Intracranial electroencephalogram cross-subject decoding method based on multi-manifold parallel movement

Abstract

The application belongs to the technical field of physiological electrical signals, and particularly relates to an intracranial electroencephalogram cross-subject decoding method based on multi-manifold parallel movement. A migration path of task state data of a new subject to a source domain and model matching degree of the new subject are established based on resting state data, and finally the migrated task state data is input into an optimal source model for decoding. The resting state data is used as a unified reference standard of different subjects, and only a small amount of resting state data of a new subject can directly reuse a decoding model of historical subjects.

Description

Technical Field

[0001] This invention belongs to the field of physiological electrical signal technology, specifically relating to a method for decoding intracranial electroencephalogram (EEG) signals across subjects based on multi-manifold parallel movement. Background Technology

[0002] During the calibration of EEG signal decoding models for new data, the source and target domains have different data spaces, meaning that the decoding model trained on the source domain cannot match the target domain. Related technologies typically employ data transfer to address this data space difference between the source and target domains.

[0003] For example, patent CN 114548166 B discloses a heterogeneous label space transfer learning method for EEG signals based on Riemannian manifolds. Based on the EEG data of the same subject, the EEG data of labels 1 and 2 in each sub-band are concatenated into new labels 1 and 2 in different orders. The concatenated EEG data of labels 1 and 2 is used as the source domain, and the EEG data of labels 3 and 4 is used as the target domain. A correspondence is established between the average covariance matrices of the source and target domains. The EEG data of the source domain is transformed according to this correspondence, and the labels of the target domain are assigned to the EEG data of the source domain according to the correspondence, resulting in the transformed source domain EEG data. Clearly, this scheme is based on the EEG data of the same subject and can ignore the data space changes caused by subject differences, solving the matching problem of the decoding model from training data to real-time data through the correspondence between data of different frequency bands. However, in brain-computer interface products, the universality of the product requires the decoding model to adapt to the EEG data of different subjects. However, in cross-subject scenarios, different frequency bands (such as low-frequency β and high-frequency High-γ) are affected to varying degrees by differences in electrode position and impedance changes, resulting in "non-uniform drift" in their statistical distribution. That is, different subjects cannot form a correspondence in different or the same frequency band data. The translation path formed by setting the correspondence in the average covariance matrix of a single subject will also mask the differences between frequency bands, leading to direct transfer failure. Summary of the Invention

[0004] This invention provides a method for decoding intracranial electroencephalogram (EEG) signals across subjects based on multi-manifold parallel movement, in order to solve the problem of cross-subject data transfer.

[0005] To address the aforementioned technical problems, this invention provides a method for decoding intracranial electroencephalogram (EEG) signals across subjects based on multi-manifold parallel movement, comprising: constructing a source library, i.e., obtaining at least one source model and its corresponding source domain based on historical data of different subjects under the same task; constructing multi-band independent manifolds, i.e., constructing a first reference matrix corresponding to the source domain, a second reference matrix corresponding to the resting-state data of the new subject, and a third reference matrix corresponding to the task-state data of the new subject in different frequency bands; calling the optimal model, i.e., setting the second reference matrix as a benchmark, comparing the differences between the second reference matrix and each of the first reference matrices one by one to obtain the model matching degree of the new subject, so as to call the optimal source model and optimal source domain from the source library; constructing a migration path by combining the second reference matrix and the first reference matrix corresponding to the optimal source domain; performing data migration, i.e., mapping the third reference matrix to the optimal source domain along the migration path; and decoding the migrated task-state data using the optimal source model.

[0006] Furthermore, constructing the migration path includes: based on the source domain corresponding to the optimal source model. Calculate the source domain matrix for each frequency band. That is, the first reference matrix; before performing the task, based on the new subject's target domain Calculate the target domain matrix for each frequency band. That is, the second reference matrix; find the connection on the independent manifold. arrive The shortest geodesic path between them is used as the migration path for the data in this frequency band.

[0007] Furthermore, the execution data migration includes: data mapping, which involves obtaining the task-state covariance matrix corresponding to the target domain task-state data during task execution. That is, the third reference matrix, which preserves the source domain. Unchanged, using the parallel translation conversion formula to... Map along the migration path to To obtain the task-state covariance matrix of the target domain after the migration. Projecting the source domain into its tangent space, we obtain the source domain covariance matrix. and will and Projecting each vector onto the tangent space yields the corresponding tangent vector. ,Right now and The weighted fusion of multi-manifold features involves weighting and concatenating tangent vectors from different frequency bands in the tangent space to obtain the translated feature vector, which represents the transferred task-state data.

[0008] Furthermore, the parallel translation conversion formula is as follows: The tangent vector Configured as ,in This represents the covariance matrix of the corresponding domain, i.e., the task-state covariance matrix of the target domain after the transition. or source domain covariance matrix , The source domain matrix represents the corresponding frequency; the transferred task-state data is characterized as... ,in Indicates the corresponding frequency band The weights are β, which represents the β band in the key frequency band of the EEG signal, and Hγ, which represents the Hγ band in the key frequency band of the EEG signal.

[0009] The migrated task-state data is characterized as ,in Indicates the corresponding frequency band The weight.

[0010] Furthermore, the difference is configured as the geodesic distance between the new subject's second reference matrix and each of the first reference matrices in the source library.

[0011] Furthermore, constructing a multi-band independent manifold includes: extracting at least one key frequency band of the EEG signal using a filter bank; and for each key frequency band... Calculate their spatial covariance matrix respectively Based on spatial covariance matrix Construct multiple independent Riemannian submanifolds, namely the first reference matrix, the second reference matrix, and the third reference matrix corresponding to each data.

[0012] The present invention also provides an implantable brain-computer interface device, including a signal sensor, a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the method and decodes the intracranial electroencephalogram (EEG) signals acquired by the signal sensor.

[0013] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method.

[0016] The beneficial effects of this invention are that the intracranial EEG cross-subject decoding method based on multi-manifold parallel movement first constructs independent manifolds of multiple frequency bands for the data, then establishes a migration path for the task-state data of the new subject to be translated to the source domain and the model matching degree of the new subject based on the resting-state data, and finally inputs the migrated task-state data into the optimal source model for decoding. This method can effectively solve the problem of non-uniform drift of multi-frequency band data across subjects using a small amount of resting-state data before the task, and directly reuse the decoding model of the historical subjects.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. To make the foregoing objects, features, and advantages of the invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See Figure 1 To address the issue of multi-band non-uniform drift in data between different subjects in intracranial brain-computer interface (BCI) technology, this embodiment provides a method for decoding intracranial EEG signals across subjects based on multi-manifold parallel migration, including: step S1, constructing a source library; step S2, constructing multi-band independent manifolds; step S3, calling the optimal model; step S4, constructing a migration path; step S5, performing data migration; and step S6, decoding the migrated task-state data using the optimal source model. The specific steps are as follows.

[0022] Step S1: Build the source library.

[0023] Step S11, Historical Data Collection and Processing. Historical intracranial electroencephalogram (iEEG) data from different subjects are collected according to the same task. This historical data will be used to train the "historical decoding model." Data features, such as the spatial covariance matrix (SCM), are extracted from the preprocessed data of each historical subject, and the data of each subject is converted into an SPD matrix.

[0024] Step S12, Construction of Source Models. At least one "source model" is trained and obtained using these historical data. These source models are pre-trained decoding models and can be considered "historical decoding models." For example, these source models can be pre-trained linear classifiers, such as Linear Discriminant Analysis (LDA) or Support Vector Machines (SVM). A decoding model is trained separately for each historical subject (or a group of similar subjects). For example, an LDA classifier is trained using the data of subject A to obtain model A.

[0025] Step S13, Determining the source domain. Each source model corresponds to the "source domain" of the historical data used for its training. The source domain S represents the data distribution space of that historical subject or subject group.

[0026] Step S14, Source Library Establishment. Through the above process, a "source library" containing one or more source models and their corresponding source domains is established. The purpose of this source library is to provide new subjects with a set of pre-trained models to choose from, reducing the influence of non-essential factors, and achieving model reuse and zero / few calibrations. The final "source library" is a set in which each element is a pair of "{source model, source domain}". For example: { (model A, reference matrix of source domain A), (model B, reference matrix of source domain B), ...}. This allows the system to have multiple pre-trained "expert models", providing new users with a pool of selectable, ready-to-use models.

[0027] Step S2: Construct independent manifolds with multiple frequency bands.

[0028] In intracranial electroencephalography (EEG), key frequency bands can be divided into high-frequency (High-gamma) and low-frequency (Beta) bands. Because these two bands exhibit different physical drift characteristics across subjects, they need to be processed separately. Specifically: multiple key frequency bands of the EEG signal are extracted using a filter bank (e.g., the β band is 15-30Hz, and the Hγ band is 50-150Hz); for each frequency band... Calculate their spatial covariance matrix respectively Construct multiple independent Riemannian submanifolds , The corresponding reference matrices for each data point are obtained. Specifically, in different frequency bands, a first reference matrix is ​​formed corresponding to the source domain, a second reference matrix to the resting-state data of the new subject, and a third reference matrix to the task-state data of the new subject. Multiple reference matrices can be formed based on the frequency band; for example, the first reference matrix can also be formed... , A second reference matrix will also be formed. , A third reference matrix will also be formed. , .

[0029] Step S3: Call the best model.

[0030] The second reference matrix can be set as the benchmark, and the differences between the second reference matrix and each first reference matrix can be compared one by one to obtain the model matching degree of the new subject, so as to call the best source model and the best source domain from the source library.

[0031] Optionally, resting-state data can be understood as non-task segment data, i.e., EEG data collected when no specific task was performed (i.e., non-task segment). This data can be labeled or identified from the raw data in the source and target domains to extract the corresponding non-task segment data. Alternatively, for the raw data, intracranial EEG (iEEG) data from multiple different historical subjects can be collected, ensuring that all these historical data were collected while performing the same or similar tasks (e.g., imagining left / right hand movements, viewing specific types of images, etc.). Standard preprocessing is performed on the raw iEEG signals, including but not limited to: filtering (removing power line noise and DC drift), artifact removal (e.g., eye movement, EMG interference), and data segmentation, dividing the continuous signal into multiple task-related "trials" or epochs. For new subjects, before performing the task, there is generally only a small amount or some resting-state data, with virtually no task-state data. Only after selecting a matching decoding model will the training task be attempted, and then the decoding model is optimized and the training scheme adjusted based on the task execution results. Choosing the wrong decoding model before the task can lead to severely distorted data decoding results, making it impossible for doctors to determine whether the problem lies in the suitability of the rehabilitation training program for the subject or the adaptability of the decoding model. Clearly, without sufficient data, results cannot be validated; conversely, with an unsuitable model, the accuracy of the results is difficult to verify, creating a dilemma in the design of rehabilitation training programs. Therefore, selecting a suitable decoding model is the most common practical problem encountered by new participants.

[0032] Therefore, in this case, model fit can be defined as a function of the geodesic distance between the second reference matrix of the new subject and the first reference matrix of each source model in the source library. The model fit is quantified by comparing the geometric distance (i.e., geodesic distance) between the source domain data of the new subject and each historical subject in the source library in the Riemannian manifold space (a curved reference matrix composed of symmetric positive definite matrices, such as the spatial covariance matrix). For example, a shorter geodesic distance indicates that the resting-state data distribution of the new subject is geometrically closer to the source domain corresponding to that source model, thus implying a higher model fit. Conversely, a longer geodesic distance indicates a lower model fit. After calculating the model fit between the new subject and each source model in the source library, the source model with the highest fit (i.e., the smallest geodesic distance) is selected as the optimal source model and used for subsequent data transfer and decoding.

[0033] Specifically, only a short segment (e.g., tens of seconds) of resting-state iEEG data needs to be collected from the new subject. Then, its data features are extracted, and the new subject's resting-state data is processed into a series of spatial covariance matrices. The Riemann mean of all resting-state covariance matrices for the new user is calculated as the second reference matrix. The first reference matrix (or the Riemann mean of the resting-state data) of the source domain corresponding to each historical model in the source library is then sequentially called. In the Riemannian manifold space, the geodesic distance between the second reference matrix of the new subject's resting state and the first reference matrix of each historical model is calculated. This distance precisely measures the "distance" between two SPD matrices in the curved reference matrix. The shorter the distance, the more geometrically similar the brain activity patterns of the two subjects are, and the higher the model fit. Based on all calculated distances, the historical model with the shortest distance is selected as the "best source model." This selected model and its corresponding "source domain" will be used for subsequent data transfer.

[0034] Preferably, the model matching degree can be evaluated independently for each frequency band, and the "best source model" can be selected based on the comprehensive results of the multi-frequency band data evaluation, for example but not limited to the following acquisition methods. (1) Weighted average of multi-frequency band distance: Considering that different frequency bands may have different effects on model performance, the geodesic distance of each frequency band can be weighted and averaged. A total distance score is calculated for each source model. For example, for model A, its total distance is Weight and The weight of w2 can be set according to the importance of different frequency bands to the decoding task. If the higher frequency is more important, the weight of w2 can be higher. The system will calculate a total distance score D(A), D(B), D(C)... for all models (A, B, C...), and then select the model with the smallest total distance score as the unique "best source model". (2) Simple averaging of multi-band distance: Averaging the geodesic distances of all frequency bands can be understood as a special case of weighted averaging, that is, all weights are equal. Similarly, the model with the smallest total score is selected. (3) The worst-case scenario is optimal, meaning the system considers the worst-performing model across all frequency bands. For model A, its worst-case performance is max( , For model B, its worst performance is max( , Then, the system selects the model that "performs relatively best even in the worst-performing frequency band." This strategy ensures that the selected model has good compatibility across all critical frequency bands, avoiding the selection of a model that is extremely mismatched in a particular frequency band. Alternatively, the source model with the smallest maximum geodesic distance can be selected across all frequency bands to ensure good compatibility across all critical frequency bands.

[0035] Step S4: Construct the migration path.

[0036] Source domain corresponding to the optimal source model The source domain matrix for each frequency band is calculated using the Fréchet Mean algorithm (also known as Karcher Mean). That is, the first reference matrix; based on the target domain of the new subject. Calculate the target domain matrix for each frequency band. That is, the second reference matrix; find the connection on the independent manifold. arrive The shortest geodesic path between them is used as the migration path for the data in this frequency band.

[0037] Step S5: Perform data migration.

[0038] Step S51, data mapping.

[0039] Unlike the "whitening to identity matrix" approach, this case employs a parallel shift transformation. First, during task execution, the task-state covariance matrix corresponding to the target domain task-state data is obtained. That is, the third reference matrix. Secondly, maintain the source domain data. Without changing the target domain, the task-state covariance matrix is ​​calculated using the parallel shift conversion formula for the frequency band data, or (which can also be understood as the migration path). "Transported" to the source domain matrix along geodesics This is equivalent to a translation of the coordinate system, not a non-linear stretching, and thus better preserves the coordinate system. The internal relative geometric structure. The transformation formula is: .

[0040] Step S52, source domain tangent space projection.

[0041] Since step S51 has already generated the task-state covariance matrix "Transported" to the source domain matrix It can be directly in the source domain matrix Constructing tangent space Using a log-map, the transferred task-state covariance matrix is... Source domain covariance matrix Projecting onto the tangent space yields the tangent vector. and , where the tangent vector .

[0042] Step S53: Weighted fusion of multi-manifold features.

[0043] For each key frequency band (e.g., β, Hγ), an independent Riemannian submanifold is constructed. On each independent submanifold, migration paths are calculated separately and parallel shifts are performed, thus establishing different "alignment strategies" for frequency bands with different drift characteristics. This is achieved instead of mixing data from all frequency bands into a single "joint covariance matrix" to construct a single manifold. The tangent vectors from different frequency bands are weighted and concatenated to obtain the translated eigenvectors. ,in Indicates the corresponding frequency band The weights, β represent the β band in the key frequency band of the EEG signal, and Hγ represent the Hγ band in the key frequency band of the EEG signal, are used... The input is a pre-trained linear classifier (such as LDA or SVM) for decoding. The weights... Cross-validation of source domain data or automatic adjustment based on Riemannian discretization of different frequency bands can be used to automatically "suppress" frequency bands with excessive drift and poor alignment across subjects using the decoding model. A "independent manifold processing + post-fusion" approach addresses the heterogeneity of different frequency bands separately, rather than using a uniform transformation for a "one-size-fits-all" solution, effectively resolving the problem of "non-uniform frequency band drift." Compared to traditional single joint covariance methods, this approach offers greater robustness and flexibility.

[0044] In summary, the method for decoding intracranial EEG signals across subjects based on multi-manifold parallel movement in this embodiment aims to solve the problem of inconsistent drift of signals in different frequency bands across subjects, and has the following advantages.

[0045] (1) Model Adaptive Calling Capability. The "model selection" problem is transformed into a "geometric distance calculation" problem in Riemann space. By "parallel shifting," the data of new patients is seamlessly aligned to the distribution space of "historical expert models," so that new patients do not need to undergo a lengthy full-process calibration (data collection). Only a few seconds of resting-state data are required to make the decoding model applicable to new subjects. Using a small amount of easily collected resting-state data of new users, the most similar "brain template" can be "intelligently" matched from the historical model library by calculating the geodesic distance between the new user and the resting-state data of each model in the source library. Compared with randomly selecting models or having users try them one by one, this method only requires a small amount of resting-state data to "intelligently" match, which significantly improves the practicality and efficiency of the system. This design enables the system to have fast, zero / few-shot calibration capability, which significantly reduces the user's usage threshold and time cost, solves a major bottleneck that has long restricted the practical application of BCI technology, and has outstanding prospects for industrial application.

[0046] (2) High geometric structure fidelity: Compared with traditional tangent space alignment (TSA) or re-centering methods, this application applies parallel transport for core alignment within a "multi-manifold" framework. This method ensures that the relative geometric structure of the data is not distorted during the "transportation" process, preserving the task-related, finely discriminative local geometric features of the data to the maximum extent. Especially in small-sample learning scenarios, it can avoid geometric distortions introduced by data transformation, thereby achieving higher decoding accuracy.

[0047] (3) Solving the problem of non-uniform frequency band drift. Unlike the existing technology that mixes multi-band signals into a single covariance matrix in a "one-size-fits-all" approach, this invention proposes a strategy of "independent multi-band manifold processing + subsequent weighted fusion". It constructs an independent reference matrix for each frequency band and performs alignment independently, fundamentally addressing the "non-uniformity". The subsequent weighted adaptive fusion mechanism can better suppress noisy frequency bands, thereby addressing the physiological phenomenon of inconsistent differences between high-frequency and low-frequency signals among different patients in complex cross-subject scenarios and improving the robustness of the system.

[0048] In some embodiments, an implantable brain-computer interface device is also provided, including a signal sensor, a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the method, including decoding the intracranial electroencephalogram (EEG) signals acquired by the signal sensor. Specifically, the implantable brain-computer interface device includes, but is not limited to, structural components and hardware circuits. The method is run on a lower-level computer, a higher-level computer, and a display module. Signal input is provided by a signal sensor, an amplifier, and an analog-to-digital converter module.

[0049] (1) Signal sensor to collect raw analog EEG signals.

[0050] (2) Lower-level computer (MCU / ARM / DSP hardware + embedded software): stores / forwards digital signals. Amplifier module (hardware analog circuit): amplifies the original analog EEG signal. Analog-to-digital converter module (hardware): converts the amplified analog EEG signal into a digital signal at a fixed sampling rate and sends it to the upper-level computer.

[0051] (3) The host computer (PC hardware + PC software) processes digital signals and uses a decoding model to decode the EEG signals into control instructions for training peripherals, and displays them to the user in conjunction with the UI workflow.

[0052] (4) Display module: presents the UI interface, user-operable UI, and sets the decoding model parameters.

[0053] (5) Training peripherals, which may be exoskeletons or hand training devices, execute control commands to complete training tasks.

[0054] In some embodiments, a computer device is also provided, including a memory, a processor, and a computer program stored on the memory. The processor executes the computer program to implement the steps of the method. The processor may be a central processing unit (CPU), an ASIC, or one or more integrated circuits configured to implement embodiments of the present invention. In specific implementations, if the memory and processor are implemented independently, they can be interconnected via a bus to communicate with each other. The bus may be an Industry Standard Architecture (ISA) bus, a PCI bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. If the memory and processor are integrated on a single chip, they can communicate with each other through an internal interface.

[0055] In some embodiments, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method. The storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0056] In some embodiments, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the method. The computer program may include program code, which includes computer operation instructions and may be stored in a computer-readable storage medium. Based on this understanding, when the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be implemented in the form of a software product or sold or used as an independent product, the computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the method described in the various embodiments of the present invention.

[0057] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0058] Based on the above-described preferred embodiments of the present invention, and through the above description, those skilled in the art can make various changes and modifications without departing from the technical concept of the present invention. That is, the technical scope of the present invention is not limited to the contents of the specification.

Claims

1. A method for decoding intracranial electroencephalography signals across subjects based on manifold-parallel translation, characterized in that, include: Constructing a source library involves obtaining at least one source model and its corresponding source domain based on historical data from different subjects under the same task. Construct independent manifolds with multiple frequency bands, that is, in different frequency bands, construct the first reference matrix corresponding to the source domain, the second reference matrix corresponding to the resting state data of the new subject, and the third reference matrix corresponding to the task state data of the new subject. The best model is called by setting the second reference matrix as the benchmark and comparing the differences between the second reference matrix and each first reference matrix one by one to obtain the model matching degree of the new subject, so as to call the best source model and the best source domain from the source library. By combining the second reference matrix with the first reference matrix corresponding to the optimal source domain, a migration path is constructed; Perform data migration, which involves mapping the third reference matrix along the migration path to the optimal source domain; The migrated task-state data is decoded using the optimal source model.

2. The method according to claim 1, characterized in that, Building the migration path includes: a source domain corresponding to the best source model computing a source domain matrix for each frequency band i.e. the first reference matrix; Before performing the task, based on a new subject's target domain Computing the target domain matrix for each frequency band i.e. the second reference matrix; Finding the shortest geodesic path between to on an independent manifold as the migration path for the band data.

3. The method according to claim 2, characterized in that, The execution data migration includes: Data mapping, in performing a task, obtaining the task state covariance matrix corresponding to the target domain task state data , that is, the third reference matrix, keeps the source domain unchanged, and the parallel movement conversion formula is used to map to the target domain along the migration path to obtain the task state covariance matrix of the target domain after migration ; Source domain tangent space projection, obtaining a source domain covariance matrix , and projecting and to the tangent space respectively to obtain corresponding tangent vectors , i.e. and ; The weighted fusion of multi-manifold features involves weighting and concatenating tangent vectors from different frequency bands in the tangent space to obtain the translated feature vector, which represents the transferred task-state data.

4. The method according to claim 3, characterized in that, The parallel movement conversion formula is: ； The tangent vector is configured to wherein denotes the covariance matrix of the corresponding domain, i.e. the task state covariance matrix of the target domain after the transfer or the source domain covariance matrix , denotes the source domain matrix for the corresponding frequency; The migrated task-state data is characterized as wherein denotes the weight of the corresponding frequency band , β denotes the β band in the key frequency band of the electroencephalogram signal, and Hγ denotes the Hγ band in the key frequency band of the electroencephalogram signal.

5. The method according to claim 1, characterized in that, The difference is configured as the geodesic distance between the new subject's second reference matrix and each of the first reference matrices in the source library.

6. The method according to claim 1, characterized in that, Constructing multi-band independent manifolds includes: At least one key frequency band of the EEG signal is extracted using a filter bank; for each critical band respectively compute their spatial covariance matrices ; Spatial covariance matrix based A plurality of independent Riemann submanifolds are constructed, i.e., the first reference matrix, the second reference matrix, and the third reference matrix corresponding to each data.

7. An implantable brain-machine interface device comprising a signal sensor, a memory, a processor, and a computer program stored on the memory, wherein the computer program comprises instructions that, when executed by the processor, cause the processor to: receive a signal from the signal sensor; determine a signal quality of the signal; and output a signal quality indicator based on the signal quality. The processor executes the computer program to perform the steps of the method as described in any one of claims 1-6, decoding the intracranial electroencephalogram (EEG) signals acquired by the signal sensor.

8. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, causes the processor to perform the method of any one of claims 1 to 7. The processor executes the computer program to implement the steps of the method as claimed in any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

Images