EEG (electroencephalogram) signal classification method based on task-oriented graph filtering and multi-scale convolution

Abstract

The invention belongs to the technical field of electroencephalogram signal processing and brain-computer interfaces, and aims at solving the problems that an existing method is insufficient in task related feature extraction and low in classification precision. The method comprises the following steps: collecting and preprocessing a multi-channel electroencephalogram signal; fusing the phase lock value, the amplitude square coherence and the Pearson's correlation coefficient to construct a multi-modal brain function connection matrix, and constructing a discriminative brain function connection diagram based on a task label; a graph Laplacian operator is used to transform an original signal, a task related mode is enhanced, and redundant information is suppressed; inputting the processed signal into a multi-scale time fusion convolution module, and adaptively extracting deep features of different time scales; and finally realizing classification through a full connection layer. According to the method, an end-to-end classification framework is formed through combination of multi-modal function connection analysis, task oriented graph signal enhancement and multi-scale deep feature learning, the brain cooperation mode is comprehensively captured, and the accuracy and robustness of electroencephalogram mode recognition are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of electroencephalogram signal processing and pattern recognition, and mainly relates to an electroencephalogram signal classification method based on task-oriented graph filtering and multi-scale convolution. BACKGROUND

[0002] Electroencephalogram (EEG) signals, as a direct reflection of brain neural activities, are important information carriers in the field of brain-computer interface. However, the signal has low signal-to-noise ratio, high dimension, and significant individual differences. How to stably and efficiently extract discriminative features related to specific tasks (such as motor imagery) from the signal has always been a technical difficulty in the field. Traditional methods such as common spatial patterns or blind source separation techniques lack pertinence in the processing procedure, and while filtering out noise, they may weaken the key task features. In recent years, graph signal processing methods based on fixed brain topology prior have been introduced, but their static graph models are difficult to adapt to the dynamic changes of brain functional connectivity under different task states, resulting in limited feature enhancement effect.

[0003] With the development of deep learning, convolutional neural networks and other models have been used for end-to-end feature learning, but they usually require a large amount of training data and fail to effectively incorporate brain functional connectivity and other neuroscience prior knowledge to guide the learning process, which poses the risk of overfitting and poor interpretability. The core deficiency common to these existing technologies is that neither the signal processing nor the feature learning stage is deeply integrated with the goal of the final classification task, and there is a lack of a mechanism that can adaptively enhance the discriminative information of specific classes driven by the task goal, thereby restricting the further improvement of the performance of electroencephalogram pattern classification. SUMMARY

[0004] To overcome the deficiencies of the prior art, the present application provides an electroencephalogram signal classification method and system based on task-oriented graph filtering and multi-scale convolution. The core idea of the method is to prioritize the prior information of the classification task to guide the construction of the graph structure, thereby achieving targeted feature enhancement. The method includes the following steps:

[0005] Step 1: Electroencephalogram signal acquisition and preprocessing: acquire multi-channel electroencephalogram signals and perform denoising and filtering preprocessing operations;

[0006] Step 2: Constructing a multi-modal functional connectivity matrix: based on the preprocessed electroencephalogram signals, calculate the phase lock value, amplitude squared coherence, and Pearson correlation coefficient between channels to construct a multi-modal functional connectivity matrix representing different dimensional dependency relationships;

[0007] Step 3: Task-oriented construction of brain functional connectivity graph: Based on the specific classification task category label, the training data is grouped by category, and a discriminative brain functional connectivity graph is constructed for each category using the multi-modal functional connectivity matrix obtained in step 2.

[0008] Step 4: Graph Laplacian transformation to enhance signal: Calculate the Laplacian matrix of the brain functional connectivity graph obtained in step 3, and left multiply the original electroencephalogram signal by the matrix to realize signal transformation based on the graph structure, so as to enhance the task-related mode and suppress redundant information.

[0009] Step 5: Multi-scale deep feature extraction: input the signal transformed in step 4 into a multi-scale time fusion convolutional neural network module to adaptively extract deep features at different time scales.

[0010] Step 6: Classification and recognition: input the deep features extracted in step 5 into the full connection layer and the classifier to complete the classification and recognition of the electroencephalogram pattern.

[0011] Further, the calculation formulas of the phase lock value, amplitude square coherence and Pearson correlation coefficient in step 2 are as follows: the phase lock value (Phase-Locking Value, PLV) is used to measure the strength of phase synchronization between two electrode channels. To calculate the phase lock value, first perform Hilbert transform on the electroencephalogram signal to extract the instantaneous phase information, and the calculation formula of the phase lock value (PLV) is defined as:

[0012]

[0013] In the formula, θ i (t) and θ j (t) are the instantaneous phases of channel i and channel j at time point t, and T is the data length. The value of PLV varies between 0 and 1, and the larger the value, the stronger the phase synchronization between channels.

[0014] The amplitude square coherence (Magnitude-Squared Coherence, MSC) is used to measure the coherence of different channels in the frequency domain. Through Fourier transform, the signal is converted to the frequency domain, and the power spectral density of the channel and the cross spectral density between channels at a given frequency are calculated, and the calculation formula of the amplitude square coherence (MSC) is defined as:

[0015]

[0016] In the formula, S ij (ω) is the cross spectral density of channel i and j at frequency ω, S ii (ω) and S jj(ω) represents the self-power spectral density of the two channels at frequency ω. The coherence coefficient varies between 0 and 1, with a value closer to 1 indicating a stronger linear relationship between the two signals at a specific frequency.

[0017] The Pearson Correlation Coefficient (PCC) measures the degree of correlation between two time series by calculating their covariance and standard deviation. The formula is defined as follows:

[0018]

[0019] In the formula, x i (t) and x j (t) represents the signal values ​​of the two channels at time t, respectively. and This represents the average value of the corresponding channel signal over a time period T. The PCC value varies between -1 and 1, and the closer its absolute value is to 1, the stronger the linear correlation between the channels.

[0020] In step 2, constructing the multimodal functional connectivity matrix specifically involves fusing three metrics: phase lock value, amplitude squared coherence, and Pearson correlation coefficient, to form a comprehensive functional connectivity matrix W, with the fusion formula being W = α·W PLV +β·W MSC +γ·W PCC , where α, β, γ are adjustable non-negative weighting coefficients.

[0021] The specific implementation of "task-oriented construction of discriminative brain functional connectivity maps" in step 3 includes: category division: grouping the EEG signal samples in the training dataset according to their respective categories based on the category labels of the specific classification task; template construction: for each category, aggregating and calculating the multimodal functional connectivity matrix of all samples under that category obtained in step 2 to generate an average brain functional connectivity map template representing that category.

[0022] In step 4, the Laplace matrix is ​​the random walk normalized Laplace matrix, and its calculation formula is L. rw =D -1 L = ID -1 W, where D is the degree matrix, W is the adjacency matrix of the brain functional connectivity map constructed in step 3, L = DW is the non-normalized Laplacian matrix, and I is the identity matrix.

[0023] In step 5, the multi-scale temporal fusion convolution module consists of multiple parallel one-dimensional temporal convolutional layers with different kernel sizes, used to extract features at different time scales and fuse the outputs of each parallel branch.

[0024] Step 6 specifically involves classification and recognition: inputting the deep features extracted in step 5 into the fully connected layer and the classifier to complete the classification and recognition of EEG patterns.

[0025] A brainwave signal classification system based on task-guided graph filtering and multi-scale convolution, used to implement the method described in claims 1 to 6, characterized in that it includes: a data acquisition and preprocessing module for acquiring and preprocessing multi-channel brainwave signals; a multimodal mapping module configured to: calculate phase lock value, amplitude squared coherence, and Pearson correlation coefficient to construct a multimodal functional connectivity matrix; and aggregate and generate a dedicated average brain functional connectivity map template for each category based on the category labels of the training data; a graph signal filtering and enhancement module configured to calculate the Laplacian matrix of the functional connectivity map template and transform the brainwave signals; a multi-scale feature extraction module including a multi-scale temporal fusion convolutional neural network for extracting deep features from the transformed signals; and a classification and recognition module including a fully connected layer and a classifier for completing classification based on the extracted features.

[0026] The advantages and beneficial effects of this invention are as follows:

[0027] This invention integrates classification objectives into the signal processing front end by constructing a dedicated graph template for each category, making the filtering process highly targeted and accurately enhancing the EEG patterns most relevant to the task. By combining multimodal brain functional connectivity analysis, graph signal enhancement, and deep learning feature learning, it achieves the mining of discriminative information from multiple angles and scales, significantly improving feature quality. Through adaptively constructed graph structures, it effectively adapts to differences in different subjects and task states, reducing dependence on fixed priors, thereby improving the model's generalization ability and robustness. Ultimately, it forms a complete end-to-end framework from signal processing to classification and recognition, characterized by clear structure and superior performance. Attached Figure Description

[0028] Figure 1 It is a brain signal classification method based on task-guided graph filtering and multi-scale convolution. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0030] Step 1: EEG signal acquisition and preprocessing: Acquire multi-channel EEG signals and perform noise reduction and filtering preprocessing operations;

[0031] Step 2: Construct a multimodal functional connectivity matrix: Based on the preprocessed EEG signals, calculate the phase lock value, amplitude squared coherence and Pearson correlation coefficient between channels to construct a multimodal functional connectivity matrix that represents the dependencies in different dimensions.

[0032] Step 3: Task-oriented construction of brain functional connectivity maps: Based on specific classification task category labels, the training data are grouped by category, and a discriminative brain functional connectivity map is constructed for each category using the multimodal functional connectivity matrix obtained in Step 2.

[0033] Step 4: Graph Laplace Transform to Enhance Signal: Calculate the Laplace matrix of the brain functional connectivity graph obtained in Step 3, and multiply the matrix on the left by the original EEG signal to realize graph-based signal transformation, thereby enhancing task-related patterns and suppressing redundant information;

[0034] Step 5: Multi-scale deep feature extraction: Input the signal transformed in step 4 into the multi-scale temporal fusion convolutional neural network module to adaptively extract deep features at different time scales;

[0035] Step 6: Classification and Recognition: Input the deep features extracted in Step 5 into the fully connected layer and the classifier to complete the classification and recognition of EEG patterns.

Claims

1. A method for classifying electroencephalogram (EEG) signals based on task-guided graph filtering and multi-scale convolution, characterized in that, Includes the following steps: Step 1: EEG signal acquisition and preprocessing: Acquire multi-channel EEG signals and perform noise reduction and filtering preprocessing operations; Step 2: Construct a multimodal functional connectivity matrix: Based on the preprocessed EEG signals, calculate the phase lock value, amplitude squared coherence and Pearson correlation coefficient between channels to construct a multimodal functional connectivity matrix that represents the dependencies in different dimensions. Step 3: Task-oriented construction of brain functional connectivity maps: Based on specific classification task category labels, the training data are grouped by category, and a discriminative brain functional connectivity map is constructed for each category using the multimodal functional connectivity matrix obtained in Step 2. Step 4: Graph Laplace Transform to Enhance Signal: Calculate the Laplace matrix of the brain functional connectivity graph obtained in Step 3, and multiply the matrix on the left by the original EEG signal to realize graph-based signal transformation, thereby enhancing task-related patterns and suppressing redundant information; Step 5: Multi-scale deep feature extraction: Input the signal transformed in step 4 into the multi-scale temporal fusion convolutional neural network module to adaptively extract deep features at different time scales; Step 6: Classification and Recognition: Input the deep features extracted in Step 5 into the fully connected layer and the classifier to complete the classification and recognition of EEG patterns.

2. The method according to claim 1, characterized in that, In step 2, the calculation formulas for the phase-locking value, amplitude squared coherence, and Pearson correlation coefficient are as follows: The phase-locking value (PLV) is used to measure the strength of phase synchronization between the two electrode channels. To calculate the phase-locking value, the EEG signal is first subjected to Hilbert transform to extract instantaneous phase information. The calculation formula for PLV is defined as follows: The calculation formula for the phase-locking value (PLV) is: In the formula, θ i (t) and θ j (t) represents the instantaneous phase of channel i and channel j at time points, respectively, and T is the data length. The value of PLV varies between 0 and 1, and the larger the value, the stronger the phase synchronization between channels. The Magnitude-Squared Coherence (MSC) is used to measure the coherence of different channels in the frequency domain. The signal is converted to the frequency domain using Fourier transform, and the power spectral density of each channel and the cross-spectral density between channels are calculated at a given frequency. The formula for calculating MSC is defined as follows: In the formula, S ij (ω) is the cross spectral density of channels i and j at frequency ω, S ii (ω) and S jj (ω) represents the self-power spectral density of the two channels at frequency ω. The coherence coefficient varies between 0 and 1, with a value closer to 1 indicating a stronger linear relationship between the two signals at a specific frequency. The Pearson Correlation Coefficient (PCC) measures the degree of correlation between two time series by calculating their covariance and standard deviation. The formula is defined as follows: In the formula, x i (t) and x j (t) represents the signal values ​​of the two channels at time t, respectively. and This represents the average value of the corresponding channel signal over a time period T. The PCC value varies between -1 and 1, and the closer its absolute value is to 1, the stronger the linear correlation between the channels.

3. The method according to claim 2, characterized in that, In step 2, constructing the multimodal functional connectivity matrix specifically involves fusing three metrics: phase lock value, amplitude squared coherence, and Pearson correlation coefficient, to form a comprehensive functional connectivity matrix W, with the fusion formula being W = α·W PLV +β·W MSC +γ·W PCC , where α, β, γ are adjustable non-negative weighting coefficients.

4. The method according to claim 1, characterized in that, The specific implementation of "task-oriented construction of discriminative brain functional connectivity maps" in step 3 includes: category division: grouping the EEG signal samples in the training dataset according to their respective categories based on the category labels of the specific classification task; template construction: for each category, aggregating and calculating the multimodal functional connectivity matrix of all samples in that category obtained in step 2 to generate an average brain functional connectivity map template representing that category.

5. The method according to claim 1, characterized in that, In step 4, the Laplace matrix is ​​the random walk normalized Laplace matrix, and its calculation formula is L. rw =D -1 L = ID -1 W, where D is the degree matrix, W is the adjacency matrix of the brain functional connectivity map constructed in step 3, L = DW is the non-normalized Laplacian matrix, and I is the identity matrix.

6. The method according to claim 1, characterized in that, In step 5, the multi-scale temporal fusion convolution module consists of multiple parallel one-dimensional temporal convolutional layers with different kernel sizes, used to extract features at different time scales and fuse the outputs of each parallel branch.

7. A brainwave signal classification system based on task-guided graph filtering and multi-scale convolution, used to implement the method described in claims 1 to 6, characterized in that, include: The data acquisition and preprocessing module is used to acquire and preprocess multi-channel EEG signals; The multimodal mapping module is configured to: calculate phase lock value, amplitude squared coherence and Pearson correlation coefficient to construct a multimodal functional connectivity matrix; and aggregate and generate a unique average brain functional connectivity map template for each category based on the category labels of the training data. The graph signal filtering and enhancement module is configured to calculate the Laplacian matrix of the functional connectivity graph template and transform the EEG signal; The multi-scale feature extraction module includes a multi-scale temporal fusion convolutional neural network, used to extract deep features from the transformed signal; The classification and recognition module, which includes a fully connected layer and a classifier, is used to perform classification based on the extracted features.

Images