Electroencephalogram signal motor imagery decoding method based on two-way second-order feature fusion

Abstract

The invention discloses an electroencephalogram signal motor imagery decoding method based on two-way second-order feature fusion, which aims at challenges of low signal-to-noise ratio, non-stationarity, cross-subject distribution difference and the like existing in motor imagery electroencephalogram signals, enhances robustness through data preprocessing, extracts multi-scale spatiotemporal features by using a multi-branch graph convolutional network, and decodes the multi-scale spatiotemporal features to obtain the motor imagery signals of the motor imagery signals of the motor imagery signals of the motor imagery signals of the motor imagery signals of the motor imagery signals. And respectively constructing global and local second-order feature structures. The global second-order feature captures an overall rhythm mode through position coding and long-range time sequence modeling, and the local second-order feature captures instantaneous dynamic change through sliding window slicing and aggregation operation. And finally, adaptively integrating double-path features in a logarithmic Euclidean space by adopting a gating fusion strategy, and realizing motor imagery category decoding through a classifier. According to the method, by combining multi-scale spatial-temporal feature extraction and two-way second-order structure modeling, the decoding stability and generalization ability are remarkably improved, and the method is suitable for scenes such as real-time brain-computer interface control, neural rehabilitation and intelligent auxiliary equipment.

Description

Technical Field

[0001] This invention belongs to the field of EEG signal decoding and motor brain-computer interface technology, specifically involving a method for decoding motor imagery based on dual-path second-order feature fusion of EEG signals. Background Technology

[0002] Brain-computer interfaces (BCIs) enable direct interaction between humans and external devices by analyzing brain electrical activity, and have significant application value in fields such as neurorehabilitation, intelligent assisted control, and human-computer interaction. As a mainstream non-invasive signal acquisition method, electroencephalography (EEG) offers advantages such as low cost, high temporal resolution, and convenient acquisition, making it particularly suitable for real-time monitoring and continuous interaction scenarios. Therefore, it has long played a crucial role in medical rehabilitation, intelligent control, and wearable devices.

[0003] Among various EEG paradigms, motor imagery (MI) has attracted considerable attention because it can induce stable cortical rhythm changes without external stimuli or actual movements. MI activates the μ and β rhythms of the sensorimotor cortex by simulating limb movements, presenting a typical ERD / ERS pattern. These neural activities are highly consistent with real movements, exhibiting good physiological interpretability and user operability. Therefore, it is widely used in rehabilitation training, prosthetic and exoskeleton control, brain-controlled wheelchairs, and virtual interaction scenarios.

[0004] Against this backdrop, accurate decoding of MI-EEG signals has become crucial for the practical application of BCI technology. Decoding performance directly impacts the accuracy of brain-controlled command recognition and the quality of system response, and is vital for the stability of rehabilitation feedback and intelligent control systems. However, MI-EEG signals generally exhibit low signal-to-noise ratios, significant non-stationarity, and cross-subject variability. They are frequently affected by interference from electromyography and eye movement artifacts. The distribution of cross-session features for the same subject also drifts with changes in state, and significant distribution shifts occur between different subjects due to differences in physiological structure. These factors make it difficult for traditional models relying on fixed spatiotemporal structures or single statistical features to maintain stable performance in real-world scenarios, limiting the generalization ability and reliability of MI decoding.

[0005] Therefore, there is an urgent need for a technical solution to address issues such as noise interference, cross-conversation drift, and cross-subject distribution differences in motor imagery EEG signals. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide a method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion.

[0007] The operation process includes: (1) building a training framework, performing data preprocessing on the original EEG signals, and setting a segment of data after random cropping and random selection of the starting point to zero in order to enhance the robustness of the model to local signal loss; (2) extracting multi-scale spatiotemporal features of EEG data based on a multi-branch graph convolutional network to obtain a spatiotemporal embedding representation for high-order statistical modeling; (3) constructing a global second-order feature branch, performing position encoding and temporal modeling on the spatiotemporal embedding features, and obtaining a global second-order structure through covariance estimation and log-Euclidean mapping; (4) constructing a local second-order feature branch, slicing the features according to a sliding window and calculating the covariance and log-mapping respectively, and forming a local second-order structure through an aggregation module; (5) performing gating fusion on global and local second-order features in log-Euclidean space, and outputting the motor imagery category through a classifier.

[0008] The specific technical solution is as follows:

[0009] A method for decoding motor imagery from EEG signals based on dual-channel second-order feature fusion includes the following steps: (1) Data preprocessing and training sample construction: The original EEG signals were randomly cropped and zeroed out at random time intervals to enhance the robustness of the model to local signal loss. Then, bandpass filtering and channel normalization were performed to form training samples. (2) Multi-branch graph convolution spatiotemporal feature extraction: The preprocessed EEG samples are input into the multi-branch graph convolutional network, and multi-scale spatiotemporal embedding features are extracted through convolutional kernels at different time scales and fused to form a unified spatiotemporal feature representation; (3) Global second-order feature construction: Learnable positional encoding is added to the spatiotemporal embedded features to perform long-range time series modeling. Global second-order features are obtained through covariance estimation and log Euclidean mapping to characterize the overall rhythmic structure. (4) Construction of local second-order features: The spatiotemporal embedding features are sliced ​​according to the sliding window, the covariance matrix is ​​calculated for each local window and logarithmic mapping is performed, and local second-order features are formed through the aggregation module to capture short-term rhythm changes; (5) Second-order feature fusion and classification decoding: In the logarithmic Euclidean space, a gating fusion strategy is used to integrate global and local second-order features, and the motion imagination category is output through a fully connected layer and a Softmax classifier.

[0010] Furthermore, the specific operation process in step (1) is as follows: S101. Random cropping: Randomly cropping fixed-length segments from continuous EEG signals; S102. Random time period zeroing: Randomly select the starting point and length in the clipped signal to zero out the data; S103. Bandpass Filter: Filters out irrelevant frequency bands. S104. Channel Normalization: Calculate the mean and standard deviation over the time dimension and perform Z-score normalization.

[0011] Furthermore, the specific operation process of step (2) is as follows: S201. Shared Graph Convolution: Extracting spatially relevant features using Chebyshev multinomial graph convolution; S202. Multi-scale temporal convolution: Short-time, medium-time and long-time temporal patterns are extracted in parallel using convolutional kernels of different lengths; S203. Feature Fusion: Perform batch normalization and activation function processing on the multi-branch outputs, and then concatenate them along the feature dimensions.

[0012] Furthermore, the specific operation process of step (3) is as follows: S301. Location-aware coding: Adding learnable location embeddings to spatiotemporal embedding features; S302. Long-range temporal modeling: Using large convolutional kernels to cover the temporal dependencies of the entire sample window; S303. Covariance Stabilization: Introducing a contraction term to enhance the robustness of the covariance matrix; S304. Log-Euclidean mapping: Maps the covariance matrix to a symmetric matrix space.

[0013] Furthermore, the specific operation process of step (4) is as follows: S401. Sliding window slice: Generates local segments along the time dimension with a preset window length and step size; S402. Local Covariance Calculation: Calculate the covariance matrix for each window and perform stabilization processing; S403. Logarithmic mapping: Transforms a local covariance matrix into a symmetric matrix; S404. Feature Aggregation: Aggregate multiple local representations through weighted averaging or attention mechanisms.

[0014] Furthermore, the second-order feature fusion in step (5) adopts a gated fusion strategy, which adaptively adjusts the fusion ratio of global and local second-order features through a learnable coefficient β. The fusion formula is as follows: , Where β can be a scalar or a vector, G2 is the global second-order feature, and L2 is the local second-order feature.

[0015] Furthermore, the classification decoding in step (5) includes expanding the fused symmetric matrix into a feature vector, mapping it through a fully connected layer, and then outputting the class probability through Softmax.

[0016] Furthermore, the length of the temporal convolution kernel in the multi-branch graph convolutional network can be adaptively adjusted according to the EEG signal sampling rate.

[0017] Furthermore, the shrinkage parameter λ in the covariance stabilization process ranges from 0 to 1 and is used to control the degree of noise suppression.

[0018] The beneficial effects of this invention are as follows: This invention significantly improves the stability and generalization ability of EEG motor imagery decoding by combining data robustness enhancement, multi-branch graph convolutional feature extraction, and dual-path second-order structure modeling. In the four-class classification task test scenario of the BCI IV 2a dataset, it achieves a classification accuracy of 89.16%, far exceeding the accuracy of similar models. It is applicable to scenarios such as real-time brain-computer interface control and neurorehabilitation. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process of the present invention;

[0020] Figure 2 This is a schematic diagram of the model structure of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the scope of protection of the present invention is not limited thereto.

[0022] like Figure 1 As shown, a method for decoding motor imagery from EEG signals based on dual-path second-order feature fusion includes the following steps: Step 1: Data Preprocessing and Training Sample Construction The raw EEG signals were preprocessed as necessary. Data robustness was enhanced by random cropping and zeroing of random time periods. The samples were then filtered and normalized to form training samples that could be used for subsequent modeling.

[0023] Step 2: Spatiotemporal Feature Extraction via Multi-Branch Graph Convolution The preprocessed EEG samples are input into a multi-branch graph convolutional network. Each branch extracts the spatiotemporal embedding features of the EEG signals under multiple time scales and channel spatial relationships through different temporal convolution receptive fields and dynamic adjacency matrix modeling strategies. Through the fusion of spatial graph convolution, channel interaction and temporal convolution, a deep spatiotemporal feature representation containing global spatial topology and local temporal dynamics is formed.

[0024] Step 3: Global Second-Order Feature Construction For the spatiotemporal embedding features of multi-branch outputs, we first add learnable positional encoding to enhance the temporal positional information representation, and then model the global dynamic characteristics of the entire experiment through a long-range temporal convolutional structure. Subsequently, we perform covariance estimation on the enhanced temporal features, obtain the global covariance matrix through stabilization processing, and map it to the symmetric matrix space using a log-Euclidean mapping to obtain global second-order features for characterizing the overall rhythmic structure.

[0025] Step 4: Construction of local second-order features The spatiotemporal embedding features are slidably sliced ​​along the time dimension according to a preset window length and step size. The covariance matrix is ​​calculated for each local window and logarithmic mapping is performed to obtain several local second-order features that reflect local instantaneous dynamics. Then, the second-order feature aggregation module is used to perform weighted or attention aggregation operations on multiple local representations to form stable local second-order feature representations, so as to fully capture the rhythmic change characteristics of short time periods.

[0026] Step 5: Second-order feature fusion and classification decoding A gating fusion strategy is used to integrate global and local second-order features in logarithmic Euclidean space, and the motion imagination category is output through a fully connected layer and a Softmax classifier.

[0027] Step 1: Data Preprocessing and Training Sample Construction Step 101: Random cropping Continuous EEG data Random cutting to obtain a fixed length Training clips: , in The starting point for cropping is the random sampling within the legal range.

[0028] Step 102: Set to zero for a random time period Randomly select a starting point from the clipped signal. And randomly generate a zero-length The data within the specified time range is zeroed out to simulate local signal loss. This step improves the model's robustness to local time loss, sudden noise, and data acquisition interruptions.

[0029] Step 103: Bandpass Filtering Perform on the data after setting it to zero Bandpass filtering removes irrelevant frequency bands: , in Let be the filtered output signal, and * denotes convolution. This represents the EEG signal after being zeroed out over a random time interval. This represents the impulse response of the bandpass filter.

[0030] Step 104: z-score channel normalization For each channel after bandpass filtering, calculate the mean and standard deviation over time: , The normalization process is as follows: .

[0031] Step 105: Training Sample Construction The normalized fragments are then formatted into the model input format to generate a training sample set. , in This represents the number of training samples obtained from the construction.

[0032] Step 2: Spatiotemporal Feature Extraction via Multi-Branch Graph Convolution Step 201: Construction of Input Features for Shared Graph Structure After preprocessing the EEG data, the normalized signal is represented as a matrix. ,in For the number of channels, This represents the number of time points. This step uses this matrix as the initial node features for the graph convolutional network to construct the input layer of the spatiotemporal network.

[0033] Since the multi-branch structure of this invention focuses on multi-timescale modeling, all branches share the same spatial topology, thereby ensuring the consistency of spatial features at different timescales.

[0034] Step 202: Extract spatially relevant features using shared graph convolution. To capture cross-electrode spatial correlation patterns in EEG signals, the input features are analyzed. Apply Chebyshev polynomial graph convolution. Graph convolution uses a shared adjacency structure. To avoid redundant calculations, spatial coding features common to all branches are generated. ChebConv can efficiently model high-order connectivity patterns between different channels, including long-range functional coupling between cortical regions.

[0035] , in For the normalized graph Laplacian matrix, For Chebyshev polynomials, For learnable parameters, The order of the polynomial

[0036] Step 203: Multi-scale temporal convolution modeling

[0037] Given that the spatial structure has been integrated, this invention uses three parallel branches. Convolutional modeling is performed at different time scales, with each branch using a temporal convolution kernel of different lengths. This allows for the extraction of time series patterns corresponding to short-term, medium-term, and long-term dynamic changes, respectively.

[0038] k = 15: Capture rapid changes, such as a sudden decrease in μ. k = 45: Capture moderate rhythms, such as the initial phase of ERD. k = 75: Captures slow, holistic beta rhythm trends.

[0039] This multi-scale modeling approach can simultaneously represent information from motion-imagined EEG at different time scales, thereby improving overall discrimination capabilities. , Features of completed spatial structure integration Conv1D(⋅) represents a one-dimensional convolution operation along the time dimension, with a kernel length of . .

[0040] Step 204: Feature normalization, activation and multi-branch fusion Batch normalization and ReLU activation are performed on the multi-scale convolution results to enhance training stability and eliminate feature scale differences caused by different convolution scales. , The temporal features are obtained by one-dimensional temporal convolution. BatchNorm(⋅) represents the batch normalization operation, and ReLU(⋅) is the linear rectified activation function.

[0041] Subsequently, the outputs of the three branches are concatenated along the feature dimension to form the final spatiotemporal embedding feature matrix. This serves as the unified input for subsequent global second-order and local second-order branches. .

[0042] Step 3: Global Second-Order Feature Construction Step 301: Learnable Positional Mapping (LPM) Since raw EEG time-series segments typically have a fixed sampling rate, but different time locations may correspond to different neural event windows, such as the ERD / ERS generation patterns before and after the motor imagery initiation point, this invention adds a learnable location mapping module (LPM) before global second-order modeling to enhance the model's sensitivity to temporal index information. This LPM generates a location embedding vector for each time point and adds it to the spatiotemporal embedding matrix. This enables the network to distinguish the physiological rhythm characteristics of the "front," "middle," and "back" segments.

[0043] The position code output by LPM is denoted as ,and Adding them along the same dimension, we get: , This is a learnable positional encoding matrix.

[0044] Step 302: Long-Range Temporal Modeling (LRTM) To capture rhythmic changes across longer time windows during motion visualization, such as slow μ-rhythmic suppression or β-bounce, this invention enhances the position-enhanced feature matrix. Long-range temporal modeling is performed using long convolutional kernels or dilated convolutions. Unlike the multi-scale convolutions in step 2, this step focuses on extracting global temporal dependencies, covering the entire sample length.

[0045] Convolution can be represented as follows: , in A larger value (such as 45–100) is usually chosen to cover the rhythmic changes of a complete 4-second window of motion visualization.

[0046] Step 303: Construction of the global covariance matrix (stabilized second-order statistics) To obtain the overall temporal statistical structure of the samples, this invention uses the feature matrix after long-range convolution. The covariance matrix is ​​calculated to characterize the second-order correlation between the feature channels. Direct calculation of the covariance is easily affected by noise and sample bias; therefore, this invention employs a covariance contraction estimation method to enhance stability.

[0047] Covariance is defined as follows: , To improve stability, a contraction term is introduced. The contraction covariance matrix is ​​obtained as follows: , in Used to control the degree of noise suppression.

[0048] Step 304: Log-Euclidean mapping (SPD → symmetric space) Since the covariance matrix belongs to a symmetric positive definite (SPD) space, its geometric structure is non-Euclidean, and direct linear operations on it would destroy geometric consistency. This invention stabilizes the covariance matrix... Mapped to the Log-Euclidean space, it is transformed into a symmetric matrix through matrix logarithm operations to maintain consistency with the structure generated by subsequent local branches.

[0049] The matrix logarithm is calculated as follows: , in That is, Global Second-Order Representation.

[0050] Step 4: Construction of local second-order features Step 401: Constructing a Local Sliding Window

[0051] The spatiotemporal embedding features obtained in step 2 Slice the data along the time dimension using a sliding window. Set the window length W and step size S to generate a series of local windows to capture short-term rhythmic features: , The spatiotemporal embedding feature matrix obtained in step 2, Let be the start time index of the i-th window, and W be the length of the sliding window. Sliding window operation Each segment has a stronger temporal resolution for the initiation, stabilization, and rebound phases of motion visualization.

[0052] Step 402: Construction of Local Fragment Covariance Matrix For each window Construct its local covariance matrix to capture the second-order statistical relationship between channels or features within a short time interval. , This matrix reflects the dynamic coupling relationships in short time segments, such as instantaneous ERD reduction, rhythm transformation, and local noise distribution.

[0053] Step 403: Local Covariance Stabilization Local windows are typically short and easily affected by noise and insufficient sample size. This invention introduces a covariance shrinkage term to improve stability.

[0054] , in Controlling the stabilization strength significantly enhances the robustness of second-order features within the local window, preventing overfitting of singular matrices to noise.

[0055] Step 404: Log-Euclidean Mapping To ensure that the local second-order structure can be integrated with the global second-order structure in the same space, this invention uses the Log-Euclidean mapping to transform the stabilized covariance into a symmetric matrix. , This step ensures that the second-order structure satisfies Euclidean operability and facilitates subsequent weighted fusion.

[0056] Step 405: Local Second-Order Feature Sequence Aggregation Since there are multiple local windows, this invention designs a local second-order aggregation module to integrate the second-order representations of multiple windows into a representative local structural feature. Aggregation methods can include averaging, weighted averaging, attention mechanisms, or one-dimensional convolutional aggregation.

[0057] The most basic mean aggregation is: , Nw represents the total number of sliding windows.

[0058] This aggregation structure can effectively compress window sequences while preserving local variation trends, and is a local representation in the dual-path second-order fusion of this invention.

[0059] Step 5: Second-order feature fusion and classification decoding Step 501: Second-order feature unification and co-domain mapping The global second-order features obtained in step 3 Compared with the local second-order features obtained in step 4 Both are derived from the covariance matrix and need to be uniformly mapped to the Log-Euclidean space to ensure that subsequent fusion operations are performed under the Euclidean structure, avoiding geometric distortion caused by direct weighting on the SPD manifold. This step uses matrix logarithmic operations to give the two second-order structures consistent algebraic properties and scale representations, facilitating subsequent fusion.

[0060] , .

[0061] Step 502: Log-domain second-order feature fusion based on gating coefficients To simultaneously preserve both the overall rhythmic structure (global features) and instantaneous dynamic changes (local features) in the final discriminative features, this invention proposes a learnable gating fusion strategy. A learnable coefficient β controls the fusion ratio of the two second-order features. β is automatically learned during training and can adaptively determine the importance distribution of global and local second-order structures under different experimental and subject conditions.

[0062] The fusion formula is as follows: , Here, β can be a scalar or a vector, allowing the model to flexibly adjust the contributions between different structures. This fusion mechanism is a key innovation of this invention, enabling the model to possess structured representation capabilities across multiple time scales.

[0063] Step 503: Classification and Decoding of Second-Order Fusion Features Fusion Matrix It belongs to the symmetric matrix form and is not suitable for direct use in classifiers. This invention expands it into feature vectors through vectorization operations. , , The input is then linearly mapped through a fully connected layer, and finally, a Softmax layer is used to obtain the probability output of the final motion imagery category. , W and b represent the weight matrix and bias term of the fully connected layer, respectively. This represents the linear response corresponding to category c; This step converts the second-order structural features into the final category output, thus decoding the motion image signal.

[0064] On the BCI IV-2a cross-session task, the overall performance of traditional shallow networks and classic deep models ranges from 72% to 79%. Among them, Shallow ConvNet, EEGNet, and LMDA-Net achieve accuracies of 73.70%, 72.27%, and 75.31%, respectively. ADF-CNN and EEG-Conformer show slight improvements, reaching 79.39% and 78.66%, respectively. The proposed SoMANet achieves the highest accuracy of 89.16%.

Claims

1. A method for decoding motor imagery from electroencephalogram (EEG) signals based on dual-path second-order feature fusion, characterized in that, Includes the following steps: (1) Data preprocessing and training sample construction: The original EEG signals were randomly cropped and zeroed out at random time intervals to enhance the robustness of the model to local signal loss. Then, bandpass filtering and channel normalization were performed to form training samples. (2) Multi-branch graph convolution spatiotemporal feature extraction: The preprocessed EEG samples are input into the multi-branch graph convolutional network, and multi-scale spatiotemporal embedding features are extracted through convolutional kernels at different time scales and fused to form a unified spatiotemporal feature representation; (3) Global second-order feature construction: Learnable positional encoding is added to the spatiotemporal embedded features to perform long-range time series modeling. Global second-order features are obtained through covariance estimation and log Euclidean mapping to characterize the overall rhythmic structure. (4) Construction of local second-order features: The spatiotemporal embedding features are sliced ​​according to the sliding window, the covariance matrix is ​​calculated for each local window and logarithmic mapping is performed, and local second-order features are formed through the aggregation module to capture short-term rhythm changes; (5) Second-order feature fusion and classification decoding: In the logarithmic Euclidean space, a gating fusion strategy is used to integrate global and local second-order features, and the motion imagination category is output through a fully connected layer and a Softmax classifier.

2. The method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion as described in claim 1, characterized in that, The specific operation process in step (1) is as follows: S101. Random cropping: Randomly cropping fixed-length segments from continuous EEG signals; S102. Random time period zeroing: Randomly select the starting point and length in the clipped signal to zero out the data; S103. Bandpass filter: Filters out irrelevant frequency bands; S104. Channel Normalization: Calculate the mean and standard deviation over the time dimension and perform Z-score normalization.

3. The method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion as described in claim 1, characterized in that, The specific operation process of step (2) is as follows: S201. Shared Graph Convolution: Extracting spatially relevant features using Chebyshev multinomial graph convolution; S202. Multi-scale temporal convolution: Short-time, medium-time and long-time temporal patterns are extracted in parallel using convolutional kernels of different lengths; S203. Feature Fusion: Perform batch normalization and activation function processing on the multi-branch outputs, and then concatenate them along the feature dimensions.

4. The method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion as described in claim 1, characterized in that, The specific operation process of step (3) is as follows: S301. Location-aware coding: Adding learnable location embeddings to spatiotemporal embedding features; S302. Long-range temporal modeling: Using large convolutional kernels to cover the temporal dependencies of the entire sample window; S303. Covariance Stabilization: Introducing a contraction term to enhance the robustness of the covariance matrix; S304. Log-Euclidean mapping: Maps the covariance matrix to a symmetric matrix space.

5. The method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion as described in claim 1, characterized in that, The specific operation process of step (4) is as follows: S401. Sliding window slice: Generates local segments along the time dimension with a preset window length and step size; S402. Local Covariance Calculation: Calculate the covariance matrix for each window and perform stabilization processing; S403. Logarithmic mapping: Transforms a local covariance matrix into a symmetric matrix; S404. Feature Aggregation: Aggregate multiple local representations through weighted averaging or attention mechanisms.

6. The method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion as described in claim 1, characterized in that, The second-order feature fusion in step (5) adopts a gated fusion strategy, which adaptively adjusts the fusion ratio of global and local second-order features through a learnable coefficient β. The fusion formula is as follows: ， Where β can be a scalar or a vector, G2 is the global second-order feature, and L2 is the local second-order feature.

7. The method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion as described in claim 1, characterized in that, The classification decoding in step (5) includes expanding the fused symmetric matrix into feature vectors, mapping them through a fully connected layer, and then outputting the class probabilities through Softmax.

8. The method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion as described in claim 1, characterized in that, The length of the temporal convolution kernel in a multi-branch graph convolutional network can be adaptively adjusted according to the sampling rate of the EEG signal.

9. A method for decoding motor imagery of EEG signals based on dual-path second-order feature fusion as described in claim 4 or 5, characterized in that, The shrinkage parameter λ in the covariance stabilization process ranges from 0 to 1 and is used to control the degree of noise suppression.

Images