MEI-TID framework under manifold-Euclidean intersection space

Abstract

The invention discloses an MEI-TID framework under a manifold-Euclidean cross space, which is used for accurately decoding topology-implicit EEG representation of an EEG signal, and comprises the following steps: S1, aggregating sub-topology features by constructing a self-adaptive cluster center strategy on the manifold space; s2, the topology-implicit representation of the EEG signal is decoded in the manifold-Euclidean cross space; and S3, establishing an adaptive calibration mechanism, and decoding. According to the MEI-TID framework provided by the invention, high-precision decoding of the topology-implicit representation of the EEG is realized in a manifold-Euclidean cross space. The framework captures an agnostic implicit topological relation and characterization of EEG signals through a dynamic cluster center strategy on a manifold space, a manifold-Euclidean cross space structure and an adaptive calibration mechanism.

Description

Technical Field

[0001] The present invention belongs to the field of signal processing technology, and in particular relates to a MEI-TID framework in a manifold-Euclidean intersection space, which is used to accurately decode the topological-implicit EEG representation of EEG signals. Background Art

[0002] Electroencephalogram (EEG) signal decoding research builds a bridge for the transmission of commands between the brain and the device. Accurate and efficient decoding of EEG plays a decisive role in the field of brain-computer interface (BCI). For example, EEG signals can be used to monitor the mental fatigue state of vehicle drivers and drone operators, and can also be used to develop devices to regulate emotions to cure patients with mental illness. The MI motor imagery system based on EEG signals can also provide rehabilitation training for stroke patients. However, the premise for the realization of these tasks is the need for accurate decoding of EEG signals. Since EEG signals are collected from multiple channels and correspond to various brain regions, they contain both brain spatial topological information and temporal information, and have high temporal resolution, strong randomness, non-stationary, nonlinear, non-Gaussian process and other characteristics, resulting in accurate decoding of EEG signals and converting complex and tiny neuronal interaction information into effective discrimination information. It has always been a difficult task and a huge challenge in this field.

[0003] In order to meet this challenge, EEG decoding technology based on deep learning (DL) models has become the most popular focus and has achieved impressive results in fatigue detection, emotion recognition and motor imagery classification. Their research mainly focuses on the following aspects: (1) Some studies tend to extract discriminative features from the perspective of time-space-frequency multi-dimensional feature fusion, and improve the decoding effect by iteratively optimizing the feature extractor (2) Some studies start from the prior knowledge of neurophysiology and design encoder-decoders by considering the local fine-grained features and global coarse-grained features of the brain separately or comprehensively, in an attempt to capture the spatial topological pattern representation of EEG signals; (3) When considering the decoding differences of individual differences, some studies consider the problem of large differences in source and target domain features and propose transfer learning schemes to narrow the differences and thus improve the generalization ability of the model and the stability of decoding. It is found that these studies are all conducted on Euclidean space. However, the flat nature of Euclidean space cannot truly represent the complexity of EEG signals, especially its complex brain topological representation, resulting in unsatisfactory overall decoding results.

[0004] In recent times, as geometric perceptual learning (GAL) has been found to be beneficial in improving the robustness of EEG signal decoding, it has attracted much attention. Some studies have found that combining DL and GAL to design a decoder can improve the decoding performance of EEG signals. And it is worth noting that Riemannian geometry (RG), which is a non-Euclidean geometry in GAL, has shown impressive decoding effects in BCI tasks. For example, Pan YT, Chou JL, Wei C S. MAtt: A manifold attention network for EEG decoding [J]. Advances in Neural Information Processing Systems, 2022, 35: 31116-31129. A manifold attention network is proposed to map EEG signals into manifold space so that the spatiotemporal information of EEG signals can be better represented, and it has been effectively verified in multiple BCI tasks. Kobler R, Hirayama J, Zhao Q, et al. SPD domain-specific batch normalization to crack interpretable unsupervised domain adaptation in EEG [J]. Advances in Neural Information Processing Systems, 2022, 35: 6219-6235 proposed a new geometric deep learning block to keep the symmetric positive definite (SPD) manifold features unchanged, thereby alleviating the problem of unsupervised domain adaptation. In addition, Yi K, Wang Y, Ren K, et al. Learning Topology-Agnostic EEG Representations with Geometry-Aware Modeling [J]. Advances in Neural Information Processing Systems, 2024, 36 introduced the idea of ​​geometric perception and designed a multi-stage pre-training strategy to communicate the spatial information between the structures of each electrode channel, and based on this, a unified topological mapping feature was obtained, and better decoding results were obtained.In addition, the method proposed by Suh YJ, Kim B H. Riemannian embedding banks for common spatial patterns with EEG-based SPD neural networks[C] / / Proceedings of the AAAI Conference on Artificial Intelligence. 2021, 35(1): 854-862, which uses Riemannian manifold space to learn common spatial patterns of multiple subdomains to extract EEG neural dynamic features, also achieved a certain improvement in the decoding results.

[0005] These GAL-based studies have improved the decoding effect of EEG signals to a certain extent, but they have ignored some important EEG features that cannot be fully extracted in the manifold space. At the same time, neurobiology shows that there are some important implicit information in the extremely complex EEG topological structure. Different topological structures will lead to differences in important discriminative representations under specific cognitive task states, which increases the difficulty of accurately decoding EEG under different specific cognitive tasks. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a MEI-TID framework in manifold-Euclidean intersection space that can effectively extract the implicit topological relationship of EEG signals, improve decoding accuracy, and achieve efficient decoding of topological-implicit EEG representations under different BCI tasks.

[0007] The objective of the present invention is achieved through the following technical solution: A MEI-TID framework in a manifold-Euclidean intersection space is used to accurately decode the topological-implicit EEG representation of EEG signals, comprising the following steps:

[0008] S1, aggregating sub-topological features by constructing an adaptive cluster center strategy on the manifold space;

[0009] S2. Topological-implicit representation of EEG signals decoded in manifold-Euclidean cross space;

[0010] S3. Establish an adaptive calibration mechanism and perform decoding.

[0011] The beneficial effects of the present invention are:

[0012] (1) Based on the significant differences in individual EEG data, a dynamic cluster center strategy on the manifold space is constructed to adaptively aggregate sub-topologies, which can represent the topological information of the brain.

[0013] (2) We perform cross-calculations on the reconstructed sub-topological structure in the joint manifold space and the sub-topological representation in the Euclidean space, and propose an adaptive calibration mechanism to unify the EEG spatiotemporal representation of the sub-topology.

[0014] (3) MEI-TID can effectively extract the implicit topological relationship of EEG signals, improve decoding accuracy, and achieve efficient decoding of topological-implicit EEG representations under different BCI tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a data processing flow chart of the MEI-TID framework in the manifold-Euclidean intersection space of the present invention;

[0016] Figure 2 It is a structural schematic diagram of the MEI-TID framework of the present invention;

[0017] Figure 3 This is the clustering effect of subjects 8-10 in data set 1;

[0018] Figure 4 Topology-agnostic representations captured by MEI-TID on three subjects in dataset 1. DETAILED DESCRIPTION

[0019] This paper proposes a new framework for EEG signal decoding in the manifold-Euclidean intersection space, called MEI-TID. Different from previous studies, we construct a dynamic cluster center strategy on the manifold space for EEG signals, and adaptively aggregate them into K sub-topologies according to node-centrality based on the differences between different individuals. In addition, based on previous studies, we found that the EEG signal features caused by the complex relationship between electrode channels are difficult to fully extract in the Euclidean space. Therefore, this paper reconstructs the high-dimensional topological structure representation by mapping these K sub-topological structures to the Riemannian manifold space. Next, we jointly construct a graph mechanism operator with the sub-topological representation in the Euclidean space, so as to achieve the decoding of topological-implicit EEG representation in the manifold-Euclidean intersection space. Finally, since the sub-topologies of interest presented by cognitive states at different time points have significant differences, we adaptively calibrate these sub-topologies to obtain a unified EEG spatiotemporal representation. We verify the proposed MEI-TID on four EEG benchmark datasets, and the results show that the decoding performance of MEI-TID is better than the existing state-of-the-art technology. Furthermore, the dissection of MEI-TID revealed its superior ability to explore implicit relationships unknowable by EEG and to capture representational differences across different cognitive states.

[0020] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0021] like Figure 1As shown, a MEI-TID framework in a manifold-Euclidean intersection space of the present invention is used to accurately decode the topological-implicit EEG representation of EEG signals, including the following steps:

[0022] S1. Aggregate sub-topological features by constructing an adaptive cluster center strategy on the manifold space; including the following steps:

[0023] S11, for the original data X in the EEG data set = {X 1 ,X 2 ,...,X N}Bandpass filtering is performed to obtain five frequency bands: Delta, Theta, Alpha, Beta, and Gamma, whose frequency bands are 0.5-4Hz, 4-8Hz, 8-12Hz, 12-30Hz, and 30-50Hz respectively;

[0024] S12, for each frequency band data, use differential entropy (DE) X f ∈R N×C×B and Pearson correlation coefficient (PCC) X P ∈R N×B×C×C Represents channel and structure information; where N, C and B represent the number of samples, channels and frequency bands respectively, and R represents the real number domain;

[0025] S13. Since the complexity of brain topological information represented by the electrodes is difficult to fully capture in Euclidean space, the calculated structural information X P Clustering is performed by transforming from Euclidean space to SPD manifold space; it is worth noting that all electrode channels are potential centers of interest, so the affinity propagation algorithm is introduced to construct a dynamic cluster center strategy in manifold space. Even for samples with the same label, their structural information will show some unknown transformations over time, so there must be differences in the cluster centers shown by different samples.

[0026] For X containing N samples P , and maps it to the SPD manifold space according to the structural information corresponding to each frequency band under a single sample. In the present invention, B is used as feature embedding and B is divided into B 1 ,B 2 ,..,B 5 ; Based on B m Calculate the covariance matrix, so that a series of covariance matrices can be obtained and recorded as CM B1 ,CM B2 ,..,CM B5 :

[0027]

[0028] Among them C i and C j represent the i-th and j-th channels respectively, and are the feature vectors of the i-th and j-th channels respectively, n represents the number of elements in the vector, They are and The lth element in ; Cov(·) represents the covariance matrix, m=1,..,5;

[0029] S14. Reconstructing the high-dimensional topological structure representation on the manifold space is achieved by tracking normalization and assigning learning factors to the diagonal elements in the matrix, denoted as

[0030]

[0031] where tr() represents the trace of the matrix, μ and E I denote the learning factor and the identity matrix respectively;

[0032] S15. For a specific object C i and other objects C j , respectively obtain C i to C j Message and C j to C i The message indicates that C j Can be used as C i The fitness of cluster centers and C i Select C j The willingness to be the cluster center; these two degrees are respectively and

[0033] Using logarithmic-Euclidean metric as C i and C j The cluster center clustering basis on the manifold space is calculated as follows:

[0034]

[0035] where ||·|| F represents the Frobenius norm, U i , U j are respectively The matrix obtained by singular value decomposition;

[0036] The log-Euclidean mean is expressed as:

[0037]

[0038] After adding the constrained weighted measure w l The weighted log-Euclidean mean can be obtained as:

[0039]

[0040] based on Iterative Updates and

[0041] for Iterate as follows:

[0042]

[0043] in Indicates that in the current state, except for object C j For other objects, select C i The willingness to be the cluster center,

[0044] Indicates that except object C j In addition, other objects and C i As the basis for clustering cluster centers on manifold space;

[0045] Represents C i Compared to C j More likely to become a cluster center;

[0046] Iterate as follows

[0047]

[0048] In order to avoid oscillation during the iteration process, the attenuation coefficient λ is introduced, and the iterative calculation process is updated as follows:

[0049]

[0050] By dynamically clustering the cluster centers of different samples through the above process, the purpose of adaptively aggregating the features at different time points into K sub-topologies based on the differences between different individuals is achieved, and the corresponding sub-topology structure is obtained. and subtopological representation set As shown below:

[0051]

[0052] S2. After completing the dynamic clustering in the manifold space, the topological-implicit representation of the EEG signal needs to be decoded in the manifold-Euclidean cross space. The specific implementation method is: based on X PThe design concept of this module is to randomly reconstruct the corresponding hidden topological structure Ω based on the original features of the sub-topological structure of each sample, and input them into a basic convolutional layer for spatial filtering, and then preliminarily obtain the essential cognitive driving structure E p and hidden cognitive driving structure H p ;

[0053]

[0054] Note that the purpose of this operation is to map different regions into similar spaces. Then, the two are mapped into the Riemann manifold space to reconstruct the cognitive driving structure:

[0055]

[0056] Represents manifold computation;

[0057] The reconstructed cognitive driving structure is integrated through the following formula:

[0058]

[0059] Then, according to the following formula, they are mapped to the similarity space through the Sigmoid function to abstract the stable essential cognitive driving structure and the hidden cognitive driving structure:

[0060]

[0061] θ [K] represents the linear layer parameter, σ sig () represents the Sigmoid function, FC() represents the fully connected layer;

[0062] The cognitive drive structure is a unidirectional connected structure. In order to reduce the impact of redundant features on the decoding results, we have developed a unidirectional drive mechanism to sparse its connectivity. The present invention has developed a unidirectional drive mechanism to sparse its connectivity. The specific calculation process is shown in the following formula:

[0063]

[0064] in and represents the weight parameter, represents the bias vector;

[0065] Then, through the graph mechanism operator, the sub-topological structure on the manifold space is updated and calculated, and then the sub-topological representation on the manifold space is obtained, and finally the essential cognitive driven representation and hidden cognitive driven representation across space are obtained:

[0066]

[0067] Among them, Graph θ () represents the graph mechanism operator, θ is the parameter of the graph mechanism operator; θ s represents the Fourier domain parameter, T s (.) represents the s-order Chebyshev polynomial, W f represents the weighting parameter.

[0068] S3. Establish an adaptive calibration mechanism and perform decoding. The specific implementation method is: after obtaining the essential cognitive driving representation and hidden cognitive driving representation of each sub-topology across space, first fuse them to obtain a unified representation of the sub-topology:

[0069]

[0070] Then K is calculated according to the following formula N The sub-topology structure is randomly generated K N calibration parameters to obtain a unified EEG spatiotemporal representation:

[0071]

[0072] Concat() represents the concatenation operation. [K] represents calibration parameters;

[0073] Next, UR f Feed this into the decoder to get the result:

[0074] UR f =FC 1×class (σ sig (BN(FC 1×C (UR f θ C )));θ class )

[0075] FC 1×class Indicates that the parameters of the fully connected layer are set to the final number of categories class, BN() represents the normalization function, FC 1×C The parameter indicating the fully connected layer is set to the number of electrode channels C.

[0076] The above steps of the present invention are implemented through three modules: Figure 2As shown. In the figure, (a) module Dynamicclustercenterstrategy represents the dynamic cluster center strategy, (b) module Manifold euclidean hybridspace structure represents the manifold Euclidean hybrid space structure, and (c) module Adaptive calibrationmechanism represents the adaptive calibration mechanism. The three modules correspond to the three steps of the present invention. In the module (a), from left to right, they are: Input: multi-dimensional topological data (Topologies for multipledimensions); Compression of a singlesubject represents compression into a single subject, that is, bandpass filtering is performed in the present invention to obtain five frequency bands; Topologies on manifold space is the topology on the manifold space, which means that the calculated structural information X P Transform from Euclidean space to SPD manifold space for clustering; finally get the clustering result. (b) In the module, Topologies of K N Brain regions are K N The brain region topology, i.e., K generated in the present invention N sub-topological structures; randomized generation of latent topologies means randomly generating hidden topological structures; latent means hidden topological structures, intrinsic means original topological structures; Unidirectional congnitive driven means unidirectional cognitive drive; Reconstructed topologies → similar space → Riemannian manifold space means reconstructed topologies → similar space → Riemannian manifold space; mapped to Riemannian manifold space; get the final cognitive connectivity structure (Final cognitive connectivity structure) In the (c) module, Paramter Sharing means parameter sharing; Joint marifold euclidean space for cross space feature representations means joint manifold-Euclidean space for cross space feature representations; Random generation of K N calibration parameters for K N subtopologies represent K NThe sub-topology randomly generates K N calibration parameters; Unified representations of fusion subtopologies represent the unified representation of fusion subtopologies; the final input UR f Enter the decoder to obtain the classification result.

[0077] In summary, the forward calculation process of the framework proposed by the present invention is shown in the following table.

[0078]

[0079]

[0080] The above model of the present invention can be trained by the process of LOSS reverse parameter update. The main training parameters are the topological representation in Euclidean space and the essential driving structure and hidden driving structure in manifold space. The loss of this framework is denoted as Then for K N X of subtopologies p and Ω, whose corresponding training parameters W P and W Ω The update process is as follows:

[0081]

[0082] It can be further expanded to:

[0083]

[0084] It is worth noting that and is the SPD matrix, expressed as therefore Their updates can be expressed as:

[0085]

[0086] The update rule of the SPD matrix is ​​further summarized as:

[0087]

[0088] Where l represents the learning rate. Then the topological representation weights can be obtained as shown below:

[0089]

[0090] in

[0091] The effect of the present invention is further verified by experiments below.

[0092] This embodiment uses four EEG datasets to verify the decoding performance of the proposed MEI-TID, which are respectively denoted as Dataset I-Dataset IV. Based on these four datasets, the most advanced decoding methods are compared with the proposed MEI-TID for different cognitive tasks. Among them, SEED (Wei-Long Zheng, and Bao-Liang Lu, Investigating Critical Frequency Bands and Channels for EEG-based Emotion Recognition with Deep Neural Networks, accepted by IEEE Transactions on Autonomous Mental Development (IEEE TAMD) 7 (3): 162-175, 2015) and SEED-IV (Wei-Long Zheng, Wei Liu, Yifei Lu, Bao-Liang Lu, and Andrzej Cichocki, EmotionMeter: A Multimodal Framework for Recognizing Human Emotions. IEEE Transactions on Cybernetics, Volume: 49, Issue: 3, March 2019, Pages: 1110-1122, DOI: 10.1109 / TCYB.2018.2797176) dataset was used to evaluate the decoding performance of MEI-TID in cross-session valence emotions and discrete emotions; the self-made fatigue dataset was used to evaluate the decoding performance of MEI-TID in cross-session fatigue (load) detection tasks; the Shu (Ma J, Yang B, Qiu W, et al. A large EEG dataset for studying cross-session variability in motorimagery brain-computer interface [J]. Scientific Data, 2022, 9 (1): 531) dataset was used to evaluate the decoding performance of MEI-TID in similar cross-session motor imagery tasks.

[0093] (1) Framework training

[0094] Considering the actual situation of different cognitive task datasets, a series of experiments are conducted on a cross-session basis to compare the decoding performance of MEI-TID and other state-of-the-art methods. When training these frameworks and models, it is necessary to use part of the subject's data for training, and then reserve part of the data as a test set. Therefore, for the dataset used in this article, we use one of the sessions as a test set each time, and the rest as a training set to train the decoder. Our experiments are implemented based on the Pytorch framework. In addition to Accuracy (Acc), we also use standard deviation (STD) to measure the volatility and stability of the accuracy indicator results. It should be emphasized that due to the relatively significant differences in the distribution of subject data features in each dataset, the convergence of MEI-TID on different datasets is different.

[0095] (2) Decoding performance comparison: MEI-TID is compared with the current state-of-the-art methods, and the results are shown in Tables 1 and 2. The results show that MEI-TID exhibits significant advantages in decoding emotion and motor imagery tasks, highlighting the superiority of the framework in terms of accuracy and stability. We also note that the decoding stability of MEI-TID on the self-made fatigue dataset (STD = 11.55%) is slightly weaker than SFTNet (STD = 9.23%) and LAG (STD = 10.17%), but there is no order of magnitude difference overall. Overall, the classification accuracy of MEI-TID is about 3%-4% higher than that of other methods on average. It is also worth noting that most studies only develop decoders for EEG signals of a single cognitive task. In comparison, the MEI-TID proposed in this paper has the ability to achieve high-accuracy decoding of EEG signals of multiple cognitive tasks and can be applied to different types of EEG data.

[0096] Table 1 Decoding performance comparison of MEI-TID and other methods on datasets I and II

[0097]

[0098] Table 2 Decoding performance comparison of MEI-TID with other methods on datasets III and IV

[0099]

[0100]

[0101] (3) Ablation experiment: This section uses ablation experiments to explore the rationality of each component of the MEI-TID framework and the contribution of each component to the final decoding result. As shown in Table 3, we first conducted a comparative study on adaptive clustering and found that the decoding effect after the adaptive clustering process was significantly better, which shows that the adaptive clustering process has captured the brain topology information represented by the mutual relationship between electrode channels to a certain extent, and increased the important feature information of the decoder. Secondly, we compared the decoding results of the conventional Euclidean space and the proposed manifold-Euclidean cross space. The experimental results show that cross-learning in the two spaces can capture the complex manifold space structure while updating the Euclidean space features, and ultimately significantly improve the decoding ability of the framework.

[0102] Table 3 Ablation study of MEI-TID on four datasets.

[0103]

[0104] NAC: non-adaptive clustering; AC: adaptive clustering; ES: Euclidean space; MECS: manifold-Euclidean cross space.

[0105] (4) Interpretable experiments: The MEI-TID framework and experimental results are further explained by combining neuroscience knowledge. Figure 3 The adaptive clustering effects of three subjects in dataset I are presented, where (a) is the clustering effect of subject 8, (b) is the clustering effect of subject 9, and (c) is the clustering effect of subject 10. In the figure, the left figure is the clustering result diagram, and the right figure is the corresponding channel distribution diagram.

[0106] It can be found that the clustering results on the manifold space have similarities and differences with the prior knowledge of neuroscience. According to the cluster distribution, these interesting patterns are found: (1) The EEG signal feature distribution of the electrodes in the prefrontal region is highly similar to that of the electrodes such as Cz that extend along the central line to the central region. Compared with the common brain functional area division research, the present invention can dig out the strong connectivity and collaboration between the prefrontal region and the front half of the central line region. (2) In general, the cluster center of the clustering results shows the characteristics of left and right hemisphere symmetry, which corresponds to the conclusion that the left and right hemispheres of the brain work together during emotion induction (Zhang G, Yu M, Liu YJ, et al. SparseDGCNN: Recognizing emotion from multichannelEEG signals [J]. IEEE Transactions on Affective Computing, 2021, 14 (1): 537-548), and it is particularly symmetrical in the front half of the brain. (3) This study found that the distribution of clustering results in the frontal, central and temporal regions of different subjects is basically consistent. On the contrary, the clustering results of the three subjects in the parietal and occipital regions are significantly different, indicating that the perception and visual impact of the brain during the emotion induction process vary from person to person. Therefore, this may also be one of the reasons for the differences between subjects. Here, we can find that the dynamic cluster center strategy can effectively distinguish the differences while obtaining the common characteristics of the subjects.

[0107] Figure 4 The unknowable implicit relationships and representations of the emotions induced by three subjects in Dataset I captured by the MEI-TID framework are presented. Figure (a) is subject 8, Figure (b) is subject 9, and Figure (c) is subject 10. From left to right, they are neutral, positive, and negative. It can be found that when the brain is not significantly stimulated, it shows a strong response in the right parietal lobe (TP8, P8, PO8), occipital lobe (Oz, O2) and central (C1, Cz, FCz).

[0108] Inspired by neurobiology, this paper proposes a MEI-TID framework in a manifold-Euclidean cross space to achieve accurate decoding of the topological-implicit EEG representation of EEG signals. By constructing an adaptive cluster center strategy on the manifold space to aggregate sub-topological features, and then combining the reconstructed sub-topological structure driven unidirectionally on the popular space and the sub-topological representation of the Euclidean space, the EEG spatiotemporal representation is finally obtained by unifying the sub-topological structure through the development of graph mechanism operators and adaptive calibration mechanisms. The experimental results on four data sets show that the decoding performance of MEI-TID is better than the most advanced decoding technology. The ablation experiment also proves that each module plays a positive role in the overall decoding ability of the framework and can improve the decoding accuracy of the MEI-TID framework. And through the interpretability experiment, it is revealed that the unknowable implicit cooperation mode between different electrode channels and the difference in cognitive state representation between individuals will affect the decoding effect to a certain extent, and once again prove the rationality of the MEI-TID framework proposed in this paper. In short, this study provides new ideas for EEG decoding research and can provide decoding support for further exploring the mysteries of neurobiology. It will be future work to further study the framework-neural mechanism of the mutual cooperation between manifold and Euclidean cross space to improve decoding results.

[0109] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A MEI-TID framework in a manifold-Euclidean intersection space for accurate decoding of topological-implicit EEG representations of EEG signals, characterized by: The following steps are involved: S1, aggregating sub-topological features by constructing an adaptive cluster center strategy on the manifold space; The steps include: S11, the original data X in the data set D = {X 1 ,X 2 ,...,X N }Bandpass filtering is performed to obtain five frequency bands: Delta, Theta, Alpha, Beta, and Gamma, whose frequency bands are 0.5-4Hz, 4-8Hz, 8-12Hz, 12-30Hz, and 30-50Hz respectively; S12. For each frequency band, use the differential entropy X f ∈R N×C×B and Pearson correlation coefficient X P ∈ R N×B×C×C Represents channel and structure information; where N, C, and B represent the number of samples, channels, and frequency bands, respectively; S13, X P Transform from Euclidean space to SPD manifold space for clustering: Use B as feature embedding and divide B into B1, B2, .., B5; Based on B m Compute the covariance matrix: Among them C i and C j denote the i-th and j-th channels respectively, and are the feature vectors of the i-th and j-th channels respectively, n represents the number of elements in the vector, They are and The lth element in ; Cov(·) represents the covariance matrix, m=1,..,5; S14. Reconstructing the high-dimensional topological structure representation on the manifold space is achieved by tracking normalization and assigning learning factors to the diagonal elements in the matrix, denoted as where tr() represents the trace of the matrix, μ and E I denote the learning factor and the identity matrix respectively; S15. For a specific object C that is dynamically initialized i and other objects C j , respectively obtain C i to C j Message and C j to C i The message indicates that C j Can be used as C i The fitness of cluster centers and C i Select C j The willingness to be the cluster center; these two degrees are respectively and Using logarithmic-Euclidean metric as C i and C j The cluster center clustering basis on the manifold space is calculated as follows: in‖·‖ F represents the Frobenius norm, U i , U j are respectively The matrix obtained by singular value decomposition; The log-Euclidean mean is expressed as: After adding the constrained weighted measure w l The weighted log-Euclidean mean can be obtained as: based on Iterative Updates and for Iterate as follows: in Indicates that in the current state, except for object C j For other objects, select C i The willingness to be the cluster center, Indicates that except object C j In addition, other objects and C i As the basis for clustering cluster centers on manifold space; Represents C i Compared to C j More likely to become a cluster center; Iterate as follows In order to avoid oscillation during the iteration process, the attenuation coefficient λ is introduced, and the iterative calculation process is updated as follows: By dynamically clustering the cluster centers of different samples through the above process, the purpose of adaptively aggregating the features at different time points into K sub-topologies based on the differences between different individuals is achieved, and the corresponding sub-topology structure and sub-topology representation set are obtained, as shown in the following formula: S2. Topological-implicit representation of EEG signals decoded in manifold-Euclidean cross space; S3. Establish an adaptive calibration mechanism and perform decoding.

2. According to the MEI-TID framework in the manifold-Euclidean intersection space of claim 1, the specific implementation method of step S2 is: based on the original features of the sub-topological structure of each sample, the corresponding hidden topological structure Ω is randomly reconstructed, and each is input into a basic convolutional layer for spatial filtering, and then the essential cognitive driving structure E is preliminarily obtained. p and hidden cognitive driving structure H p ; Then, we map the two into the Riemann manifold space to reconstruct the cognitive driving structure: Represents manifold computation; The reconstructed cognitive driving structure is integrated through the following formula: Then, according to the following formula, they are mapped to the similarity space through the Sigmoid function to abstract the stable essential cognitive driving structure and the hidden cognitive driving structure: definition Represents the calculation process; θ [K] represents the linear layer parameters; A unidirectional driving mechanism is developed to sparse its connectivity. The specific calculation process is shown in the following formula: in and represents the weight parameter, represents the bias vector; Then, through the graph mechanism operator, the sub-topological structure on the manifold space is updated and calculated, and then the sub-topological representation on the manifold space is obtained, and finally the essential cognitive driven representation and hidden cognitive driven representation across space are obtained: where θ represents the Fourier domain parameter, T s (.) represents the s-order Chebyshev polynomial, W f represents the weighting parameter.

3. According to the MEI-TID framework in the manifold-Euclidean intersection space of claim 1, the specific implementation method of step S3 is: after obtaining the essential cognitive driving representation and the hidden cognitive driving representation of each sub-topology across the space, first merge them to obtain the unified representation of the sub-topology: Then K is calculated according to the following formula N The sub-topology structure is randomly generated K N calibration parameters to obtain a unified EEG spatiotemporal representation: Concat() represents the concatenation operation. [K] represents the calibration parameters; Next, UR f Feed this into the decoder to get the result: UR f =FC 1×class (s sig (BN(FC 1×C (UR f ;θ C )));θ class ) FC 1×class Indicates that the parameters of the fully connected layer are set to the final number of categories class, BN() represents the normalization function, FC 1 ×C The parameter indicating the fully connected layer is set to the number of electrode channels C.

Images