject's personal calibration data as template signals to train spatial filters to improve the signal-to-noise ratio of the collected SSVEP signals, thereby improving the decoding efficiency of SSVEP signals. However, such methods require a large amount of calibration data to calculate effective spatial filters and template signals, and require cumbersome signal acquisition processes, which leads to high time costs and visual fatigue of the subjects, thus reducing the practicality of SSVEP-BCI.

[0003] Xi'an Jiaotong University disclosed in its patent application "A method for accurate target identification of SSVEP short-time window signals" (patent application number: CN 2023102667752, application publication number: CN 116269445A) a method of using a signal extension model to extend the SSVEP signal to achieve SSVEP accurate target identification. The implementation steps of this method are to construct a signal extension model DP-MAFD-SEM (Signal Extension Model Based on Multichannel Adaptive Fourier Decomposition with Different Phase) using the SSVEP signals corresponding to one or more stimulation targets, extend the SSVEP short signal through the signal extension model, increase the length of the SSVEP signal, and then use the canonical correlation analysis method SE-CCA (Signal-Extension Canonical Correlation Analysis) based on signal extension to complete the identification and classification of characteristic frequencies; increase the length of the SSVEP short-time window signal through signal extension to improve the reliability of the canonical correlation analysis CCA (Canonical Correlation Analysis) based on the covariance matrix estimation method, and achieve a higher recognition accuracy. Although this method can achieve a higher recognition accuracy with a shorter signal, it further improves the information transmission rate of SSVEP-BCI. However, the method still has the disadvantage that it still needs to collect a large number of short-time subject SSVEP signals, making it difficult for new users to quickly use the brain-computer interface system.

[0004] Yan et al. published a cross-subject transfer method for reducing the calibration workload in the SSVEP recognition process in their paper "Cross-subject spatial filter transfer methodfor SSVEP-EEG feature recognition" (10.1088 / 1741-2552 / ac6b57 / J NeuralEng.2022.35483331). The implementation steps of this method are: first, select data with good SSVEP response from the existing user data as the transfer object, and superimpose and average the corresponding stimulation frequencies of multiple trials. Then, the CCA / FBCCA algorithm is used to calculate the spatial projection vector of the superimposed and averaged data, and the vector is transferred to the data of the new user; finally, the feature discrimination coefficient obtained from the new user test data is calculated, and the feature discrimination coefficient is integrated using the transfer classification model and the new user classification model for target recognition. This method uses transfer learning to appropriately reduce the calibration workload for new users using the brain-computer interface system. However, the method still has the disadvantage that the performance improvement is limited when the data between the original subjects and the target subjects are significantly different. Therefore, the transfer learning method cannot fundamentally solve the problem of large brain-computer interface calibration workload. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a supervised domain adaptation cross-subject migration method for SSVEP-BCI, so as to solve the problem that the prior art still needs to collect SSVEP signals of a large number of subjects, thereby causing visual fatigue of the subjects and poor usability of the brain-computer interface, and to solve the problem that in the existing methods, the difference between the source domain and the target domain is too large during the migration process, affecting the migration effect and reducing the recognition accuracy.

[0006] The technical idea to achieve the purpose of the present invention is that the present invention designs a new data enhancement method, which can extract effective spatial filters and template signals when the original data is insufficient, averages the original data for multiple trials to obtain the template signal, then slides the window according to the phase-locked time-shift step size to obtain the phase-locked time-shifted data set, and calculates the time filter corresponding to each stimulus source to perform time filtering on the phase-locked time-shifted data set. The original data set, the template signal calculated from the original data set, and the phase-locked time-shifted data set after time filtering are combined to generate an expanded training data set, which lays the foundation for subsequent domain adaptation and inter-subject migration. The present invention learns the domain-invariant features of SSVEP and generalizes the domain-invariant features to an unknown test domain. The generated invariant spatial filters within the domain and the invariant spatial filters between domains can perform spatial filtering on the samples in the training set respectively to obtain invariant template signals within the domain and invariant template signals between domains. In this way, the frequency domain features of the target can be enhanced through the invariant template signals within the domain, and the correlation between different subjects can be maximized through the invariant signals between domains, thereby enhancing generalization, thereby solving the problem of incomplete feature extraction of user data in the prior art. The present invention transfers the spatial filter of the source subject to the SSVEP data of the new user through correlation maximization and the conversion matrix between the source domain and the target domain to complete the inter-domain migration. The target subject data is filtered by two types of spatial filters obtained from the source subject, and the three correlation coefficients between the filtered data and the template signal and the data converted by the conversion matrix are calculated to construct the feature vector to complete the target recognition task, thereby reducing the tedious and time-consuming calibration work of the subject in the SSVEP detection, and further improving the accuracy of SSVEP recognition through multiple correlation coefficients.

[0007] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0008] Step 1, setting the phase-locked time-shift step length according to the period corresponding to the stimulation frequency, and generating a phase-locked time-shift data set for each stimulation source;

[0009] Step 2, calculate the time filter;

[0010] Step 3, generate an expanded training set;

[0011] Step 4, using the in-domain invariant spatial filter, calculate the in-domain invariant template signal for each source subject in the training set under each stimulus;

[0012] Step 5, using the inter-domain invariant spatial filter, calculate the inter-domain invariant template signal for each source subject in the training set under each stimulus;

[0013] Step 6, calculate the domain adaptation transformation matrix to transfer the spatial filter of the source subject to the SSVEP data of the new user;

[0014] Step 7, calculate the three correlation coefficients corresponding to each stimulus target for the SSVEP signal to be identified, take the average value of the three correlation coefficients as the combined correlation coefficient, and take the stimulus target corresponding to the maximum correlation coefficient as the identification result of the SSVEP signal.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] First, the data enhancement method designed by the present invention can dynamically intercept the original SSVEP signal based on the sliding window strategy, and only a small amount of collected data is needed to achieve the effect of multiple collections and extract effective spatial filters and template signals. At the same time, the temporal local covariance of the signal is calculated to capture the intrinsic temporal structure information of the SSVEP signal, improve the signal-to-noise ratio of the SSVEP, and overcome the problem of limited performance improvement of traditional transfer learning when data is scarce and there are significant differences between domains, so that the present invention further improves the efficiency of SSVEP decoding. Therefore, the accuracy of subsequent target recognition based on the SSVEP signal can be completed efficiently and accurately.

[0017] Second, the intra-domain invariant template signal and inter-domain invariant template signal calculated in the present invention can not only enhance the frequency domain characteristics of the target through the intra-domain invariant template signal, but also maximize the correlation between different subjects through the inter-domain invariant signal, thereby avoiding the shortcomings of incomplete extraction of user data features in existing algorithms, and enabling the present invention to improve the generalization ability of brain-computer interfaces.

[0018] Third, the domain adaptation cross-subject migration method proposed in the present invention migrates the generalized spatial filter and template signal learned by the source domain subject to the target domain subject without accessing the SSVEP data of the target subject. From the user's perspective, the problem of visual fatigue caused by a large amount of calibration work when new users use the brain-computer interface is overcome, so that the present invention improves the usability of the brain-computer interface. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the present invention;

[0020] Figure 2 It is a comparison result diagram of the simulation experiment of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] Reference Figure 1 , further describing the implementation steps of the embodiment of the present invention.

[0023] Step 1: Set the phase-locked time-shift step size according to the period corresponding to the stimulation frequency to generate a phase-locked time-shift data set for each stimulation source.

[0024] The steps for generating a phase-locked time-shift dataset for each stimulus source are as follows:

[0025] In the first step, the average value of the SSVEP of all acquisition trials corresponding to each stimulus source is taken, and this average value is used as the template signal of the stimulus source.

[0026] In the second step, the inverse of the period corresponding to the stimulation frequency of each stimulation source is set as the phase-locked time shift step.

[0027] The third step is to slide the time window once for each template signal every phase-locked time-shift step to obtain N intercepted data segments, and all data segments are combined into the phase-locked time-shift data set of the stimulus source, where the value range of N is h represents the length of the training time window, and its value increases by 0.2 in the interval [0.2, 2].

[0028] The fourth step is to intercept a segment of data of t seconds from the start time of each template signal, and use the data segment as the original template signal of the template signal, where t=h.

[0029] Step 2, the steps to calculate the time filter are as follows:

[0030] The first step is to construct the adjacency matrix of the SSVEP signal. The matrix is ​​a two-dimensional matrix with the number of rows and columns equal to the number of sampling points of the SSVEP signal.

[0031] The element values ​​of the adjacency matrix are obtained by the following formula:

[0032]

[0033] Where A represents the adjacency matrix of the SSVEP signal, t1 and t2 represent the number of rows and columns of the adjacency matrix, K represents Tukey's tricube weighting function, and ε represents the weight coefficient, which is set to N s represents the number of sampling points of the SSVEP signal, and f represents the frequency of the stimulus source.

[0034] The second step is to calculate four time-local covariance matrices:

[0035]

[0036] Among them, C xx Represents the original template signal The temporal local covariance matrix, C rr represents the time local covariance matrix of the sine and cosine reference signals, C xr represents the time local covariance matrix between the original template signal and the sine and cosine reference signals, C rxrepresents the temporal local covariance matrix between the sine and cosine reference signals and the original template signal, L represents the Laplacian matrix, L = B–A, B represents the diagonal matrix composed of the diagonal elements of the adjacency matrix A, R n It represents the sine and cosine reference signals corresponding to the nth stimulus source, and the superscript T represents the transposition operation.

[0037] In the third step, the temporal filter is calculated as follows:

[0038]

[0039] Among them, ψ n represents the temporal filter corresponding to the nth stimulus source, It represents the value of ψ when the objective function reaches its maximum value. n Value, the superscript -1 indicates the matrix inversion operation.

[0040] Step 3, generate an extended training set. The extended training set is composed of the original data set, the phase-locked time-shifted data set after time filtering, and the original template signal, wherein the original data set is obtained by intercepting the data segment of the training time window length from the start time of all trial SSVEP signals corresponding to each stimulus.

[0041] Step 4, using the in-domain invariant spatial filter, calculate the in-domain invariant template signal for each source subject in the training set under each stimulus.

[0042] The in-domain invariant spatial filter is obtained by the following objective function:

[0043]

[0044] in, represents the domain-invariant spatial filter corresponding to the nth stimulus target of the mth source subject, m represents the index of the source subject in the training set, m = 1, 2, ..., M, M represents the total number of source domain subjects, Indicates dimension N c ×1 real number set, N c Indicates the number of channels of SSVEP data, set to 9, Indicates the w corresponding to the maximum value of the objective function m,n value, represents the SSVEP data corresponding to the nth stimulus target of the mth source subject, express The SSVEP data corresponding to the lth neighboring stimulus is the SSVEP data corresponding to the lth neighboring stimulus. The neighboring stimulus refers to the stimulus targets adjacent to each stimulus target in the horizontal and vertical directions. The “—” above indicates the mean operation, cov indicates the covariance operation, k indicates the number of K-1 times of data for each sample in the training set, and Nt Represents the number of trials, N t =1,2,...,N l , N l Represents the number of neighboring stimuli of a single stimulus source.

[0045] The domain invariant template signal is obtained by the following objective function:

[0046]

[0047] Among them, T m,n Represents the in-domain invariant template signal corresponding to the nth stimulus target of the mth source subject.

[0048] Step 5, using the inter-domain invariant spatial filter, calculate the inter-domain invariant template signal for each source subject in the training set under each stimulus.

[0049] The inter-domain invariant spatial filter is obtained by the following formula:

[0050]

[0051] Among them, v n represents the inter-domain invariant spatial filter corresponding to the nth stimulus target of all source domain subjects, m1 and m2 represent two different source subject indices, respectively. represents the SSVEP data corresponding to the nth stimulus of the m1th source subject, represents the SSVEP data corresponding to the nth stimulus of the m2th source subject in the expanded training set, Indicates the v corresponding to the maximum value of the objective function n value.

[0052] The inter-domain invariant template signal is obtained by the following formula:

[0053]

[0054] Among them, Z n represents the inter-domain invariant template signal corresponding to the nth stimulus target of all source domain subjects, Represents the mean of the SSVEP data of all source domain subjects at the nth stimulation target.

[0055] Step 6, calculate the domain adaptation transformation matrix to transfer the spatial filter of the source subject to the SSVEP data of the new user.

[0056] The domain adaptation transformation matrix is ​​as follows:

[0057] D=w left w rightT

[0058] K=v left v rightT

[0059] Among them, D and K represent the intra-domain adaptation transformation matrix and the inter-domain adaptation transformation matrix respectively, and w left represents the left singular vector obtained by singular value decomposition of the invariant spatial filter in the domain, w right represents the right singular vector obtained by singular value decomposition of the invariant spatial filter in the domain; v left represents the left singular vector obtained by singular value decomposition of the inter-domain invariant spatial filter, v right Represents the right singular vectors obtained by singular value decomposition of the invariant spatial filter in the domain.

[0060] The method to transfer the spatial filter of the source subject to the SSVEP data of the new user is as follows:

[0061] E=DZ,F=KZ

[0062] Among them, Z represents the new user SSVEP to be identified, E represents the new user SSVEP after intra-domain adaptive conversion, and F represents the new user SSVEP after inter-domain adaptive conversion.

[0063] Step 7, calculate the three correlation coefficients corresponding to each stimulus target for the SSVEP signal to be identified, take the average value of the three correlation coefficients as the combined correlation coefficient, and take the stimulus target corresponding to the maximum correlation coefficient as the identification result of the SSVEP signal.

[0064] The three correlation coefficients are calculated as follows:

[0065]

[0066] in, It represents the correlation coefficient between the average SSVEP of all subjects corresponding to the nth stimulus target after inter-domain filtering and the SSVEP to be identified after inter-domain adaptation 1; The correlation coefficient between the SSVEP of the nth stimulus of the mth subject after in-domain filtering and the SSVEP to be identified after in-domain adaptation is 2; represents the correlation coefficient between the adaptation data corresponding to the nth stimulus source and the SSVEP to be identified after filtering by the adaptation data filter3, Y n represents the adaptation data corresponding to the nth stimulus, selected from the source domain, and ρ() represents the Pearson correlation coefficient.

[0067] The average value of the three correlation coefficients is taken as the combined correlation coefficient, and the stimulation target corresponding to the maximum correlation coefficient is taken as the recognition result of the SSVEP signal.

[0068] The implementation process of the present invention is further described below in conjunction with simulation experiments:

[0069] 1. Simulation conditions:

[0070] The hardware test platform of the simulation experiment of the present invention is: the CPU is Intel (R) Core (TM) i7-8700, the main frequency is 3.2GHz, the memory is 16GB, and the GPU is NVIDIA GeForce GT 710.

[0071] The software testing platform of the simulation experiment of the present invention is: Widows11 operating system and mathematical software MATLABR2022b.

[0072] The simulation experiment of the present invention uses the following three data sets, namely:

[0073] Dataset 1, a publicly available SSVEP dataset of 12 stimulus targets from the University of California. The stimulus targets are evenly arranged in 4 rows and 3 columns. The targets generate sine signals through joint frequency phase modulation to produce flickering stimuli. The frequency range is 9.25-14.75 Hz, the step length is 0.5 Hz, and the phase range is 0-1.5π, the step length is 0.5π. A total of 10 subjects' SSVEP data were collected, with a sampling frequency of 256 Hz. The subjects conducted 15 experiments on each stimulus target.

[0074] Dataset 2, a benchmark dataset published by Tsinghua University that contains 40 stimulus targets. The stimulus targets are evenly arranged in 5 rows and 8 columns. The targets generate sine signals through joint frequency phase modulation to produce flickering stimuli. The frequency range is 8-15.8Hz, the step length is 0.2Hz, and the phase range is 0-1.5π, with a step length of 0.5π. A total of 35 subjects' SSVEP data were collected with a sampling frequency of 250Hz. The subjects conducted 6 experiments on each stimulus target.

[0075] Dataset 3, a beta dataset of 40 stimulus targets published by Tsinghua University. The stimulus targets are arranged according to the graphical interface of the traditional QWERT keyboard (including 10 numbers, 26 letters and 4 non-alphanumeric symbols). The frequency and initial phase of each target are encoded using joint frequency and phase modulation. The frequency range is 8-15.8Hz, the step length is 0.2Hz, and the phase range is 0-1.5π, the step length is 0.5π. A total of 70 subjects' SSVEP data were collected, with a sampling frequency of 250Hz, and the subjects conducted 4 experiments on each stimulus target.

[0076] 2. Simulation content and result analysis:

[0077] The simulation experiment of the present invention adopts the method of the present invention and an existing technology to identify the SSVEP data of three data sets respectively.

[0078] The existing technology used refers to a method for identifying SSVEP signals proposed by Chen Xiaogang et al. in their published paper “Filter bank canonical correlation analysis for implementing a high-speed SSVEP-based brain-computer interface” (JNE-100668.R1).

[0079] The process of identifying the SSVEP data in each data set in the simulation experiment of the present invention is as follows:

[0080] In the first step, the SSVEP data in the dataset are averaged over multiple trials, and the window is slid according to the phase-locked time-shift step to obtain the intercepted data in each window to generate a phase-locked time-shift dataset.

[0081] In the second step, the time filter is calculated to perform time filtering on the phase-locked time-shifted data set, and the original data set, the original template signal and the phase-locked time-shifted data set after time domain filtering are merged to generate an expanded training data set.

[0082] In the third step, the SSVEP data in the training set is used to obtain the intra-domain invariant spatial filter and template signal, and the inter-domain invariant spatial filter and template signal.

[0083] In the fourth step, the single data in the training set is subjected to in-domain invariant spatial filtering, and the correlation between the filtered data and the in-domain invariant template signal is maximized. At the same time, the single data is subjected to inter-domain invariant spatial filtering, and the correlation between the filtered data and the inter-domain invariant template signal is maximized to obtain the target domain filter. The target domain filter is subjected to singular value decomposition of the product of the in-domain invariant filter and the inter-domain invariant filter, and the domain adaptation transformation matrix is ​​calculated.

[0084] The fifth step is to calculate three different types of correlation coefficients between the target domain adapted test data and the migration template signal, and between the spatially filtered test data and the domain adapted data. The SSVEP data of the new user is identified using the trained filters and template signals.

[0085] In order to verify the effect of the present invention, the recognition accuracy is calculated by the following formula:

[0086]

[0087] Among them, P represents the recognition accuracy.

[0088] Figure 2 It is a curve diagram of the accuracy of the method of the present invention and the FBCCA algorithm on three different data sets, wherein the abscissa represents the SSVEP length in seconds (s), and the ordinate represents the accuracy. Figure 2 The solid line in represents the accuracy curve of the method of the present invention, and the dotted line represents the accuracy curve of the FBCCA algorithm.

[0089] Figure 2 (a) is the accuracy curve of the method of the present invention and the FBCCA algorithm on the benchmark data set. Figure 2 The solid curve in (a) is always higher than the dotted curve when the horizontal coordinate is 0.2 to 2, which shows that the accuracy of the method of the present invention is better than the FBCCA algorithm when identifying all signal lengths.

[0090] Figure 2 (b) is the accuracy curve of the method of the present invention and the FBCCA algorithm on the UCSD data set. Figure 2 The solid curve in (b) is always higher than the dotted curve when the horizontal coordinate is 0.2 to 1.8, and the two are close at 2s. This shows that the accuracy of the method of the present invention is better than that of the FBCCA algorithm when identifying signal lengths other than 2s, mainly because the FBCCA algorithm only uses the sine and cosine reference signals that do not contain the personal characteristics of the subject.

[0091] Figure 2 (c) is the accuracy curve of the method of the present invention and the FBCCA algorithm on the beta data set. Figure 2 The solid curve in (c) is always higher than the dotted curve when the horizontal coordinate is 0.2 to 1.4, and the dotted curve is slightly higher than the solid line after the horizontal coordinate is 1.4. This shows that the accuracy of the method of the present invention is better than that of the FBCCA algorithm when identifying signal lengths other than 1.2s, and the trends of the accuracy of the two are similar after 1.2s. This is mainly because the beta data set contains the least number of acquisition trials, which makes it impossible for the method of the present invention to obtain a domain-invariant filter and template signal with sufficient features. However, a signal length that is too long will greatly reduce the efficiency of recognition. Therefore, a high recognition accuracy rate for short-term SSVEP signals is more realistic.

[0092] In summary, the test accuracy of the method designed by the present invention on the three data sets is better than that of the comparison method FBCCA. At the same time, as the signal length increases, the recognition accuracy is also continuously improved. The present invention enables the target subject to directly use the brain-computer interface system without pre-training or calibration. At the same time, it can also expand the data set that retains the original features when there is less data. It solves the problem of incomplete extraction of user data features in the prior art, which easily causes visual fatigue of users and cannot face the problem of significant differences between source domain and target domain data. It is a very practical brain-computer interface recognition method.

Claims

1. A supervised domain adaptation cross-subject transfer method for SSVEP-BCI, characterized in that: Based on the sliding window strategy, the original SSVEP signal is dynamically intercepted, the intra-domain invariant template signal and the inter-domain invariant template signal are calculated respectively, and the spatial filter of the source subject is transferred to the SSVEP data of the new user to complete the inter-domain cross-subject transfer; the steps of the target recognition method are as follows: Step 1, setting the phase-locked time-shift step length according to the period corresponding to the stimulation frequency, and generating a phase-locked time-shift data set for each stimulation source; Step 2, calculate the time filter; Step 3, generate an expanded training set; Step 4, using the in-domain invariant spatial filter, calculate the in-domain invariant template signal for each source subject in the training set under each stimulus; Step 5, using the inter-domain invariant spatial filter, calculate the inter-domain invariant template signal for each source subject in the training set under each stimulus; Step 6, calculate the domain adaptation transformation matrix to transfer the spatial filter of the source subject to the SSVEP data of the new user; Step 7, calculate the three correlation coefficients corresponding to each stimulus target for the SSVEP signal to be identified, take the average value of the three correlation coefficients as the combined correlation coefficient, and take the stimulus target corresponding to the maximum correlation coefficient as the identification result of the SSVEP signal.

2. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 1, characterized in that: The steps to generate a phase-locked time-lapse dataset for each stimulus source described in step 1 are as follows: The first step is to take the average value of the SSVEP of all the acquisition trials corresponding to each stimulus source, and use this average value as the template signal of the stimulus source; The second step is to set the inverse of the period corresponding to the stimulation frequency of each stimulation source as the phase-locked time shift step; The third step is to slide the time window once for each template signal every phase-locked time-shift step to obtain N intercepted data segments, and all data segments are combined into the phase-locked time-shift data set of the stimulus source, where the value range of N is h represents the length of the training time window, and its value increases by 0.2 in the interval [0.2,2]; The fourth step is to intercept a segment of data of t seconds from the start time of each template signal, and use the data segment as the original template signal of the template signal, where t=h.

3. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 2, characterized in that: The steps for calculating the temporal filter described in step 2 are as follows: The first step is to construct the adjacency matrix of the SSVEP signal. The matrix is ​​a two-dimensional matrix with the number of rows and columns equal to the number of sampling points of the SSVEP signal. The element value of the matrix is ​​obtained by the following formula: Where A represents the adjacency matrix of the SSVEP signal, t1 and t2 represent the number of rows and columns of the adjacency matrix, K represents Tukey's tricube weighting function, and ε represents the weight coefficient, which is set to N s represents the number of sampling points of the SSVEP signal, and f represents the frequency of the stimulus; The second step is to calculate four time-local covariance matrices: Among them, C xx Represents the original template signal The temporal local covariance matrix, C rr represents the time local covariance matrix of the sine and cosine reference signals, C xr represents the time local covariance matrix between the original template signal and the sine and cosine reference signals, C rx represents the temporal local covariance matrix between the sine and cosine reference signals and the original template signal, L represents the Laplacian matrix, L = B–A, B represents the diagonal matrix composed of the diagonal elements of the adjacency matrix A, R n represents the sine and cosine reference signals corresponding to the nth stimulus source, and the superscript T indicates the transposition operation; In the third step, the temporal filter is calculated as follows: Among them, ψ n represents the temporal filter corresponding to the nth stimulus source, It represents the value of ψ when the objective function reaches its maximum value. n Value, the superscript -1 indicates the matrix inversion operation.

4. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 1, characterized in that: The expanded training set described in step 3 is composed of the original data set, the phase-locked time-shifted data set after time filtering, and the original template signal, wherein the original data set is obtained by intercepting a data segment of the training time window length from the start time of all trial SSVEP signals corresponding to each stimulus.

5. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 3, characterized in that: The domain-invariant spatial filter in step 4 is obtained by the following objective function: in, represents the domain-invariant spatial filter corresponding to the nth stimulus target of the mth source subject, m represents the index of the source subject in the training set, m = 1, 2, ..., M, M represents the total number of source domain subjects, Indicates dimension N c ×1 real number set, N c Indicates the number of channels of SSVEP data, set to 9, Indicates the w corresponding to the maximum value of the objective function m,n value, represents the SSVEP data corresponding to the nth stimulus target of the mth source subject, express The SSVEP data corresponding to the lth neighboring stimulus is the SSVEP data corresponding to the lth neighboring stimulus. The neighboring stimulus refers to the stimulus targets adjacent to each stimulus target in the horizontal and vertical directions. The "—" above indicates the mean operation, cov indicates the covariance operation, k indicates the number of K-1 times of data for each sample in the training set, and N t Indicates the number of trials, N l Represents the number of neighboring stimuli of a single stimulus source.

6. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 5, characterized in that: The domain-invariant template signal described in step 4 is obtained by the following objective function: Among them, T m,n Represents the in-domain invariant template signal corresponding to the nth stimulus target of the mth source subject.

7. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 6, characterized in that: The inter-domain invariant spatial filter described in step 5 is obtained by the following formula: Among them, v n represents the inter-domain invariant spatial filter corresponding to the nth stimulus target of all source domain subjects, m1 and m2 represent two different source subject indices, respectively. represents the SSVEP data corresponding to the nth stimulus of the m1th source subject, represents the SSVEP data corresponding to the nth stimulus of the m2th source subject in the expanded training set, Indicates the v corresponding to the maximum value of the objective function n value.

8. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 7, characterized in that: The inter-domain invariant template signal described in step 5 is obtained by the following formula: Among them, Z n represents the inter-domain invariant template signal corresponding to the nth stimulus target of all source domain subjects, Represents the mean of the SSVEP data of all source domain subjects at the nth stimulation target.

9. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 8, characterized in that: The domain adaptation transformation matrix described in step 6 is as follows: Among them, D and K represent the intra-domain adaptation transformation matrix and the inter-domain adaptation transformation matrix respectively, and w left represents the left singular vector obtained by singular value decomposition of the invariant spatial filter in the domain, w right represents the right singular vector obtained by singular value decomposition of the invariant spatial filter in the domain; v left represents the left singular vector obtained by singular value decomposition of the inter-domain invariant spatial filter, v right Represents the right singular vectors obtained by singular value decomposition of the invariant spatial filter in the domain.

10. The supervised domain adaptation cross-subject transfer method for SSVEP-BCI according to claim 9, characterized in that: The method for transferring the spatial filter of the source subject to the SSVEP data of the new user described in step 6 is as follows: E=DZ,F=KZ Among them, Z represents the new user SSVEP to be identified, E represents the new user SSVEP after intra-domain adaptive conversion, and F represents the new user SSVEP after inter-domain adaptive conversion.

Images