A speech imagery EEG decoding method based on source domain transferability and adaptive learning

By constructing a source domain transferable adaptive learning model in verbal imagination electroencephalopathy, the problem of EEG signal difference across subjects and different states is solved, and a speech imaginary brain with higher accuracy and reliability is achieved Electrodecoding effect.

CN117725367BActive Publication Date: 2024-11-29BEIJING TIANJI ZHICHENG TECHNOLOGY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410087934.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-01-22
Publication Date
2024-11-29
Estimated Expiration
2044-01-22

Smart Images

  • Figure CN117725367B_ABST
    Figure CN117725367B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for speech imagery EEG decoding based on adaptive learning of source domain transferability. The specific steps of the method include: firstly, collecting EEG data of subjects and decoding objects under different speech imagery tasks, secondly, preprocessing the data to obtain a sample matrix, and then iterating the speech imagery EEG decoding model based on adaptive learning of source domain transferability to obtain a label matrix, and finally completing speech imagery EEG decoding. The present invention minimizes the difference between domains to the greatest extent by mapping the source domain and the target domain to subspaces, and the dual mapping matrix is ​​continuously optimized during the iteration of the objective function to reduce the difference between domains; at the same time, the present invention introduces a sample weight coefficient matrix to quantitatively represent the transferability of source domain samples. During the iteration of the objective function, the source domain high transferability samples are prompted to play a greater role in the model learning process, solving the problem of inaccurate speech imagery EEG decoding due to the existence of negative samples in the field of speech imagery EEG decoding.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of electroencephalogram signal processing, and in particular relates to a speech imagery electroencephalogram decoding method of source domain transferable adaptive learning. Background Art

[0002] Language dialogue is a high-level cognitive behavior and the main way people communicate. Verbal imagination refers to the process of constructing language expression in the mind, that is, simulating, conceiving, or recalling the use of language through internal mental activities, rather than actually producing sounds or words. This kind of thinking activity usually includes the formation, organization, and interpretation of language elements. In the brain, language processing involves multiple areas, including Broca's area and the superior temporal gyrus, which play an important role in the production and understanding of speech. When an individual imagines speech, electrical activity occurs in the cerebral cortex. The physiological signal generated by the electrical activity of the human cerebral cortex, the electroencephalogram (EEG), provides us with a data source for EEG decoding of speech imagination. EEG decoding of speech imagination is a multidisciplinary research field involving multiple disciplines such as computer science, psychology, neurology, and brain cognitive science. It aims to study and develop theories, methods, and systems that can identify, interpret, process, and simulate human speech imagination. EEG decoding of speech imagination is an emerging branch in the field of EEG decoding of speech imagination, and its research object is human EEG signals. Different from common input signals including videos, pictures (body movements, facial expressions, etc.), voice, text, etc., human EEG signals are physiological signals and are not easy to disguise. Therefore, EEG decoding of speech imagination based on EEG signals can better reflect the individual's actual speech imagination task type.

[0003] How to obtain task type information about an individual's internal speech imagery in the EEG is an important goal of speech imagery EEG decoding research. In the field of cross-subject speech imagery EEG decoding, individual differences are inevitable. There are distribution differences in the EEG signals of a person in different states and at different times, as well as between people. This makes it difficult to generalize machine learning methods based on the assumption of independent and identical distribution. Therefore, the concept of transfer learning was proposed. Its core idea is to reduce the domain difference between the source domain and the target domain. However, common transfer learning methods such as the maximum mean difference method and deep neural network transfer cannot intuitively quantify the transferability of samples, and lack the interpretability of the intermediate process of transfer learning.

[0004] In the field of transfer learning, the maximum mean difference method is a common method, which maps the source domain and target domain data to a shared subspace to reduce the difference between the two domains and achieve a basically approximate distribution state. In the field of graph learning, graph-based semi-supervised learning methods generally construct graph association matrices through the similarity between samples, where sample similarity can be measured by methods such as Gaussian kernel and Euclidean distance. Graph-based semi-supervised label propagation can be understood as the process of propagating a finite number of labeled samples to a large number of unlabeled samples. When constructing a structured bipartite graph, the EEG signal that has undergone data preprocessing and feature extraction needs to be used as input. Figure 1 One side is the source domain sample (with labels), and the other side is the target domain sample (without labels). There is no connection within the layer, and the layers are fully connected. Summary of the invention

[0005] The purpose of the present invention is to provide a speech imagery EEG decoding method with source domain transferability and adaptive learning to address the lack of universality of the existing speech imagery EEG decoding method. The present invention performs a double mapping matrix P = [P s ;P t ], weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t By performing joint iterative optimization, a dual mapping matrix that can minimize the difference between domains, a weight coefficient matrix that can quantitatively represent the transferability of source domain samples, and a structured bipartite graph association matrix that well describes the degree of association between source domain samples and target domain samples are obtained, thereby improving the accuracy of speech imagery EEG decoding. In the present invention, on the one hand, the target domain label matrix F t The label of the target domain sample is obtained, which is the specific type of speech imagination of the decoding object corresponding to the sample at the moment of EEG acquisition; on the other hand, the transferability of each source domain sample can be quantified by the sample weight coefficient matrix Θ, thereby finding the source domain sample with the greatest influence in the transfer learning process, and promoting the source domain samples with high transferability to play a greater role in the model learning process.

[0006] A speech imagery EEG decoding method for source domain transferable adaptive learning comprises the following steps:

[0007] Step 1: Under different speech imagery tasks, collect labeled EEG data of one or more subjects, and unlabeled EEG data of the decoding subjects.

[0008] Step 2: Preprocess and extract features of the EEG data collected in step 1 to obtain the source domain sample matrix X s And the target domain sample matrix X t .

[0009] Step 3: Construct a sample weight coefficient matrix Θ that describes the transferability of samples in each source domain and introduce it into the dual-mapping active domain adaptation model; combine the dual-mapping active domain adaptation model and the semi-supervised label propagation model into a unified framework to obtain a speech imagery EEG decoding model.

[0010] Step 4: According to the objective function established in step 3, the parameters in the speech imagery EEG decoding model are jointly iteratively optimized to obtain the optimized target domain label matrix F t .

[0011] Step 5: Calculate the target domain label matrix F obtained in step 4 t Obtain the specific type of verbal imagery of the decoding subject during EEG acquisition.

[0012] Preferably, the objective function of the speech imagery EEG decoding model in step 3 is as follows:

[0013]

[0014]

[0015] Where X = [X s ;X t ], P s represents the source domain mapping matrix; P t represents the target domain mapping matrix; Θ represents the sample weight coefficient matrix; is the soft label of the augmented source domain sample; is the soft label of the augmented target domain sample; N s 、N t are all diagonal matrices, and the i-th diagonal element is the inverse of the number of samples in the i-th source domain and target domain; S is the optimal structured bipartite graph association matrix; B is the original structured bipartite graph association matrix; H represents the center scattering matrix; Tr(.) represents the trace operation of the matrix; L is the Laplace matrix; F = [F s ; F t ] is the label matrix; c is the number of speech imagery task types; a and λ are regularization parameters; ns represents the number of source domain samples; nt represents the number of target domain samples; Represents the calculation of the Frobenius norm.

[0016] In step 4, the double mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t Perform joint iterative optimization.

[0017] Preferably, the specific process of constructing the speech imagery EEG decoding model in step 3 is:

[0018] Step 3.1: Use the maximum mean difference strategy as the metric for the distribution difference between the source domain and the target domain in the subspace to establish a dual-mapping active domain adaptation model:

[0019]

[0020] Step 3.2: Establish a semi-supervised label propagation model based on structured bipartite graph:

[0021]

[0022] Step 3.3: Combine the two models constructed in step 3.1 and step 3.2 to obtain the objective function of the speech imagery EEG decoding model.

[0023] Preferably, the specific process of the joint iterative optimization in step 4 is:

[0024] Step 4.1: Use Euclidean distance to calculate the similarity between samples, initialize the structured bipartite graph association matrix B, and initialize the target domain label matrix F. t , so that the matrix F t All elements of are equal to

[0025] Step 4.2, update the double mapping matrix P = [P s ;P t ], fixed sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t , as follows:

[0026]

[0027] Solve equation (3) to get the updated mapping matrix P = [P s ;P t ];

[0028] Step 4.3, update the sample weight coefficient matrix Θ, fix the double mapping matrix P = [P s ;P t ], structured bipartite graph association matrix S and target domain label matrix F t , as follows:

[0029]

[0030] Solve equation (4) to obtain the updated sample weight coefficient matrix Θ;

[0031] Step 4.4: Update the structured bipartite graph association matrix S and fix the bimapping matrix P = [Ps ;P t ], sample weight coefficient matrix Θ and target domain label matrix F t , as follows:

[0032]

[0033] Solve equation (5) to obtain the updated structured bipartite graph association matrix S;

[0034] Step 4.5: Update the target domain label matrix F t , fixed double mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ and structured bipartite graph association matrix S, as follows:

[0035]

[0036] Step 4.6: Update and repeat steps 4.2, 4.3, 4.4 and 4.5 alternately until the objective function converges and the dual mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t Joint iterative optimization.

[0037] Preferably, in step 4.2 and step 4.5, the source domain mapping matrix P is solved by the Lagrangian function. s , target domain mapping matrix P t And the target domain label matrix F t .

[0038] Preferably, in step 4.3, the objective function is first simplified by a Lagrangian function and then solved by a row-by-row solution strategy to obtain the sample weight coefficient matrix Θ.

[0039] Preferably, in step 4.4, the structured bipartite graph association matrix S is obtained by solving the structured bipartite graph using a row-by-row solution strategy.

[0040] Preferably, the process of preprocessing and feature extraction in step 2 is: sampling the EEG data recorded at 1024 Hz with a sampling rate of 256 Hz, and then filtering out noise and artifacts related to EMG activity through a fifth-order Butterworth bandpass filter of 1 Hz to 70 Hz, extracting the state characteristics of the EEG signal in different frequency bands, and calculating the differential entropy of the EEG signal as a sample matrix.

[0041] In a second aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the aforementioned source domain transferable adaptive learning method for speech imagery EEG decoding.

[0042] In a third aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the aforementioned speech imagery EEG decoding method for source domain transferability adaptive learning.

[0043] The present invention has the following beneficial effects:

[0044] 1. The speech imagery EEG decoding method based on adaptive learning of source domain transferability proposed in the present invention is different from the existing speech imagery EEG decoding method which only focuses on improving the performance of speech imagery EEG decoding. The present invention focuses on mining the transferability of the EEG data of the subjects. The present invention introduces the sample weight coefficient matrix Θ describing the transferability of each source domain sample into the dual mapping active domain adaptation model, thereby clearly quantifying the transferability of the source domain samples and promoting the source domain high transferability samples to play a greater role in the model learning process.

[0045] 2. The speech imagery EEG decoding method based on adaptive learning of source domain transferability proposed in the present invention provides an effective tool with high accuracy and high reliability. By establishing a mathematical model to continuously iteratively optimize the labels of target domain samples, it can accurately identify the speech imagery task type of the decoding object based on the EEG data.

[0046] 3. The present invention aims at the difficult cross-subject EEG decoding problem in the field of EEG research. The constructed dual mapping matrix greatly reduces the inter-domain differences and promotes the application of richer source domain knowledge in state recognition in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A framework diagram of a speech imagery EEG decoding method with adaptive learning of source domain transferability.

[0048] Figure 2 Flowchart of a speech imagery EEG decoding method with source domain transferability and adaptive learning. DETAILED DESCRIPTION

[0049] The present invention is further described below in conjunction with the accompanying drawings.

[0050] A speech imagery EEG decoding method with source domain transferability and adaptive learning is used to decode EEG samples with unknown EEG speech imagery task labels, so as to identify the language intentions of people with speech disorders (such as patients with cerebral infarction). In this embodiment, the identified speech imagery types include the need to drink water, the need to eat, the need to exercise, and the need to rest. In addition, the identified speech imagery types can also be letters with different pronunciations; by imagining letters with different pronunciations and mapping them to specific tasks, the task content includes but is not limited to drinking water, eating, exercising, and resting.

[0051] In cross-subject speech imagery EEG decoding, inaccurate recognition is often caused by large differences between subjects. Therefore, this embodiment hopes to reduce the difference between domains by constructing a mapping matrix to map the source domain and the target domain to the subspace. At the same time, this embodiment introduces a sample weight coefficient matrix in the process of reducing the difference between domains, quantitatively represents the transferability of source domain samples, and promotes the source domain high transferability samples to play a greater role in the model learning process, solving the problem of inaccurate speech imagery EEG decoding caused by individual differences in the field of cross-subject speech imagery EEG decoding.

[0052] like Figure 1 and Figure 2 As shown, a speech imagery EEG decoding method for source domain transferable adaptive learning specifically includes the following steps:

[0053] Step 1: Collect labeled EEG data of multiple subjects under different speech imagery tasks; and collect unlabeled EEG data of decoding subjects under different speech imagery tasks.

[0054] 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 the noise and artifacts related to EMG activity are filtered out through a fifth-order Butterworth bandpass filter of 1Hz to 70Hz. Signal processing technology is used to extract the state characteristics of the EEG signal in different frequency bands. Specifically, differential entroy (DE) is used as a feature, and the differential entropy of the EEG signal is calculated as the sample matrix. d is the number of features, n is the total number of samples; select the sample matrix X of the subject and the decoding object as the source domain sample matrix x s And the target domain sample matrix x t .

[0055] Step 3: Construct a model of speech imagery EEG decoding method based on adaptive learning of source domain transferability. Construct a sample weight coefficient matrix that describes the transferability of each source domain sample and introduce it into the dual-mapping active domain adaptation model; combine the dual-mapping active domain adaptation model and the semi-supervised label propagation model into a unified framework to obtain a speech imagery EEG decoding model. Iterate and update the objective function to obtain the sample weight coefficient matrix and the target domain label matrix; the sample weight coefficient matrix quantitatively represents the transferability of the source domain samples, and the target domain label matrix obtains the target domain label to complete speech imagery EEG decoding.

[0056] Step 3.1: Use the maximum mean difference strategy as the metric for the distribution difference between the source domain and the target domain in the subspace and incorporate the idea of ​​active learning into it to establish a dual-mapping active domain adaptation model:

[0057]

[0058] In formula (1), X s represents the source domain sample, X t represents the target domain sample; P s represents the source domain mapping matrix; P t represents the target domain mapping matrix; Θ represents the sample weight coefficient matrix; is the soft label of the augmented source domain sample; is the soft label of the augmented target domain sample; Where N s 、N t are all diagonal matrices, and the i-th diagonal element is the inverse of the number of samples in the i-th source domain and target domain (i=1, 2, ..., c); c is the number of types of verbal imagination; n=ns+nt represents the total number of samples, where ns represents the number of samples in the source domain and nt represents the number of samples in the target domain; H represents the central scattering matrix; Ip represents the unit matrix; is a constraint on the double mapping matrix; θ = diag(θ), θ ≥ 0, θ T 1 ns =1 constrains the weight coefficient matrix Θ to be a diagonal matrix and the sum of its diagonal elements is 1.

[0059] Step 3.2: Establish a semi-supervised label propagation model based on structured bipartite graph:

[0060]

[0061] In formula (2), S is the optimal structured bipartite graph association matrix describing the degree of association between source domain samples and target domain samples, s ijrepresents the element of the i-th row and j-th column of the matrix S, which is used to characterize the degree of association between the i-th source domain sample and the j-th target domain sample; B is the original structured bipartite graph association matrix describing the degree of association between the source domain sample and the target domain sample; Tr(·) represents the trace operation of the matrix; L is the Laplace matrix; is the label matrix used for label propagation; a and λ are regularization parameters, whose values ​​are in {10 -3 , 10 -2 , ..., 10 3}Select; Indicates the calculation of the Frobenius norm; S ≥ 0, S1 ns =1 nt The elements in the matrix S are constrained to be non-negative and the sum of the elements in each row of the matrix S is 1.

[0062] Step 3.3: Combine the two models constructed in step 3.1 and step 3.2, integrate the dual-map active domain adaptation model and the semi-supervised label propagation model based on structured bipartite graph into a unified framework for joint optimization, and the obtained objective function is as follows:

[0063]

[0064] Step 4: According to the objective function established in step 3, the double mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t Perform joint iterative optimization.

[0065] Step 4.1: Use Euclidean distance to calculate the similarity between samples, initialize the structured bipartite graph association matrix B, and initialize the target domain label matrix F. t , so that the matrix F t All elements of are equal to

[0066] Step 4.2, update the double mapping matrix P = [P s ;P t ], fixed sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t , as follows:

[0067]

[0068] In order to simplify the calculation when solving equation (4), let P = [P s ;P t ]、 Formula (4) can be simplified as follows:

[0069]

[0070] By solving equation (5) using the Lagrange multiplier method, we can obtain the objective function about P as follows:

[0071]

[0072] In equation (6), Φ is the Lagrange multiplier. Taking the partial derivative of P and assigning the result to zero, we can simplify it to get the following equation:

[0073]

[0074] Formula (7) is a typical eigenvalue decomposition problem. After eigenvalue decomposition, the updated mapping matrix P can be obtained as follows: s ;P t ].

[0075] Step 4.3, update the sample weight coefficient matrix Θ, fix the double mapping matrix P = [P s ;P t ], structured bipartite graph association matrix S and target domain label matrix F t , as follows:

[0076]

[0077] Formula (8) can be converted into the following form:

[0078]

[0079] Applying the Lagrange multiplier method to solve equation (9) can obtain the updated sample weight coefficient matrix Θ.

[0080] Step 4.4: Update the structured bipartite graph association matrix S and fix the bimapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ and target domain label matrix F t , as follows:

[0081]

[0082] Formula (10) can be converted into the following form:

[0083]

[0084] definition Formula (11) can be simplified into the following form:

[0085]

[0086] By solving equation (12) row by row, we can obtain the updated structured bipartite graph association matrix S.

[0087] Step 4.5: Update the target domain label matrix F t , fixed double mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ and structured bipartite graph association matrix S, as follows:

[0088]

[0089] Formula (13) can be converted into the following form:

[0090]

[0091] Apply the Lagrange multiplier method and the row-by-row solution method to solve equation (14) and obtain the updated target domain label matrix F t .

[0092] Step 4.6: Update and repeat steps 4.2, 4.3, 4.4 and 4.5 alternately until the objective function converges and the dual mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t Joint iterative optimization.

[0093] Step 5: Analyze the target domain label matrix F obtained in step 4 t , according to the target domain label matrix F t Obtain the EEG state of the decoding object during EEG acquisition, F t Each column of corresponds to a speech imagery type, and each row corresponds to the EEG speech imagery type prediction of a sample. t For each row of , the type corresponding to the largest element in the row is the speech imagination type of the decoding object corresponding to the unlabeled sample at the moment of EEG acquisition.

[0094] Step 6. By analyzing the sample weight coefficient matrix Θ obtained in step 4, it can be seen that the sample weight coefficient matrix Θ quantitatively represents the transferability of the source domain samples. During the iteration of the objective function, the source domain samples with high transferability are encouraged to play a greater role in the model learning process.

Claims

1. A speech imagery EEG decoding method with source domain transferability and adaptive learning, characterized by: The following steps are involved: Step 1: Under different speech imagery tasks, collect labeled EEG data of one or more subjects and decode unlabeled EEG data of the subjects; Step 2: Preprocess and extract features of the EEG data collected in step 1 to obtain the source domain sample matrix X s And the target domain sample matrix X t ; Step 3: construct a sample weight coefficient matrix Θ that describes the transferability of samples in each source domain and introduce it into the dual-mapping active domain adaptation model; combine the dual-mapping active domain adaptation model and the semi-supervised label propagation model into a unified framework to obtain a speech imagery EEG decoding model; The objective function of the speech imagery EEG decoding model is as follows: Where X = [X s ;X t ], P s represents the source domain mapping matrix; P t represents the target domain mapping matrix; Θ represents the sample weight coefficient matrix; is the soft label of the augmented source domain sample; is the soft label of the augmented target domain sample; N s 、N t are all diagonal matrices, and the i-th diagonal element is the inverse of the number of samples in the i-th source domain and target domain; S is the optimal structured bipartite graph association matrix; B is the original structured bipartite graph association matrix; H represents the center scattering matrix; Tr(·) represents the trace operation of the matrix; L is the Laplace matrix; F = [F s ; F t ] is the label matrix; c is the number of EEG state types; α and λ are regularization parameters; ns represents the number of source domain samples; nt represents the number of target domain samples; Represents the calculation of the Frobenius norm; The specific process of constructing the speech imagery EEG decoding model is as follows: Step 3.1: Use the maximum mean difference strategy as the metric for the distribution difference between the source domain and the target domain in the subspace to establish a dual-mapping active domain adaptation model: Step 3.2: Establish a semi-supervised label propagation model based on structured bipartite graph: Step 3.3, combining the two models constructed in step 3.1 and step 3.2 to obtain the objective function of the speech imagery EEG decoding model; Step 4: According to the objective function established in step 3, the parameters in the speech imagery EEG decoding model are jointly iteratively optimized to obtain the optimized target domain label matrix F t ; The joint iterative optimization is to double the mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t Conduct joint iterative optimization; Step 5: Calculate the target domain label matrix F obtained in step 4 t Obtain the specific type of verbal imagery of the decoding subject during EEG acquisition.

2. The method for speech imagery EEG decoding based on source domain transferability and adaptive learning according to claim 1, characterized in that: The specific process of joint iterative optimization in step 4 is: Step 4.1: Use Euclidean distance to calculate the similarity between samples, initialize the structured bipartite graph association matrix B, and initialize the target domain label matrix F. t , so that the matrix F t All elements of are equal to Step 4.2, update the double mapping matrix P = [P s ;P t ], fixed sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t , as follows: Solve equation (3) to get the updated mapping matrix P = [P s ;P t ]; Step 4.3, update the sample weight coefficient matrix Θ, fix the double mapping matrix P = [P s ;P t ], structured bipartite graph association matrix S and target domain label matrix F t , as follows: Solve equation (4) to obtain the updated sample weight coefficient matrix Θ; Step 4.4: Update the structured bipartite graph association matrix S and fix the bimapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ and target domain label matrix F t , as follows: Solve equation (5) to obtain the updated structured bipartite graph association matrix S; Step 4.5: Update the target domain label matrix F t , fixed double mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ and structured bipartite graph association matrix S, as follows: Step 4.6: Update and repeat steps 4.2, 4.3, 4.4 and 4.5 alternately until the objective function converges and the dual mapping matrix P = [P s ;P t ], sample weight coefficient matrix Θ, structured bipartite graph association matrix S and target domain label matrix F t Joint iterative optimization.

3. The method for speech imagery EEG decoding based on source domain transferability and adaptive learning according to claim 2, characterized in that: In steps 4.2 and 4.5, the source domain mapping matrix P is solved by the Lagrangian function s , target domain mapping matrix P t And the target domain label matrix F t .

4. The method for speech imagery EEG decoding with source domain transferability and adaptive learning according to claim 2, characterized in that: In step 4.3, the objective function is first simplified by the Lagrangian function and then solved by row-by-row to obtain the sample weight coefficient matrix Θ.

5. The method for speech imagery EEG decoding with source domain transferability and adaptive learning according to claim 2, characterized in that: In step 4.4, the structured bipartite graph association matrix S is obtained by solving the problem row by row.

6. The method for speech imagery EEG decoding with source domain transferability and adaptive learning according to claim 1, characterized in that: The process of preprocessing and feature extraction in step 2 is as follows: the EEG data recorded at 1024 Hz are sampled with 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 of 1 Hz to 70 Hz, the state characteristics of the EEG signal in different frequency bands are extracted, and the differential entropy of the EEG signal is calculated as the sample matrix.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The memory stores computer programs; the processor executes the speech imagery EEG decoding method for source domain transferability adaptive learning as described in any one of claims 1-6.

8. A readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, it is used to implement the speech imagery EEG decoding method for source domain transferability adaptive learning as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Electroencephalogram migration emotion recognition method combining semi-supervised regression and graph label propagation

    CN114818822A

  • Method, apparatus, device, medium and program product for motor imagery decoding

    CN115456011A