Deep learning brain power positioning method based on independent component analysis

Abstract

The invention relates to a deep learning brain power supply positioning method based on independent component analysis, which comprises the following steps: carrying out independent component reconstruction processing on electroencephalogram / magnetoencephalogram signal data to obtain a plurality of reconstructed electroencephalogram / magnetoencephalogram signals; performing primary source positioning on the reconstructed electroencephalogram / magnetoencephalogram signal to obtain a plurality of primary source positioning results; and carrying out positioning recovery on the initial source positioning result to obtain a final source positioning result. According to the method, each independent component is used for reconstructing an electroencephalogram / magnetoencephalogram signal, and the reconstructed signal is used for source positioning to obtain a single potential activation source. The process of solving a single source by using each independent component is not only beneficial to improving the precision of source positioning, but also can provide a clearer source moving image. And performing combination based on single sources obtained by each independent division to form a multi-source solution. The independent component source solving method not only can effectively reduce the complexity of the problem, but also can reduce the error caused by signal mixing, and improves the overall positioning precision and robustness.

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface and deep learning technology, and in particular to a deep learning brain source localization method based on independent component analysis. Background Art

[0002] sLORETA (Low-Resolution Electromagnetic Tomography) is a technique for solving electromagnetic inverse problems, which is used to infer the location of the source of neural activity inside the brain from electroencephalogram (EEG) or magnetoencephalogram (MEG) signals. The core challenge of the electromagnetic inverse problem is that there are usually infinite possible solutions to derive the source activity inside the brain from the limited electric or magnetic field data on the brain surface, which makes the inverse problem ill-posed and difficult to solve. In order to overcome this difficulty, regularization methods and physical constraints must be used to narrow the solution space and improve the stability of the solution.

[0003] The technical core of sLORETA is to regularize the inverse problem using the least squares criterion. It assumes that the current density of adjacent source points is highly correlated. This assumption makes the solution of source localization smooth, thereby stabilizing the solution of the inverse problem. Through this smooth constraint, sLORETA is able to generate relatively stable source activity maps from EEG or MEG data without the need for a large amount of complex prior information. Although its assumption leads to lower spatial resolution and makes it difficult to accurately distinguish the activities of adjacent brain regions, sLORETA provides a reliable solution in the presence of noise and insufficient data.

[0004] sLORETA technology has a wide range of applications in the field of source localization, especially in neuroscience and clinical research. It is used to explore the functional activities of the brain, such as the localization of cognitive processes and emotional responses, and can also help study the pathological characteristics of neurological diseases, such as the localization of epileptic foci. Compared with other source localization methods, the advantage of sLORETA is that its implementation is relatively simple and computationally efficient.

[0005] Limitations of sLORETA technology:

[0006] 1. Since sLORETA assumes that the current density of adjacent source points is highly correlated, the result is a smoothed brain power source activity map, which leads to low spatial resolution and makes it difficult to accurately distinguish different brain power sources that are close in space.

[0007] 2. Although the regularization and spatial smoothing process improves the stability of the solution, it will cause underestimation or distortion of the numerical intensity and cannot truly reflect the intensity of the original source.

[0008] 3. EEG signals are easily interfered by environmental noise and artifacts (such as eye movements or muscle activities), especially when the signal-to-noise ratio is not high, which will affect the source localization accuracy of sLORETA. Summary of the invention

[0009] The purpose of the present invention is to propose a deep learning brain power source localization method based on independent component analysis to solve the problems existing in the above-mentioned prior art.

[0010] To achieve the above object, the present invention provides the following solutions:

[0011] A deep learning brain source localization method based on independent component analysis, comprising:

[0012] Performing independent component reconstruction processing on the EEG / MEG signal data to obtain a number of reconstructed EEG / MEG signals;

[0013] Performing preliminary source localization on the reconstructed EEG / MEG signals to obtain several preliminary source localization results; wherein the preliminary source localization results are: initial current intensities of m sources in three directions at t time points;

[0014] The preliminary source positioning result is positioned and recovered to obtain a number of sparse source positioning results, and the sparse source positioning results are added and combined to obtain a final source positioning result; wherein the final source positioning result is: the precise current intensity of m sources in three directions at t time points.

[0015] Optionally, performing independent component reconstruction processing on the EEG / MEG signal data includes:

[0016] Centralized processing of EEG / MEG signal data;

[0017] Performing whitening processing on the centralized EEG / MEG signal;

[0018] Iteratively solve the independent components of the whitened EEG / MEG signals;

[0019] Based on each of the independent components, a plurality of reconstructed EEG / MEG signals are obtained.

[0020] Optionally, performing whitening processing on the centralized EEG / MEG signal includes:

[0021] Calculating the mean of each variable in the EEG / MEG signal data;

[0022] The mean is subtracted from each variable so that all variables have a mean of zero.

[0023] Optionally, performing whitening processing on the centralized EEG / MEG signal includes:

[0024] Performing singular value decomposition on the centralized EEG / MEG signal to obtain eigenvectors and eigenvalues;

[0025] The decomposed EEG / MEG signals are transformed using eigenvectors and eigenvalues ​​to obtain whitened data.

[0026] Optionally, iteratively solving the independent components of the whitened EEG / MEG signal includes:

[0027] Select a non-Gaussianity measure;

[0028] Randomly initialize the demixing matrix;

[0029] Based on a preset formula, using the non-Gaussian metric, iterating each weight vector independent component in the demixing matrix;

[0030] Orthogonalize the iterated weight vector;

[0031] Determine whether the orthogonal weight vector meets the convergence condition. If not, continue to iterate the weight vector until it converges to obtain independent components.

[0032] The EEG / MEG signals are reconstructed using the independent components.

[0033] Optionally, the preset formula is:

[0034]

[0035] Among them, w i represents the i-th column vector of the demixing matrix W, i.e. the i-th weight vector, represents the iteratively updated w i , w i The transpose of , Ε represents the mathematical expectation, Y whitened represents the whitened matrix, g represents the derivative of the non-Gaussian measure, and g' represents the derivative of g;

[0036] The independent components are:

[0037] S=WY whitened

[0038] Where W is the mixing matrix after the independent components are orthogonalized;

[0039] The reconstructed EEG / MEG signal is:

[0040]

[0041] Among them, s i is the ith independent component of S, Y recon is the EEG / MEG signal reconstructed from the i-th independent component, Indicates i , Y represents the original observed EEG / MEG signal.

[0042] Optionally, the method for performing preliminary source localization on the reconstructed EEG / MEG signal is: using the sLORETA method.

[0043] Optionally, performing positioning recovery on the preliminary source positioning result includes:

[0044] Sequentially performing encoding processing, intermediate feature processing, decoding processing, and skip connection processing on the preliminary source positioning result to obtain a single source positioning of each independent component;

[0045] Adding the single-source positioning combinations to form a multi-source solution;

[0046] The encoding process includes: extracting potential feature vectors in the preliminary source positioning;

[0047] The intermediate feature processing is: processing the potential feature vector;

[0048] The decoding process is: reconstructing the processed potential feature vector;

[0049] The jump connection process is used to concatenate the features of the encoder part with the features of the corresponding layer of the decoder.

[0050] The beneficial effects of the present invention are:

[0051] The present invention utilizes each independent component to reconstruct the EEG signal, and utilizes the reconstructed EEG signal to perform source localization to obtain a single potential activation source. The process of solving a single source using each independent component not only helps to improve the accuracy of source localization, but also provides a clearer source activity image. Finally, the single sources of all independent components are combined to form a multi-source solution. This method of solving the source with independent components can not only effectively reduce the complexity of the problem, but also reduce the error caused by signal mixing, and improve the overall positioning accuracy and robustness. In addition, in practical applications, EEG signals are often interfered by various noises and artifacts, such as eye movements, electromyographic interference, etc. Through the independent components separated by independent component analysis (ICA), these interference signals can be effectively identified and removed, thereby obtaining a purer source signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0053] Figure 1 Schematic diagram of the overall architecture of deep learning brain power source localization based on independent component analysis according to an embodiment of the present invention;

[0054] Figure 2 Schematic diagram of the specific structure of the neural network source positioning and recovery module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1 As shown, this embodiment proposes a deep learning brain power source localization method based on independent component analysis, including:

[0058] Performing independent component reconstruction processing on the EEG signal data to obtain a number of reconstructed EEG signals;

[0059] Performing preliminary source localization on the reconstructed EEG / MEG signals to obtain several preliminary source localization results; wherein the preliminary source localization results are: initial current intensities of m sources in three directions at t time points;

[0060] The preliminary source positioning result is positioned and recovered to obtain a number of sparse source positioning results, and the sparse source positioning results are added and combined to obtain a final source positioning result; wherein the final source positioning result is: the precise current intensity of m sources in three directions at t time points.

[0061] Specifically, this embodiment uses independent component analysis (ICA) to solve independent components, and uses different independent components to reconstruct EEG signals for solving a single activation source. Then the individual sources of all independent components are combined to obtain a sparse multi-source solution. This embodiment can improve the accuracy of source intensity recovery: by introducing a deep learning model, the intensity distortion caused by sLORETA regularization is reduced, and the accurate recovery of source current intensity is improved. Improve the position accuracy of source localization: The deep learning model can learn complex spatial features, gradually optimize the source space positioning through residual iteration, solve the sLORETA low resolution problem, and achieve more accurate source localization. Improve the ability of multi-source localization: Use Independent Component Analysis (ICA) to solve independent components, and use different independent components to reconstruct EEG signals to improve the performance of sparse multi-source localization.

[0062] Furthermore, performing independent component reconstruction on the multivariate EEG signal data includes:

[0063] Centralize the EEG signal data;

[0064] Performing whitening processing on the centralized EEG signal;

[0065] Iteratively solve the independent components of the whitened EEG signal;

[0066] Based on each of the independent components, a plurality of reconstructed EEG / MEG signals are obtained.

[0067] Furthermore, performing whitening processing on the centralized EEG / MEG signal includes:

[0068] Calculate the mean of each variable in the EEG signal data; where the variable is the value of each electrode in the EEG signal matrix over a period of time. In FastICA, the centering operation is performed on the time dimension, that is, the data of each electrode is subtracted from its mean at all time points.

[0069] The mean is subtracted from each variable so that all variables have a mean of zero.

[0070] Furthermore, performing whitening processing on the centralized EEG signal includes:

[0071] Performing singular value decomposition on the centralized EEG signal to obtain eigenvectors and eigenvalues;

[0072] The decomposed EEG signal is transformed using the eigenvectors and eigenvalues ​​to obtain the whitened data.

[0073] Furthermore, the independent components of the whitened EEG signal are iteratively solved including:

[0074] Select a non-Gaussianity measure;

[0075] Randomly initialize the demixing matrix; wherein the demixing matrix is ​​the inverse matrix of the mixing matrix, and the demixing matrix is ​​randomly initialized and generated to provide an initial point for the algorithm so that the correct demixing matrix can be gradually found through subsequent iterations;

[0076] Based on the preset formula, each weight vector independent component in the demixing matrix is ​​iterated using non-Gaussian measurement;

[0077] Orthogonalize the iterated weight vector;

[0078] Determine whether the orthogonal weight vector meets the convergence condition. If not, continue to iterate the weight vector until it converges to obtain independent components.

[0079] Reconstruct EEG signals using independent components.

[0080] Specifically, in this embodiment, an independent component reconstruction module is used to perform independent component reconstruction processing on multivariate EEG signal data;

[0081] Independent component analysis (ICA) signal reconstruction: ICA is used to decompose EEG signals from multiple sources and extract independent components. By treating the observed signal as a linear combination of independent source signals, the ICA algorithm is used to decompose the mixed signal into multiple independent components, and then the EEG signal is reconstructed from these independent components to achieve the separation and reconstruction of a single activation source. This step obtains independent components through processes such as centering, whitening, and iterative solution. In particular, the FastICA algorithm is used to estimate independent components by maximizing non-Gaussianity.

[0082] The independent component reconstruction module aims to separate the independent components of the signal through independent component analysis, and reconstruct the electroencephalogram (EEG) signal based on these independent components, so as to analyze each activation source separately. Independent component analysis is a powerful multivariate data processing method, whose goal is to decompose the observed multivariate signal into independent components. The core concept of this method is to regard the mixed signal as a linear superposition of several independent source signals, separate these independent source signals through algorithms, and restore their original form. The main reason for choosing ICA is that it can ensure that each set of reconstructed EEG signals contains only one source signal as much as possible through independent component decomposition. This assumption is based on the independence between source signals. In source signal localization, reducing the number of source signals (ideally a single source signal) can significantly reduce the localization complexity in the case of multiple sources, thereby converting a complex source localization problem into multiple relatively simple sub-problems.

[0083] The basic assumption of ICA is that the observed signal is a linear mixture of several independent source signals. Assuming there is an observed signal matrix Y, it can be expressed as the product of the source signal matrix S and the unknown mixing matrix A, that is: Y = AS. Where Y is the observed mixed signal, A is the unknown mixing matrix, and S is the independent source signal. The goal of ICA is to estimate a demixing matrix W so that the separated signal S is as independent as possible. In other words, S = WX, and the source signal is restored by estimating W.

[0084] The algorithm flow for solving independent components can be divided into the following main steps: centering, whitening, and iterative solving of independent components. First, the centering step calculates the mean of each variable and subtracts the mean from the data so that the mean of all variables is zero. This step simplifies the subsequent calculations, as shown in formula (1).

[0085] Y centered =Y-Ε[Y] (1)

[0086] Where Y is the observation data matrix, each column represents an observation variable, and each row represents an observation sample. E[Y] is the mean of the data. Next, whitening is performed, which linearly transforms the data into data with zero mean and a covariance matrix of the unit matrix. This step eliminates the linear correlation between variables, making it easier to solve the subsequent independent components. The whitening step is usually implemented through principal component analysis (PCA), specifically, performing singular value decomposition (SVD) on the centered data, and then transforming the data using its eigenvectors and eigenvalues, as shown in formula (2).

[0087] Y whitened =VD -1 / 2 U T Y centered (2)

[0088] Among them, U and V are the eigenvectors of the covariance matrix of the observed data, and D is its eigenvalue matrix. FastICA is an efficient ICA algorithm that uses fixed point iteration to estimate independent components by maximizing non-Gaussianity. The specific steps are as follows.

[0089] (1) Selecting non-Gaussian metrics: Commonly used non-Gaussian metrics include Negentropy and Kurtosis. A common choice is to use the Negentropy approximation, as shown in formula (3):

[0090] G(y)=log(cosh(y)) (3)

[0091] (2) Initialize the demixing matrix: Randomly initialize the demixing matrix W.

[0092] (3) Iterative update: For each independent component w i , and perform the following iterations:

[0093]

[0094] Wherein, g is a nonlinear function, such as g(y)=tanh(y), and g' is the derivative of g.

[0095] Independent component w i , is a column vector in the demixing matrix, regarded as a weight vector. The demixing matrix is ​​a weight matrix used to extract independent components S from the observed data EEG.

[0096] Then for w i Normalize:

[0097]

[0098] (4) Orthogonalization: Ensure that the column vectors of the demixing matrix are orthogonal (for more than one independent component):

[0099] W=(WW T ) -1 / 2 W

[0100] (5) Convergence determination: Check whether the convergence condition is met, i.e., w i Is the change of is less than the set threshold. If not converged, return to step (3) to continue iterating. When the demixing matrix W is obtained, the independent components can be obtained by the following formula (6):

[0101] S=WY whitened (6)

[0102] Where S is the estimated independent component matrix. After obtaining the independent component S, each independent component is used to reconstruct the EEG signal, as shown in formula (7).

[0103]

[0104] Among them, s i is the ith independent component of S, Y recon[i] The EEG signal reconstructed for the i-th independent component. ICA does not require supervisory information and can autonomously learn and extract independent components from the data. It is applicable to various types of EEG data. By decomposing and reconstructing EEG signals, complex multi-source signals can be separated into multiple simple independent components, reducing the complexity of subsequent multi-source localization and improving the interpretability of EEG signals.

[0105] Further, for one of the reconstructed EEG / MEG or MEG data Y recon[i],The method for preliminary source localization is : using the sLORETA method.

[0106] Specifically, in this embodiment, the sLORETA preliminary source localization module is used to perform preliminary source localization on the reconstructed EEG signal. The sLORETA preliminary source localization is performed using multiple reconstructed EEG signals of the previous module to obtain multiple preliminary localization results, and the preliminary localization results are input into the sLORETA preliminary source localization module.

[0107] Among them, sLORETA achieves source localization by building a forward model between EEG signals and brain source activity using a linear inverse algorithm. To solve the multi-solution problem, sLORETA introduces a spatial smoothness regularization constraint, assumes that the neural activity of adjacent voxels is similar, constructs a regularization matrix to limit the multi-solution, and normalizes the result to be proportional to the true current density, thereby achieving preliminary source localization.

[0108] Further, performing positioning recovery on the preliminary source positioning result includes:

[0109] Sequentially performing encoding processing, intermediate feature processing, decoding processing, and skip connection processing on the preliminary source positioning result to obtain a single source positioning of each independent component;

[0110] The single-source positioning is combined to form a multi-source solution.

[0111] Specifically, in this embodiment, a neural network source location recovery module is used to recover the preliminary source location result; the neural network source location recovery module uses the U-Net model as the main structure to achieve reconstruction from sLORETA smooth source to sparse source. It consists of four parts: encoding part, intermediate feature processing part, decoding part and jump connection, such as Figure 2 shown.

[0112] Coding processing: used to extract the potential feature vector in the preliminary source localization. The coding processing part includes: convolution layer, activation function and pooling layer. The convolution layer is used to extract the features in the current source localization; the activation function is used to learn the nonlinear features of the current source localization; the pooling layer is used to reduce the feature dimension and obtain the potential feature vector.

[0113] Intermediate feature processing: used to process the latent feature vector. 1. The latent feature vector is processed by the GRU network. 2. The covariance matrix of the current electromagnetic signal is obtained. (That is, the EEG covariance information is flattened by the flatten layer, and then reshaped to the same dimension as the latent feature after passing through multiple linear layers and activation functions) 3. The current covariance matrix is ​​used to obtain features with the same dimension as the latent feature vector using a neural network. 4. The two features are fused to obtain the processed feature vector.

[0114] Decoding: used to reconstruct the processed latent feature vector and obtain the current sparse source. The decoding part includes: a number of transposed convolution layers and activation functions. The transposed convolution layer is used to upsample the processed latent feature vector to restore it to its original size (i.e., the size of the current source location). The transposed convolution has learnable parameters, and the optimal upsampling method can be obtained through network learning. The activation function is used to learn the nonlinear characteristics of the processed latent feature vector.

[0115] The jump connection part is to splice the features of the encoder part with the features of the corresponding layer of the decoder. The functions are: 1. Retain and transmit high-resolution detail information. 2. Fuse shallow features with deep features, which not only alleviates the problem of gradient disappearance, but also enriches the feature representation, so that the model can better learn different levels of brain power information.

[0116] After completing the single-source localization of each independent component, combining these single-source solutions to form a multi-source solution is the last step in brain source localization. The overall algorithm of deep learning brain source localization based on independent component analysis is shown in the following algorithm.

[0117]

[0118] This embodiment uses a combination of the sLORETA method and a deep learning model, uses sLORETA for preliminary source localization, and provides a relatively rough brain source localization result. Then the results of sLORETA are repaired using a neural network to restore the sparse sources at the precise location and restore the intensity of the activated source as much as possible. The advantage of this combined method is that it combines the fast calculation and preliminary positioning capabilities of sLORETA with the high-precision repair capabilities of the neural network, thereby achieving more accurate and stable results in source localization.

[0119] The neural network source location recovery module uses the U-Net model as the main structure to achieve reconstruction from sLORETA smooth source to sparse source, such as Figure 2 As shown in the figure. The intermediate feature processing part of this module is mainly based on feature fusion and extraction of feature time series information. The EEG covariance matrix is ​​used as prior information to fuse with potential features, providing prior knowledge of the correlation between EEG signals, which can improve the positioning ability of the model. GRU is integrated in the U-Net model to process intermediate features and enhance the model's ability to capture time series information.

[0120] The process of reconstructing EEG signals using independent components is of great significance. ICA is different from principal component analysis (PCA), which focuses on maximizing the variance of the signal rather than independence. This embodiment reconstructs the EEG signal using each independent component, and uses the reconstructed EEG signal to perform source localization to obtain a single potential activation source. The process of solving a single source using each independent component not only helps to improve the accuracy of source localization, but also provides a clearer source activity image. Finally, the single sources of all independent components are combined to form a multi-source solution. This method of solving the source with independent components can not only effectively reduce the complexity of the problem, but also reduce the errors caused by signal mixing, and improve the overall positioning accuracy and robustness. In addition, in practical applications, EEG signals are often interfered by various noises and artifacts, such as eye movements, electromyographic interference, etc. Through the independent components separated by ICA, these interference signals can be effectively identified and removed, thereby obtaining a purer source signal.

[0121] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A deep learning method for localizing brain power sources based on independent component analysis, characterized in that: include: Performing independent component reconstruction processing on the EEG / MEG signal data to obtain a number of reconstructed EEG / MEG signals; Performing preliminary source localization on the reconstructed EEG / MEG signals to obtain several preliminary source localization results; wherein the preliminary source localization results are: initial current intensities of m sources in three directions at t time points; The preliminary source positioning result is positioned and recovered to obtain a number of sparse source positioning results, and the sparse source positioning results are added and combined to obtain a final source positioning result; wherein the final source positioning result is: the precise current intensity of m sources in three directions at t time points.

2. According to the deep learning method for locating brain power sources based on independent component analysis in claim 1, it is characterized in that: The independent component reconstruction processing of EEG / MEG signal data includes: Centralized processing of EEG / MEG signal data; Performing whitening processing on the centralized EEG / MEG signal; Iteratively solve the independent components of the whitened EEG / MEG signals; Based on each of the independent components, a plurality of reconstructed EEG / MEG signals are obtained.

3. According to claim 2, the deep learning method for locating brain power sources based on independent component analysis is characterized in that: The whitening process of the centralized EEG / MEG signal includes: Calculating the mean of each variable in the EEG / MEG signal data; The mean is subtracted from each variable so that all variables have a mean of zero.

4. According to claim 2, the deep learning method for locating brain power sources based on independent component analysis is characterized in that: The whitening process of the centralized EEG / MEG signal includes: Performing singular value decomposition on the centralized EEG / MEG signal to obtain eigenvectors and eigenvalues; The decomposed EEG / MEG signals are transformed using eigenvectors and eigenvalues ​​to obtain whitened data.

5. According to claim 2, the deep learning method for locating brain power sources based on independent component analysis is characterized in that: The independent components of the whitened EEG / MEG signals are iteratively solved including: Select a non-Gaussianity measure; Randomly initialize the demixing matrix; Based on a preset formula, using the non-Gaussian metric, iterating each weight vector independent component in the demixing matrix; Orthogonalize the iterated weight vector; Determine whether the orthogonal weight vector meets the convergence condition. If not, continue to iterate the weight vector until it converges to obtain independent components. The EEG / MEG signals are reconstructed using the independent components.

6. The deep learning brain source localization method based on independent component analysis according to claim 5 is characterized in that: The preset formula is: Among them, w i represents the i-th column vector of the demixing matrix W, i.e. the i-th weight vector, represents the iteratively updated w i , w i The transpose of , Ε represents the mathematical expectation, Y whitened represents the whitened matrix, g represents the derivative of the non-Gaussian measure, and g' represents the derivative of g; The independent components are: S=WY whitened Where W is the mixing matrix after the independent components are orthogonalized; The reconstructed EEG / MEG signal is: Among them, s i is the ith independent component of S, Y recon is the EEG / MEG signal reconstructed from the i-th independent component, Indicates i , Y represents the original observed EEG / MEG signal.

7. The deep learning brain source localization method based on independent component analysis according to claim 1 is characterized in that: The method for performing preliminary source localization on the reconstructed EEG / MEG signal is: using the sLORETA method.

8. The deep learning brain source localization method based on independent component analysis according to claim 1 is characterized in that: Performing positioning recovery on the preliminary source positioning result includes: Sequentially performing encoding processing, intermediate feature processing, decoding processing, and skip connection processing on the preliminary source positioning result to obtain a single source positioning of each independent component; Adding the single-source positioning combinations to form a multi-source solution; The encoding process includes: extracting potential feature vectors in the preliminary source positioning; The intermediate feature processing is: processing the potential feature vector; The decoding process is: reconstructing the processed potential feature vector; The jump connection process is used to concatenate the features of the encoder part with the features of the corresponding layer of the decoder.

Images