A cross-subject speech imagery EEG decoding method with privacy protection function

Abstract

The present invention relates to a cross-subject speech imagination EEG decoding method with a privacy protection function. The EEG data of the subjects under the speech imagination task are first collected and preprocessed to generate a target sample matrix. Then, the model is trained using known source domain samples, and the model is used to predict the target domain samples. According to the prediction results, some target domain samples are screened out to construct an intermediate domain. Finally, the source domain and the target domain are aligned through a domain adaptation method to achieve the goal of protecting the privacy of the source domain data and improving the decoding accuracy of the target domain data. This method uses a power average index to screen the target domain samples for constructing the intermediate domain, and better screens the target domain samples that are most similar to the source domain under the premise of protecting the privacy of the source domain data, so that these high transfer samples play a greater role in the transfer learning process.

Description

Technical Field

[0001] The invention relates to a cross-subject speech imagination EEG decoding method with a privacy protection function, and belongs to the technical field of EEG signal processing. Background Art

[0002] As a high-level cognitive behavior, language dialogue is the core means of communication in human society. As an important part of language processing, speech imagination refers to the construction, simulation or recall of language expression by individuals at the mental level, without involving actual pronunciation or writing behavior. This internal psychological process involves the coordinated work of multiple functional areas of the brain, especially Broca's area and superior temporal gyrus, which play a key role in speech production and comprehension. When individuals engage in speech imagination, the corresponding neural activity will produce detectable potential changes in the cerebral cortex, which are recorded as electroencephalography (EEG). As a physiological signal that directly reflects the activity of the cerebral cortex, EEG provides a valuable source of biological data information for the study and analysis of speech imagination. Speech imagination decoding is an interdisciplinary research direction that integrates computer science, psychology, neuroscience and brain cognitive science, aiming to develop a theoretical framework, methodological system and technical platform that can effectively process, analyze and identify human speech imagination task information. Among them, EEG speech imagination decoding is an emerging research branch in this field, which focuses on analyzing and interpreting EEG signals related to speech imagination. Compared with traditional input signals such as video, images, voice and text, EEG signals have unique physiological characteristics and are difficult to disguise artificially. Therefore, speech imagination decoding based on EEG signals can more accurately capture and reveal the individual's true inner speech imagination content.

[0003] Cross-subject EEG decoding of speech imagery is one of the most practical directions in EEG decoding. This is because the cross-subject EEG decoding model can extract universal features from the EEGs of different subjects, thereby creating a generalized decoder model that can be applied to unseen subjects; in practical applications such as neurorehabilitation, users hope to use the EEG decoding system directly without long-term personalized training, and the cross-subject decoding model can effectively solve this demand.

[0004] In the process of cross-subject EEG decoding, we must also consider protecting the privacy and security of the source subject's data. Neural signal data such as electroencephalogram (EEG) are highly sensitive personal information, which contains private information such as an individual's thoughts, emotions, and even potential health conditions. As a cutting-edge scientific research field, cross-subject speech imagery EEG signal decoding must strictly comply with relevant laws and regulations in data processing to avoid causing international disputes and legal liabilities.

[0005] As for personal privacy, an individual's EEG activity is directly related to his thoughts and inner world, and is one of the most private information. When decoding EEG signals of speech imagination across subjects, collecting, analyzing, and disseminating such information without consent or taking appropriate protective measures seriously infringes on the individual's privacy rights. With the advancement of technology, EEG signal data can theoretically be used to reconstruct the content of personal thoughts, and may even be used for remote control or psychological analysis in the future. The lack of effective privacy protection measures may lead to malicious use of an individual's thoughts, emotions, or behavior patterns, resulting in serious consequences such as identity theft and psychological manipulation. For scientific research projects, ensuring that participants feel assured of the security of their personal information is a prerequisite for attracting volunteers to participate and obtaining high-quality data. If privacy protection is insufficient, it will lead to public distrust of scientific research activities, affecting the progress of scientific research and social acceptance.

[0006] However, achieving accurate EEG speech imagery decoding faces many challenges, especially in cross-subject scenarios. Due to significant biological differences between individuals and the dynamic changes of EEG signals of the same subject under different states and times, the distribution of EEG signals shows obvious non-independent and identically distributed characteristics, which seriously hinders the generalization performance of traditional machine learning algorithms that rely on the independent and identically distributed assumption in cross-subject speech imagery decoding tasks. To solve this problem, the concept of transfer learning was introduced, and its core goal is to narrow the domain gap between the source domain (labeled data set) and the target domain (subjects to be decoded) and improve the model's cross-subject adaptability.

[0007] Although existing transfer learning methods such as the maximum mean difference method and deep neural network transfer have alleviated the domain difference problem to a certain extent, they still have some limitations. On the one hand, these methods often find it difficult to intuitively quantify the transferability of samples, which affects the transparency and interpretability of the transfer learning process. On the other hand, when dealing with cross-subject EEG speech imagery decoding tasks, they may ignore the subtle differences in EEG features between individuals and their impact on the decoding effect. Summary of the invention

[0008] In order to overcome the technical deficiencies of the existing EEG speech imagination decoding methods, the present invention provides a cross-subject speech imagination EEG decoding method with privacy protection function. The method proposed in the present invention aims to combine the sample selection strategy in the field of transfer learning, and by constructing an intermediate domain, effectively utilize the source domain sample information to screen the target domain samples. The present invention constructs an intermediate domain by proposing a power mean exponent and a set threshold to screen the target domain samples that are similar to the source domain. Under the premise of protecting privacy, it can also better predict the unlabeled target domain samples and improve the prediction accuracy; after obtaining the intermediate domain, a discriminant transfer regression method of joint low-rank and sparse subspace learning is constructed to obtain the transformation matrix and projection matrix to align the intermediate domain with the target domain, which can minimize the difference between the source domain and the target domain, and thereby improve the accuracy of EEG state recognition.

[0009] The specific steps of a cross-subject speech imagery EEG decoding method with privacy protection function include:

[0010] Step 1: Collect EEG data of the target subjects under different speech imagery tasks;

[0011] Step 2: Preprocess and extract features of the EEG data collected in step 1 to obtain the sample matrix of the subject d is the characteristic number, n is t is the total number of samples, as the target domain sample matrix; the source domain sample matrix and the corresponding label matrix [X s ; Y s ] It is known that the local source domain model is obtained by training the support vector machine model, and the local source domain model is used to train the target domain sample matrix X t Make a prediction to get the target domain pseudo label probability

[0012] Step 3: Construct a speech imagery EEG signal decoding method for cross-subject transfer learning, whose objective function is as follows; According to the objective function established in and the set threshold, the target domain samples are screened to construct an intermediate domain [X tmp ; Y tmp ];

[0013] Step 4: According to the intermediate domain [X tmp ; Y tmp ] and the collected target domain samples [X t ; Y t ], construct a discriminative transfer regression method for joint low-rank and sparse subspace learning, whose objective function is as follows:

[0014]

[0015] stAP T Xt =AP T X tmp Z+E,M≥0,Z≥0,Y=[Y tmp , Y t ]

[0016] in and are two projection matrices, d represents the characteristic dimension of the projection subspace. The size of the original transformation matrix B is m×c, and the sizes of the projection matrix P and matrix A are m×d and c×d respectively. X=[X tmp ;X t ], is the label matrix of the intermediate domain samples, represents the label relaxation matrix, ||Z|| * is the nuclear norm of the matrix Z, Z is the reconstruction coefficient matrix, Z = [z ij ], α, γ represent regularization parameters, ||E||1 represents the l1 norm of the matrix E, is the noise matrix, represents the i-th sample feature of the middle domain, Represents the jth sample feature of the target domain sample. represents the spectral norm of the matrix;

[0017] Step 5: According to the objective function established in step 4, transform matrix A, projection matrix P, label relaxation matrix M, reconstruction coefficient matrix Z, noise matrix E, and target domain label matrix Y t Joint iterative optimization of

[0018] Step 6: According to the target domain label matrix Y t Obtain the verbal imagery task of the target subject during EEG acquisition.

[0019] The step 2 specifically includes:

[0020] Step 2.1: Preprocessing and feature extraction: The EEG data recorded at 1024 Hz are sampled at a sampling rate of 256 Hz, and then the noise and artifacts related to EMG activity are filtered out through a fifth-order Butterworth bandpass filter with a frequency range of 1 Hz to 70 Hz. The state characteristics of the EEG signal in different frequency bands are extracted, and the spectral entropy of the EEG signal is calculated as the target domain sample matrix X t ;

[0021] Step 2.2: Build a support vector machine model to train the source domain sample matrix and label matrix to obtain a local source domain model, and use the source domain model to predict the target domain sample matrix to obtain the target domain pseudo label probability

[0022] The specific process of constructing the speech imagery EEG signal decoding method for cross-subject transfer learning in step 3 is as follows:

[0023] Step 3.1: Use the power average exponential strategy as a metric to measure the sample similarity between the source domain and the target domain, as shown in Formula;

[0024]

[0025] Step 3.2: According to the calculated power mean index Mp of each target domain sample and the set threshold, select the set of qualified samples and pseudo labels as the intermediate domain [X tmp ; Y tmp ].

[0026] The specific solution process of constructing the discriminant transfer regression method of joint low-rank and sparse subspace learning in step 4 is:

[0027] Step 4.1: Convert the objective function of the formula into a Lagrangian function for solution;

[0028] Step 4.2: Fix other variables and update the transformation matrix A as follows:

[0029]

[0030]

[0031] stAP T X t =AP T X tmp Z+E, Z=Z1, M≥0, Z≥0, Y=[Y tmp ; Y t ]

[0032] Solve the equation to get the updated transformation matrix A; iteratively update each variable by fixing other variables;

[0033] Step 4.3: Fix other variables and update the projection matrix P as follows:

[0034]

[0035] stAP T X t =AP T X tmp Z+E, Z=Z1, M≥0, Z≥0, Y=[Y tmp ; Y t ]

[0036] Solve the equation to get the updated projection matrix P; solve the optimal solution of the transformation matrix A and the projection matrix P through iterative updating;

[0037] Step 4.4: Fix other variables and update the label relaxation matrix M as follows:

[0038]

[0039] stM≥0

[0040] Solve the equation and get the updated label relaxation matrix M;

[0041] Step 4.5: Fix other variables and update the reconstruction coefficient matrix Z as follows:

[0042]

[0043] stAP T X t =AP T X tmp Z+E, Z≥0

[0044] Solve the formula and get the updated reconstruction coefficient matrix Z;

[0045] Step 4.6: Fix other variables and update the noise matrix E:

[0046]

[0047] stAP T X t =AP T X tmp Z+E

[0048] Solve the equation and get the updated noise matrix E;

[0049] Step 4.7: Fix other variables and update the target domain label matrix Y t :

[0050]

[0051] stY=[Y tmp , Y t ],Y t ≥0, Y t 1=1

[0052] Solve the formula and get the updated target domain label matrix Y t ;

[0053] Step 4.8: Repeat steps 4.2, 4.3, 4.4, 4.5, 4.6 and 4.7 alternately until the objective function converges, and complete the transformation matrix A, projection matrix P, label relaxation matrix M, reconstruction coefficient matrix Z, noise matrix E, and target domain label matrix Y. t Joint iterative optimization.

[0054] The power mean exponent characterizes the similarity between target domain samples and source domain samples during the transfer learning process, and the threshold quantifies whether each target domain sample is transferable.

[0055] The similarity between the source domain and the target domain is measured using the power mean exponent by constructing the intermediate domain sample matrix and its corresponding label matrix.

[0056] In step 4.1, according to the definition of Eq., the optimal problem is transformed into a Lagrangian function;

[0057] In steps 4.2 to 4.6, each variable is iteratively updated by fixing the other variables.

[0058] In step 4.7, the target domain sample label matrix Y is solved by iterative updating and updating convergence. t The weight coefficient matrix Z obtained in step 4 is used to achieve effective alignment of intermediate domain and target domain samples.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] In cross-subject EEG state recognition experiments, inaccurate recognition often occurs due to large differences between subjects. Therefore, it is expected to reduce inter-domain differences by constructing a mapping matrix to map the source domain and target domain to a subspace. The present invention uses a discriminant transfer regression method that combines low-rank and sparse subspace learning to perform feature transformation, and maps the intermediate domain and target domain with high consistency with the source domain to subspaces, which greatly reduces inter-domain differences. At the same time, the present invention introduces a weight coefficient matrix in the process of reducing inter-domain differences, which quantitatively represents the transferability of source domain samples, and promotes source domain samples with high transferability to play a greater role in the transfer learning process, solving the problem of inaccurate EEG state recognition caused by individual differences in the field of cross-subject EEG state recognition.

[0061] The present invention first collects the EEG data of the subjects under the speech imagination task, and preprocesses it to generate a target sample matrix. Then, the model is trained using known source domain samples, and the model is used to predict the target domain samples, and some target domain samples are screened out according to the prediction results to construct an intermediate domain. Finally, the source domain and the target domain are aligned by a domain adaptation method to achieve the goal of protecting the privacy of the source domain data and improving the decoding accuracy of the target domain data. This method uses a power average index to screen the target domain samples for constructing the intermediate domain, and better screens the target domain samples that are most similar to the source domain under the premise of protecting the privacy of the source domain data, so that these high transfer samples play a greater role in the transfer learning process. At the same time, a discriminant transfer regression method for joint low-rank and sparse subspace learning is introduced. It can not only eliminate the influence of outliers and noise, avoid the migration of negative information to the target domain, but also solve the problem of biased learning results that may be caused by category imbalance, and consider the local geometric structure of the data, so as to better capture the intrinsic characteristics and relationships of the data, and improve the robustness and generalization of the model. Thereby more effectively solving the domain offset problem in the decoding process and improving the accuracy of decoding.

[0062] The cross-subject speech imagery EEG decoding method with privacy protection function proposed in the present invention provides an effective tool with high accuracy and high reliability. By establishing a mathematical model to screen target domain samples and constructing a discriminant transfer regression method to narrow the domain difference between the source domain and the target domain, the speech imagery task information of the target subject can be accurately identified based on the EEG data.

[0063] The present invention aims at the difficult problem of cross-subject state recognition in the field of EEG research. The power mean exponential and the discriminant transfer regression method of joint low-rank and sparse subspace learning constructed in the present invention greatly reduce the inter-domain differences and promote the application of richer source domain knowledge in state recognition in the target domain.

[0064] The present invention introduces a discriminant transfer regression method that combines low-rank and sparse subspace learning. It can eliminate the influence of outliers and noise, avoid the migration of negative information to the target domain, and solve the problem of biased learning results that may be caused by category imbalance. At the same time, it takes into account the local geometric structure of the data, so as to better capture the intrinsic characteristics and relationships of the data and improve the robustness and generalization of the model. In this way, the domain shift problem in the decoding process is more effectively solved and the decoding accuracy is improved.

[0065] This innovative decoding method makes full use of source domain knowledge. By quantifying sample transferability and optimizing the domain adaptation process, it improves the accuracy and robustness of EEG speech imagery decoding across subjects. It provides a more sophisticated and efficient tool for understanding and analyzing individual speech imagery processes, and is of great value in promoting related scientific research, clinical applications, and the development of human-computer interaction technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0067] Figure 1 A schematic diagram of the cross-subject speech imagery EEG decoding method with a privacy protection function of the present invention;

[0068] Figure 2 It is a flow chart of the cross-subject speech imagery EEG decoding method with privacy protection function of the present invention; DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] A cross-subject speech imagery EEG decoding method with privacy protection function is used to identify EEG samples with unknown EEG speech imagery task labels.

[0071] like Figure 1 As shown, a cross-subject speech imagery EEG decoding method with privacy protection function specifically includes the following steps:

[0072] Step 1: Collect EEG data from multiple subjects under different speech imagery tasks.

[0073] Step 2: Preprocess and extract features from the EEG data collected in step 1. The EEG data recorded at 1024Hz is sampled at a sampling rate of 256Hz, and then filtered through a fifth-order Butterworth bandpass filter of 1Hz to 70Hz to remove noise and artifacts related to EMG activity. Signal processing technology is used to extract the state characteristics of the EEG signal in different frequency bands, specifically using spectral entropy (SE) as a feature, thereby obtaining the target domain sample matrix X t , the source domain sample matrix and the corresponding label matrix [X s ; Y s ] is known.

[0074] The process of extracting spectral entropy is to segment the filtered EEG signal according to a predetermined time window, with each window being 1 second. The power spectral density (PSD) of the EEG signal in each time window is calculated, and the power spectrum of the signal is obtained using the Fast Fourier Transform (FFT). The power spectral density of each frequency component is normalized to form a probability distribution. The specific method is to divide the power spectral value of each frequency component by the sum of the total power spectral values, as shown in the following formula:

[0075]

[0076] Among them, f i represents the i-th frequency component, PSD(f i ) is the power spectrum value corresponding to this frequency component.

[0077] Then calculate the spectral entropy of the signal in each time window as follows:

[0078]

[0079] The local source domain model is obtained by training the support vector machine model, and the target domain sample X is trained using the local source domain model. t Make a prediction to get the target domain pseudo label probability

[0080] Step 3: Construct a speech imagery EEG signal decoding model for cross-subject transfer learning. Use the source domain data to train the support vector machine model and use the model to predict the target domain data to obtain the predicted probability label matrix. Use the power average method to evaluate the predicted probability label matrix, and select a part of the target domain samples with good performance as the intermediate domain according to the set threshold. At the same time, this part of the samples is more representative of the source domain samples, realizing the transfer learning of the source domain privacy protection.

[0081] Step 3.1: Use the power average index strategy as the measurement criterion for the distribution difference between the intermediate domain and the target domain, as shown in Formula;

[0082] Step 3.2: According to the calculated power mean index (M p ) and the set threshold, select the set of qualified samples and pseudo labels as the intermediate domain [X tmp ; Y tmp ].

[0083] Step 4: According to the intermediate domain [X tmp ; Y tmp ] and the target domain sample [X t ; Y t], a speech imagery EEG decoding method with cross-subject domain adaptive learning was constructed. The main purpose was to perform feature transformation, learn the shared subspace of samples in the intermediate domain and the target domain, and improve the classification accuracy.

[0084] Step 4.1: To facilitate the solution, separate the variables in the optimal problem and introduce the auxiliary variable Z1 into the problem, that is, formula:

[0085]

[0086] stAP T X t =AP T X tmp Z+E, Z=Z1, M≥0, Z≥0, Y=[Y tmp ; Y t ]

[0087] The augmented Lagrange multiplier function L corresponding to the formula is as follows:

[0088]

[0089] Where: μ represents the penalty parameter, Y1, Y2 are Lagrange multipliers.

[0090] Step 4.2: Since there are multiple variables in the formula, iteratively update each variable by fixing other variables:

[0091] Fix other variables and update A:

[0092] make T2=X t -X tmp Z

[0093]

[0094] in, D s , D t Both represent diagonal matrices, whose entries are the row sums or column sums of Z respectively.

[0095] The optimal problem of A can be described as:

[0096]

[0097] Taking the derivative of the formula with respect to A and setting the derivative to 0 gives:

[0098]

[0099] Fix other variables and update P:

[0100]

[0101] Taking the derivative of the formula with respect to P and setting the derivative to 0, we can get:

[0102]

[0103] This equation satisfies the Sylvester equation form, and the optimal solution can be obtained directly using Gaussian elimination method.

[0104] Fix other variables and update M:

[0105]

[0106] stM≥0

[0107] Let N = AP T XY, V = 2Y-1, we can get:

[0108]

[0109] sM ij ≥0

[0110] Among them, M ij The optimal solution is:

[0111] M ij =max(N ij V ij ,0)

[0112] Therefore, the optimal solution of M is:

[0113] M=max(N⊙V,0)

[0114] where ⊙ represents the Hadamard product operator between two matrices.

[0115] Fix other variables and update Z:

[0116] Solve the optimization problem of Z according to the formula:

[0117]

[0118] make Setting the derivative with respect to Z to 0, we get:

[0119]

[0120] Let T7 = (AP T X tmp ) T (AP T X tmp ),

[0121] The optimal solution for Z is:

[0122]

[0123] Fix other variables and update Z1:

[0124]

[0125] Fix other variables and update E:

[0126] In the formula, all variables except E are fixed, we can get:

[0127]

[0128] The optimal solution of the formula is as follows:

[0129]

[0130] Among them, the shrink function is used for soft threshold processing to shrink the first input parameter.

[0131] Fix other variables and update the Lagrange multipliers Y1, Y2 and parameter μ:

[0132]

[0133] Fix other variables and update Y t :

[0134]

[0135] Y tmp 1=1,Y t ≥0, Y=[Y tmp , Y t ]

[0136] This formula can be simplified to:

[0137]

[0138] Y t 1=1,Y t ≥0, Y=[Y tmp , Y t ]

[0139] For Y t Each row of y i It is decomposed into separate optimization tasks, so for the i-th row, the optimization problem becomes:

[0140]

[0141] We make b i =APT x i +m i , a i =2m i +1, the optimization problem becomes:

[0142]

[0143]

[0144] The Lagrangian functions introduced are:

[0145]

[0146] y i The optimal solution of can be obtained through the KKT condition.

[0147] The algorithm flow is as follows:

[0148] Algorithm 1: Solution process of the discriminant transfer regression method with joint low-rank and sparse subspace learning

[0149] 1. Input: intermediate domain sample set X tmp and the corresponding label Y tmp , target domain sample set X t , parameters α, λ, γ, maximum number of iterations maxiter;

[0150] 2. Initialization: Use the PCA projection matrix to initialize the matrix P,

[0151] M=1, Z=Z1=0, E=0, Y1=Y2=0,

[0152] iter=1,μ=0.1,μ max =10 7 ,ρ=1.01,ξ=10 -7 .

[0153] 3. When iter<maxiter and does not converge:

[0154] Utilize the update A;

[0155] Update P using formula;

[0156] Update M using formula;

[0157] Update Z using formula;

[0158] Update Z1 using formula;

[0159] Utilize the update E;

[0160] Update Y1, Y2, μ using the formula;

[0161] Utilization Update Y t ;

[0162] iter = iter + 1;

[0163] Utilize||AP T X t -AP T X tmp ZE|| F <ξ Check convergence;

[0164] 4. Output: Target domain sample label matrix Y t .

[0165] This method learns a nonlinear transformation matrix and projection matrix by solving the optimization problem, mapping the data of the intermediate domain and the target domain into a common feature space, making their distribution in the space closer, thereby improving the generalization ability of the model in the target domain. Using the information of known labels and relaxation matrices, it ensures that when optimizing the model parameters, the target domain label matrix Y t Inference is performed to obtain the optimal label prediction through iterative optimization while keeping the model complexity under control and label smoothness.

[0166] Step 5: Analyze the intermediate domain sample matrix constructed in step 3 and its corresponding label matrix. By measuring the similarity between the source domain and the target domain with the power mean exponent, the difference between the source subject domain and the target subject domain is smoothed. Analyze the weight coefficient matrix Z obtained in step 4. The weight coefficient matrix Z realizes the effective alignment of the intermediate domain and target domain samples, enhances the robustness of the model, and reduces the degradation of model performance due to distribution differences.

[0167] Step 6: Analyze the target domain label matrix Y obtained in step 4 t , according to the target domain label matrix, the EEG state of the target subject during EEG acquisition is obtained, Y t Each column of Y corresponds to an EEG state category, and each row corresponds to the EEG state prediction of a sample. t For each row of , the category corresponding to the largest element in the row is the EEG state of the target subject corresponding to the unlabeled sample at the time of EEG acquisition.

[0168] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A cross-subject speech imagery EEG decoding method with privacy protection function, characterized by: The following steps are involved: Step 1: Collect EEG data of the target subjects under different speech imagery tasks; Step 2: Preprocess and extract features of the EEG data collected in step 1 to obtain the sample matrix of the subject d is the characteristic number, n is t is the total number of samples, as the target domain sample matrix; the source domain sample matrix and the corresponding label matrix [X s ; Y s ] It is known that the local source domain model is obtained by training the support vector machine model, and the local source domain model is used to train the target domain sample matrix X t Make a prediction to get the target domain pseudo label probability Step 3: Use the power average index strategy as the metric for the distribution difference between the intermediate domain and the target domain. According to the power average index M of each target domain sample, P And the set threshold, select the set of qualified samples and pseudo labels as the intermediate domain; Step 4: According to the intermediate domain [X tmp ; Y tmp ] and the collected target domain samples [X t ; Y t ], construct a discriminative transfer regression method for joint low-rank and sparse subspace learning, whose objective function is as follows: stAP T X t =AP T X tmp Z+E, M≥0, Z≥0, Y, [Y tmp ,AND t ] in and are two projection matrices, d represents the characteristic dimension of the projection subspace, the size of the original transformation matrix B is m×c, the sizes of the projection matrix P and matrix A are m×d and c×d respectively, X=[X tmp ;X t ], is the label matrix of the intermediate domain samples, represents the label relaxation matrix, ||Z|| * is the nuclear norm of the matrix Z, Z is the reconstruction coefficient matrix, Z = [z ij ], α, γ represent regularization parameters, ||E||1 represents the l1 norm of the matrix E, is the noise matrix, represents the i-th sample feature of the middle domain, represents the jth sample feature of the target domain sample, represents the spectral norm of the matrix; Step 5: According to the objective function established in step 4, transform matrix A, projection matrix P, label relaxation matrix M, reconstruction coefficient matrix Z, noise matrix E, and target domain label matrix Y t Joint iterative optimization of Step 6: According to the target domain label matrix Y t Obtain the verbal imagery task of the target subject during EEG acquisition.

2. The method for cross-subject speech imagery EEG decoding with privacy protection function according to claim 1, characterized in that: The step 2 specifically includes: Step 2.1: Preprocessing and feature extraction: The EEG data recorded at 1024 Hz are sampled at a sampling rate of 256 Hz, and then the noise and artifacts related to EMG activity are filtered out through a fifth-order Butterworth bandpass filter with a frequency range of 1 Hz to 70 Hz. The state characteristics of the EEG signal in different frequency bands are extracted, and the spectral entropy of the EEG signal is calculated as the target domain sample matrix X t ; Step 2.2: Build a support vector machine model to train the source domain sample matrix and label matrix to obtain a local source domain model, and use the source domain model to predict the target domain sample matrix to obtain the target domain pseudo label probability 3. The method for cross-subject speech imagery EEG decoding with privacy protection function according to claim 1, characterized in that: The specific process of constructing the speech imagery EEG signal decoding method for cross-subject transfer learning in step 3 is as follows: Step 3.1: Use the power average exponential strategy as the metric to measure the sample similarity between the source domain and the target domain, as shown in formula (2); Among them, n t represents the number of samples in the target domain, and k represents the number of categories of samples in the target domain; represents the target domain pseudo-label probability matrix; p represents the power average exponent; Step 3.2: According to the calculated power average exponent M of each target domain sample p And the set threshold, select the qualified samples and pseudo-label sets as the intermediate domain [X tmp ; Y tmp ].

4. The method for cross-subject speech imagery EEG decoding with privacy protection function according to claim 1, characterized in that: The specific solution process of constructing the discriminant transfer regression method for joint low-rank and sparse subspace learning in step 4 is: Step 4.1: Convert the objective function of formula (1) into a solution function; Step 4.2: Fix other variables and update the transformation matrix A as follows: stAP T X t =AP T X tmp Z+E,Z=Z1,M≥0,Z≥0,Y=[Y tmp ;AND t ] Solve equation (3) to obtain the updated transformation matrix A; Step 4.3: Fix other variables and update the projection matrix P as follows: s.tAP T X t =AP T X tmp Z+E,Z=Z1,M≥0,Z≥0,Y=[Y tmp ;Y t ] Solve equation (4) to obtain the updated projection matrix P; Step 4.4: Fix other variables and update the label relaxation matrix M as follows: stM≥0 Solve equation (5) to obtain the updated label relaxation matrix M; Step 4.5: Fix other variables and update the reconstruction coefficient matrix Z as follows: st.AP T X t =AP T X tmp Z+E,Z≥0 Solve equation (6) to obtain the updated reconstruction coefficient matrix Z; Step 4.6: Fix other variables and update the noise matrix E: stAP T X t =AP T X tmp Z+E Solve equation (7) to obtain the updated noise matrix E; Step 4.7: Fix other variables and update the target domain label matrix Y t : stY=[Y tmp ,AND t ],AND t ≥0,Y t 1=1 Solve equation (8) to get the updated target domain label matrix Y t ; Step 4.8: Repeat steps 4.2, 4.3, 4.4, 4.5, 4.6 and 4.7 alternately until the objective function converges, and complete the transformation matrix A, projection matrix P, label relaxation matrix M, reconstruction coefficient matrix Z, noise matrix E, and target domain label matrix Y. t Joint iterative optimization.

5. The method for cross-subject speech imagery EEG decoding with privacy protection function according to claim 4, characterized in that: The transformation matrix A, projection matrix P, and weight coefficient matrix Z obtained in step 4 are used to achieve effective alignment of the intermediate domain and target domain samples.

Images