Methods, apparatus, devices, media, and programs for decoding motion imagery.

Abstract

A method, apparatus, device, medium, and program product for decoding motor imagery are disclosed. The method includes: acquiring target EEG data of a target subject, historical EEG data of multiple existing subjects, and multiple base classification models corresponding one-to-one with the multiple existing subjects, each base classification model being trained based on the historical EEG data of its corresponding existing subject; determining the feature matrices of the target EEG data and the historical EEG data of the multiple existing subjects respectively, to obtain the target feature matrix of the target subject and the source domain feature matrices of the multiple existing subjects; using the target feature matrix as the target domain, performing one-to-one single-source domain transfer learning with the multiple source domain feature matrices respectively, to obtain multiple target feature transformation matrices corresponding one-to-one with the multiple existing subjects; inputting the multiple target feature transformation matrices into the corresponding base classification models respectively, to obtain multiple base classification results; and determining the motor imagery decoding information of the target subject based on the multiple base classification results.

Description

Technical Field

[0001] This disclosure relates to the fields of brain-computer interfaces and motor imagery, and more particularly to a method, apparatus, electronic device, storage medium, and computer program product for decoding motor imagery. Background Technology

[0002] Motor imagery (MI), as a typical brain-computer interface (BCI) paradigm, has received widespread attention in neuroscience, rehabilitation medicine, and human-computer interaction due to its advantages such as requiring no external device stimulation and allowing for active control. It provides a new way for people with movement disorders to interact with the outside world, particularly in the rehabilitation of stroke patients. Among various brain signals, electroencephalography (EEG) is widely used in motor imagery decoding research due to its advantages of high temporal resolution, non-invasiveness, and low cost. Motor imagery decoding based on EEG signals involves feature extraction and task recognition for an individual's imagined motor task. In subsequent applications, the information obtained from motor imagery decoding can be used to convert an individual's motor intentions into output commands, establishing an organic information link between motor intentions and interface devices.

[0003] With the rapid development of artificial intelligence, especially machine learning, machine learning models are widely used in motor imagery decoding tasks. For example, classification models in machine learning can be used to classify the EEG data of target subjects (i.e., test subjects) to predict the task type corresponding to the EEG data. However, due to the non-stationary nature of EEG and the data distribution differences caused by large individual differences in brain signals, the efficiency of motor imagery decoding (MI) is reduced, affecting the use of brain imagery decoding (BCI). Transfer learning (TL) is an effective method to solve the data distribution differences. It can use the existing EEG data of subjects to find the similarity between it and the EEG data of target subjects, and use the knowledge obtained from the source domain to help improve the performance of target subjects in motor imagery decoding tasks.

[0004] In related technologies, when using transfer learning to improve the performance of motor imagery decoding tasks, the common practice is to fuse the EEG data of multiple existing subjects into a single source domain, and then perform single-source domain transfer learning with the target subject's EEG data. However, this process ignores individual differences between different existing subjects, resulting in significant differences between the target subject's and the fused existing subject's EEG data. The consistency of the EEG data distribution is poor, leading to limited transfer effects and failing to obtain ideal motor imagery decoding results. Summary of the Invention

[0005] To address the aforementioned problems, this disclosure is proposed. Embodiments of this disclosure provide a method, apparatus, electronic device, storage medium, and computer program product for motion image decoding.

[0006] According to one aspect of the present disclosure, a method for decoding motor imagery is provided, comprising: acquiring target EEG data of a target subject, historical EEG data of multiple existing subjects, and multiple base classification models corresponding one-to-one with the multiple existing subjects, wherein each base classification model is trained based on the historical EEG data of the corresponding existing subject; determining feature matrices of the target EEG data and the historical EEG data of the multiple existing subjects respectively by covariance matrix centroid alignment and extraction of tangent space features, thereby obtaining a target feature matrix of the target subject and source domain feature matrices of the multiple existing subjects; using the target feature matrix as the target domain, performing one-to-one single-source domain transfer learning with the source domain feature matrices of the multiple existing subjects respectively, thereby obtaining multiple target feature transformation matrices corresponding one-to-one with the multiple existing subjects; based on the base classification models and target feature transformation matrices corresponding to the multiple existing subjects respectively, inputting the multiple target feature transformation matrices into the corresponding base classification models respectively, and having the multiple base classification models predict the task type of the motor task imagined by the target subject respectively, thereby obtaining multiple base classification results of the target subject; and determining the motor imagery decoding information of the target subject based on the multiple base classification results.

[0007] In some embodiments, determining the motion imagery decoding information of the target subject based on multiple base classification results includes: fusing the multiple base classification results using weighted voting based on the fusion weights corresponding to each of the multiple base classification models to obtain motion imagery decoding information; or, processing the multiple base classification results using a majority voting algorithm to determine the motion imagery decoding information.

[0008] In some embodiments, the method further includes the step of obtaining multiple base classification models: taking the source domain feature matrices of each of the multiple existing subjects as sample feature matrices, and determining the sample label of each sample feature matrix to obtain sample data corresponding to each of the multiple existing subjects, wherein the sample data includes the corresponding sample feature matrix and its sample label; constructing an initial base classification model for the sample data corresponding to each existing subject, and training the initial base classification model using the sample data corresponding to the existing subject to obtain multiple base classification models corresponding to each of the multiple existing subjects as multiple base classification models.

[0009] In some embodiments, the method further includes the following steps for determining fusion weights: extracting a predetermined number of EEG data from the target EEG data, and labeling the extracted EEG data according to the task type to obtain calibration data; determining the calibration feature matrix of the calibration data; performing one-to-one single-source-domain transfer learning on the calibration feature matrix based on the source domain feature matrices and base classification models corresponding to multiple existing subjects, and inputting the matrix after transfer learning into the corresponding base classification model to obtain the prediction results of multiple base classification models; determining the calibration parameters of multiple base classification models based on the labeling results of the calibration data and the prediction results of multiple base classification models, wherein the calibration parameters characterize the accuracy of the base classification models; and determining the fusion weights of the base classification models based on the calibration parameters.

[0010] In some embodiments, determining the fusion weight of the base classification model based on the calibration parameter includes: if the calibration parameter is greater than a preset threshold, determining the calibration parameter as the fusion weight of the corresponding base classification model; if the calibration parameter is not greater than the preset threshold, determining the fusion weight of the base classification model corresponding to the calibration parameter as 0.

[0011] In some embodiments, obtaining historical EEG data of multiple existing subjects includes: obtaining historical EEG data of multiple existing subjects from public datasets or private datasets collected by individuals.

[0012] According to another aspect of the present disclosure, an apparatus for decoding motor imagery is provided, comprising: an acquisition unit configured to acquire target EEG data of a target subject, historical EEG data of multiple existing subjects, and multiple base classification models corresponding one-to-one with the multiple existing subjects, wherein each base classification model is trained based on the historical EEG data of the corresponding existing subject; and a calculation unit configured to determine feature matrices of the target EEG data and the historical EEG data of the multiple existing subjects respectively by covariance matrix centroid alignment and extraction of tangent space features, thereby obtaining a target feature matrix of the target subject and source domain feature matrices of the multiple existing subjects. The transfer unit is configured to use the target feature matrix as the target domain and perform one-to-one single-source domain transfer learning with the source domain feature matrices of multiple existing subjects to obtain multiple target feature transformation matrices that correspond one-to-one with the multiple existing subjects; the classification unit is configured to input the multiple target feature transformation matrices into the corresponding base classification models based on the base classification models corresponding to the multiple existing subjects, and the multiple base classification models predict the task type of the motion task imagined by the target subject to obtain multiple base classification results of the target subject; the decoding unit is configured to determine the motion imagination decoding information of the target subject based on the multiple base classification results.

[0013] In another aspect of this disclosure, an electronic device is provided, comprising: a memory for storing a computer program product; and a processor for executing the computer program product stored in the memory, wherein when the computer program product is executed, it implements the method in any of the above embodiments.

[0014] In another aspect of this disclosure, a computer-readable storage medium is provided that stores computer program instructions thereon, which, when executed by a processor, implement the methods in any of the above embodiments.

[0015] In another aspect of this disclosure, a computer program product is provided, including computer program instructions that, when executed by a processor, implement the methods in any of the above embodiments.

[0016] Based on the method, apparatus, medium, and electronic device for motor imagery decoding provided in the above embodiments of this disclosure, after acquiring the target EEG data of the target subject, the historical EEG data of multiple existing subjects, and their corresponding base classification models, the target feature matrix of the target EEG data and the source domain feature matrix of the historical EEG data are determined through covariance matrix centroid alignment and tangent space feature extraction. Then, the target feature matrix is ​​used as the target domain and subjected to one-to-one single-source domain transfer learning with multiple source domain feature matrices, so that the target EEG data is aligned with the historical EEG data of multiple existing subjects, resulting in target feature transformation matrices corresponding to multiple existing subjects. Subsequently, the multiple target feature transformation matrices are input into their respective base classification models, and the target feature transformation matrices are processed by the multiple base classification models to predict the task type of the motor task imagined by the target subject, obtaining multiple base classification results. Finally, the multiple base classification results are fused to determine the motor imagery decoding information of the target subject. Since each of the multiple base classification models is trained based on the historical EEG data of an existing subject, different base classification models can reflect individual differences among different subjects, making each base classification model more targeted. Then, through one-to-one single-source domain transfer learning, the target subject's EEG data is aligned with the historical EEG data corresponding to the base classification model. This alignment-then-decision fusion approach addresses the problem of ignoring differences among multiple subjects in the source domain during multi-source transfer learning, reducing data distribution differences caused by individual EEG variations and EEG data non-stationarity. It fully utilizes the targeting of the base classification models, helping to improve the accuracy of motor imagery decoding and thus expanding its application prospects in various fields. For example, it can more accurately determine the intentions of people with movement disorders or stroke patients, thereby facilitating interaction for people with movement disorders and rehabilitation treatment for stroke patients.

[0017] The technical solutions of this disclosure will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] The above and other objects, features, and advantages of this disclosure will become more apparent from the more detailed description of the embodiments thereof in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the disclosure and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0019] Figure 1 This is a scene illustration of the method for motion image decoding disclosed herein;

[0020] Figure 2 A flowchart of one embodiment of the method for motion image decoding disclosed herein;

[0021] Figure 3 This is yet another scenario illustration of the method for motion image decoding disclosed herein;

[0022] Figure 4 This is a flowchart illustrating the acquisition of a base classification model in one embodiment of the method for decoding motion imagery disclosed herein;

[0023] Figure 5 This is a flowchart illustrating the determination of fusion weights in one embodiment of the method for motion image decoding disclosed herein;

[0024] Figure 6 This is a schematic diagram of one embodiment of the apparatus for acquiring the state of a target disclosed herein;

[0025] Figure 7 A structural diagram of an electronic device provided for an exemplary embodiment of this application. Detailed Implementation

[0026] Example embodiments according to this disclosure will now be described in detail with reference to the accompanying drawings. It is obvious that the described embodiments are merely some embodiments of this disclosure, and not all embodiments of this disclosure, and it should be understood that this disclosure is not limited to the example embodiments described herein.

[0027] It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of this disclosure.

[0028] Those skilled in the art will understand that the terms "first," "second," etc., in the embodiments of this disclosure are only used to distinguish different steps, devices, or modules, and do not represent any specific technical meaning, nor do they indicate a necessary logical order between them.

[0029] It should also be understood that in the embodiments disclosed herein, "a plurality of" may refer to two or more, and "at least one" may refer to one, two or more.

[0030] It should also be understood that any component, data or structure mentioned in the embodiments of this disclosure can generally be understood as one or more unless expressly defined or given to the contrary in the context.

[0031] Furthermore, the term "and / or" in this disclosure is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this disclosure generally indicates that the preceding and following related objects have an "or" relationship.

[0032] It should also be understood that the description of the various embodiments in this disclosure emphasizes the differences between the various embodiments, and the similarities or similarities can be referred to each other. For the sake of brevity, they will not be described in detail.

[0033] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0034] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use.

[0035] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0036] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0037] The embodiments of this disclosure can be applied to electronic devices such as terminal devices, computer systems, and servers, and can operate with a wide range of other general-purpose or special-purpose computing system environments or configurations. Examples of well-known terminal devices, computing systems, environments, and / or configurations suitable for use with electronic devices such as terminal devices, computer systems, or servers include, but are not limited to: personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments that include any of the above systems, etc.

[0038] Electronic devices such as terminal devices, computer systems, and servers can be described in the general context of computer system executable instructions (such as program modules) executed by a computer system. Typically, program modules can include routines, programs, object programs, components, logic, data structures, etc., which perform specific tasks or implement specific abstract data types. Computer systems / servers can be implemented in a distributed cloud computing environment. In a distributed cloud computing environment, tasks can be executed by remote processing devices linked through a communication network. In a distributed cloud computing environment, program modules can reside on local or remote computing system storage media, including storage devices.

[0039] Application Overview

[0040] In practice, individual differences often exist among different subjects, such as mental state, cognitive level, EEG differences, and brain anatomy. Consequently, there are often differences in the EEG data of different subjects. Related technologies typically involve aggregating historical EEG data from multiple existing subjects and then performing single-source domain transfer learning with the EEG data of the target subject to improve the alignment between the target subject's and existing subjects' EEG data.

[0041] In the process of realizing this disclosure, the inventors discovered that the above-mentioned single-source domain transfer learning ignores the individual differences between subjects from different source domains, resulting in limited transfer learning effects. It is difficult to guarantee the consistency of the EEG data distribution after the target subject and existing subjects are fused, thus affecting the accuracy of running imagination decoding.

[0042] Exemplary Overview

[0043] The following is combined Figure 1 The method for motion image decoding disclosed herein is illustrated by way of example. Figure 1 This is a schematic diagram of a scene to which the method for motion image decoding disclosed herein applies. Figure 1 In the scenario shown, electronic device 100 can be a terminal computer or a server, on which pre-trained base classification models 110, 120, and 130 are installed. The base classification models can be classification models from deep learning models, such as support vector machines (SVMs) or K-nearest neighbor classifiers. Each base classification model is trained based on the historical EEG data of an existing subject. For example, base classification model 110 is a support vector machine, and its training samples only include the historical EEG data 170 of subject a; base classification model 120 can be a support vector machine or a K-nearest neighbor classifier, and its training samples only include the historical EEG data 180 of subject b; base classification model 130's training samples only include the historical EEG data 190 of subject c.

[0044] Electronic device 100 can acquire target EEG data 140 from the target subject through various methods, including accessing a preset network location. It can also communicate with an EEG testing instrument, allowing the electronic device 100 to directly acquire the target EEG data 140 when the instrument is testing the target subject. Subsequently, the electronic device 100 can perform covariance matrix centroid alignment and extract tangent space features from the target EEG data 140, historical EEG data 170, historical EEG data 180, and historical EEG data 190, generating a target feature matrix 150 for the target EEG data 140, as well as source domain feature matrices 171, 181, and 191. Then, the electronic device 100 performs one-to-one single-source-domain transfer learning on the target feature matrix 150 with the source domain feature matrices 171, 181, and 191, respectively, to obtain target feature transformation matrices 141, 142, and 143. Next, the target feature transformation matrices 141, 142, and 143 are input into the corresponding base classification models 110, 120, and 130, respectively, to obtain base classification results 111, 121, and 131 output by each base classification model. These base classification results represent the movement type of the motor task imagined by the target subject, as predicted by the base classification model. Then, the electronic device 100 can fuse the base classification results 111, 121, and 131, for example, using a majority voting algorithm or a weighted voting algorithm, to determine the target subject's motor imagery decoding information 160. This motor imagery decoding information may include the task type of the motor task imagined by the target subject during the acquisition of the target EEG data. In practical applications, decoding information through motor imagery can reveal the motor intentions of the subjects (test subjects), which can facilitate interaction and rehabilitation treatment for people with motor disorders or stroke patients.

[0045] Exemplary methods

[0046] Figure 2 This is a flowchart of one embodiment of the method for motion image decoding disclosed herein. Figure 2 As shown, the process includes the following steps:

[0047] Step 210: Obtain the target EEG data of the target subject, the historical EEG data of multiple existing subjects, and multiple base classification models that correspond one-to-one with the multiple existing subjects.

[0048] Each base classification model is trained based on the historical EEG data of the corresponding existing subjects.

[0049] In this embodiment, the target subject refers to the test subject, and the target EEG data refers to the data obtained by the EEG testing instrument when performing an EEG test on the target subject. The method for decoding motor imagery disclosed herein processes the target EEG data to predict the type of motor task imagined by the target subject during the EEG testing phase, thereby obtaining the target subject's motor imagery decoding information.

[0050] Existing subjects refer to historical test subjects, and historical EEG data refers to the EEG data collected from existing subjects during testing. Each existing subject can correspond to one or more historical EEG data sets. Historical EEG data can represent the EEG data obtained from conducting EEG tests on existing subjects. As an example, during the EEG testing phase, subjects can be instructed to imagine a specific motor task while simultaneously acquiring EEG data. For instance, subjects can be instructed to imagine "movement to the left," and their EEG data can be acquired simultaneously; then "movement to the left" can serve as the label for that segment of EEG data.

[0051] The following example illustrates the EEG testing. After connecting the EEG testing instrument to the target subject, bandpass filtering (e.g., 8Hz-30Hz) can be used to periodically sample the subject's EEG data. For example, a 3-second data segment can be collected each time, resulting in a sampling record composed of multiple data segments arranged in chronological order. Repeating this process yields multiple sampling records, which together constitute the target subject's EEG data. Independent component analysis can then be used to remove physiological artifacts such as electrooculography (EOG) and electromyography (EMG) from the EEG recordings. The target EEG data is then represented in the form of a matrix group. Each matrix in the matrix group represents a sampling record. The number of rows in the matrix represents the number of electrodes on the EEG testing instrument, i.e., the number of EEG channels, and the number of columns represents the number of sample points (i.e., the number of sampling times). Each value corresponds to a data segment in the corresponding sampling record.

[0052] In some optional implementations of this embodiment, the executing entity may obtain historical EEG data of multiple existing subjects from a public dataset or a privately collected dataset. For example, a public dataset may be BCICompetition IV dataset1.

[0053] In this embodiment, the base classification model represents the correspondence between the target feature matrix and the task type, and the base classification result indicates the task type predicted by the base classification model. The training data for each base classification model includes only the historical EEG data of one existing subject.

[0054] As an example, task types can be represented using encoding, such as the number "1" representing "left hand movement," the number "2" representing "right hand movement," and the number "3" representing "both feet movement." Assuming the target EEG data contains three sampling records, the target feature matrix contains three slicing vectors. After the executing agent inputs the target feature matrix into each base classification model, each base classification model can predict the task type corresponding to the three slicing vectors and output the base classification result in vector form. For example, a base classification result of vector (1, 3, 2) indicates that when the first sampling record was collected, the target subject imagined the task type as "left hand movement"; when the second sampling record was collected, the target subject imagined the task type as "both feet movement"; and when the third sampling record was collected, the target subject imagined the task type as "right hand movement."

[0055] The training process of a base classification model is explained below with an example. First, multiple initial base classification models can be constructed according to actual needs, such as support vector machines or K-nearest neighbor classifiers. Then, the same number of historical EEG data from existing subjects as the initial base classification models are obtained and labeled to construct training samples. Finally, using a one-to-one correspondence, one initial base classification model is trained using one training sample from an existing subject, thus obtaining multiple base classification models.

[0056] Step 220: By aligning the centroids of the covariance matrix and extracting tangent space features, the feature matrices of the target EEG data and the historical EEG data of multiple existing subjects are determined respectively, thus obtaining the target feature matrix of the target subject and the source domain feature matrix of multiple existing subjects.

[0057] In this embodiment, the target feature matrix can characterize the feature information of the target EEG data, and the source domain feature matrix can represent the feature information of the historical EEG data of the existing subjects.

[0058] The following uses the target feature matrix as an example to illustrate the processing of EEG data. The executing entity (e.g., ...) Figure 1 The electronic device 100 shown can determine the target feature matrix of the target EEG data in the following manner. First, the covariance matrix of each matrix in the target EEG data is calculated using the following formula (1), and the resulting positive semi-definite matrix is ​​regarded as a point in the Riemannian manifold space. Then, the Euclidean mean of each matrix is ​​determined using formula (2). Then, the centroid of the covariance matrix is ​​aligned using formula (3). After that, the tangent space features are extracted using the following formula (4), and the centroid-aligned covariance matrix is ​​mapped into a vector of the tangent space. The tangent vectors corresponding to each matrix included in the target EEG data can then form the target feature matrix.

[0059]

[0060]

[0061]

[0062] x i =upper(log) M (P′ i (4)

[0063] Where i = 1, 2, ..., n, n represents the number of matrices included in the target EEG recording; t represents the number of sample points; P i The covariance matrix representing the i-th matrix; Represents the matrix corresponding to the i-th sampling record; c represents the number of EEG channels; log M () denotes the Riemann logarithm operation; upper() denotes taking the upper triangular portion of the c×c dimensional covariance matrix; x i Indicates the i-th matrix in the tangent space In the vector, d represents the feature dimension of the tangent space, and d = c(c+1) / 2 represents the feature dimension of the tangent space.

[0064] Optionally, the executing entity can also process the target EEG data and historical EEG data using Euclidean alignment to obtain a target feature matrix and a source domain feature matrix. Each existing subject corresponds to a source domain feature matrix, which includes feature information from all historical EEG data of the existing subject.

[0065] Step 230: Using the target feature matrix as the target domain, perform one-to-one single-source domain transfer learning with the source domain feature matrices of each of the multiple existing subjects to obtain multiple target feature transformation matrices that correspond one-to-one with the multiple existing subjects.

[0066] In this embodiment, the executing entity can reduce the data distribution differences between the target EEG data and the historical EEG data of multiple existing subjects through one-to-one single-source domain transfer learning. For example, a transfer joint matching algorithm or a balanced distribution adaptive algorithm can be used to achieve one-to-one single-source domain transfer learning between the target domain and the feature matrices of multiple source domains.

[0067] As an example, the implementing entity can first merge the target feature matrix with an existing source domain feature matrix of a subject to obtain the input matrix for single-source domain transfer learning. h = h s +h T Where l represents the number of rows in the matrix, h represents the number of columns in the matrix, and h s h represents the number of columns in the sample feature matrix of the source domain. T Let X represent the number of columns in the target feature matrix. Then, map the matrix X to a high-dimensional reproducing kernel Hilbert space. The kernel mapping can be represented as... The kernel matrix can be represented as Next, the migration mapping matrix can be calculated. In the dimensionality reduction process, feature fitting and source domain sample reweighting can be used together to reduce inter-domain differences. The matrix obtained after mapping the input matrix X can be represented as Z = A. T K and matrix Z are the target feature transformation matrices corresponding to the existing subjects. By performing the above steps on the target feature matrix and each source domain feature matrix respectively, the target feature transformation matrices corresponding to each existing subject can be obtained.

[0068] Step 240: Based on the base classification models and target feature transformation matrices corresponding to multiple existing subjects, input the multiple target feature transformation matrices into the corresponding base classification models, and have the multiple base classification models predict the task type of the motion task imagined by the target subject, thereby obtaining multiple base classification results of the target subject.

[0069] In this embodiment, each existing subject corresponds to a target feature transformation matrix and a base classification model, thereby determining the correspondence between the input feature matrix and the base classification model. The executing entity can input the target feature transformation matrix obtained in step 230 into the corresponding base classification model.

[0070] Further reference Figure 3 , Figure 3 A scene illustration of the method for motion image decoding disclosed herein is shown. Figure 3 As shown. The base classification models trained using the historical EEG data of existing subjects 310, 320, 330, 340, and 350 are 311, 321, 331, 341, and 351. The executing agent can perform one-to-one single-source-domain transfer learning on the target feature matrix of the target subject 360 and the source domain feature matrices corresponding to existing subjects 310, 320, 330, 340, and 350, respectively, to obtain the target feature transformation matrices 312, 322, 332, 342, and 352 corresponding to each existing subject. Then, the executing entity inputs the target feature transformation matrices 312, 322, 332, 342, and 352 into the base classification models 311, 321, 331, 341, and 351 respectively, to obtain the base classification results 313, 323, 333, 343, and 353. Then, the executing entity can fuse the base classification results 313, 323, 333, 343, and 353 to obtain the final motion imagery decoding information 370.

[0071] Step 250: Based on multiple base classification results, determine the motor imagery decoding information of the target subject.

[0072] In this embodiment, the motor imagery decoding information represents the task type of the motor task corresponding to the target EEG data.

[0073] The executing entity can fuse multiple base classification results based on a preset fusion strategy to determine the final motion image decoding information. For example, majority voting algorithms or weighted voting algorithms can be used.

[0074] The method for decoding motor imagery provided in this embodiment acquires the target EEG data of the target subject, historical EEG data of multiple existing subjects, and their corresponding base classification models. Then, through covariance matrix centroid alignment and tangent space feature extraction, the target feature matrix of the target EEG data and the source domain feature matrix of the historical EEG data are determined. Next, the target feature matrix is ​​used as the target domain and subjected to one-to-one single-source domain transfer learning with multiple source domain feature matrices, aligning the target EEG data with the historical EEG data of multiple existing subjects, resulting in target feature transformation matrices corresponding to multiple existing subjects. Subsequently, the multiple target feature transformation matrices are input into their respective base classification models, which process the target feature transformation matrices to predict the task type of the motor task imagined by the target subject, obtaining multiple base classification results. Finally, the multiple base classification results are fused to determine the motor imagery decoding information of the target subject. Since each of the multiple base classification models is trained based on the historical EEG data of an existing subject, different base classification models can reflect individual differences between different subjects, making each base classification model more targeted. Then, through one-to-one single-source domain transfer learning, the EEG data of the target subject is aligned with the historical EEG data corresponding to the base classification model. The method of alignment before decision fusion is used to solve the problem of ignoring the differences among multiple subjects in the source domain during multi-source transfer, reducing the data distribution differences caused by individual EEG differences and the non-stationarity of EEG data. The targeting of the base classification model can be fully utilized, which helps to improve the accuracy of motor imagery decoding, and thus expands the application prospects of motor imagery decoding in various application fields. For example, it can more accurately understand the intentions of people with movement disorders or stroke patients, thereby facilitating the interaction of people with movement disorders and the rehabilitation treatment of stroke patients.

[0075] In some optional implementations of this embodiment, the historical EEG data of multiple existing subjects can also be obtained in the following way: obtaining the historical EEG data of multiple existing subjects from a public dataset or a privately collected dataset.

[0076] As an example, a public dataset could be BCI Competition IV dataset1. A private dataset could be any EEG dataset obtained by any individual or institution.

[0077] In this embodiment, the executing entity can interact with public or private datasets via a network, thereby more flexibly acquiring historical EEG data from multiple existing subjects.

[0078] The inventors also designed comparative experiments to verify the effectiveness of the multi-source transfer fusion scheme (i.e., including "one-to-one single-source domain transfer learning" and "decision fusion") in this embodiment. Multiple transfer learning methods were simultaneously applied to two base classification models: support vector machine and K-nearest neighbor classifier. These included single-source non-transfer, multi-source non-transfer, multi-source weighted non-transfer, and the multi-source transfer fusion scheme in this embodiment. Then, each of these processing methods was used to decode motor imagery on the public datasets BCICompetition IV dataset1 (which includes EEG data from 7 subjects) and / or dataset2a (which includes EEG data from 9 subjects), and the accuracy of each processing method was calculated. Experimental results show that among the above processing methods, the multi-source transfer fusion method provided in this embodiment has the highest accuracy, followed by the multi-source weighted non-transfer method, then the single-source transfer method, while the single-source non-transfer method and the multi-source non-transfer method have essentially the same, lowest accuracy. Therefore, it can be seen that the combination of "one-to-one single-source domain transfer learning" and "decision fusion" in this embodiment can significantly improve the accuracy of motion imagery decoding.

[0079] Furthermore, based on the publicly available BCI Competition IV dataset1, the inventors fixed the optimal parameters (regularization parameter λ = 0.1, source domain number of participants N = 6, subspace dimension = 390), and used the multi-source transfer fusion method in this embodiment to sequentially decode the EEG data of each participant to obtain the accuracy rate for different participants. The accuracy rate was then used to evaluate the multi-source transfer fusion method in this embodiment. Experimental results show that the accuracy rates of different participants are basically consistent, indicating that the multi-source transfer fusion method in this embodiment has relatively stable performance when dealing with EEG data from different participants.

[0080] In some optional implementations of this embodiment, step 250 can be implemented in the following ways: based on the fusion weights corresponding to each of the multiple base classification models, the results of multiple base classification models are fused using weighted voting to obtain motion imagery decoding information; or, the results of multiple base classification models are processed using a majority voting algorithm to determine motion imagery decoding information.

[0081] In this embodiment, the fusion weights can be set based on experience or based on the parameters of the base classification model, such as the accuracy or F-score of the base classification model.

[0082] Continuing with the example from step 220, let's illustrate the weighted voting algorithm. Assume the fusion weights of the three base classification models are 0.9, 0.3, and 0.4, respectively, and the corresponding base classification results are (1, 3, 2), (2, 1, 2), and (2, 3, 1), respectively. For the prediction result of the first sample record, the weight of "1" is 0.9, and the weight of "2" is 0.7 (i.e., the sum of 0.3 and 0.4). Therefore, the "1" with the larger weight can be determined as the prediction result of the first sample record. Similarly, the prediction result of the second sample record can be determined as "3", and the prediction result of the third sample record can be determined as "2". The final motor imagery decoding information is (1, 3, 2), that is, the motor tasks imagined by the target subject during the EEG test are "left hand movement", "two feet movement", and "right hand movement" in sequence.

[0083] Continuing with the example from step 220, let's explain the majority voting algorithm. Assume the three base classification results are (1, 3, 2), (2, 1, 2), and (2, 3, 1). The prediction result of the first sample record includes two "2"s and one "1". According to the majority voting principle, the "2" with more votes can be determined as the prediction result of the first sample record. Similarly, the prediction result of the second sample record can be determined as "3", and the prediction result of the third sample record can be determined as "2". The final motor imagery decoding information is (2, 3, 2), that is, the motor tasks imagined by the target subject during the EEG test are "left hand movement", "two feet movement", and "right hand movement" in sequence.

[0084] In this embodiment, multiple base classification results are fused through majority voting or weighted voting to determine motion imagery decoding information. This not only utilizes the performance differences of each base classification model in different dimensions, but also allows for the fusion of multiple base classification results from a global perspective, which helps to improve the accuracy of motion imagery decoding.

[0085] Next, refer to Figure 4 , Figure 4 A flowchart illustrating the acquisition of a base classification model in one embodiment of the method for decoding motion imagery of this disclosure is shown, as follows: Figure 3 As shown, the process includes the following steps:

[0086] Step 410: Use the source domain feature matrices of each of the multiple existing subjects as sample feature matrices, and determine the sample label of each sample feature matrix to obtain the sample data corresponding to each of the multiple existing subjects.

[0087] The sample data includes the corresponding sample feature matrix and its sample labels.

[0088] In this embodiment, the executing entity can use the source domain feature matrix obtained in step 220 as the sample feature matrix, and then label the sample feature matrix according to the task type of the motor task imagined by the subject when the historical EEG data was collected, so as to generate sample data. In this way, each sample data corresponds to an existing subject.

[0089] As an example, sample labels can be in vector form. Suppose an existing subject's historical EEG data contains m sampling records, then the corresponding target feature matrix can be represented as follows: Among them, X j This represents the j-th sampling record; the sample label can be represented as... Among them, y j This represents the task type of the j-th sampling record; the final sample data can be represented as...

[0090] Step 420: Construct an initial base classification model for each existing subject's corresponding sample data, and train the initial base classification model using the sample data corresponding to the existing subject to obtain multiple base classification models corresponding to each existing subject, which serve as multiple base classification models.

[0091] As an example, assuming there are subjects a, b, and c, the implementing entity can construct three initial base classification models A, B, and C. Then, using the sample data corresponding to subject a, initial base classification model A is trained; using the sample data corresponding to subject b, initial base classification model B is trained; and using the sample data corresponding to subject c, initial base classification model C is trained. This yields the base classification models corresponding to subjects a, b, and c respectively. The training process may include: inputting the sample feature matrix into the initial base classification model, using the sample labels as the desired output, iteratively optimizing the parameters of the initial base classification model until the training termination condition is met, such as the loss function converging or the number of iterations reaching a preset number, resulting in the trained base classification model.

[0092] Figure 4 The illustrated embodiment constructs an initial base classification model for each existing subject's historical EEG data, and trains this initial base classification model using the sample data of that existing subject, resulting in multiple base classification models corresponding to each existing subject. Since the training samples for each base classification model only include the historical EEG data of one existing subject, different base classification models can learn the feature information of different existing subjects. This allows for a more targeted prediction strategy, and enables multiple base classification models to preserve individual differences among different existing subjects.

[0093] Next reference Figure 5 , Figure 5A flowchart illustrating the determination of fusion weights in one embodiment of the terminology-based anti-motion imagery decoding of this disclosure is shown, as follows: Figure 5 As shown, the process includes the following steps:

[0094] Step 510: Extract a preset number of EEG data from the target EEG data, and label the extracted EEG data according to the task type to obtain calibration data.

[0095] In this embodiment, the calibration data is a portion of the target EEG data. The preset quantity is usually smaller than the amount of data contained in the target EEG data, and its size can be set according to actual needs or experience.

[0096] As an example, assuming the target EEG data includes 100 sampling records, the executing entity can extract 10 sampling records from them, and then label the EEG data included in the sampling records according to the task type corresponding to the sampling records to obtain calibration data.

[0097] Step 520: Determine the calibration feature matrix of the calibration data.

[0098] In this embodiment, the executing entity can perform covariance matrix centroid alignment and extract tangent space features on the calibration data to determine the calibration feature matrix of the calibration data.

[0099] Step 530: Based on the source domain feature matrices and base classification models corresponding to multiple existing subjects, perform one-to-one single source domain transfer learning on the calibration feature matrices respectively, and input the matrix after transfer learning into the corresponding base classification model to obtain the prediction results of multiple base classification models.

[0100] In this embodiment, the prediction result represents the task type predicted by the base classification model based on the calibration data. The executing entity can perform one-to-one single-source-domain transfer learning between the calibration feature matrix and each source domain feature matrix to reduce the distributional differences between the calibration data and the historical EEG data of each existing subject. Then, the matrices obtained after transfer learning are input into the base classification models corresponding to each source domain to obtain the prediction results of each base classification model.

[0101] Step 540: Based on the labeling results of the calibration data and the prediction results of each of the multiple base classification models, determine the calibration parameters of each of the multiple base classification models.

[0102] Among them, the calibration parameters characterize the accuracy of the base classification model.

[0103] As an example, the implementing entity can determine the accuracy or F-score of each base classification model based on the labeling results and prediction results, and then use the accuracy or F-score as the calibration model parameter of the base classification model.

[0104] For example, the executing entity can also calculate the distribution distance between the labeling results and each prediction result separately, and determine the calibration parameters of each base classification model based on the distribution distance. In this case, the calibration parameters are negatively correlated with the distribution distance.

[0105] Step 550: Determine the fusion weights of the base classification model based on the calibration parameters.

[0106] As an example, the implementing entity can directly determine the calibration parameters as the fusion weights of the base classification models. Alternatively, it can sort the calibration parameters by their numerical values ​​and assign different fusion weights to each base classification model according to the sorting order.

[0107] In some optional implementations of this embodiment, the fusion weight can also be determined in the following way: if the calibration parameter is greater than a preset threshold, the calibration parameter is determined as the fusion weight of the corresponding base classification model; if the calibration parameter is not greater than the preset threshold, the fusion weight of the base classification model corresponding to the calibration parameter is determined as 0.

[0108] In this implementation, the preset threshold represents the expected accuracy of the base classification model. If the calibration parameter is not greater than the preset threshold, it indicates that the accuracy of the base classification model does not meet the expectation, and its fusion weight is set to 0. This prevents the base classification result from being introduced again when fusing the results of various base classifications. In this way, the interference of low-accuracy base classification results with motion imagery decoding information can be avoided, which helps to further improve the accuracy of motion imagery decoding.

[0109] In practice, when using brain-computer interface systems to decode motor imagery in new subjects, personalized calibration is often required. This calibration process necessitates collecting a large amount of EEG data from the new subjects and using this data to retrain the deep learning model. This allows the deep learning model to learn the subject's characteristic information, thereby improving the accuracy of motor imagery decoding.

[0110] exist Figure 5 In the illustrated embodiment, the executing entity only needs to obtain a small amount of calibration data of the target subject from the target EEG data and use it to determine the fusion weight of each base classification model to effectively decode EEG motor imagery. Under the premise of ensuring the accuracy of motor imagery decoding, the calibration time of the brain-computer interface can be greatly shortened, realizing the efficient application of brain-computer interface devices.

[0111] Exemplary device

[0112] Figure 6 This is a schematic diagram illustrating the structure of one embodiment of the apparatus for motion image decoding disclosed herein. The apparatus of this embodiment can be used to implement the corresponding method embodiments of this disclosure. Figure 6The apparatus shown includes: an acquisition unit 610 configured to acquire target EEG data of a target subject, historical EEG data of multiple existing subjects, and multiple base classification models corresponding one-to-one with the multiple existing subjects, wherein each base classification model is trained based on the historical EEG data of the corresponding existing subject; a calculation unit 620 configured to determine the feature matrices of the target EEG data and the historical EEG data of the multiple existing subjects respectively by aligning the centroids of the covariance matrix and extracting tangent space features, thereby obtaining the target feature matrix of the target subject and the source domain feature matrices of the multiple existing subjects; and a transfer unit 630 configured to transfer the target EEG data to the target EEG data. The feature matrix, as the target domain, is subjected to one-to-one single-source domain transfer learning with the source domain feature matrices of multiple existing subjects, resulting in multiple target feature transformation matrices corresponding one-to-one with the multiple existing subjects. The classification unit 640 is configured to input the multiple target feature transformation matrices into the corresponding base classification models based on the base classification models corresponding to the multiple existing subjects, and the multiple base classification models predict the task type of the motion task imagined by the target subject, thereby obtaining multiple base classification results of the target subject. The decoding unit 650 is configured to determine the motion imagination decoding information of the target subject based on the multiple base classification results.

[0113] In one embodiment, the decoding unit 650 includes: a majority voting module configured to fuse multiple base classification results using weighted voting based on the fusion weights corresponding to each of the multiple base classification models to obtain motion imagery decoding information; or, a weighted voting module configured to process multiple base classification results using a majority voting algorithm to determine motion imagery decoding information.

[0114] In one embodiment, the device further includes a model training unit configured to: use the source domain feature matrices of multiple existing subjects as sample feature matrices, and determine the sample label of each sample feature matrix to obtain sample data corresponding to each of the multiple existing subjects, wherein the sample data includes the corresponding sample feature matrix and its sample label; construct an initial base classification model for each sample data corresponding to an existing subject, and train the initial base classification model using the sample data corresponding to the existing subject to obtain base classification models corresponding to multiple existing subjects, which serve as multiple base classification models.

[0115] In one embodiment, the device further includes a weight determination unit, which comprises: a data acquisition module configured to extract a preset number of EEG data from the target EEG data and label the extracted EEG data according to the task type to obtain calibration data; a data processing module configured to determine the calibration feature matrix of the calibration data; a calibration module configured to perform one-to-one single-source-domain transfer learning operations on the calibration feature matrix based on the source domain feature matrices and base classification models corresponding to multiple existing subjects, and input the matrix after transfer learning into the corresponding base classification model to obtain the prediction results of multiple base classification models; a parameter determination module configured to determine the calibration parameters of multiple base classification models based on the labeling results of the calibration data and the prediction results of multiple base classification models, wherein the calibration parameters characterize the accuracy of the base classification models; and a weight determination module configured to determine the fusion weights of the base classification models based on the calibration parameters.

[0116] In one embodiment, the weight determination module is further configured to: if the calibration parameter is greater than a preset threshold, determine the calibration parameter as the fusion weight of the corresponding base classification model; if the calibration parameter is not greater than the preset threshold, determine the fusion weight of the base classification model corresponding to the calibration parameter as 0.

[0117] In one embodiment, the acquisition unit 610 further includes a historical EEG data acquisition module, configured to acquire historical EEG data of multiple existing subjects from a public dataset or a privately collected dataset.

[0118] Exemplary electronic devices

[0119] The following is for reference. Figure 7 To describe an electronic device according to embodiments of the present disclosure. Figure 7 A block diagram of an electronic device according to an embodiment of the present disclosure is shown. (As follows) Figure 7 As shown, the electronic device 700 includes one or more processors 710 and memory 720.

[0120] The processor 710 may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 700 to perform desired functions.

[0121] The memory 720 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, and flash memory. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 710 may execute the program instructions to implement the methods for motion image decoding and / or other desired functions described in the various embodiments of this disclosure above.

[0122] In one example, the electronic device 700 may further include an input device 730 and an output device 740, etc., these components being interconnected via a bus system and / or other forms of connection mechanisms (not shown). Furthermore, the input device 730 may also include, for example, a keyboard, a mouse, etc. The output device 740 can output various information to the outside. The output device 740 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0123] Of course, for the sake of simplicity, Figure 7 Only some of the components of the electronic device 700 relevant to this disclosure are shown, omitting components such as buses, input / output interfaces, etc. In addition, the electronic device 700 may include any other suitable components depending on the specific application.

[0124] Exemplary computer program products and computer-readable storage media

[0125] In addition to the methods and apparatus described above, embodiments of this disclosure may also be computer program products comprising computer program instructions that, when executed by a processor, cause the processor to perform the steps in the methods for motion image decoding according to various embodiments of this disclosure as described in the "Exemplary Methods" section above.

[0126] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this disclosure. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on a user's computing device, partially on a user's computing device, as a standalone software package, partially on a user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0127] Furthermore, embodiments of this disclosure may also be computer-readable storage media storing computer program instructions that, when executed by a processor, cause the processor to perform the steps in the methods for motion image decoding according to various embodiments of this disclosure as described in the "Exemplary Methods" section above.

[0128] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples (not an exhaustive list) of a readable storage medium may include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0129] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0130] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they largely correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0131] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0132] The methods and apparatus of this disclosure may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, this disclosure may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the methods according to this disclosure. Thus, this disclosure also covers recording media storing programs for performing the methods according to this disclosure.

[0133] It should also be noted that in the apparatus, devices, and methods of this disclosure, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions to this disclosure.

[0134] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0135] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for motor imagery decoding, comprising: obtaining target electroencephalogram (EEG) data of a target subject, historical EEG data of a plurality of existing subjects respectively, and a plurality of base classification models corresponding to the plurality of existing subjects respectively, wherein each base classification model is trained based on the historical EEG data of the corresponding existing subject; determining feature matrices of the target EEG data and the historical EEG data of the plurality of existing subjects respectively by covariance matrix centroid alignment and extraction of tangent space features, to obtain a target feature matrix of the target subject and source domain feature matrices of the plurality of existing subjects respectively; performing one-to-one single-source domain transfer learning between the target feature matrix as a target domain and the source domain feature matrices of the plurality of existing subjects respectively, to obtain a plurality of target feature transformation matrices corresponding to the plurality of existing subjects respectively; inputting the plurality of target feature transformation matrices into the base classification models corresponding to the plurality of existing subjects respectively based on the base classification models and the target feature transformation matrices, and predicting task types of motor tasks imagined by the target subject by the plurality of base classification models respectively, to obtain a plurality of base classification results of the target subject; determining motor imagery decoding information of the target subject based on the plurality of base classification results.

2. The method of claim 1, wherein, The determination of the motor imagery decoding information of the target subject based on the plurality of base classification results comprises: fusing the plurality of base classification results by weighted voting based on fusion weights corresponding to the plurality of base classification models respectively, to obtain the motor imagery decoding information; or processing the plurality of base classification results by majority voting algorithm to determine the motor imagery decoding information.

3. The method of claim 2, wherein, The method further comprises the steps of obtaining the plurality of base classification models: taking the source domain feature matrices of the plurality of existing subjects respectively as sample feature matrices, and determining sample labels of each sample feature matrix, to obtain sample data corresponding to the plurality of existing subjects respectively, wherein the sample data comprises the corresponding sample feature matrix and the sample label thereof; constructing an initial base classification model for the sample data corresponding to each existing subject, and training the initial base classification model by the sample data corresponding to the existing subject, to obtain the base classification models corresponding to the plurality of existing subjects respectively as the plurality of base classification models.

4. The method of claim 3, further comprising the steps of determining the fusion weights as follows: extracting a preset number of EEG data from the target EEG data, and labeling the extracted EEG data based on the task types respectively to obtain calibration data; determining a calibration feature matrix of the calibration data; performing the one-to-one single-source domain transfer learning on the calibration feature matrix based on the source domain feature matrices and the base classification models corresponding to the plurality of existing subjects respectively, and inputting the matrix after transfer learning into the corresponding base classification model to obtain prediction results of the plurality of base classification models respectively. determine calibration parameters of the plurality of base classification models respectively based on the calibration data and the prediction results of the plurality of base classification models, the calibration parameters representing accuracy of the base classification models; determine fusion weights of the base classification models based on the calibration parameters.

5. The method of claim 4, wherein, determine fusion weights of the base classification models based on the calibration parameters, including: if the calibration parameter is greater than a preset threshold, determine the calibration parameter as the fusion weight of the corresponding base classification model; if the calibration parameter is not greater than the preset threshold, determine the fusion weight of the base classification model corresponding to the calibration parameter as 0.

6. The method according to any one of claims 1 to 5, wherein, obtain historical electroencephalogram data of the plurality of existing subjects respectively, including: obtain historical electroencephalogram data of the plurality of existing subjects respectively from a public data set or a private data set collected by an individual.

7. An apparatus for motor imagery decoding, comprising: an obtaining unit configured to obtain target electroencephalogram data of a target subject, historical electroencephalogram data of a plurality of existing subjects respectively, and a plurality of base classification models corresponding to the plurality of existing subjects one by one, wherein each base classification model is trained based on historical electroencephalogram data of a corresponding existing subject; a computing unit configured to determine feature matrices of the target electroencephalogram data and the historical electroencephalogram data of the plurality of existing subjects respectively by covariance matrix centroid alignment and extraction of tangent space features, to obtain a target feature matrix of the target subject and source domain feature matrices of the plurality of existing subjects respectively; a migration unit configured to perform one-to-one single-source domain transfer learning between the target feature matrix as a target domain and the source domain feature matrices of the plurality of existing subjects respectively, to obtain a plurality of target feature transformation matrices corresponding to the plurality of existing subjects one by one; a classification unit configured to input the plurality of target feature transformation matrices into corresponding base classification models based on the base classification models corresponding to the plurality of existing subjects respectively and the target feature transformation matrices, to predict task types of a motor task imagined by the target subject by the plurality of base classification models respectively, to obtain a plurality of base classification results of the target subject; a decoding unit configured to determine motor imagery decoding information of the target subject based on the plurality of base classification results.

8. An electronic device, comprising: including: a memory for storing a computer program product; a processor for executing the computer program product stored in the memory, and when the computer program product is executed, the method of any one of claims 1-6 is implemented.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1-6.

10. A computer program product comprising computer program instructions, characterised in that, The computer program instructions are executed by the processor to implement the method of any one of claims 1-6. The computer program instructions are executed by the processor to implement the method of any one of claims 1-6.

Images