Dynamic brain network clustering method and system based on nonlinear electroencephalogram signal feature fusion

Abstract

The application provides a dynamic brain network clustering method and system based on nonlinear electroencephalogram signal feature fusion, and belongs to the cross field of neuroscience and information technology. The specific process of the method is as follows: data acquisition and processing: the electroencephalogram signals of a subject are collected by using multiple electrodes; dynamic cognitive network construction: the multidimensional electroencephalogram signals of each frequency band are divided by using a sliding window, the phase lock-in values of two groups of different channel electroencephalogram signals of the same frequency band in the same time window are obtained, a dynamic functional brain network is constructed, and a multi-frequency band multi-channel dynamic functional connection matrix is obtained; dynamic network clustering: each connection matrix is regarded as an independent cluster, the dynamic network clustering is realized by merging the two clusters with the shortest distance to form a new cluster, and a non-complete binary tree is formed; for the non-complete binary tree, starting from the topmost cluster, the clusters are accessed from top to bottom, the clusters are merged and removed, so that the clusters are within a set range interval, and finally the dynamic brain network clustering is realized.

Description

Technical Field

[0001] This invention proposes a dynamic brain network clustering method and system based on the fusion of nonlinear electroencephalogram (EEG) signal features. It is applicable to the classification of brain network states during continuous tasks and belongs to the interdisciplinary field of neuroscience and information technology. Background Technology

[0002] Electroencephalogram (EEG) signals are crucial for brain network analysis. By recording the interactions between brain regions under different states, the working mechanisms of the brain can be effectively revealed. Extracting accurate and reliable EEG signals is fundamental to constructing brain networks. Given the high complexity and nonlinearity of EEG signals, traditional Fourier transform-based decomposition methods struggle to capture the dynamics of the nervous system, such as phase synchronization and chaotic phenomena. Consequently, the interactions between multiple brain regions are also nonlinear. Empirical mode decomposition (EMD) is often used for nonlinear signal extraction; however, it can cause mode aliasing.

[0003] Brain functional networks are divided into resting-state and task-oriented networks. Current brain network research largely focuses on static studies of resting-state or task-specific networks. Considering that higher-order cognitive activities often stem from the highly dynamic and complex nature of brain networks, some researchers use sliding window correlation to reveal the connectivity of resting-state networks in diseases, while others use recursive dynamic connectivity to study sub-networks in schizophrenia. The sliding window method is currently the mainstream approach in dynamic brain network research, but it faces the dilemma of choosing between window length and resolution. To accurately quantify the dynamic characteristics of neural rhythms while smoothing out noise effects, this invention proposes setting the window length to 5 to 10 times the corresponding neural rhythm cycle and the window overlap rate to 30% to avoid truncation effects caused by small overlap rates and state aliasing caused by large overlap rates.

[0004] Brain network clustering analysis is an important method for analyzing brain functional connectivity patterns. During cognition, the brain exhibits diverse connectivity patterns that demonstrate high stability and consistency. In dynamic functional connectivity studies of diseases, researchers have used K-means clustering to identify subtypes of Parkinson's disease patients. Other researchers have clustered dynamic brain networks during the onset of mental illness to distinguish specific connectivity patterns at different stages of the illness in patients with major depressive disorder and schizophrenia. However, traditional clustering methods suffer from the following dilemmas: k-means algorithms, which use Manchester or Euclidean distance as distance metrics, lack practical physiological support, have large errors, leading to uninterpretable results, and the "curse of dimensionality" drastically reduces algorithm efficiency as data dimensionality increases. On the other hand, spectral clustering and density-based clustering (DBSCAN) methods require defining matrix similarity, also facing the "curse of dimensionality." While current dimensionality reduction algorithms can significantly reduce data dimensionality, they easily lead to varying degrees of loss of functional connectivity, compromising data quality. Furthermore, in studies that focus on strong connections by setting thresholds, there is still no unified conclusion on how to control the threshold during the clustering process and eliminate spurious connections. Summary of the Invention

[0005] In view of this, the present invention provides a dynamic brain network clustering method and system based on nonlinear EEG signal feature fusion, which uses mask empirical mode decomposition technology to accurately extract EEG signal components of different frequency bands by adding mask signals.

[0006] The technical solution for implementing the present invention is as follows:

[0007] In a first aspect, the present invention provides a dynamic brain network clustering method based on the fusion of nonlinear electroencephalogram (EEG) signal features, the specific process of which is as follows:

[0008] Data acquisition and processing: Electroencephalogram (EEG) signals of the subjects were acquired using multiple electrodes; the acquired EEG signals were preprocessed; for the EEG signals acquired by each electrode, the frequency bands of the EEG signals were extracted using the mask empirical mode decomposition method.

[0009] Construction of a dynamic cognitive network: A sliding window is used to divide the multidimensional EEG signal into each frequency band. Two sets of EEG signals from different channels within the same frequency band are extracted within the same time window. The instantaneous phase time series of the EEG signals between each pair of channels are calculated, and the phase-locked values ​​are obtained. Using the channels from which the data is collected as network nodes and the phase-locked values ​​as the connection edges between nodes, a dynamic functional brain network is formed, resulting in a multi-band, multi-channel dynamic functional connectivity matrix.

[0010] Dynamic network clustering: Each connection matrix is ​​treated as an independent cluster. Dynamic network clustering is achieved by merging the two clusters with the shortest distance to form a new cluster, forming an incomplete binary tree. For the incomplete binary tree, starting from the top-level cluster, the clusters are visited from top to bottom to merge and remove clusters, so that the clusters are within a set range, and finally dynamic brain network clustering is achieved.

[0011] Furthermore, when the present invention uses multiple electrodes to collect the subject's electroencephalogram (EEG) signals, the subject is in a specific emotional state or is engaged in task learning.

[0012] Furthermore, the preprocessing described in this invention includes: eliminating errors caused by the measurement of the original reference electrode pair; performing filtering using a bandpass filter; and separating eye-tracking signal interference.

[0013] Furthermore, this invention employs a masked empirical mode decomposition method to extract the frequency bands of EEG signals of interest. The specific process is as follows:

[0014] 1. Select the mask signal m(t,θ) based on the frequency characteristics of the collected EEG signals;

[0015] 2. The mask signal is superimposed on the original EEG signal x(t) to form a new signal x. m (t)=x(t)+a·m(t,θ), where a is the amplitude coefficient of the mask signal;

[0016] 3. For the new signal x superimposed with the mask signal m (t) Applying empirical mode decomposition, a series of intrinsic mode functions (IMFs) are obtained. The specific process is as follows:

[0017] (1) Determine the new signal x m (t) All local maxima and minima are interpolated locally between the extreme points, and all points are connected to construct the upper envelope e. up (t) and lower envelope e down (t);

[0018] (2) Calculate the average value of the envelope n(t) = (e up (t)+e down (t)) / 2, calculate the deviation d(t) from the average value of the new signal. m (t)-n(t);

[0019] (3) Determine whether d(t) satisfies the conditions of IMF. If it does, treat d(t) as an IMF. If not, return to step (1) as a new signal until the conditions are met. Each IMF should satisfy two basic conditions: the number of extreme points and the number of zero crossings in the entire dataset must be equal or differ by at most one; the local average must be zero at any point.

[0020] 4. Repeat the above process until signal x is obtained. m (t) is completely decomposed into a series of IMFs and a residual term;

[0021] 5. The IMFs corresponding to each group after decomposition are averaged to obtain the final series of IMFs, which are the EEG signals of different frequency bands.

[0022] Furthermore, in the present invention, when using a sliding window to divide the multidimensional EEG signal into each frequency band, the sliding window is set to 5-10 times the maximum period of the selected frequency band signal, and the window overlap rate is set to 30%.

[0023] Furthermore, the phase-locked value of the present invention is calculated according to the following formula:

[0024]

[0025] Where, Δφ rel (t n ) indicates that signals X and Y are in time t n The phase difference at time, where N represents the total number of sample points within the signal duration.

[0026] Furthermore, the distance between the two clusters A and B in this invention is:

[0027]

[0028] Among them, A i With B j It is a sub-cluster of clusters A and B, n A With n B Indicates the number of subclusters. and Finger cluster A i With B j The number of matrices contained in it, N A and N B d(A) represents the number of matrices contained in clusters A and B. i B j ) represents sub-cluster A i With B j The distance between them.

[0029] Furthermore, in this invention, P and Q are two adjacent m×m connection matrices, and the connection strength similarity d between P and Q is calculated. strength for:

[0030]

[0031] in, This refers to the number of connections shared by P and Q, where p(i,j) and q(i,j) are the elements in the i-th row and j-th column of the corresponding P and Q matrices.

[0032] Furthermore, in this invention, P and Q are two adjacent m×m connection matrices, and the connection structure similarity d between P and Q is calculated. structure for:

[0033]

[0034] Among them, e p (i,j) and e q (i,j) represents the corresponding E P and E q The element in the i-th row and j-th column, E P and E q This indicates a connection unique to P and Q.

[0035] In a first aspect, the present invention provides a dynamic brain network clustering system based on nonlinear electroencephalogram (EEG) signal feature fusion, comprising:

[0036] The data acquisition and processing module contains multiple electrodes for acquiring the subject's electroencephalogram (EEG) signals; preprocesses the acquired EEG signals; and extracts the frequency bands of the EEG signals for each electrode using the mask empirical mode decomposition method.

[0037] The dynamic cognitive network construction module uses a sliding window to divide the multidimensional EEG signals of each frequency band, extracts two sets of EEG signals from different channels of the same frequency band within the same time window, calculates the instantaneous phase time series of the EEG signals between each pair of channels and obtains the phase-locked value; with the channels of the collected data as network nodes and the phase-locked value as the connection edge between nodes, a dynamic functional brain network is formed, and a multi-frequency multi-channel dynamic functional connection matrix is ​​obtained.

[0038] The dynamic network clustering module treats each connection matrix as an independent cluster and achieves dynamic network clustering by merging the two clusters with the shortest distance to form a new cluster, thus forming an incomplete binary tree. For the incomplete binary tree, starting from the top-level cluster, the module visits from top to bottom, merging and removing clusters to ensure that the clusters are within a set range, ultimately achieving dynamic brain network clustering. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 The flowchart shows a dynamic brain network clustering method based on nonlinear EEG signal feature fusion.

[0041] Figure 2 Here is a flowchart of the clustering algorithm;

[0042] Figure 3 Here is a flowchart for matrix distance calculation;

[0043] Figure 4 Here is a flowchart of the clustering post-processing algorithm;

[0044] Figure 5 The distances between each point within a cluster and the cluster center are given by two clustering methods: (a) k-means algorithm, (b) innovative algorithm.

[0045] Figure 6 The distances between matrices within the six clusters formed by the two clustering methods are: (a) k-means algorithm, (b) innovative algorithm.

[0046] Figure 7 The network topology of 6 clusters formed by two clustering methods is shown below: (a) Clustering coefficients of each cluster under the k-means algorithm; (b) Clustering coefficients of each cluster under the innovative algorithm; (c) Global efficiency of each cluster under the k-means algorithm; (d) Global efficiency of each cluster under the innovative algorithm. Detailed Implementation

[0047] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0048] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0049] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will 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 set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0050] To fully construct a dynamic brain network, this application proposes a dynamic brain network clustering method based on the fusion of nonlinear electroencephalogram (EEG) signal features. Steps one to three are data acquisition and preprocessing modules, steps four and five are dynamic cognitive network construction modules, and steps six to eight are dynamic network clustering modules. Figure 1 As shown, the detailed process is as follows:

[0051] Step 1: Collect the subject's electroencephalogram (EEG) signals using multiple electrodes at a sampling frequency of fs.

[0052] In this embodiment, EEG signal data of subjects under specific emotional states induced by stimuli such as pictures and video clips, or EEG signal data of subjects receiving different learning tasks, are collected at a sampling frequency of fs.

[0053] Step two involves preprocessing the acquired EEG signals to remove measurement errors and interference.

[0054] In this embodiment, the preprocessing is as follows: using the average value of whole-brain electrodes as a reference, the error caused by the original reference electrode is eliminated; a 50Hz notch filter is used to eliminate power frequency interference, and a bandpass filter from 0.5Hz to 150Hz is used to extract neural signals in the corresponding frequency band; finally, independent component analysis is used to compare the separated independent components with the recorded eye movement signals to separate eye movement signal interference.

[0055] Step 3: For the EEG signals acquired by each electrode, the masked empirical mode decomposition method is used to extract the frequency bands of interest. These typically include the following frequency bands: delta (0.5-4Hz), theta (4-8Hz), alpha (8-12Hz), beta (12-30Hz), low-gamma (30-75Hz), and high-gamma (75-150Hz). The specific steps are as follows:

[0056] 1. Select the mask signal m(t,θ) based on the frequency characteristics of the collected EEG signals.

[0057] 2. The mask signal is superimposed on the original EEG signal x(t) to form a new signal x. m (t)=x(t)+a·m(t,θ), where a is the amplitude coefficient of the mask signal and θ is the phase.

[0058] 3. For the new signal x superimposed with the mask signal m (t) Applying empirical mode decomposition, a series of intrinsic mode functions (IMFs) are obtained. The specific process is as follows:

[0059] (1) Determine the new signal x m(t) All local maxima and minima are interpolated locally between the extreme points, and all points are connected to construct the upper envelope e. up (t) and lower envelope e down (t).

[0060] (2) Calculate the average value of the envelope n(t) = (e up (t)+e down (t)) / 2, calculate the deviation d(t) from the average value of the new signal. m (t)-n(t).

[0061] (3) Determine whether d(t) satisfies the conditions of IMF. If it does, treat d(t) as an IMF. If not, return to step (1) as a new signal until the conditions are met. Each IMF should satisfy two basic conditions: the number of extreme points and the number of zero crossings in the entire dataset must be equal or differ by at most one; the local average must be zero at any point.

[0062] 4. Repeat the above process until signal x is obtained. m (t) is completely decomposed into a series of IMFs and a residual term.

[0063] 5. For all masked signals, the IMF corresponding to the sequence number of each group after decomposition is averaged to obtain the final series of IMFs, which are the EEG signals of different frequency bands.

[0064] Step four: Use a sliding window to divide the multidimensional EEG signal into each frequency band.

[0065] In this embodiment, to ensure that each window can accurately describe the corresponding frequency information, the window length is determined by the frequency band of interest and is set to 5-10 times the maximum period of the signal in that frequency band. The window overlap rate is set to 30% to avoid the truncation effect caused by a small overlap rate and the state aliasing caused by a large overlap rate.

[0066] Step 5: Extract two sets of EEG signals from different channels within the same frequency band at the same time window, calculate the instantaneous phase time sequence of the EEG signals between each pair of channels, and calculate the phase-locked value based on the EEG sequence according to the following formula:

[0067]

[0068] Phase-locked signal value (PLS) is a quantitative value used to measure phase synchronization, representing the degree of synchronization between two different EEG channels within a certain time period. Specifically, Δφ... rel (t n )=Δφ x (t n )-Δφ y (tn ), indicating that signals X and Y are in time t n The phase difference at time N is the total number of sample points during the signal duration. The value range of PLV is 0 to 1. The larger the value, the stronger the synchronization between the two signals. The minimum value of PLV is 0, which means that the two signals are completely out of sync. The maximum value of PLV is 1, which means that the two signals are completely in phase.

[0069] The data acquisition channels are selected as network nodes, and the phase-locked loop values ​​are selected as the connection edges between network nodes. Thus, a multi-band, multi-channel dynamic functional connection matrix is ​​constructed to form a dynamic functional brain network. Let the number of channels be m and the number of time windows be L. A dynamic functional brain network with m nodes and L time windows constitutes an m×m×L dynamic functional connection matrix.

[0070] Step 6: For any set of multi-channel dynamic functional connection matrices, perform dynamic network clustering based on the "new distance". The clustering algorithm flowchart is shown below. Figure 2 The specific steps are as follows:

[0071] 1. A dynamic functional brain network with m nodes and L time windows forms an m×m×L dynamic functional connectivity matrix. These L adjacent connectivity matrices are considered as L independent clusters, and the distance between any two clusters is calculated. The specific calculation method for the distance between two clusters A and B is as follows:

[0072]

[0073] Among them, A i With B j It is a subcluster of clusters A and B (i.e., a non-empty subset of a cluster; for the initial L independent clusters, its subcluster is itself), n A With n B Indicates the number of subclusters. and Finger cluster A i With B j The number of matrices contained in it, N A and N B d(A) represents the number of matrices contained in clusters A and B. i B j ) represents a sub-cluster A of clusters A and B. i With B j The distance between them, when n A =n B When = 1, both clusters A and B have only one sub-cluster, indicating that the sub-clusters of the two clusters have not been merged. Since the initial L clusters at the bottom layer are L adjacent connection matrices, d represents the distance between the two matrices. Figure 3 As shown, the calculation method is as follows:

[0074] (1) Let P and Q be two adjacent m×m connection matrices, and their corresponding binary matrices be P. * and Q * E P and E q E represents a connection unique to P and Q. p =P∩(1-Q) * E q =Q∩(1-P * ), C pq A connection representing the coexistence of P and Q: C pq =P∩Q

[0075] (2) Calculate the connection strength similarity d between P and Q. structure for:

[0076]

[0077] Where, N Cpq This refers to the number of connections shared by P and Q, where p(i,j) and q(i,j) are the elements in the i-th row and j-th column of the corresponding P and Q matrices. Connection strength similarity d strength The value range is [0,1]. The smaller the value, the more similar the connection strength between the two matrices.

[0078] (3) Calculate the connection structure similarity d between P and Q. structure for:

[0079]

[0080] Among them, e p (i,j) and e q (i,j) represents the corresponding E P and E q The element in the i-th row and j-th column has a connection structure similarity d. structure The value range is [0,1]. The smaller the value, the more similar the connection structure of the two matrices.

[0081] (4) The distance d(P,Q) between the two matrices is calculated by the following formula:

[0082]

[0083] The matrix distance d(P,Q) ranges from [0,1], and the smaller the value, the more similar the two matrices are.

[0084] 2. Merge the two clusters with the shortest distance to form a new cluster. The new cluster is regarded as the "root" of the two merged clusters. The two clusters are also called sub-clusters of the new cluster, and the distance between the clusters is updated.

[0085] 3. Determine if only one cluster remains (i.e., the final cluster). If so, the clustering ends; otherwise, return 1. After clustering, a non-complete binary tree is formed, with the final cluster at the top and the initial L independent clusters at the bottom.

[0086] Step seven: For the incomplete binary tree formed in step six, starting from the top-level cluster, visit the clusters from top to bottom, merging and removing clusters to ensure that the clusters are within the defined range. For example... Figure 4 As shown, the specific steps are as follows:

[0087] 1. Set the range of clusters to control the structure and size of clusters. Visit from top to bottom starting from the root node (final cluster), where C(i,j) represents the j-th cluster at level i. sp Let p represent a sub-cluster of C(i,j), where p = 1, 2.

[0088] 2. Sort the i-th level clusters from largest to smallest, starting from the first unvisited cluster: Determine if the size of the current cluster C(i,j) meets the requirements. If it is greater than the preset range, proceed to step 3. If it is less than the preset range, proceed to step 4. If it meets the range, traverse the next cluster and proceed to step 6.

[0089] 3. Visit subclusters C(i,j) of cluster C(i,j) s1 Proceed to step 2.

[0090] 4. Calculate the two subclusters C(i,j) of cluster C(i,j) and the cluster C(i,j+1) at the same level. s1 and C(i,j+1) s2 The distance between the two clusters is used to merge the cluster with the nearest sub-cluster. If the size of the merged cluster meets the requirements, the size of the sub-cluster is modified, the process returns to the upper level, and the merged cluster is removed. If the size of the cluster is larger than the range, the process proceeds to step 5.

[0091] 5. Visit the two subclusters C(i+1,k) of the nearest subcluster (let's call it C(i+1,k)). s1 and C(i+1,k) s2 Calculate the clusters C(i,j) and subclusters C(i+1,k) that need to be merged. s1 and C(i+1,k) s2 If the cluster size meets the requirements, adjust the sub-cluster size, return to the upper level, and remove the clusters that need to be merged. If the cluster size is larger than the range, proceed to step 5 again.

[0092] 6. Determine whether the clusters in the current layer have been traversed. If the traversal is complete, proceed to step 7.

[0093] 7. Determine if the current layer is the top layer. If it is, complete the clustering post-processing. If not, return to the previous layer and proceed to step 2.

[0094] After completing steps one through seven, the results of clustering a set of dynamic functional connectivity matrices can be obtained. These results can intuitively show the hierarchical relationship between each category and have strong classification accuracy.

[0095] This invention can be used in brain network research for cognitive behavioral tasks or the identification and classification of neurological diseases. In cognitive behavioral tasks, accurate emotion classification has significant applications in education and traffic safety. For example, in education, prolonged negative emotions can affect students' interpersonal relationships and learning efficiency; emotion classification can help identify these negative emotional states earlier. Similarly, anger experienced by drivers can easily lead to road rage, seriously endangering the lives of drivers and other road users; emotion classification can help detect this phenomenon in a timely manner and prevent accidents. In the identification and classification of neurological diseases, this method can provide a new and effective approach for cognitive assessment of patients with depression and Alzheimer's disease, offering potential for early diagnosis and treatment of neurological diseases and providing a new perspective for the development of early diagnostic tools for neurological diseases.

[0096] Example:

[0097] Taking a self-built dataset used to study the mechanism of emotion's influence on cognition as an example, in the experiment, five subjects performed a reading task after receiving positive emotional stimuli. 485 time windows in the alpha frequency band were clustered, and the results were compared with the commonly used k-means method based on Manchester distance in the same field. Figure 5-7 As shown, the following findings were made:

[0098] (1) The present invention can comprehensively consider the connection and connection strength of the brain's physiological structure and form clusters with distinct hierarchical structure and controllable size range through clustering.

[0099] (2) From the perspective of the distance from each point in the cluster to the cluster center, the method selected in this invention can effectively maintain similar structural distances and is better optimized in terms of cluster connection strength distance.

[0100] (3) From the perspective of the distance between each pair of points within the cluster, the clusters formed by the method selected in this invention are more compact.

[0101] (4) From the analysis of brain network topology, the cluster formed by the present invention has more compact points in terms of global efficiency and clustering coefficient, indicating that the functional structure within the cluster is more consistent.

[0102] The above results demonstrate that the algorithm and system proposed in this invention can be effectively applied to the classification of dynamic functional networks in cognitive behavior or neurological diseases, providing a reliable tool for revealing human cognitive activities and brain network research.

[0103] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic brain network clustering method based on nonlinear electroencephalogram (EEG) signal feature fusion, characterized in that, The specific process is as follows: Data acquisition and processing: Electroencephalogram (EEG) signals of the subjects were acquired using multiple electrodes; the acquired EEG signals were preprocessed; for the EEG signals acquired by each electrode, the frequency bands of the EEG signals were extracted using the mask empirical mode decomposition method. Construction of a dynamic cognitive network: A sliding window is used to divide the multidimensional EEG signal into each frequency band. Two sets of EEG signals from different channels within the same frequency band are extracted within the same time window. The instantaneous phase time series of the EEG signals between each pair of channels are calculated, and the phase-locked values ​​are obtained. Using the channels from which the data is collected as network nodes and the phase-locked values ​​as the connection edges between nodes, a dynamic functional brain network is formed, resulting in a multi-band, multi-channel dynamic functional connectivity matrix. Dynamic network clustering: Each connection matrix is ​​treated as an independent cluster. Dynamic network clustering is achieved by merging the two clusters with the shortest distance to form a new cluster, forming an incomplete binary tree. For the incomplete binary tree, starting from the top-level cluster, the clusters are visited from top to bottom to merge and remove clusters, so that the clusters are within a set range, and finally dynamic brain network clustering is achieved. The distance between two clusters A and B is: Among them, A i With B j It is a sub-cluster of clusters A and B, n A With n B Indicates the number of subclusters. and Finger cluster A i With B j The number of matrices contained in it, N A and N B d(A) represents the number of matrices contained in clusters A and B. i B j ) represents sub-cluster A i With B j The distance between them, d(A) i B j Based on A i With B j The distance between the matrices contained in the matrix is ​​obtained by calculation. The distance d(P,Q) between two matrices P and Q is calculated by the following formula: Where, d strength Let d be the connection strength similarity between matrices P and Q. structure The similarity of the connection structures between P and Q.

2. The dynamic brain network clustering method based on nonlinear EEG signal feature fusion according to claim 1, characterized in that, When collecting EEG signals from subjects using multiple electrodes, the subjects are in a specific emotional state or are engaged in task learning.

3. The dynamic brain network clustering method based on nonlinear EEG signal feature fusion according to claim 1, characterized in that, The preprocessing includes: eliminating errors caused by the measurement of the original reference electrode pair; filtering using a bandpass filter; and separating eye-tracking signal interference.

4. The dynamic brain network clustering method based on nonlinear EEG signal feature fusion according to claim 1, characterized in that, The masked empirical mode decomposition method is used to extract the frequency bands of EEG signals of interest. The specific process is as follows:

1. Select the mask signal m(t,θ) based on the frequency characteristics of the collected EEG signals; 2. The mask signal is superimposed on the original EEG signal x(t) to form a new signal x. m (t)=x(t)+a·m(t,θ), where a is the amplitude coefficient of the mask signal; 3. For the new signal x superimposed with the mask signal m (t) Applying empirical mode decomposition, a series of intrinsic mode functions (IMFs) are obtained. The specific process is as follows: (1) Determine the new signal x m (t) All local maxima and minima are interpolated locally between the extreme points, and all points are connected to construct the upper envelope e. up (t) and lower envelope e down (t); (2) Calculate the average value of the envelope n(t) = (e up (t)+e down (t)) / 2, calculate the deviation d(t) from the average value of the new signal. m (t)-n(t); (3) Determine whether d(t) satisfies the conditions of IMF. If it does, treat d(t) as an IMF. If not, return to step (1) as a new signal until the conditions are met. Each IMF should satisfy two basic conditions: the number of extreme points and the number of zero crossings in the entire dataset must be equal or differ by at most one; the local average must be zero at any point.

4. Repeat the above process until signal x is obtained. m (t) is completely decomposed into a series of IMFs and a residual term; 5. The IMFs corresponding to each group after decomposition are averaged to obtain the final series of IMFs, which are the EEG signals of different frequency bands.

5. The dynamic brain network clustering method based on nonlinear EEG signal feature fusion according to claim 1, characterized in that, When using a sliding window to divide the multidimensional EEG signal into each frequency band, the sliding window is set to 5-10 times the maximum period of the selected frequency band signal, and the window overlap rate is set to 30%.

6. The dynamic brain network clustering method based on nonlinear EEG signal feature fusion according to claim 1, characterized in that, The phase-locked value is calculated according to the following formula: Where, Δφ rel (t n ) indicates that signals X and Y are in time t n The phase difference at time, where N represents the total number of sample points within the signal duration.

7. The dynamic brain network clustering method based on nonlinear EEG signal feature fusion according to claim 1, characterized in that, Let P and Q be two adjacent m×m connection matrices. Calculate the connection strength similarity d between P and Q. strength for: in, This refers to the number of connections shared by P and Q, where p(i,j) and q(i,j) are the elements in the i-th row and j-th column of the corresponding P and Q matrices.

8. The dynamic brain network clustering method based on nonlinear EEG signal feature fusion according to claim 1, characterized in that, Let P and Q be two adjacent m×m connection matrices. Calculate the connection structure similarity d between P and Q. structure for: Among them, e p (i,j) and e q (i,j) represents the corresponding E P and E q The element in the i-th row and j-th column, E P and E q This indicates a connection unique to P and Q.

9. A dynamic brain network clustering system based on nonlinear electroencephalogram (EEG) signal feature fusion, characterized in that, include: Data acquisition and processing module, dynamic cognitive network construction module, and dynamic network clustering module The data acquisition and processing module contains multiple electrodes for acquiring the subject's electroencephalogram (EEG) signals; The collected EEG signals are preprocessed; For the EEG signals acquired by each electrode, the frequency bands of the EEG signals are extracted using the mask empirical mode decomposition method; The dynamic cognitive network construction module uses a sliding window to divide the multidimensional EEG signals of each frequency band, extracts two sets of EEG signals from different channels of the same frequency band within the same time window, calculates the instantaneous phase time series of the EEG signals between each pair of channels and obtains the phase-locked value; with the channels of the collected data as network nodes and the phase-locked value as the connection edge between nodes, a dynamic functional brain network is formed, and a multi-frequency multi-channel dynamic functional connection matrix is ​​obtained. The dynamic network clustering module is used to treat each connection matrix as an independent cluster. Dynamic network clustering is achieved by merging the two clusters with the shortest distance to form a new cluster, forming an incomplete binary tree. For the incomplete binary tree, starting from the top-level cluster, the nodes are visited from top to bottom to merge and remove clusters, so that the clusters are within a set range, and finally dynamic brain network clustering is achieved. The distance between two clusters A and B is: Among them, A i With B j It is a sub-cluster of clusters A and B, n A With n B Indicates the number of subclusters. and Finger cluster A i With B j The number of matrices contained in it, N A and N B d(A) represents the number of matrices contained in clusters A and B. i B j ) represents sub-cluster A i With B j The distance between them, d(A) i B j Based on A i With B j The distance between the matrices contained in the matrix is ​​obtained by calculation. The distance d(P,Q) between two matrices P and Q is calculated by the following formula: Where, d strength Let d be the connection strength similarity between matrices P and Q. structure The similarity of the connection structures between P and Q.

Images