A multi-view EEG speech imagery intention recognition method based on adaptive graph learning

Through the multi-view EEG speech imaginative intention recognition method learned by adaptive graph, the joint optimization objective function is used to solve the problem of failing to fully utilize the multi-view data of EEG signals in the prior art, and the recognition efficiency and accuracy are improved. Key views and frequency bands are selected in EEG acquisition and feature extraction, which enhances the recognition ability of the model.

CN118551257BActive Publication Date: 2024-11-29BEIJING TIANJI ZHICHENG TECHNOLOGY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410708542.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-03
Publication Date
2024-11-29
Estimated Expiration
2044-06-03

AI Technical Summary

Technical Problem

Existing EEG speech imagination recognition methods fail to make full use of multi-view data of EEG signals and cannot adaptively learn the importance of views and features, resulting in insufficient recognition efficiency and accuracy.

Method used

Using the multi-view EEG speech imaginative intention recognition method with adaptive graph learning, the objective function is constructed to improve the discriminant ability and generalization of the model by jointly optimizing the unlabeled continuous label matrix, similarity matrix, feature weight matrix and view weight vector. ability and adaptively adjust views, channels and frequency bands during training.

Benefits of technology

It improves the accuracy and efficiency of verbal imagination intention recognition, and can select key views, channels and frequency bands during EEG acquisition, preprocessing and feature extraction, enhancing the recognition ability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118551257B_ABST
    Figure CN118551257B_ABST
Patent Text Reader

Abstract

The present invention discloses a multi-view EEG speech imagery intention recognition method based on adaptive graph learning. The method is as follows: 1. Collect EEG data of multiple subjects in different speech imagery task scenarios. 2. Preprocess the collected EEG and extract the features of different channels and frequency bands under different views. 3. Establish a multi-view EEG speech imagery intention recognition model based on adaptive graph learning. 4. Jointly iteratively optimize the objective function. 5. Obtain the predicted label of the unlabeled sample and complete the speech imagery EEG recognition task. The present invention utilizes the complementary information available in multiple views to learn a better similarity matrix and an unlabeled sample label matrix, thereby improving the ability to recognize speech imagery intention through EEG data; the present invention can adaptively learn the contribution of each view and feature to model recognition, and the importance of views and features can guide the selection of channels and frequency bands during EEG acquisition, preprocessing and feature extraction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram signal processing, and specifically designs a multi-view electroencephalogram speech imagery intention recognition method based on adaptive graph learning. Background Art

[0002] Verbal dialogue is a basic form of human communication, playing an important role in information exchange, feedback, emotional expression, intention communication, and the establishment and maintenance of social relationships. Verbal imagination is a psychological phenomenon that involves the process of simulating language production in the brain without actual language output. During the process of verbal imagination, the cerebral cortex will produce a physiological signal - electroencephalography (EEG). This is a signal spontaneously generated by the human brain during activity. Even when a person cannot speak, as long as the brain performs a verbal imagination task, an EEG signal will be generated. Therefore, by identifying the EEG of verbal imagination, it can effectively help people with speech disorders communicate with the outside world and express their intentions.

[0003] How to obtain the individual's speech imagination content from EEG signals is an important goal of speech imagination intention recognition research. In most existing EEG speech imagination recognition studies, a single view is used, and the feature data of different dimensions of EEG signals are not fully mined and utilized. At the same time, the importance of each view in the model under multiple views also needs to be different. In addition, the importance of different channels and different frequency bands of EEG signals in the process of individual speech imagination is different. In the research process, the feature data of different channels and frequency bands cannot be simply used directly. It is reasonable to give them different weights. In actual application scenarios, it is often impossible to obtain labels for all data. Semi-supervised learning can complete the recognition task with only a small amount of label information. In the field of graph learning, graph-based semi-supervised learning methods usually use the similarity between samples to construct a similarity matrix, and use the similarity matrix to achieve the effect of aggregation of the same category, thereby identifying unlabeled sample labels. Taking the above considerations into consideration, the present invention proposes a multi-view EEG speech imagination intention recognition model for adaptive graph learning. The model can make full use of the multi-view data of speech imagery EEG data, and learn the similarity between samples through semi-supervised graph learning to complete the intention recognition task of speech imagery. At the same time, the model can also adaptively learn the importance of views and features in the process of graph learning to effectively improve the importance of key views, channels, and frequency bands. And according to the contribution of views and features in the model, more valuable views, channels, and frequency bands can be selected during EEG acquisition, preprocessing, and feature extraction, thereby improving recognition efficiency. Summary of the invention

[0004] The present invention aims to propose a multi-view EEG speech imagery intention recognition method based on adaptive graph learning to address the shortcomings of existing EEG speech imagery recognition methods. u , similarity matrix S, feature weight matrix Θ v The final model can effectively distinguish the labels of unlabeled samples; obtain a similarity matrix that well describes the degree of association between samples; and obtain a quantitative description of the weight coefficient of the view and feature importance. In the present invention, on the one hand, EEG data can be mined and utilized as much as possible through multiple views to improve the model's discrimination and generalization capabilities; on the other hand, the speech imagination recognition method can adaptively obtain the feature weight matrix Θ during the training process. v and the view weight vector α, which can effectively improve the contribution of important views, frequency bands and channels to speech imagery recognition, and can guide the selection of channels and frequency bands during EEG acquisition, preprocessing and feature extraction, thereby improving recognition accuracy and efficiency.

[0005] In a first aspect, the present invention provides a multi-view EEG speech imagery intention recognition method based on adaptive graph learning, which comprises the following steps:

[0006] Step 1: Collect EEG data of multiple subjects in different speech imagery task scenarios.

[0007] Step 2: pre-process the EEG data collected in step 1 and the EEG data for which speech imagination intention needs to be identified, and extract the features of different channels and frequency bands of the pre-processed EEG data in different views to obtain a sample set. The EEG data for which speech imagination intention needs to be identified is obtained by collecting EEG signals from the subject.

[0008] Step 3: Combine the view weight vector α and the feature weight matrix Θ for each view v , the objective function is constructed as follows:

[0009]

[0010] Among them, X v represents a sample set; sample set X v Including the corresponding known discrete label matrix Y l A subset of labeled samples and the corresponding unlabeled discrete label matrix Y u A subset of unlabeled samples F=[F l ; F u ] is a continuous label matrix; F l =Y l ; F u With the unlabeled discrete label matrix Yu Corresponding. α=[α 1 , α 2 , …, α p ] T represents the weight vector of p views in the model; S is the similarity matrix; s ij is the similarity between the i-th sample and the j-th sample; represents the weight of the kth feature of the vth view; L s is the Laplace matrix; d v Represents the number of features for the v-th view.

[0011] Step 4: According to the objective function constructed in step 3, the unlabeled continuous label matrix F u , similarity matrix S, feature weight matrix Θ v and the view weight vector α are jointly iteratively optimized.

[0012] Step 5: According to the unlabeled continuous label matrix F u Determine the subject's verbal imagery type during EEG data collection.

[0013] As a preference, after step 5 is executed, according to the feature weight matrix θ v And the view weight vector α is used to weight different channels and frequency bands, determine the frequency bands and channels that play a key role in speech imagery EEG recognition, and adjust the selection of channels and frequency bands during EEG acquisition, preprocessing and feature extraction.

[0014] As a preference, the channel importance ω i The expression is in, Indicates the mth view in the vth view i The weight of the feature. i Indicates that the feature corresponding to the i-th channel is in θ v The subscript in the frequency band importance ψ j The expression is in, Indicates the mth view in the vth view j The weight of the feature; m j Indicates that the feature corresponding to the jth frequency band is in θ v For all channels, the channel importance ω i Sort from large to small; for all frequency bands, sort by frequency band importance ψ j Sort from large to small; take channel importance ω i The largest channels are selected as channels; the importance of the frequency band is taken as j The largest frequency bands are selected as the frequency bands.

[0015] Preferably, the specific process of preprocessing and feature extraction of EEG data in step 2 is: after downsampling the EEG data, the noise and electrooculogram artifacts are filtered out through a bandpass filter. The filtered data is divided into 5 frequency bands, namely Delta (1-4Hz), Theta (4-8Hz), Alpha (8-14Hz), Beta (14-31Hz) and Gamma (31-50Hz). Based on this data, the spectral entropy and power spectral density features of different channels and frequency bands are extracted as two views of the speech imagination recognition method.

[0016] As a preference, the specific process of constructing the multi-view EEG speech imagery intention recognition method of adaptive graph learning in step 3 is: using Euclidean distance to measure the similarity between samples, the distance between samples and the similarity are negatively correlated, and considering the view weight and feature weight, a similarity construction model can be obtained:

[0017]

[0018] In order to prevent the similarity matrix S and view weights from having trivial solutions and adding semi-supervised label propagation information, regular term constraints can be obtained, and finally the objective function formula (1) is obtained.

[0019] Preferably, the method of joint iterative optimization in step 4 is:

[0020] Step 4.1: Initialize the similarity matrix S, view weight vector α, and feature weight matrix Θ v and the unlabeled continuous label matrix F u .

[0021] Step 4.2: Update the unlabeled continuous label matrix F u , fixed similarity matrix S, feature weight matrix Θ v and the view weight vector α, we get F u The objective function is:

[0022]

[0023] Solving equation (3) can update the unlabeled continuous label matrix F u .

[0024] Step 4.3: Update the view weight vector α, fix the similarity matrix S and feature weight matrix Θ v and the unlabeled continuous label matrix F u , the objective function of α is obtained as:

[0025]

[0026] Solving equation (4) can update the view weight vector α.

[0027] Step 4.4: Update the similarity matrix S, fix the view weight vector α and feature weight matrix Θ v and the unlabeled continuous label matrix F u , the objective function of S is obtained as:

[0028]

[0029] Solving equation (5) can update the similarity matrix S.

[0030] Step 4.5: Update the feature weight matrix Θ v , fix the view weight vector α, the similarity matrix S and the unlabeled continuous label matrix F u , we get Θ v The objective function is:

[0031]

[0032] Solving equation (6) can update the feature weight matrix Θ v .

[0033] Step 4.6, repeat steps 4.2, 4.3, 4.4, and 4.5 until the objective function converges, and finally complete the similarity matrix S, view weight vector α, and feature weight matrix Θ v and the unlabeled continuous label matrix F u Joint iterative optimization.

[0034] As a preference, the feature weight matrix Θ v is a diagonal matrix, the diagonal elements represents the weight of the k-th feature of the v-th view.

[0035] Preferably, the Laplace matrix D is a diagonal matrix, and the i-th diagonal element is defined as

[0036] 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 multi-view EEG speech imagery intention recognition method based on adaptive graph learning.

[0037] 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 multi-view EEG speech imagery intention recognition method based on adaptive graph learning.

[0038] The present invention has the following beneficial effects:

[0039] 1. The multi-view EEG speech imagery intention recognition method based on adaptive graph learning proposed in the present invention can obtain information from multiple views, and use the complementary information available in multiple views to learn a better similarity matrix and unlabeled sample label matrix, thereby improving the intention recognition ability of EEG data for speech imagery tasks.

[0040] 2. The multi-view EEG speech imagery intention recognition method based on adaptive graph learning proposed in this invention not only focuses on the recognition task of EEG data, but also attaches importance to the contribution of each view and feature to model recognition. Feature weight matrix Θ v The view weight vector α is continuously optimized during the iteration of the model, and the weights of each view and feature are adaptively adjusted to strengthen the contribution of key views, channels, and frequency bands to model recognition, thereby improving recognition accuracy. At the same time, the combination of view importance and feature importance helps to quantitatively find important channels and important frequency bands in the view that play a key role in the recognition of speech imagery intentions, and helps to selectively extract the corresponding key channels and frequency bands based on the obtained quantitative data during EEG acquisition, preprocessing, and feature extraction, thereby optimizing the accuracy and efficiency of model recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Framework diagram of a multi-view EEG speech imagery intention recognition method based on adaptive graph learning. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings.

[0043] like Figure 1 As shown in FIG, a multi-view EEG speech imagery intention recognition method using adaptive graph learning specifically includes the following steps:

[0044] Step 1: Collection of speech imagery EEG data.

[0045] The subjects were asked to perform speech imagination tasks when wearing EEG caps and given relevant prompts such as voice and visual prompts. The EEG signals generated by the subjects during speech imagination were collected as raw data. The EEG signals of speech imagination were labeled accordingly.

[0046] Step 2: Preprocess and extract features of speech imagery EEG data.

[0047] First, the EEG data collected in step 1 and the EEG data that need to identify speech imagination intention are downsampled to a sampling frequency of 256Hz, and then the noise and electrooculogram artifacts are filtered out through a fifth-order Butterworth bandpass filter of 1Hz to 70Hz. The filtered data is divided into 5 frequency bands, namely Delta (1-4Hz), Theta (4-8Hz), Alpha (8-14Hz), Beta (14-31Hz) and Gamma (31-50Hz). Based on this data, the spectral entropy and power spectral density features of different channels and frequency bands are extracted as two views of the speech imagination recognition method. The EEG data that needs to identify speech imagination intention is obtained by collecting EEG signals from the subjects.

[0048] Step 3: Combine the view weight vector α and the feature weight matrix Θ for each view v ,construct a multi-view EEG speech imagery intention recognition model with adaptive graph learning.

[0049] The Euclidean distance is used to measure the similarity between samples. The distance between samples is negatively correlated with the similarity. The distance calculation takes into account that each feature has different weights. At the same time, the distance between samples of multiple views will affect the similarity calculation, and the weight of each view will also be different. Combining the above considerations, we get the similarity construction model:

[0050]

[0051] in, represents the sample set, n represents the total number of samples, n=l+u, l represents the number of samples with labels, u represents the number of samples without labels, d v represents the number of features of the vth view, v = 1, 2, ..., p, p is the total number of views; X v It consists of two subsets, one of which is l labeled samples. The corresponding discrete label matrix c represents the number of categories; the other subset is u samples without labels The corresponding discrete label matrix is is unknown; α=[α 1 , α 2 , …, α p ] T Represents the weight of each of the p views in the model; represents the weight of the kth feature of the vth view, Θ v By θ v The diagonal matrix formed. is a similarity matrix that describes the similarity between samples; s ijis the data in row i and column j of S, indicating the similarity between the i-th sample and the j-th sample; is the Laplace matrix, D is a diagonal matrix, and the i-th diagonal element is defined as

[0052] Considering that the similarity matrix s may have a trivial solution, that is, the similarity between the sample itself and itself is 1, and the similarity with the rest of the samples is 0. In order to prevent this kind of situation, the Frobenius norm of the similarity matrix needs to be added on the basis of formula (7) to constrain S to make its distribution more uniform. For the same view weight vector α, the coefficient of the view that contributes most to the speech imagination recognition task is 1, and the weights of the rest of the views are 0. In order to prevent this kind of situation, the 2-norm of α is added on the basis of formula (7). Since the model is trained in a semi-supervised mode, the propagation information of the semi-supervised label needs to be added to formula (7). The final objective function is as follows:

[0053]

[0054] in represents the continuous label matrix, F l =Y l is the continuous label matrix of known label samples, F u It is a continuous label matrix of unlabeled samples, which is the data that needs to be predicted during the training process of the model.

[0055] Step 4: According to the objective function of step 3, the unlabeled continuous label matrix F u , similarity matrix S, feature weight matrix Θ v and the view weight vector α are jointly iteratively optimized.

[0056] Step 4.1: Before optimizing the objective function, all these data need to be initialized. The unlabeled continuous label matrix F u Initially all elements are The similarity matrix S is obtained by preliminarily calculating the similarity between samples based on the Euclidean distance; the feature weight matrix Θ v The corresponding feature weight matrix Θ of the vth view v The diagonal elements of When the view weight vector α is initialized, all elements are p represents the total number of views.

[0057] Step 4.2: Update the unlabeled continuous label matrix F u , fixed similarity matrix S, feature weight matrix Θ v and the view weight vector α. The objective function can be transformed into

[0058]

[0059] Using the Lagrangian method to solve equation (9), we get

[0060]

[0061] So we get F u Update formula:

[0062] F u =-L uu -1 L ul Y l . (11)

[0063] Step 4.3: Update the view weight vector α, fix the similarity matrix S and feature weight matrix Θ v and the unlabeled continuous label matrix F u The objective function can be transformed into

[0064]

[0065] make And α T The 2-norm expansion of

[0066]

[0067] Through the formula, formula (13) is equivalent to

[0068]

[0069] where h is given by h v Solving equation (14) can obtain the updated view weight vector α.

[0070] Step 4.4: Update the similarity matrix S, fix the view weight vector α and feature weight matrix Θ v and the unlabeled continuous label matrix F u The objective function can be transformed into

[0071]

[0072] Known make And expand the F norm of S, and because we need to solve each row s i are independent of each other, so for each row s i The optimization can be written as

[0073]

[0074] Re-order d i is d ij The vector composed of is equivalent to

[0075]

[0076] Solving equation (17) can obtain the updated similarity matrix S.

[0077] Step 4.5: Update the feature weight matrix Θ v , fix the view weight vector α, the similarity matrix S and the unlabeled continuous label matrix F u The objective function can be transformed into

[0078]

[0079] Let Q v =X v Θ v , and The conversion traces are

[0080]

[0081] Each Θ v They are independent of each other, so we need to v are updated, for each Θ v The objective function can be written as

[0082]

[0083] Let M v =X vT L s X v , Indicates M v The i-th diagonal element of M v Substitute into formula (20). Because Θ v is a diagonal matrix, so expanding the trace form gives the objective function:

[0084]

[0085] By solving equation (21) using the Lagrange multiplier method, we can obtain the function

[0086]

[0087] right Taking the partial derivative and setting it equal to zero gives

[0088]

[0089] According to the condition θvT 1=1, then we can solve for η and substitute it into formula (23) to get

[0090]

[0091] Step 4.6, repeat steps 4.2, 4.3, 4.4, and 4.5 until the objective function converges, and finally complete the similarity matrix S, view weight vector α, and feature weight matrix Θ v and the unlabeled continuous label matrix F u Joint iterative optimization.

[0092] Step 5: Analyze the unlabeled continuous label matrix F obtained in step 4 u The labels of these samples are predicted to realize the recognition of speech imagination intention. The view weight vector α obtained in step 4 can quantitatively represent the contribution of each view, that is, the importance of spectral entropy and power spectral density in the process of speech imagination intention recognition. The view weight vector α obtained in step 4 is combined with the feature weight matrix Θ v , the importance of different channels and frequency bands to speech imagery recognition can be quantitatively calculated, where the importance formula of the i-th channel is:

[0093]

[0094] Where m i Indicates that the feature corresponding to the i-th channel is in θ v The subscript in Indicates the mth view in the vth view i The weight of the feature. Similarly, the importance formula of the jth frequency band is

[0095]

[0096] Where m j Indicates that the feature corresponding to the jth frequency band is in θ v The subscript in Indicates the mth view in the vth view j By sorting the importance of the obtained channels and frequency bands, the key channels and frequency bands in speech imagery intention recognition can be obtained, which can optimize the selection of channels and frequency bands during EEG acquisition, preprocessing and feature extraction, thereby improving the accuracy and efficiency of speech imagery intention recognition.

Claims

1. A multi-view EEG speech imagery intention recognition method based on adaptive graph learning, Features: The following steps are involved: Step 1: Collect EEG data of multiple subjects in different speech imagery task scenarios; Step 2: preprocess the EEG data collected in step 1 and the EEG data for which speech imagination intention needs to be identified, and extract features of different channels and frequency bands of the preprocessed EEG data in different views to obtain a sample set; Two views are extracted, namely spectral entropy and power spectral density features; Step 3: Combine the view weight vector α and the feature weight matrix Θ for each view υ , the objective function is constructed as follows: Among them, X υ represents a sample set; sample set X υ Including the corresponding known discrete label matrix Y l A subset of labeled samples and the corresponding unlabeled discrete label matrix Y u A subset of unlabeled samples F=[F l ; F u ] is a continuous label matrix; F l =Y l ; F u With the unlabeled discrete label matrix Y u Corresponding; α=[α 1 , α 2 , …, α p ] T represents the weight vector of p views in the model; S is the similarity matrix; s ij is the similarity between the i-th sample and the j-th sample; represents the weight of the k-th feature of the υ-th view; L s is the Laplace matrix; d υ represents the number of features of the υth view; Step 4: According to the objective function constructed in step 3, the unlabeled continuous label matrix F u , similarity matrix S, feature weight matrix Θ υ and view weight vector α for joint iterative optimization; Step 5: According to the unlabeled continuous label matrix F u Determine the subject's verbal imagery type during EEG data collection.

2. According to the method of multi-view EEG speech imagery intention recognition based on adaptive graph learning in claim 1, Features: After step 5 is executed, according to the feature weight matrix Θ υ And the view weight vector α is used to weight different channels and frequency bands, determine the frequency bands and channels that play a key role in speech imagery EEG recognition, and adjust the selection of channels and frequency bands during EEG acquisition, preprocessing and feature extraction.

3. According to the method of multi-view EEG speech imagery intention recognition based on adaptive graph learning in claim 2, Features: Channel Importanceω i The expression is in, represents the mth i The weight of the feature; m i Indicates that the feature corresponding to the i-th channel is in θ υ Subscript in; frequency band importance ψ j The expression is in, represents the mth j The weight of the feature; m j Indicates that the feature corresponding to the jth frequency band is in θ υ Subscript in; for all channels according to the channel importance ω i Sort from large to small; for all frequency bands, sort by frequency band importance ψ j Sort from large to small; take channel importance ω i The largest number of channels are used as key channels; the frequency band importance ψ j The largest frequency bands serve as key frequency bands.

4. According to the method of multi-view EEG speech imagery intention recognition based on adaptive graph learning in claim 1, Features: The specific process of preprocessing and feature extraction of EEG data in step 2 is as follows: after downsampling the EEG data, the noise and electrooculogram artifacts are filtered out through a bandpass filter; the filtered data is divided into 5 frequency bands, namely Delta, Theta, Alpha, Beta and Gamma; based on this data, the spectral entropy and power spectral density features of different channels and frequency bands are extracted as two views of the speech imagery recognition method.

5. According to the method of multi-view EEG speech imagery intention recognition based on adaptive graph learning in claim 1, Features: The specific process of constructing the multi-view EEG speech imagery intention recognition method of adaptive graph learning in step 3 is: using Euclidean distance to measure the similarity between samples, there is a negative correlation between the distance and similarity between samples, and considering the view weight and feature weight, the similarity construction model can be obtained: In order to prevent the similarity matrix S and view weights from having trivial solutions and adding semi-supervised label propagation information, regular term constraints can be obtained, and finally the objective function formula (1) is obtained.

6. The multi-view EEG speech imagery intention recognition method based on adaptive graph learning according to claim 1, Features: The method of joint iterative optimization in step 4 is: Step 4.1: Initialize the similarity matrix S, view weight vector α, and feature weight matrix Θ υ and the unlabeled continuous label matrix F u ; Step 4.2: Update the unlabeled continuous label matrix F u , fixed similarity matrix S, feature weight matrix Θ υ and the view weight vector α, we get F u The objective function is: Solving equation (3) can update the unlabeled continuous label matrix F u ; Step 4.3: Update the view weight vector α, fix the similarity matrix S and feature weight matrix Θ υ and the unlabeled continuous label matrix F u , the objective function of α is obtained as: Solving equation (4) can update the view weight vector α; Step 4.4: Update the similarity matrix S, fix the view weight vector α and feature weight matrix Θ υ and the unlabeled continuous label matrix F u , the objective function of S is obtained as: Solving equation (5) can update the similarity matrix S; Step 4.5, update the feature weight matrix Θ", fix the view weight vector α, the similarity matrix S and the unlabeled continuous label matrix F u , we get Θ υ The objective function is: Solving equation (6) can update the feature weight matrix Θ υ ; Step 4.6, repeat steps 4.2, 4.3, 4.4, and 4.5 until the objective function converges, and finally complete the similarity matrix S, view weight vector α, and feature weight matrix Θ υ and the unlabeled continuous label matrix F u Joint iterative optimization.

7. The multi-view EEG speech imagery intention recognition method based on adaptive graph learning according to claim 1, Features: Feature weight matrix Θ υ is a diagonal matrix, the diagonal elements represents the weight of the k-th feature of the υ-th view.

8. The multi-view EEG speech imagery intention recognition method based on adaptive graph learning according to claim 1, Features: Laplacian Matrix D is a diagonal matrix, and the i-th diagonal element is defined as 9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, Features: The memory stores computer programs; the processor executes the multi-view EEG speech imagery intention recognition method based on adaptive graph learning as described in any one of claims 1-8.

10. A readable storage medium storing a computer program; Features: When the computer program is executed by a processor, it is used to implement the multi-view EEG speech imagery intention recognition method based on adaptive graph learning as described in any one of claims 1 to 8.

Citation Information

Patent Citations

  • Electroencephalogram signal identification method based on adaptive multi-view deep learning framework

    CN115607169A

  • Speech imagination electroencephalogram decoding method for source domain mobility adaptive learning

    CN117725367A