Population Classification Method Based on EEG Multi-Band Feature Fusion

Abstract

In an embodiment of the present disclosure, a population classification method based on electroencephalogram multi-band feature fusion is provided, belonging to the technical field of data recognition. Specifically, it includes: Step 1, after preprocessing the collected electroencephalogram data, extract different multiple bands respectively; Step 2, extract the frequency-domain features and time-domain features of each band; Step 3, group the electroencephalogram data of each band, perform multi-scale characterization on the frequency-domain features and time-domain features of each group, and use the group attention mechanism to perform feature fusion from within the group and between groups respectively to obtain multi-wave multi-scale fusion features; Step 4, repeat Steps 1 to 3 to obtain the multi-wave multi-scale fusion features of multiple target persons and obtain the potential graph structure of the target persons through an adaptive graph learning method, and then input the potential graph structure and the multi-wave multi-scale fusion features into a graph convolutional neural network to obtain a classification result. Through the solution of the present disclosure, the recognition efficiency and accuracy are improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the technical field of data recognition, and in particular to a population classification method based on the fusion of multi-band features of electroencephalogram (EEG). Background Art

[0002] Electroencephalogram (EEG) is an electrophysiological monitoring method for recording brain waves. Specifically, electrodes are placed on the scalp to record the voltage fluctuations generated by the ionic currents of brain neurons, and what is measured is the synchronous sum of the postsynaptic potentials when numerous pyramidal cells are excited. From the perspective of signal processing, EEG generates a representation of a potential, interacting, non-linear, non-stationary, and multivariate neural circuit. At the same time, since EEG is a method with high temporal resolution and non-invasive, it is often applied in the fields of biomedical intelligent Internet of Things (IoT) and brain-computer interface (BCI). At the same time, since EEG reflects the superposition of the electromagnetic fields generated by the interactions between cortical neurons at the macroscopic level. Therefore, the behaviors of potential neuron populations can all be indirectly studied through EEG.

[0003] EEG records the electroencephalographic activities of the subjects. The EEG brain waves are divided into 4 types, namely (delta (1 - 4HZ), theta (4 - 8HZ), alpha (8 - 12HZ), and beta (12 - 30HZ)). According to the research of modern electroencephalophysiological neuropsychology, different types of brain waves have their unique operation modes, providing characteristics of different functional states of the brain. Among them, the frequency and amplitude content of each type of brain wave will vary with factors such as the physical and mental states, age, and diseases of the subjects.

[0004] It can be seen that there is an urgent need for a population classification method based on the fusion of multi-band features of electroencephalogram that can automatically, efficiently, and accurately identify the types of electroencephalogram. Summary of the Invention

[0005] In view of this, the embodiments of the present disclosure provide a population classification method based on the fusion of multi-band features of electroencephalogram, which at least partially solves the problems of poor recognition efficiency and accuracy in the prior art.

[0006] The embodiments of the present disclosure provide a population classification method based on the fusion of multi-band features of electroencephalogram, including:

[0007] Step 1, after preprocessing the collected electroencephalogram data, extract different multiple bands respectively;

[0008] Step 2, extract the frequency-domain features and time-domain features of each band;

[0009] Step 3: Group the electroencephalogram data of each band, perform multi-scale characterization on the frequency-domain features and time-domain features of each group, and use the group attention mechanism to perform feature fusion within and between groups respectively to obtain multi-wave multi-scale fusion features;

[0010] Step 4: Repeat Step 1 to Step 3 to obtain the multi-wave multi-scale fusion features of multiple target persons and use the adaptive graph learning method to obtain the potential graph structure of the target persons. Then, input the potential graph structure and the multi-wave multi-scale fusion features into the graph convolutional neural network to obtain the classification result.

[0011] According to a specific implementation manner of the embodiment of the present disclosure, the preprocessing operation includes electrode positioning, noise elimination, and electrooculogram artifact removal. Finally, the electroencephalogram frequencies are divided into different bands in descending order.

[0012] According to a specific implementation manner of the embodiment of the present disclosure, the frequency-domain features include power spectral density and relative power spectral density, and the time-domain features include the mobility, complexity, and sample entropy of Hjorth parameters.

[0013] According to a specific implementation manner of the embodiment of the present disclosure, the formula for calculating the mobility of the Hjorth parameter is

[0014]

[0015] where s(t) represents the electroencephalogram time series of n channels, var(.) represents the variance operation, d represents the differential operation, and dt represents the differential with respect to time t;

[0016] The formula for calculating the complexity is

[0017]

[0018] Under the condition that the time series contains a finite number of sampling points, the formula for calculating the sample entropy is

[0019]

[0020] where m represents the embedding dimension, r represents the neighborhood radius, Cq(m,r) represents the probability that two sequences match m points under the similarity tolerance r, and Cq(m + 1,r) represents the probability that two sequences match m + 1 points.

[0021] According to a specific implementation manner of the embodiment of the present disclosure, Step 3 specifically includes:

[0022] Group the electroencephalogram data of each band. Under the guidance of the Group Lasso least angle regression algorithm, each group extracts features of different scales through multiple convolution kernels of different sizes, calculates the weights of different kernels, uses the L1 norm to constrain the multi-scale feature fusion within the group, and selects the optimal convolution kernel size for each group; for the multi-scale features fused in each group, use the L2 norm to constrain the fusion between different bands, and finally, through the sub-branch network, guide the learning of the multi-wave multi-scale attention weight feature map to obtain the multi-wave multi-scale fusion features.

[0023] According to a specific implementation manner of the embodiments of the present disclosure, the expression of the potential graph structure is

[0024] A ij = Sim(O i , O j ) = cos(W A O i , W A O j )

[0025] where W A ∈R U×U , W A is a self-learning weight matrix, U is the number of target persons, A ij is the weighted cosine similarity between target persons i and j, Sim(.) represents the similarity measure, O i represents the multi-wave multi-scale fusion feature value of target person i, and O j represents the multi-wave multi-scale fusion feature value of target person j.

[0026] The population classification scheme based on electroencephalogram multi-band feature fusion in the embodiments of the present disclosure includes: Step 1, after preprocessing the collected electroencephalogram data, extract different multiple bands respectively; Step 2, extract the frequency domain features and time domain features of each band; Step 3, group the electroencephalogram data of each band, perform multi-scale characterization on the frequency domain features and time domain features of each group, and use the group attention mechanism to perform feature fusion from within the group and between groups respectively to obtain multi-wave multi-scale fusion features; Step 4, repeat Steps 1 to 3 to obtain the multi-wave multi-scale fusion features of multiple target persons and obtain the potential graph structure of the target persons through the adaptive graph learning method, and then input the potential graph structure and the multi-wave multi-scale fusion features into the graph convolutional neural network to obtain the classification result.

[0027] The beneficial effects of the embodiments of the present disclosure are as follows: Through the scheme of the present disclosure, the electroencephalogram characteristics of different bands are fused and the hidden associations between target persons are found, and the classification result of the electroencephalogram data is obtained based on the features of different bands of the electroencephalogram data and the implicit connections of the subjects, improving the recognition efficiency and accuracy. Brief Description of the Drawings

[0028] To more clearly illustrate the technical solutions of the embodiments of the present disclosure, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present disclosure. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0029] Figure 1 It is a schematic flowchart of a population classification method based on electroencephalogram multi-band feature fusion provided by an embodiment of the present disclosure;

[0030] Figure 2 It is a schematic data processing flowchart of a population classification method based on electroencephalogram multi-band feature fusion provided by an embodiment of the present disclosure;

[0031] Figure 3 It is a schematic data processing flowchart of an SMR and GAM module provided by an embodiment of the present disclosure. Detailed implementation manners

[0032] The embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0033] The following uses specific specific examples to illustrate the implementation manners of the present disclosure. Those skilled in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in this specification. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all the embodiments. The present disclosure can also be implemented or applied through other different specific implementation manners. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present disclosure without creative efforts belong to the scope of protection of the present disclosure.

[0034] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is illustrative only. Based on the present disclosure, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, this device and / or this method can be implemented using other structures and / or functions in addition to one or more of the aspects described herein.

[0035] It should also be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present disclosure in a schematic manner. The diagrams only show the components related to the present disclosure, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and ratio of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0036] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0037] Although current research has widely applied EEG data to disease diagnosis, current methods all study the overall EEG data. Rarely are the EEG data considered in different frequency bands. For the methods that divide the frequency bands, they also tend to study a single frequency band, or simply linearly connect all the band features. The combined correlation between different bands is not considered, the characteristics of different EEG bands are not fused, and the implicit connections between different subjects are also ignored. Therefore, fully considering the characteristics of different EEG bands, performing multi-band fusion, and capturing the internal correlation map between subjects is a challenging task. On the one hand, these existing inventions rarely consider the relationship between different brain regions in different bands; on the other hand, these inventions generally manually design the patient graph adjacency matrix in a multi-stage framework, and rarely consider the higher-order relationships between patients.

[0038] The embodiments of the present disclosure provide a population classification method based on the fusion of multi-band features of electroencephalogram, and the method can be applied to the analysis process of neurological diseases in a medical scenario.

[0039] See Figure 1 , which is a schematic flowchart of a population classification method based on the fusion of multi-band features of electroencephalogram provided by the embodiments of the present disclosure. As Figure 1 and Figure 2 shown, the method mainly includes the following steps:

[0040] Step 1, after preprocessing the collected electroencephalogram data, extract different multiple bands respectively; further, the preprocessing operation includes electrode positioning, noise elimination, and removal of electrooculogram artifacts, and finally divide the electroencephalogram frequencies from high to low into different bands in sequence.

[0041] In specific implementation, the patient can be used as the target person. For the collected EEG data, the first step of the preprocessing process is to locate the electrode channels. This step uses the python-mne toolkit. According to the international 10-20 system montage map, there are a total of 21 electrode patches (Fp1, Fp2, F3, F4, C3, C4, P3, P4, F7, F8, T3, T4, T5, T6, Fz, Cz, Pz, A1, A2, O1, and O2). With the A1 and A2 earlobe electrodes as references, the physical positions of each brain channel are located, enabling the brain channels to have spatial information. The second step of the preprocessing process is filtering. This step uses band-pass filtering (1-55HZ) in the mne package to eliminate high-frequency noise and notch filtering (50HZ) to remove power interference, performing simple cleaning on the original EEG data. The third step of the preprocessing process is to use independent component analysis to remove artifacts. All electromyogram activities, eye movements, and other artifacts will be marked and discarded from the analysis. The fourth step of the preprocessing process is to use a band-pass filter to separate four types of brain waves, namely alpha waves (8-12HZ), beta waves (12-30HZ), theta waves (4-8HZ), and delta waves (1-4HZ), from the artifact-free EEG signals. Finally, the preprocessing process segments each type of brain wave into non-overlapping band segments with a duration of 2 seconds.

[0042] Step 2, extract the frequency-domain features and time-domain features of each band;

[0043] Optionally, the frequency-domain features include power spectral density and relative power spectral density, and the time-domain features include the mobility, complexity, and sample entropy of the Hjorth parameters.

[0044] Furthermore, the formula for calculating the mobility of the Hjorth parameter is

[0045]

[0046] where s(t) represents the electroencephalogram time series of n channels, var(.) represents the variance operation, d represents the differential operation, and dt represents the differential with respect to time t;

[0047] The formula for calculating the complexity is

[0048]

[0049] Under the condition that the time series contains a finite number of sampling points, the formula for calculating the sample entropy is

[0050]

[0051] Among them, m represents the embedding dimension, r represents the neighborhood radius, Cq(m,r) represents the probability that two sequences match m points under the similarity tolerance r, and Cq(m + 1,r) represents the probability that two sequences match m + 1 points.

[0052] In specific implementation, considering the complexity and length of EEG band signals, this method extracts high-order features based on the prior knowledge of brain science and cognitive science, represents the overall data with a small number of dimensional indicators, and conducts subsequent analysis.

[0053] First, perform time-domain analysis on each segment. Among them, the Hjorth parameters help to understand the dynamic characteristics of the signal, and the mobility and complexity of the Hjorth parameters can be used to measure and compare the characteristics of slow signal changes and fast-changing trends.

[0054] The Hjorth Mobility measures the change in slope, and the calculation formula is as follows:

[0055]

[0056] Among them, s(t) is the electroencephalogram time series of n channels, and var(.) is the variance operation.

[0057] The Hjorth Complexity measures how many standard slopes there are in an amplitude. This parameter compares the similarity between the original signal and a pure sine wave. If the signal is more similar, the value converges to 1. The calculation formula is as follows:

[0058]

[0059] Sample Entropy (SampEn) is a new and improved measurement method to evaluate the complexity and certain self-similarity of time series. It measures the complexity of time series by measuring the probability of generating new patterns in the signal. The calculation formula is as follows:

[0060]

[0061] Among them, m is the embedding dimension, r is the radius of the neighborhood, Cq(m,r) is the probability that two sequences match m points under the similarity tolerance r, and Cq(m + 1,r) is the probability that two sequences match m + 1 points.

[0062] Furthermore, Fourier transform is performed on each segment, and frequency domain analysis is carried out. The power spectral density (PSD) represents the signal power within the unit frequency band of different EEG signals and is one of the most commonly used EEG features. We use the most widely applied Welch periodogram method to calculate the PSD of the EEG time series of each channel in different frequency bands. The relative spectral density (RSD) represents the ratio of the PSD of a specific frequency band to the total band PSD, indicating the energy ratio distribution within the frequency band of different brain waves. Finally, the average value of the 5 features extracted from all segments of each target person is taken as the final feature of the target person.

[0063] Step 3: Group the EEG data of each frequency band, perform multi-scale characterization on the frequency domain features and time domain features of each group, and use the group attention mechanism to perform feature fusion within and between groups respectively to obtain multi-wave multi-scale fusion features.

[0064] Based on the above embodiments, the specific steps of step 3 include:

[0065] Group the EEG data of each frequency band. Under the guidance of the Group Lasso least angle regression algorithm, each group extracts features of different scales through multiple convolution kernels of different sizes, calculates the weights of different kernels, constrains the multi-scale feature fusion within the group with the L1 norm, and selects the optimal convolution kernel size for each group; for the multi-scale features fused in each group, the fusion between different frequency bands is constrained with the L2 norm, and finally, the learning of the multi-wave multi-scale attention weight feature map is guided through a sub-branch network to obtain multi-wave multi-scale fusion features.

[0066] Specifically, first, different frequency bands can be grouped. For multi-wave fusion, it is necessary to capture the correlation between different brain waves. As an effective method, the attention mechanism is widely used to explore the relationship between elements at different positions in sequential data. The attention mechanism can also be used to strengthen the expression of the most informative features while suppressing the expression of invalid features, enabling the model to adaptively focus on important regions in the context. Secondly, for various manual features of EEG, simply concatenating the features of different brain channels cannot learn deeper information. Clinically, there may be some common characteristics between brain regions that are physically close in space. Selecting a small visual field space may lead to the loss of complementary information between different brain regions, but paying attention to too many brain channels may also lead to learning interference information and affecting the performance of the model. Therefore, it is necessary to explore the brain region information within different visual field ranges. This method embeds an operator capable of multi-scale feature extraction in a convolutional neural network (CNN) to form a multi-scale pyramid convolutional structure, learning multi-scale features from large to small and integrating the feature information of brain regions within different ranges.

[0067] For multi-wave and multi-scale feature fusion, this method mainly considers from two aspects. The intra-group attention mechanism is used for multi-scale fusion, and the inter-group attention mechanism is used for multi-wave fusion. At the same time, the Group Lasso grouped least angle regression algorithm is used to enhance the prediction accuracy and interpretability of this module.

[0068] For each patient u, a multi-wave feature X u ∈R C×H×W is used as the input, where C is the number of bands, H is the brain channels, and W is the number of handcrafted features.

[0069] First, by splitting the multi-scale feature representation (SMR) module, grouped multi-scale feature extraction and fusion are achieved. The specific process of the SMR module is as Figure 3 shown. The multi-wave feature X u is split into S groups, denoted by , where For each group, M different-sized convolutional kernels are used to learn different-scale spatial information in parallel and independently:

[0070]

[0071] where, represents the original feature of the i-th group of patient u, represents the feature map extracted by the j-th convolutional kernel of patient u in the i-th group. The number of convolutional kernels can be selected from multiple numbers from 0 to 9, and the size of the convolutional kernel ranges from 5X5 to 21X21, representing fusing each feature and fusing each brain channel respectively. The relationship between the convolutional kernel k j and the number of convolutional kernels M is:

[0072] k j = 2×(j + 2)+1, j = 0, 1, …, M - 1

[0073] The multiple different-scale feature maps learned within each group are used to form an intra-group attention mechanism using the Group Lasso-based group attention (GAM) module to obtain the attention weights for each kernel

[0074] The main process of the GAM module is as Figure 2 shown. This module consists of two parts: squeezing and excitation. Among them, squeezing generates channel statistical information through an adaptive global average pooling layer, which is used to embed global spatial information into the channel descriptor to encode global information:

[0075]

[0076] Excitation is used to adaptively recalibrate the relationship in the channel direction to obtain the attention weights for different channels:

[0077]

[0078] Among them, σ represents the ReLU activation function, represents the fully connected (FC) layer. Through two fully connected layers, the linear information between channels can be combined more effectively, which helps the interaction of information with high and low channel dimensions.

[0079] Multiply the group kernel attention weights and the multi-scale feature map and then take the average to obtain the multi-scale fusion feature of each group Concatenate them to obtain the multi-scale fusion feature maps of different groups:

[0080]

[0081] Use the GAM module to extract the attention weight information between different groups from the multi-scale fusion feature map, form the inter-group attention mechanism, and obtain the attention weight vector P of each group u .

[0082]

[0083] Among them, is the connection operator, is 's attention value.

[0084] Finally, multiply the attention weight vector of each group by the corresponding multi-scale fusion feature map and then concatenate them to obtain a refined feature map with richer multi-wave multi-scale feature information.

[0085]

[0086]

[0087] Among them, ⊙ represents matrix multiplication.

[0088] To guide the learning of intra-group and inter-group attention weights, this method adopts the Group Lasso grouped least angle regression algorithm. First, use the L1 norm for the attention matrix within each group to ensure intra-group sparsity, and then use the L2 norm for the inter-group attention matrix to ensure inter-group smoothness:

[0089] L gl = L ib + L ob

[0090] Step 4: Repeat Steps 1 to 3 to obtain the multi-wave and multi-scale fusion features of multiple target persons, and use the adaptive graph learning method to obtain the potential graph structure of the target persons. Then, input the potential graph structure and the multi-wave and multi-scale fusion features into the graph convolutional neural network to obtain the classification result.

[0091] Based on the above embodiments, the expression of the potential graph structure is

[0092] A ij =Sim(O i ,O j )=cos(W A O i ,W A O j )

[0093] where W A ∈R U×U , W A is a self-learning weight matrix, U is the number of target persons, A ij is the weighted cosine similarity between target persons i and j, Sim(.) represents the similarity measure, O i represents the multi-wave and multi-scale fusion feature value of target person i, and O j represents the multi-wave and multi-scale fusion feature value of target person j.

[0094] In specific implementation, considering the correlation between target persons, a simple but effective learnable metric function is proposed by learning the similarity measure of nodes to obtain the potential graph structure between target persons:

[0095] A ij =Sim(O i ,O j )=cos(W A O i ,W A O j ),

[0096] W A ∈R U×U , W A is a self-learning weight matrix, U is the number of target persons, A ij is the weighted cosine similarity between target persons i and j, Sim(.) represents the similarity measure, O i represents the multi-wave and multi-scale fusion feature value of target person i, and O j represents the multi-wave and multi-scale fusion feature value of target person j. Usually, the adjacency matrix in real life is generally non-negative. Therefore, A is scaled to the interval [0,1] through the sigmoid function. And the learned A is a symmetric matrix, meeting the expectation of the real patient group graph.

[0097] Meanwhile, the graph convolutional neural network is very sensitive to the graph structure, and the learned graph is closely related to the performance correctly recognized by the graph convolution. Therefore, the learned graph should have the characteristics of sparsity, connectivity, and smoothness. To this end, three loss functions are designed as constraints to assist in adaptive graph learning. The formula for the sparsity constraint is as follows:

[0098]

[0099] To avoid all elements of the learned A being 0, the formula for the connectivity constraint is as follows:

[0100]

[0101] Referring to the mean filtering method, smoothing is directly performed in the spatial domain, and the formula for the smoothness constraint is as follows:

[0102]

[0103] Among them, the graph signal set is {O1, O2,... O U}. The overall constraint formula for adaptive graph learning is as follows:

[0104] L g (A, O) = αL smooth (A, O) + βL con (A) + γL sparse (A)

[0105] Among them, α, β, and γ are three hyperparameters for balancing the losses.

[0106] Taking the learned adjacency matrix A and the feature map signal O of multi-wave and multi-scale fusion as the input of the graph convolution, information aggregation of the fusion features is performed by modeling the adjacency relationship between the target persons, and the target persons are predicted in an inductive manner:

[0107]

[0108] Compared with the ordinary multi-layer perceptron (MLP), the graph convolutional network is more suitable for obtaining the connections between patients and obtaining high-order representations. In the present invention, the three modules are trained together, and the total loss function during training is as follows:

[0109] L Total = L GCN + L sup + L g (A, O) + L gl

[0110] Among them, L supTo stabilize the learning process, the refined feature map with multi-wave and multi-scale feature information obtained is input into a two-layer MLP for prediction to assist in deep feature fusion. L sup and L GCN both belong to the cross-entropy loss. L g (A, O) is the defined graph loss. L gl is the loss that guides intra-group and inter-group attention learning.

[0111] The population classification method based on electroencephalogram multi-band feature fusion provided in this embodiment groups the bands, uses multiple convolution kernels of different sizes for each group of original features to extract feature representations of different scales, and fuses the spatial information between multi-scale brain electrodes; designs a group attention method, inputs the extracted multi-scale electroencephalogram spatial information into the GAM module to form an intra-group attention mechanism to guide the fusion of multi-scale information within the same group, and integrates the associations between brain regions within different ranges. Cascades the multi-scale features after fusion of each group and inputs them into the GAM module to form an inter-group attention mechanism, extracts the band attention between groups, establishes long-term band dependencies, and finally multiplies them with the multi-scale depth features to guide the fusion between different bands. At the same time, uses the Group Lasso least angle regression algorithm to guide the feature fusion between groups and within groups; designs an adaptive graph learning method to obtain a potential graph structure to flexibly match the graph convolution prediction task and reveals the association relationships between target persons.

[0112] The units involved in the embodiments described in this disclosure can be implemented in software or in hardware.

[0113] It should be understood that the various parts of this disclosure can be implemented by hardware, software, firmware, or a combination thereof.

[0114] As described above, the above are only the specific embodiments of this disclosure, but the protection scope of this disclosure is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by this disclosure should be covered by the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be subject to the protection scope of the claims.

Claims

1. A population classification method based on the fusion of multi-band features of electroencephalogram, characterized in that, Including: Step 1: After preprocessing the collected electroencephalogram (EEG) data, extract multiple different bands respectively. Step 2: Extract the frequency-domain features and time-domain features of each band. Among them, the frequency-domain features include power spectral density and relative power spectral density, and the time-domain features include the mobility, complexity, and sample entropy of Hjorth parameters. Step 3: Group the EEG data of each band, perform multi-scale characterization on the frequency-domain features and time-domain features of each group, and use the group attention mechanism to perform feature fusion within and between groups respectively to obtain multi-wave multi-scale fusion features. The specific content of Step 3 includes: Group the EEG data of each band. Under the guidance of the Group Lasso least angle regression algorithm, each group extracts features of different scales through multiple convolution kernels of different sizes, calculates the weights of different kernels, constrains the multi-scale feature fusion within the group with the L1 norm, and selects the optimal convolution kernel size for each group. For the multi-scale features fused in each group, constrain the fusion between different bands with the L2 norm, and finally, through the sub-branch network, guide the learning of the multi-wave multi-scale attention weight feature map to obtain multi-wave multi-scale fusion features. Step 4: Repeat Step 1 to Step 3 to obtain the multi-wave multi-scale fusion features of multiple target persons and obtain the potential graph structure of the target persons through the adaptive graph learning method. Then, input the potential graph structure and the multi-wave multi-scale fusion features into the graph convolutional neural network to obtain the classification result.

2. The method according to claim 1, wherein , The preprocessing operation includes electrode positioning, noise elimination, and removal of electrooculogram artifacts. Finally, the EEG frequencies are divided into different bands in descending order.

3. The method according to claim 1, wherein , The calculation formula for the mobility of the Hjorth parameter is Among them, \(s(t)\) represents the electroencephalogram time series of \(n\) channels, and \(var(.)\) represents the variance operation. represents the differential operation. represents the derivative with respect to time \(t\). The calculation formula for the complexity is Under the condition that the time series contains a finite number of sampling points, the calculation formula for the sample entropy is where m represents the embedding dimension and r represents the neighborhood radius, represents the probability that two sequences match m points under the similarity tolerance r, represents the probability that two sequences match m + 1 points.

4. The method according to claim 3, wherein , The expression of the potential graph structure is Among them, , is a self-learning weight matrix, U is the number of target personnel, is the weighted cosine similarity between target personnel i and j, represents the similarity measure, represents the multi-wave and multi-scale fusion eigenvalue of target personnel i, represents the multi-wave and multi-scale fusion eigenvalue of target personnel j.

Images