阅读说明：本技术 一种基于源定位和脑网络的自然动作脑电识别方法 (Natural action electroencephalogram recognition method based on source localization and brain network ) 是由 徐宝国 邓乐莹 汪逸飞 王欣 宋爱国 于 2020-11-11 设计创作，主要内容包括：本发明公开了一种基于源定位和脑网络的自然动作脑电识别方法,包括：(1)对自然动作进行多通道脑电测量；(2)对采集到的EEG信号进行预处理,提取MRCP、θ波、α波、β波和γ波；(3)确定信号的导程场矩阵,利用L1正则化约束求出源的初值解,利用逐次超松弛法迭代求解得到源定位结果；(4)以源为节点,采用短时滑动窗逐时间点计算每对源之间的PLV,构建脑网络；(5)逐时间点计算网络邻接矩阵和5个脑网络指标,将这些特征送入分类器进行训练和测试,对脑网络指标进行统计检验。本发明采用T-wMNE算法和逐次超松弛法结合的方式对传统源定位方法进行了改进,并以源为节点构建脑网络,有利于提高自然动作脑电的解码精度和揭示人体的神经运作机制。(The invention discloses a natural action electroencephalogram identification method based on source localization and brain network, which comprises the following steps: (1) performing multi-channel electroencephalogram measurement on natural actions; (2) preprocessing the collected EEG signals, and extracting MRCP, theta waves, alpha waves, beta waves and gamma waves; (3) determining a lead field matrix of the signal, solving an initial value solution of a source by utilizing L1 regularization constraint, and obtaining a source positioning result by utilizing successive ultra-relaxation iterative solution; (4) calculating PLV between each pair of sources by using the sources as nodes and adopting a short-time sliding window to construct a brain network; (5) and calculating the network adjacency matrix and 5 brain network indexes by time points, sending the characteristics into a classifier for training and testing, and carrying out statistical test on the brain network indexes. The invention improves the traditional source positioning method by combining the T-wMNE algorithm and the successive super relaxation method, constructs a brain network by taking the source as a node, and is beneficial to improving the decoding precision of the natural action electroencephalogram and revealing the neural operation mechanism of the human body.)

1. A natural motion electroencephalogram identification method based on source localization and brain network is characterized by comprising the following steps:

(1) performing multi-channel electroencephalogram measurement on natural actions;

(2) preprocessing the collected EEG signals, removing artifacts, and extracting MRCP, theta wave, alpha wave, beta wave and gamma wave;

(3) determining a lead field matrix of a signal, solving an initial value solution of a source by utilizing L1 regularization constraint, then iterating the initial value solution by utilizing a successive ultra-relaxation method, and taking the latest solution vector as a final estimation result of source positioning after iteration is finished;

(4) taking a source as a node, calculating phase synchronization PLV between each pair of sources at each time point by adopting a short-time sliding window, constructing an edge between the two sources when the PLV value is greater than a set threshold value, and taking a normalized value of the PLV value as the weight of the edge;

(5) calculating the characteristic path length, the clustering coefficient, the node average strength, the average betweenness, the efficiency and the network adjacency matrix at each time point, sending the characteristics into a classifier for training and testing, carrying out statistical test on the first 5 characteristics, and analyzing the difference of the characteristics corresponding to different actions on time or frequency bands.

2. The source localization and brain network-based natural motion electroencephalogram recognition method according to claim 1, wherein the step (2) specifically comprises the following sub-steps:

(a1) pre-filtering the acquired EEG signal;

(a2) eliminating data channels with abnormal kurtosis, performing spherical interpolation, and using the average value of four channels closest to the interpolated channel as the channel value;

(a3) finding and removing EOG and EMG components in the EEG using a blind source separation algorithm;

(a4) segment and baseline correction of EEG;

(a5) removing the test times with extreme values larger than 200 μ V and abnormal joint probability or abnormal kurtosis, wherein the threshold value of the two is 5 times of the standard deviation;

(a6) co-averaging the EEG references;

(a7) and respectively carrying out zero-phase Butterworth band-pass filtering on the heavily-referenced EEG at 0.3-3 Hz, 4-8 Hz, 8-13 Hz, 13-30 Hz and 30-45 Hz, and respectively extracting MRCP, theta wave, alpha wave, beta wave and gamma wave.

3. The natural motion electroencephalogram recognition method based on source localization and brain network according to claim 1, characterized in that in the step (3), the following sub-steps are included:

(b1) selecting a head model;

(b2) solving a positive problem to obtain a lead field matrix L;

(b3) determining a time point to be analyzed, and setting an iteration error epsilon and a maximum iteration number K;

(b4) and (3) solving the initial solution of the source vector by utilizing a T-wMNE algorithm:

wherein W ═ diag (| | l)₁||,||l₂||,…,||l_NI) is a weighting matrix, N is the number of sources, here equal to the number of electrodes, s_tRepresenting the source vector at a point in time t, v_tRepresents the electrode potential at a point in time t, λ being the regularization coefficient;

(b5) and (c) iterating the initial value solution obtained in the step (b4) by using a successive super relaxation method:

wherein s is_i,tRepresents the value of the ith source at time t, i ═ 1,2, …, N; v. of_j,tThe potential of the jth electrode at the time point t is represented, j is 1,2, …, N, ω is a relaxation factor, and k represents the number of iterations;

(b6) when | | | s_t ^(k+1)-s_t ^(k)Epsilon or k is less than or equal to | |>And K, finishing iteration, and taking the latest solution vector as a final estimation result of source positioning, otherwise, continuing the iteration.

4. The method for electroencephalogram recognition of natural actions based on source localization and brain network as claimed in claim 1, wherein in the step (4), the source vector of a single trial is first subjected to Hilbert transform at each time point to obtain the phase of the source vector at each time point, and then the PLV value of each pair of sources at each time point is calculated:

Δφ_ij＝φ_i-φ_j

wherein M is 1,2, …, M represents the mth trial;

when PLV_ijAnd when the value is greater than the threshold value, constructing an edge between the source i and the source j, and normalizing the value to be used as the weight of the edge:

5. the source localization and brain network based natural motion electroencephalogram recognition method according to claim 1, wherein in the step (5), the statistical test method is t-test, the classifier is an sLDA classifier, and the classifier is trained and tested by adopting 10 times of five-fold cross validation.

6. The natural motion electroencephalogram recognition method based on source localization and brain network of claim 3, wherein in the step (b5), the optimal relaxation factor makes | | | v after 10 iterations by selecting step length of 0.01 between (1,2)_t-Ls_tThe smallest ω of | is obtained.

Technical Field

The invention belongs to the field of biological signal processing, relates to an electroencephalogram signal identification method, in particular to a natural action electroencephalogram identification method based on source localization and a brain network, and provides a technical means for electroencephalogram decoding of natural actions.

Background

The brain-computer interface (BCI) is a means for directly communicating and controlling with the outside through electroencephalogram signals, and is also a research hotspot in the technical field of rehabilitation medical engineering and nerve engineering in recent years. In recent years, BCI-based rehabilitation training relies mainly on repeated imagination of basic motor tasks, such as performing a hand-holding glass operation by using repeated foot motor imagination as a control signal, which brings unnatural and uncoordinated operation experience to users. In order to get a better operational experience for the user, the imagined movements should be made as close as possible to the actions actually performed. However, the natural actions are complex, and there are many joints, such as holding, pinching, rotating, inserting and pulling, and many times these actions activate the same motor brain area, so that the traditional electroencephalogram identification method cannot achieve a good distinguishing effect on the natural actions.

Brain electrical source localization and brain network analysis are one of the research hotspots in the field of brain-computer interfaces in recent years. The source localization includes source feature reconstruction and source location localization, and the brain network can construct functional or causal connections between nodes. The research on the signals of natural motion on the source space is helpful to overcome the volume conduction effect, thereby improving the decoding precision, and the brain network constructed by taking the source as the object is helpful to reveal the nerve operation mechanism of the human body.

Disclosure of Invention

In order to solve the problems, the invention discloses a natural action electroencephalogram identification method based on source localization and a brain network.

In order to achieve the above purpose, the invention adopts the following technical scheme: a natural motion electroencephalogram identification method based on source localization and brain network specifically comprises the following steps:

(1) performing multi-channel electroencephalogram measurement on natural actions;

(2) preprocessing the collected EEG signals, removing artifacts, and extracting MRCP, theta wave, alpha wave, beta wave and gamma wave;

(3) determining a lead field matrix of a signal, solving an initial value solution of a source by utilizing L1 regularization constraint, then iterating the initial value solution by utilizing a successive ultra-relaxation method, and taking the latest solution vector as a final estimation result of source positioning after iteration is finished;

(4) taking a source as a node, calculating phase synchronization PLV between each pair of sources at each time point by adopting a short-time sliding window, constructing an edge between the two sources when the PLV value is greater than a set threshold value, and taking a normalized value of the PLV value as the weight of the edge;

(5) calculating the characteristic path length, the clustering coefficient, the node average strength, the average betweenness, the efficiency and the network adjacency matrix at each time point, sending the characteristics into a classifier for training and testing, carrying out statistical test on the first 5 characteristics, and analyzing the difference of the characteristics corresponding to different actions on time or frequency bands.

In the step (2), the method comprises the following steps:

(a1) pre-filtering the acquired EEG signal;

(a2) eliminating data channels with abnormal kurtosis, performing spherical interpolation, and using the average value of four channels closest to the interpolated channel as the channel value;

(a3) finding and removing EOG and EMG components in the EEG using a blind source separation algorithm;

(a4) segment and baseline correction of EEG;

(a5) removing the test times with extreme values larger than 200 μ V and abnormal joint probability or abnormal kurtosis, wherein the threshold value of the two is 5 times of the standard deviation;

(a6) co-averaging the EEG references;

(a7) and respectively carrying out zero-phase Butterworth band-pass filtering on the heavily-referenced EEG at 0.3-3 Hz, 4-8 Hz, 8-13 Hz, 13-30 Hz and 30-45 Hz, and respectively extracting MRCP, theta wave, alpha wave, beta wave and gamma wave.

In the step (3), the method comprises the following steps:

(b1) selecting a head model;

(b2) solving a positive problem to obtain a lead field matrix L;

(b3) determining a time point to be analyzed, and setting an iteration error epsilon and a maximum iteration number K;

(b4) and (3) solving the initial solution of the source vector by utilizing a T-wMNE algorithm:

wherein W ═ diag (| | l)₁||,||l₂||,…,||l_NI) is a weighting matrix, N is the number of sources, here equal to the number of electrodes, s_tRepresenting the source vector at a point in time t, v_tRepresents the electrode potential at a point in time t, λ being the regularization coefficient;

(b5) and (c) iterating the initial value solution obtained in the step (b4) by using a successive super relaxation method:

wherein s is_i,tRepresents the value of the ith source at time t, i ═ 1,2, …, N; v. of_j,tThe potential of the jth electrode at the time point t is represented, j is 1,2, …, N, ω is a relaxation factor, and k represents the number of iterations;

(b6) when | | | s_t ^(k+1)-s_t ^(k)Epsilon or k is less than or equal to | |>And K, finishing iteration, and taking the latest solution vector as a final estimation result of source positioning, otherwise, continuing the iteration.

Particularly, the optimal value of omega in (b5) is that 0.01 is used as a step length between (1 and 2), and the value is obtained after 10 iterations of selection_t-Ls_tThe smallest ω.

In the step (4), hilbert transform is firstly performed on the source vector of a single trial at each time point to obtain the phase of the source vector at each time point, and then PLV values of each pair of sources at each time point are calculated:

Δφ_ij＝φ_i-φ_j

wherein M is 1,2, …, M represents the mth trial;

when PLV_ijAnd when the value is greater than the threshold value, constructing an edge between the source i and the source j, and normalizing the value to be used as the weight of the edge:

in the step (5), the statistical test method is t-test, the classifier is an sLDA classifier, and the classifier is trained and tested by adopting 10 times of five-fold cross validation.

The invention has the beneficial effects that:

(1) according to the method, a T-wMNE algorithm and a successive relaxation method are combined to carry out source positioning on the electroencephalogram of natural actions, and compared with a traditional source positioning method, the method has the advantages that the solution sparsity and the robustness of the calculation process are guaranteed, and meanwhile, the solution precision and speed are improved;

(2) compared with the traditional method for constructing the brain network by taking the electrode channel as the node, the method can visually see the difference of the dynamic change process of the source corresponding to different natural actions, and is favorable for revealing the nerve operation mechanism of the human body;

(3) the invention respectively carries out source positioning and brain network analysis on a plurality of frequency bands, and can visually see the difference of different natural actions on frequency band activation and the root cause of the MRCP which has identifiability for different natural actions.

Drawings

FIG. 1 is a flow chart of the natural action electroencephalogram identification method based on source localization and brain network.

FIG. 2 is a flow chart of EEG signal preprocessing in a natural motion electroencephalogram identification method based on source localization and brain network according to the present invention.

FIG. 3 is a flow chart of EEG source localization in a natural action electroencephalogram recognition method based on source localization and brain network according to the present invention.

Detailed Description

The present invention will be further illustrated with reference to the accompanying drawings and specific embodiments, which are to be understood as merely illustrative of the invention and not as limiting the scope of the invention.

The invention designs a natural action electroencephalogram identification method based on source localization and brain network, as shown in figure 1, the steps are as follows:

(1) performing multi-channel electroencephalogram measurement on natural actions;

(2) preprocessing the collected EEG signals, removing artifacts, and extracting action-related cortical potential (MRCP), theta waves, alpha waves, beta waves and gamma waves;

(3) determining a lead field matrix of a signal, solving an initial value solution of a source by utilizing L1 regularization constraint, then iterating the initial value solution by utilizing a successive ultra-relaxation method, and taking the latest solution vector as a final estimation result of source positioning after iteration is finished;

(4) taking a source as a node, calculating phase synchronization PLV between each pair of sources at each time point by adopting a short-time sliding window, constructing an edge between the two sources when the PLV value is greater than a set threshold value, and taking a normalized value of the PLV value as the weight of the edge;

(5) calculating the characteristic path length, the clustering coefficient, the node average strength, the average betweenness, the efficiency and the network adjacency matrix at each time point, sending the characteristics into a classifier for training and testing, carrying out statistical test on the first 5 characteristics, and analyzing the difference of the characteristics corresponding to different actions on time or frequency bands.

As shown in fig. 2, step (2) includes the following sub-steps:

(a1) pre-filtering the acquired EEG signal;

(a2) eliminating data channels with abnormal kurtosis, performing spherical interpolation, and using the average value of four channels closest to the interpolated channel as the channel value;

(a3) finding and removing EOG and EMG components in the EEG using a blind source separation algorithm;

(a4) segment and baseline correction of EEG;

(a5) removing the test times with extreme values larger than 200 μ V and abnormal joint probability or abnormal kurtosis, wherein the threshold value of the two is 5 times of the standard deviation;

(a6) co-averaging reference (CAR) to EEG;

(a7) and respectively carrying out zero-phase Butterworth band-pass filtering on the heavily-referenced EEG at 0.3-3 Hz, 4-8 Hz, 8-13 Hz, 13-30 Hz and 30-45 Hz, and respectively extracting MRCP, theta wave, alpha wave, beta wave and gamma wave.

As shown in fig. 3, step (3) includes the following sub-steps:

(b1) a head model is selected.

(b2) The scalp is divided into N smaller 3D grids, 3 current dipoles with dipole moments in the directions of X, Y and the Z-axis are placed in each grid, the vector sum of which is equivalent to one possible current dipole, and a lead field matrix L is determined according to the following equation:

where the ith column of L represents the potential distribution produced by the ith current dipole source at each electrode location, r_i ^*Representing the position vector of the current dipole, r_jPosition vector representing the measured scalp electrode, s ═ se_iRepresenting the dipole moment (s is magnitude, e) of the current dipole_iIs direction), v (r)_jAnd t) representsThe potential of the j-th electrode at time t, i 1, …, N indicates N current dipoles, j 1, …, N indicates N measuring electrodes.

(b3) And determining a time point to be analyzed, and setting an iteration error epsilon and a maximum iteration number K.

(b4) And (3) solving the initial solution of the source vector by utilizing a T-wMNE algorithm:

wherein W ═ diag (| | l)₁||,||l₂||,,||l_NI) is a weighting matrix, N is the number of sources, here equal to the number of electrodes, s_tRepresenting the source vector at a point in time t, v_tDenotes the electrode potential at the time point t, and λ is the regularization coefficient.

(b5) And (c) iterating the initial value solution obtained in the step (b4) by using a successive super relaxation method:

wherein, omega epsilon (1,2) is used as a relaxation factor, and the step length of 0.01 on (1,2) is selected to lead to | v after 10 times of iteration_t-Ls_tThe smallest omega of | is used as the optimal relaxation factor, s_i,tDenotes the value of the ith source at time t, i ═ 1,2, …, N, v_j,tThe potential of the jth electrode at time t is shown, j is 1,2, …, N, k indicates the number of iterations.

(b6) When | | | s_t ^(k+1)-s_t ^(k)Epsilon or k is less than or equal to | |>And K, finishing iteration, and taking the latest solution vector as a final positioning estimation result of the source, otherwise, continuing the iteration.

In the step (4), hilbert transform is firstly performed on the source vector of a single trial at each time point to obtain the phase of the source vector at each time point, and then PLV values of each pair of sources at each time point are calculated:

Δφ_ij＝φ_i-φ_j

where M is 1,2, …, M indicates the mth trial.

When PLV_ijAnd when the value is greater than the set threshold value, constructing an edge between the source i and the source j, and normalizing the value to serve as the weight of the edge:

in the step (5), the statistical test method is t-test, the classifier is an sLDA classifier, and the classifier is trained and tested by adopting 10 times of five-fold cross validation.