Meta-learning method for brain-computer interface, EEG signal recognition method and device

Through the cross-domain gradient update and freezing network layer, the problem of poor performance of the brain-computer interface model on new users and privacy leakage is solved, and better model generalization and privacy protection is achieved, suitable for brains of multi-source domain data machine interface system.

CN114879838BActive Publication Date: 2024-11-22HUAZHONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210420053.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-20
Publication Date
2024-11-22
Estimated Expiration
2042-04-20

AI Technical Summary

Technical Problem

The existing brain-computer interface model requires the storage of source domain user data, resulting in user privacy leakage, and poor performance on new users and difficulty in calibration.

Method used

Using meta-learning method, through cross-domain gradient update optimization method, domain unchange knowledge is learned on multiple source domains, only model parameters are saved, domain offset parameters that do not affect the model are eliminated, network layer is frozen, and model is optimized to improve generalization Capacity and privacy protection.

Benefits of technology

Without destroying the privacy of the source domain data, better source domain models are trained to improve model generalization capabilities, reduce computing costs, avoid user privacy leakage, and effectively transfer learning on the target domain.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114879838B_ABST
    Figure CN114879838B_ABST
Patent Text Reader

Abstract

The invention discloses a meta-learning method for brain-computer interface, an electroencephalogram signal recognition method and a device, which belong to the field of brain-computer interface and machine learning. The method comprises: S1, obtaining a data set containing M source domain user data, M>2; S2, randomly combining the source domain user data in the data set in pairs to obtain M / 2 source domain user data groups, for each source domain user data group, using one of the source domain user data to update the model parameters of the brain-computer interface model, and calculating the loss of the updated brain-computer interface model on the other source domain user data; S3, calculating the average loss of each loss obtained in S2, and updating the model parameters of the brain-computer interface model before updating in S2 according to the average loss; S4, repeatedly executing S2-S3 until the loss function of the updated brain-computer interface model converges, and outputting the brain-computer interface model obtained by the last update. Protect the privacy of source domain data and improve the quality of source domain models.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of brain-computer interface and machine learning, and more specifically, to a meta-learning method for brain-computer interface, an electroencephalogram signal recognition method and device. Background Art

[0002] Brain-computer interface is a new form of human-computer interaction. It uses sensors to detect and analyze brain neural activity, which is usually recorded by electroencephalogram (EEG), to achieve direct communication between the user's brain and external devices, thereby realizing assisted interaction and rehabilitation training. It is one of the important supporting technologies for the future metaverse.

[0003] The following problems exist in brain-computer interfaces. First, in order to enable new users to use the brain-computer interface system with less or no calibration, cross-subject testing is a common strategy. Due to individual differences between each user, EEG signals will also have different patterns. Therefore, models trained on old user data often perform poorly on new users. Second, most existing cross-subject methods use the idea of ​​transfer learning and are based on features or instances. They require access to source data or features. EEG signals contain various private information, and the privacy protection performance of existing methods is limited. Third, the EEG signal itself is small in magnitude and cannot provide a large amount of training data. The difficulty of calibration also leads to a relatively smaller magnitude of calibration data on new users. The privacy protection issue also requires the system to try to give up saving the original data and only save the model parameters, so that the model can no longer access the training data of old users on new users.

[0004] Meta-learning technology can alleviate such problems of the model by upgrading the learning object from data to learning tasks. How to train a better model based on meta-learning technology without saving source domain user data and thus not leaking user privacy is an urgent problem to be solved in the field of brain-computer interface. Summary of the invention

[0005] In view of the defects of the prior art and the need for improvement, the present invention provides a meta-learning method, an EEG signal recognition method and a device for brain-computer interface, which aims to solve the problem that the existing brain-computer interface model needs to save source domain user data, resulting in user privacy leakage.

[0006] To achieve the above-mentioned purpose, according to one aspect of the present invention, a meta-learning method for brain-computer interface is provided, comprising: S1, obtaining a data set comprising M source domain user data, each of the source domain user data comprising multiple labeled samples, M>2; S2, randomly combining the source domain user data in the data set in pairs to obtain M / 2 groups of source domain user data groups, for each group of source domain user data groups, using one of the source domain user data to update the model parameters of the brain-computer interface model, and calculating the loss of the updated brain-computer interface model on the other source domain user data; S3, calculating the average loss of each loss obtained in S2, and updating the model parameters of the brain-computer interface model before updating in S2 according to the average loss; S4, repeatedly executing S2-S3 until the loss function of the updated brain-computer interface model converges, and outputting the brain-computer interface model obtained by the last update.

[0007] Furthermore, when the target domain subject calibration data is obtained before S4, S4 also includes: using the target domain subject calibration data to re-adjust the brain-computer interface model obtained by the last update; the final output of S4 is the re-adjusted brain-computer interface model.

[0008] Furthermore, the brain-computer interface model obtained by further adjustment is:

[0009]

[0010] Among them, θ′ MDMAML ,θ MDMAML They are the model parameters of the brain-computer interface model before and after re-adjustment, is θ MDMAML The gradient of , γ is the first learning rate of the brain-computer interface model, To re-adjust the previous brain-computer interface model, for The loss function of .

[0011] Furthermore, S2 also includes: for any source domain user data group, when the loss of the updated brain-computer interface model on another source domain user data is greater than the loss of the brain-computer interface model before the update on another source domain user data, the loss corresponding to the source domain user data group obtained in S2 is eliminated.

[0012] Furthermore, the use of one of the source domain user data to update the model parameters of the brain-computer interface model in S2 includes: freezing the parameters of the model network layers in the brain-computer interface model whose cosine similarity before and after the domain offset is higher than a set threshold, and using one of the source domain user data to update the parameters of the remaining model network layers in the brain-computer interface model.

[0013] Furthermore, the update operation in S3 is:

[0014]

[0015]

[0016]

[0017] Among them, θ is the brain-computer interface model f θ The model parameters, β is the second learning rate of the brain-computer interface model, is the gradient of θ, L MDMAML is the average loss obtained in S3, S i , S j is the i-th and j-th source domain user data in the dataset S, For the intermediate model In S j The loss calculated above, θ′ i is the brain-computer interface model f θ In S i The intermediate model obtained after the update The model parameters, α is the third learning rate of the brain-computer interface model, is the brain-computer interface model f θ In S i The loss calculated above; S2-S3 are repeatedly executed in S4 to optimize θ until L MDMAML Converge and output the brain-computer interface model obtained by the last update

[0018] According to another aspect of the present invention, a method for recognizing electroencephalogram signals is provided, comprising: obtaining data to be recognized of new users in the target domain and inputting the data into the brain-computer interface model finally obtained by the meta-learning method for brain-computer interface as described above, to obtain corresponding recognition results.

[0019] According to another aspect of the present invention, a meta-learning device for a brain-computer interface is provided, comprising: an acquisition module, used to acquire a data set containing M source domain user data, each of the source domain user data containing multiple labeled samples, M>2; a cross-domain gradient update module, used to randomly combine the source domain user data in the data set in pairs to obtain M / 2 groups of source domain user data groups, for each group of source domain user data groups, using one of the source domain user data to update the model parameters of the brain-computer interface model, and calculating the loss of the updated brain-computer interface model on the other source domain user data; an invariant domain update module, used to calculate the average loss of each loss obtained in the cross-domain gradient update module, and update the model parameters of the brain-computer interface model before updating in the cross-domain gradient update module according to the average loss; a repeated update and output module, used to repeatedly execute the cross-domain gradient update module and the invariant domain update module until the loss function of the updated brain-computer interface model converges, and output the brain-computer interface model obtained by the last update.

[0020] According to another aspect of the present invention, there is provided an electronic device, comprising: a processor; and a memory storing a computer executable program, wherein when the program is executed by the processor, the processor executes the meta-learning method for brain-computer interface as described above, or the processor executes the electroencephalogram signal recognition method as described above.

[0021] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0022] (1) A cross-domain gradient update optimization method is adopted on multiple source domains to learn effective underlying domain invariant knowledge in high-noise low-level data, improve the generalization ability of the model, and train a better source domain model without destroying the privacy of the source domain data. Since the above steps are all performed on the source domain, only the model parameters need to be recorded, and there is no risk of source domain data being stored in the device and leaked, thus avoiding user privacy leakage. Only the gradient update method is improved, which is independent of the model architecture itself. It is applicable to almost all artificial neural network models using gradient descent optimization method and has strong applicability.

[0023] (2) For all source domain data, first randomly remove a source domain, and use the source domain splicing method to train the model in a traditional way on the remaining source domain until the loss function converges, and then train again on the removed source domain to find the model parameter layer that is almost unaffected in the latter stage of training, and determine that the above parameter layer has nothing to do with cross-domain knowledge; based on this, in the actual training process, when updating the model parameters to obtain the intermediate model, there is no need to update the parameters of these layers, that is, freeze the network layers with small changes before and after the domain shift, and train other network layers that are more sensitive to cross-domain knowledge, so that the model can better learn domain invariant knowledge and at the same time achieve the effect of reducing the computational cost;

[0024] (3) During the transfer learning process where the model learns the knowledge of source domain data and applies it to the target domain, check whether there is indeed an improvement in the target domain. If there is no improvement, it is negative transfer. The loss corresponding to the negative transfer is eliminated and not included in the update. This can prevent the model from negatively transferring in the source domain during the training process, thereby achieving the effect of training a better source domain model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flow chart of a meta-learning method for a brain-computer interface provided by an embodiment of the present invention;

[0026] Figure 2 A model optimization process for cross-user meta-learning using MDMAML provided in an embodiment of the present invention;

[0027] Figure 3 A block diagram of a meta-learning system for a brain-computer interface provided by an embodiment of the present invention;

[0028] Figure 4 A block diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0030] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the drawings are used to distinguish similar objects but not necessarily to describe a specific order or sequence.

[0031] The meta-learning method for brain-computer interface provided by the embodiment of the present invention includes a framework of a multi-source domain cross-domain meta-learning method based on an artificial neural network optimization algorithm, a layer freezing method, and a method for reducing negative transfer, such as Figure 2Based on two commonly used artificial neural network structures in EEG, online and offline scenarios are considered respectively, and experiments are conducted on two motor imagery datasets and two event-related potential datasets. The results show that the model trained under the Multi-DomainModel-Agnostic Meta-Learning (MDMAML) framework performs better than traditional methods and other cutting-edge methods.

[0032] Figure 1 A flow chart of a meta-learning method for a brain-computer interface provided by an embodiment of the present invention. Figure 1 , combined with Figure 2 , the meta-learning method for brain-computer interface in this embodiment is described in detail, and the method includes operations S1-S4.

[0033] Operation S1: obtaining a data set including M source domain user data, each source domain user data including a plurality of labeled samples, M>2.

[0034] The data set obtained is S. S i is the source domain user data in the dataset S. Each source domain user data S i Contains multiple labeled samples The source domain user data S i The nth sample and label in the source domain S i The number of samples in .

[0035] In this embodiment, before model training, the model parameters f of the brain-computer interface model need to be initialized. θ , input the learning rates α, β, γ of the brain-computer interface model, and input the loss function L of the brain-computer interface model.

[0036] Operation S2 randomly combines the source domain user data in the data set in pairs to obtain M / 2 groups of source domain user data. For each group of source domain user data, one of the source domain user data is used to update the model parameters of the brain-computer interface model, and the loss of the updated brain-computer interface model is calculated on the other source domain user data.

[0037] The source domain user data in the data set are randomly combined in pairs to obtain M / 2 groups of source domain user data. Each source domain user data group (S i , S j ) as a pair of domains for cross-domain learning. For any source domain user data set (S i , S j ), using the source domain user data S i Update the model parameters of the brain-computer interface model to obtain the intermediate model Model parameters: In the source domain user data S j Calculate the loss of the intermediate model

[0038] According to an embodiment of the present invention, in operation S2, updating the model parameters of the brain-computer interface model using one of the source domain user data includes: freezing the parameters of the model network layer in the brain-computer interface model whose cosine similarity before and after the domain shift is higher than a set threshold, and using the one of the source domain user data to update the parameters of the remaining model network layers in the brain-computer interface model. The threshold is set to 0.9, for example, and the frozen model network layer is not included in the update.

[0039] According to an embodiment of the present invention, operation S2 further includes: for any source domain user data group, when the loss of the updated brain-computer interface model on another source domain user data is greater than the loss of the brain-computer interface model before the update on another source domain user data, the loss corresponding to the source domain user data group obtained in operation S2 is eliminated. Specifically, for example, when the intermediate model In the source domain user data S j Loss on Larger than the previous brain-computer interface model In the source domain user data S j Loss on Such losses shall not be accumulated.

[0040] Operation S3, calculating the average loss of the losses obtained in operation S2, and updating the model parameters of the brain-computer interface model before being updated in operation S2 according to the average loss.

[0041] It should be noted that what is calculated in operation S3 is the average loss of each loss after the elimination operation in operation S2, and the average loss obtained is:

[0042]

[0043]

[0044] Furthermore, according to the average loss L MDMAML The model parameters of the brain-computer interface model before being updated in operation S2 are updated, and the updating operation is:

[0045]

[0046] Among them, θ is the brain-computer interface model f θ The model parameters, β is the second learning rate of the brain-computer interface model, is the gradient of θ, L MDMAML is the average loss obtained in S3, S i , S jis the i-th and j-th source domain user data in the dataset S, For the intermediate model In S j The loss calculated above, θ′ i is the brain-computer interface model f θ In S i The intermediate model obtained after the update The model parameters, α is the third learning rate of the brain-computer interface model, is the brain-computer interface model f θ In S i The loss calculated above.

[0047] Operation S4, repeatedly perform operations S2-S3 until the loss function of the updated brain-computer interface model converges, and output the brain-computer interface model obtained by the last update.

[0048] According to an embodiment of the present invention, in operation S4, operations S2 to S3 are repeatedly performed to optimize θ until L MDMAML Converge and output the brain-computer interface model obtained by the last update

[0049] When there is no calibration data of the target domain subjects, the last updated brain-computer interface model is directly output in operation S4. And directly use the brain-computer interface model in offline scenarios Perform EEG signal recognition.

[0050] When there is calibration data of the target domain subject (that is, when the calibration data of the target domain subject is obtained before operation S4), refer to Figure 2 Operation S4 also includes: using the target domain subject calibration data to adjust the brain-computer interface model obtained by the last update The final output in operation S4 is the brain-computer interface model obtained by re-adjustment. And directly use the brain-computer interface model in offline scenarios Perform EEG signal recognition.

[0051] The brain-computer interface model obtained by re-adjustment for:

[0052]

[0053] Among them, θ′ MDMAML ,θ MDMAML They are the model parameters of the brain-computer interface model before and after re-adjustment, is θ MDMAML The gradient of , γ is the first learning rate of the brain-computer interface model, To re-adjust the previous brain-computer interface model, for The loss function of .

[0054] The embodiments of the present invention conduct cross-user experiments (retaining one user for verification) on the BNCI-Horizon-001-2014 (MI1), BNCI-Horizon-002-2014 (MI2), BNCI-Horizon-009-2014 (ERP1), and Kaggle-Inria-BCI-Challenge (ERP2) datasets, and test the effect of MDMAML when the amount of calibration data of the target new user changes from small to large. Among them, k-shot means using k labeled samples for calibration; Baseline is a baseline method using source domain splicing; MAML is a model-agnostic meta-learning method; ANIL (Almost No Inner Loop) is a meta-learning method of MAML variant; ProtoNets (Prototypical Networks) is a small sample method of prototype networks; MDMAML is a method in the embodiments of the present invention; EEGNet and Shallow ConvNet are two commonly used architectures of artificial neural network models for EEG signals.

[0055] The experimental results of the mean ± standard deviation of the online cross-subject small sample calibration classification accuracy on the MI1 dataset are shown in Table 1.

[0056] Table 1

[0057]

[0058] The experimental results of the mean ± standard deviation of the online cross-subject small sample calibration classification accuracy on the MI2 dataset are shown in Table 2.

[0059] Table 2

[0060]

[0061] The experimental results of the mean ± standard deviation of the online cross-subject small sample calibration classification AUC values ​​on the ERP1 dataset are shown in Table 3.

[0062] Table 3

[0063]

[0064] The experimental results of the mean ± standard deviation of the online cross-subject small sample calibration classification AUC values ​​on the ERP2 dataset are shown in Table 4.

[0065] Table 4

[0066]

[0067] Experimental results show that the effect of MDMAML is better than that of traditional optimization methods and other cutting-edge algorithms. It achieves an improvement in the optimization method of artificial neural network models in multi-source domains with privacy protection, and achieves the effect of improving the performance of brain-computer interface systems in this scenario.

[0068] The advantage of the meta-learning method for brain-computer interface in the embodiment of the present invention is that it makes better use of source domain data and trains a better source domain model. Compared with the traditional splicing method that does not distinguish between multi-source domain data, the MDMAML method's meta-learning process in multi-source domains makes the model familiar with the cross-domain learning process, learns better about domain-invariant knowledge, learns more in the model parameter layer that is more sensitive to cross-domain knowledge, and is less affected by negative transfer during the learning process. This solution only improves the model optimization method, so it can be applied to almost all artificial neural network models that use the gradient descent update method; at the same time, because only the source domain model is saved, and the source domain data is not used by new subjects in the target domain, the privacy of the source domain data is also protected.

[0069] The embodiment of the present invention also provides a method for recognizing an EEG signal, comprising: obtaining the target domain new user data to be recognized and inputting the data into: Figure 1-Figure 2 The corresponding recognition results are obtained in the brain-computer interface model finally obtained by the meta-learning method for brain-computer interface in the illustrated embodiment.

[0070] Figure 3 A block diagram of a meta-learning system for a brain-computer interface provided by an embodiment of the present invention. Figure 3 The meta-learning system 300 for brain-computer interface includes an acquisition module 310, a cross-domain gradient update module 320, an invariant domain update module 330, and a repeated update and output module 340.

[0071] The acquisition module 310 performs operation S1, for example, to acquire a data set including M source domain user data, each source domain user data including a plurality of labeled samples, M>2.

[0072] The cross-domain gradient update module 320, for example, performs operation S2, which is used to randomly combine the source domain user data in the data set in pairs to obtain M / 2 groups of source domain user data groups. For each group of source domain user data groups, one of the source domain user data is used to update the model parameters of the brain-computer interface model, and the loss of the updated brain-computer interface model is calculated on the other source domain user data.

[0073] The invariant domain update module 330, for example, performs operation S3, which is used to calculate the average loss of each loss obtained in the cross-domain gradient update module, and update the model parameters of the brain-computer interface model before updating in the cross-domain gradient update module according to the average loss.

[0074] The repeated update and output module 340, for example, executes operation S4, which is used to repeatedly execute the cross-domain gradient update module and the invariant domain update module until the loss function of the updated brain-computer interface model converges, and outputs the brain-computer interface model obtained by the last update.

[0075] The meta-learning system 300 for brain-computer interface is used to perform the above Figure 1-Figure 2 The meta-learning method for brain-computer interface in the embodiment shown. For details not covered in this embodiment, please refer to the aforementioned Figure 1-Figure 2 The meta-learning method for brain-computer interface in the illustrated embodiment will not be described in detail here.

[0076] The embodiment of the present disclosure also shows an electronic device, such as Figure 4 As shown, the electronic device 400 includes a processor 410 and a readable storage medium 420. The electronic device 400 can execute the above Figure 1-Figure 2 The meta-learning method for brain-computer interface described in ; alternatively, the electronic device 400 can execute the EEG signal recognition method described above.

[0077] Specifically, the processor 410 may include, for example, a general-purpose microprocessor, an instruction set processor and / or a related chipset and / or a dedicated microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 410 may also include an onboard memory for cache purposes. The processor 410 may be a processor for executing reference Figure 1-Figure 2 The described method flow according to the embodiment of the present disclosure may be a single processing unit or multiple processing units for different actions.

[0078] The readable storage medium 420 may be, for example, any medium capable of containing, storing, conveying, propagating, or transmitting instructions. For example, the readable storage medium may include, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, component, or propagation medium. Specific examples of readable storage media include: a magnetic storage device, such as a magnetic tape or a hard disk (HDD); an optical storage device, such as a compact disk (CD-ROM); a memory, such as a random access memory (RAM) or flash memory; and / or a wired / wireless communication link.

[0079] The readable storage medium 420 may include a computer program 421, which may include code / computer executable instructions, which when executed by the processor 410 causes the processor 410 to perform, for example, the above-mentioned combination Figure 1-Figure 2 The described method sequences and any variations thereof.

[0080] Computer program 421 may be configured to have, for example, computer program code including computer program modules. For example, in an exemplary embodiment, the code in computer program 421 may include one or more program modules, for example, 421A, module 421B, ... It should be noted that the division method and number of modules are not fixed, and those skilled in the art may use appropriate program modules or program module combinations according to actual conditions. When these program module combinations are executed by processor 410, processor 410 may execute, for example, the above combination of Figure 1-Figure 2 The described method sequences and any variations thereof.

[0081] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A meta-learning method for brain-computer interface, characterized in that: include: S1, obtaining a data set including M source domain user data, each of the source domain user data including multiple labeled samples, M>2; S2, randomly combining the source domain user data in the data set in pairs to obtain M / 2 groups of source domain user data, for each group of source domain user data, using one of the source domain user data to update the model parameters of the brain-computer interface model, and calculating the loss of the updated brain-computer interface model on the other source domain user data; S3, calculating the average loss of the losses obtained in S2, and updating the model parameters of the brain-computer interface model before being updated in S2 according to the average loss; S4, repeatedly executing S2-S3 until the loss function of the updated brain-computer interface model converges, and outputting the brain-computer interface model obtained by the last update; When the target domain subject calibration data is obtained before S4, S4 further includes: using the target domain subject calibration data to adjust the brain-computer interface model updated last time; The final output in S4 is the brain-computer interface model obtained by re-adjustment; The brain-computer interface model obtained by adjusting again is: Among them, θ′ MDMAML ,θ MDMAML They are the model parameters of the brain-computer interface model before and after re-adjustment, is θ MDMAML The gradient of , γ is the first learning rate of the brain-computer interface model, To re-adjust the previous brain-computer interface model, for The loss function of The updating of the model parameters of the brain-computer interface model using one of the source domain user data in S2 includes: freezing the parameters of the model network layers in the brain-computer interface model whose cosine similarity before and after the domain shift is higher than a set threshold, and updating the parameters of the remaining model network layers in the brain-computer interface model using one of the source domain user data; The update operation in S3 is: Among them, θ is the brain-computer interface model f θ The model parameters, β is the second learning rate of the brain-computer interface model, is the gradient of θ, L MDMAML is the average loss obtained in S3, S i , S j is the i-th and j-th source domain user data in the dataset S, For the intermediate model In S j The loss calculated above, θ′ i is the brain-computer interface model f θ In S i The intermediate model obtained after the update The model parameters, α is the third learning rate of the brain-computer interface model, is the brain-computer interface model f θ In S i The loss calculated above; In S4, S2-S3 are repeatedly executed to optimize θ until L MDMAML Converge and output the brain-computer interface model obtained by the last update 2. The meta-learning method for brain-computer interface according to claim 1, characterized in that: The S2 also includes: for any source domain user data group, when the loss of the updated brain-computer interface model on another source domain user data is greater than the loss of the brain-computer interface model before the update on another source domain user data, the loss corresponding to the source domain user data group obtained in S2 is eliminated.

3. A method for recognizing an electroencephalogram signal, characterized in that: include: The data of new users to be identified in the target domain is obtained and input into the brain-computer interface model finally obtained by the meta-learning method for brain-computer interface as described in any one of claims 1-2 to obtain the corresponding identification result.

4. A meta-learning device for brain-computer interface, characterized in that: include: An acquisition module is used to acquire a data set including M source domain user data, each of which includes multiple labeled samples, M>2; A cross-domain gradient update module, used for randomly combining the source domain user data in the data set in pairs to obtain M / 2 groups of source domain user data, for each group of source domain user data, using one of the source domain user data to update the model parameters of the brain-computer interface model, and calculating the loss of the updated brain-computer interface model on the other source domain user data; An invariant domain update module, used to calculate the average loss of each loss obtained in the cross-domain gradient update module, and update the model parameters of the brain-computer interface model before updating in the cross-domain gradient update module according to the average loss; A repeated update and output module, used for repeatedly executing the cross-domain gradient update module and the invariant domain update module until the loss function of the updated brain-computer interface model converges, and outputting the brain-computer interface model obtained by the last update; When the target domain subject calibration data is acquired before the repeated updating and outputting module performs the operation, the repeated updating and outputting module is further used to: use the target domain subject calibration data to adjust the brain-computer interface model obtained by the last update again; The repeated updating and output module finally outputs the brain-computer interface model obtained by re-adjustment; The brain-computer interface model obtained by adjusting again is: Among them, θ ′ MDMAML ,θ MDMAML They are the model parameters of the brain-computer interface model before and after re-adjustment, is θ MDMAML The gradient of , γ is the first learning rate of the brain-computer interface model, To re-adjust the previous brain-computer interface model, for The loss function of The cross-domain gradient update module uses one of the source domain user data to update the model parameters of the brain-computer interface model, including: freezing the parameters of the model network layer in the brain-computer interface model whose cosine similarity before and after the domain shift is higher than a set threshold, and using one of the source domain user data to update the parameters of the remaining model network layers in the brain-computer interface model; The update operation of the unchanged domain update module is: Among them, θ is the brain-computer interface model f θ The model parameters, β is the second learning rate of the brain-computer interface model, is the gradient of θ, L MDMAML is the average loss obtained in the invariant domain update module, S i , S j is the i-th and j-th source domain user data in the dataset S, For the intermediate model In S j The loss calculated above, θ′ i is the brain-computer interface model f θ In S i The intermediate model obtained after the update The model parameters, α is the third learning rate of the brain-computer interface model, is the brain-computer interface model f θ In S i The loss calculated above; The repeated update and output module repeatedly executes the cross-domain gradient update module and the invariant domain update module to optimize θ until L MDMAML Converge and output the brain-computer interface model obtained by the last update 5. An electronic device, characterized in that: include: processor; A memory storing a computer executable program, which, when executed by the processor, enables the processor to execute the meta-learning method for brain-computer interface as described in claims 1-2, or enables the processor to execute the electroencephalogram signal recognition method as described in claim 3.