Motor imagery electroencephalogram signal enhancement method based on graph attention

Abstract

The invention requests to protect a motor imagery electroencephalogram signal enhancement method based on graph attention. The method comprises the following steps: firstly, carrying out multi-stage preprocessing on an original electroencephalogram signal, including band-pass filtering and baseline drifting, and removing artifacts in combination with independent component analysis, then, extracting time information of each channel by utilizing a one-dimensional convolutional network so as to capture time sequence characteristics in a motor imagery process, on the basis, constructing a graph attention network, taking the electroencephalogram signal channels as nodes, and taking the electroencephalogram signal channels as the nodes; the method comprises the following steps of: dynamically modeling and optimizing the internal relation between channels, strengthening motor imagery related channels such as C3, C4 and Cz by combining priori knowledge, further introducing a time attention network to identify and enhance key time slices rich in discrimination information, and finally, forming a high-dimensional feature matrix by fusing space-enhanced and time-enhanced electroencephalogram features, so as to realize the recognition and enhancement of the motor imagery related channels. And inputting into a classifier for pattern recognition.

Description

Technical Field

[0001] The present invention belongs to a pattern recognition and signal processing method in a brain-computer interface, and is a data enhancement based on the combination of graph attention and prior knowledge. Background Art

[0002] As a cutting-edge technology system, brain-computer interface (BCI) aims to build a direct communication bridge between the human brain and computers or other external devices, enabling individuals to control external devices with their mental intentions. This feature not only greatly expands the scope of human-computer interaction, but also provides new possibilities for many application scenarios such as auxiliary medical care, neuroscience research, and virtual reality. Currently, a variety of invasive and non-invasive methods are available for recording physiological signals related to brain activity. Among these technologies, electroencephalography (EEG) has become one of the most advantageous options for current research and practical applications due to its non-invasive nature, affordability, and high portability.

[0003] Among brain-computer interface technologies, motor imagery paradigms are particularly widely used. They typically rely on multi-channel electroencephalography (EEG) to capture physiological phenomena related to brain activity. This technology is particularly well-suited for motor imagery tasks, decoding motor intentions in the brain to enable interaction with external devices. However, while increasing the number of EEG channels theoretically allows for the capture of more information, practical implementation presents a series of challenges. While multi-channel EEG provides rich data, it also introduces a significant amount of redundant information and noise. This redundant information not only increases the complexity of data processing but can also obscure the true EEG signal characteristics, thereby reducing the performance of the BCI system.

[0004] Therefore, a method is needed to automatically and efficiently extract and process EEG data features so that important EEG features can be enhanced to facilitate subsequent classification and recognition. Summary of the Invention

[0005] This invention aims to address the redundant information and noise caused by excessive channels in the field of motor imagery brain-computer interfaces. A method for enhancing motor imagery EEG signals based on graph attention is proposed. The technical solution of this invention is as follows:

[0006] A method for enhancing motor imagery EEG signals based on graph attention, comprising the following steps:

[0007] Step 1: After obtaining the raw EEG data of motor imagery, the raw signal is first preprocessed to ensure the signal quality;

[0008] Step 2: Extract temporal features from all channels of the preprocessed EEG data, keeping the original number of channels unchanged;

[0009] Step 3: Use the graph attention network to construct the EEG signal after step 2, with each channel as a node, capture node similarity, and strengthen the C3, C4 and Cz channels;

[0010] Step 4: The EEG data from step 2 is passed through the temporal attention network to enhance the relevant temporal features;

[0011] Step 5: The features enhanced in steps 3 and 4 are fused using the spatiotemporal feature fusion method of Riemannian manifold to form a new feature.

[0012] Step 6: Input the features from step 5 into the classifier to obtain the motor imagery classification results;

[0013] Furthermore, step 1: multi-stage preprocessing optimization is performed on the original EEG signal, power frequency interference and baseline drift are eliminated through a zero-phase Butterworth bandpass filter 1-40 Hz, and independent component analysis combined with a template matching strategy is used to separate eye movement artifacts and electromyographic noise, followed by spatial normalization processing based on a common average reference, and finally the continuous signal is segmented into time series segments containing 4 seconds of motor imagery based on event markers.

[0014] Furthermore, in step 2: targeting the intrinsic timing characteristics reflected in the motor imagery process, one-dimensional convolution is used to capture the temporal information in the original EEG signal of each channel by performing a one-dimensional convolution operation on each independent channel.

[0015]

[0016] Where k represents the convolution kernel.

[0017] The one-dimensional convolution includes a 50ms window short-term fine feature extraction kernel and a 1000ms window long-term rhythm feature extraction kernel, and maintains the topological invariance of the feature space through residual connection.

[0018] Furthermore, in step 3, an adaptive brain network optimization system under biophysical constraints is constructed, which constructs a graph of EEG signals with channels as nodes. On the one hand, the relationship between nodes is effectively captured by the graph attention network. This network aims to dynamically model and optimize the intrinsic connections between different channels in EEG signals through an adaptive learning mechanism. The node attention weight formula in the graph attention network is as follows:

[0019]

[0020] Where N represents the neighboring nodes of a node, Represents node v i and v j The similarity coefficient formula is as follows:

[0021] eij =LeakyReLu(a([Wh i ||Wh j ])),j∈N

[0022] Where W is a shared parameter, || is the concatenation of features. a is a single-layer feedforward neural network.

[0023] The update formula of the node after using multi-head attention is as follows:

[0024]

[0025] Where σ represents the activation function and K represents K attention heads.

[0026] On the other hand, the physiological prior knowledge of the brain regions related to motor imagery (C3 / C4 / Cz) is embedded in the self-attention weights, and the feature discriminability and neural interpretability are simultaneously improved through a multi-objective optimization function. The formula is as follows:

[0027]

[0028] Where loss is the cross entropy loss, y i The true label, z i Represents the predicted label of the valid channel.

[0029] Furthermore, step 4: a temporal attention network is used to identify the critical periods rich in discriminative information in step 2 and enhance the information. Specifically, a fully connected network is used to aggregate signals and a ReLU activation function is used to mask unimportant time.

[0030] Furthermore, in step 5, the two enhanced feature matrices are converted into symmetric positive definite matrices (SPD) respectively to construct a Riemannian manifold space, and the formula is as follows:

[0031]

[0032] Where X is the enhanced feature matrix, T is the time point, and I is the identity matrix.

[0033] Determine the reference point P in the common tangent space through the Riemann mean ref :

[0034]

[0035] Project it into the tangent space through the mapping formula:

[0036]

[0037] The final fusion formula is as follows:

[0038] Sfuse =αS G +βS T

[0039] α, β are weight coefficients.

[0040] Finally, the fused tangent space features are projected back to the Riemann manifold to obtain the feature fusion matrix:

[0041]

[0042] Furthermore, in step 6, a feature fusion matrix is ​​obtained through the previous steps, and the fused features are input into various classifiers for processing.

[0043] The advantages and beneficial effects of the present invention are as follows:

[0044] The present invention achieves feature enhancement and accurate discrimination of motor imagery EEG signals through a series of carefully designed preprocessing and feature enhancement steps. Subsequently, a one-dimensional convolution is used to extract intrinsic temporal features on each channel. A graph attention network, which structures each channel as a node, adaptively captures and models the complex relationships between nodes. Simultaneously, with the help of a multi-head attention mechanism and cross-entropy constraints on the key channels C3, C4, and Cz, the most discriminative information during motor imagery is effectively enhanced. Finally, a temporal attention network further identifies and enhances time periods rich in key information and concatenates the spatially and temporally enhanced EEG matrices, significantly improving the recognition accuracy and robustness of subsequent classifiers in motor imagery tasks. This fully demonstrates the application potential of the present invention in the field of brain-computer interfaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a method for enhancing motor imagery EEG signals based on graph attention provided by the present invention. DETAILED DESCRIPTION

[0046] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.

[0047] The technical solution of the present invention to solve the above technical problems is:

[0048] As shown in the figure, this embodiment provides a method for enhancing motor imagery EEG signals based on graph attention, including the following steps:

[0049] Step 1: Multi-stage preprocessing optimization was performed on the raw EEG signals. A zero-phase Butterworth bandpass filter (1-40 Hz) was used to eliminate power frequency interference and baseline drift. Independent component analysis combined with a template matching strategy was used to separate eye movement artifacts from myoelectric noise. Spatial normalization was then performed using a common average reference. Finally, the continuous signal was segmented into 4-second time segments containing motor imagery based on event markers.

[0050] Step 2: Aiming at the intrinsic timing characteristics reflected in the motor imagery process, one-dimensional convolution is used to capture the temporal information in the original EEG signal of each channel by performing a one-dimensional convolution operation on each independent channel.

[0051]

[0052] Where k represents the convolution kernel.

[0053] The one-dimensional convolution includes a 50ms window short-term fine feature extraction kernel and a 1000ms window long-term rhythm feature extraction kernel, and maintains the topological invariance of the feature space through residual connection.

[0054] Step 3: Build an adaptive brain network optimization system under biophysical constraints. Graph EEG signals are graphed with channels as nodes. A graph attention network is used to effectively capture the relationships between nodes. This network aims to dynamically model and optimize the intrinsic connections between different channels in EEG signals through an adaptive learning mechanism. The node attention weight formula in the graph attention network is as follows:

[0055]

[0056] Where N represents the neighboring nodes of a node, Represents node v i and v j The similarity coefficient formula is as follows:

[0057] e ij =LeakyReLu(a([Wh i ||Wh j ])),j∈N

[0058] Where W is a shared parameter, || is the concatenation of features. a is a single-layer feedforward neural network.

[0059] The update formula of the node after using multi-head attention is as follows:

[0060]

[0061] Where σ represents the activation function and K represents K attention heads.

[0062] On the other hand, the physiological prior knowledge of the brain regions related to motor imagery (C3 / C4 / Cz) is embedded in the self-attention weights, and the feature discriminability and neural interpretability are simultaneously improved through a multi-objective optimization function. The formula is as follows:

[0063]

[0064] Where loss is the cross entropy loss, y i The true label, z i Represents the predicted label of the valid channel.

[0065] Step 4: A temporal attention network aims to identify the critical periods rich in discriminative information in step 2 and enhance the information. Specifically, a fully connected network is used to aggregate signals and the ReLU activation function is used to mask unimportant time.

[0066] Step 5: Convert the two enhanced feature matrices into symmetric positive definite matrices (SPD) respectively and construct the Riemannian manifold space. The formula is as follows:

[0067]

[0068] Where X is the enhanced feature matrix, T is the time point, and I is the identity matrix.

[0069] Determine the reference point P in the common tangent space through the Riemann mean ref :

[0070]

[0071] Project it into the tangent space through the mapping formula:

[0072]

[0073] The final fusion formula is as follows:

[0074] S fuse =αS G +βS T

[0075] α, β are weight coefficients.

[0076] Finally, the fused tangent space features are projected back to the Riemann manifold to obtain the feature fusion matrix:

[0077]

[0078] Step 6: Obtain a feature fusion matrix through the previous steps and input the fused features into various classifiers for processing.

[0079] The above embodiments should be understood as merely illustrating the present invention and not as limiting the scope of protection of the present invention. After reading the contents of the present invention, technicians may make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for enhancing EEG signals of motor imagery based on graph attention, characterized in that: The following steps are involved: Step 1: After obtaining the raw EEG data of motor imagery, the raw signal is first preprocessed to ensure the signal quality; Step 2: Extract temporal features from all channels of the preprocessed EEG data, keeping the original number of channels unchanged; Step 3: Use the graph attention network to construct the EEG signal after step 2, with each channel as a node, capture node similarity, and strengthen the C3, C4 and Cz channels; Step 4: The EEG data from step 2 is passed through the temporal attention network to enhance the relevant temporal features; Step 5: The features enhanced in steps 3 and 4 are fused using the spatiotemporal feature fusion method of Riemannian manifold to form a new feature. Step 6: Input the features of step 5 into the classifier to obtain the motor imagery classification results.

2. The method for enhancing motor imagery EEG signals based on graph attention according to claim 1, characterized in that: The step 1 is to perform multi-stage preprocessing optimization on the original EEG signal, eliminate power frequency interference and baseline drift through a zero-phase Butterworth bandpass filter 1-40 Hz, and use independent component analysis combined with a template matching strategy to separate eye movement artifacts and myoelectric noise. Then, spatial normalization is performed based on a common average reference, and finally the continuous signal is segmented into time series segments containing 4 seconds of motor imagery based on event markers.

3. The method for enhancing motor imagery EEG signals based on graph attention according to claim 1, characterized in that: Step 2: In view of the intrinsic timing characteristics reflected in the motor imagery process, one-dimensional convolution is used to capture the temporal information in the original EEG signal of each channel. By performing a one-dimensional convolution operation on each independent channel, Where k represents the convolution kernel, The one-dimensional convolution includes a 50ms window short-term fine feature extraction kernel and a 1000ms window long-term rhythm feature extraction kernel, and maintains the topological invariance of the feature space through residual connection.

4. The method for enhancing motor imagery EEG signals based on graph attention according to claim 1, characterized in that: Step 3: Construct an adaptive brain network optimization system under biophysical constraints, graphing EEG signals with channels as nodes. On the one hand, the graph attention network effectively captures the relationship between nodes. This network aims to dynamically model and optimize the intrinsic connections between different channels in EEG signals through an adaptive learning mechanism. The node attention weight formula in the graph attention network is as follows: Where N represents the neighboring nodes of a node, Represents node v i and v j The similarity coefficient formula is as follows: e ij =LeakyReLu(a([Wh i ||Wh j ])),j∈N Where W is a shared parameter, || is the concatenation of features. a is a single-layer feedforward neural network, The update formula of the node after using multi-head attention is as follows: Where σ represents the activation function, K represents K attention heads, On the other hand, the physiological prior knowledge of the brain regions related to motor imagery (C3 / C4 / Cz) is embedded in the self-attention weights, and the feature discriminability and neural interpretability are simultaneously improved through a multi-objective optimization function. The formula is as follows: Where loss is the cross entropy loss, y i The true label, z i Represents the predicted label of the valid channel.

5. The method for enhancing motor imagery EEG signals based on graph attention according to claim 1, characterized in that: The step 4: a temporal attention network, which aims to identify the critical period rich in discriminative information in step 2 and strengthen the information, specifically through a fully connected network to aggregate signals and ReLU activation function to mask unimportant time.

6. The method for enhancing motor imagery EEG signals based on graph attention according to claim 1, characterized in that: Step 5: Convert the two enhanced feature matrices into symmetric positive definite matrices (SPD) respectively to construct a Riemannian manifold space. The formula is as follows: Where X is the enhanced feature matrix, T is the time point, I is the identity matrix, Determine the reference point P in the common tangent space through the Riemann mean ref : Project it into the tangent space through the mapping formula: The final fusion formula is as follows: S fuse =αS G +βS T α, β are weight coefficients, Finally, the fused tangent space features are projected back to the Riemann manifold to obtain the feature fusion matrix:

7. The method for enhancing motor imagery EEG signals based on graph attention according to claim 1, characterized in that: Step 6: Obtain a feature fusion matrix through the previous steps. The fused features are input into various classifiers for processing.

Images