An EEG recognition method that integrates structural information between brain regions

Abstract

The present invention provides an electroencephalogram signal recognition method that fuses structural information between brain regions, comprising the following steps: First, preprocess the electroencephalogram (EEG) signal and divide the EEG signal according to different brain regions; then use a deep neural network to learn the EEG depth features of local and global brain regions, and stack the EEG depth features of local and global brain regions into a feature matrix; finally, learn the structural information between different brain regions by introducing a kernel norm regularization term. The present invention can use a deep neural network to learn the EEG depth features of local and global brain regions, and at the same time introduce a kernel norm regularization term to learn the structural information between brain regions based on the EEG feature matrix, realizing the accurate recognition of different types of EEG signals and ensuring the safe and reliable operation of the EEG-based BCI system.

Description

Technical Field

[0001] The present invention relates to an electroencephalogram (EEG) signal recognition method that fuses structural information between brain regions, and belongs to the field of EEG signal recognition. Background Art

[0002] Electroencephalogram (EEG) signals have the advantages of high temporal resolution, non-invasive acquisition process, and high portability, and are widely used to record the dynamic activity process of the brain. Establishing an accurate mapping relationship between EEG signals and limb movement intentions using artificial intelligence methods can help disabled patients control assistive robots and wheelchairs to complete daily tasks, and can also be used in aspects such as video games and military services. In recent years, deep learning methods have been successfully applied to the extraction and classification of EEG signal features, and good results have been achieved. However, functions such as movement and emotion of the brain are the result of the interaction between different brain regions that are spatially distributed but functionally related to each other. Currently, most deep learning methods ignore the relevant information existing between different brain regions, which will have a certain impact on the classification results. To address the above problems, researchers have proposed an EEG recognition method that fuses structural information between brain regions. First, the EEG signals are divided according to different brain regions, and deep neural networks are used to learn the EEG depth features of local and global brain regions, and the local and global brain region EEG depth feature vectors are stacked into a feature matrix. Finally, by introducing a kernel norm regularization term to learn the structural information between brain regions based on the EEG feature matrix, accurate recognition of EEG signals is achieved. Summary of the Invention

[0003] The present invention provides an electroencephalogram (EEG) signal recognition method that fuses structural information between brain regions, uses deep neural networks to learn the EEG depth features of local and global brain regions, stacks the local and global brain region EEG depth feature vectors into a feature matrix, and at the same time introduces a kernel norm regularization term to learn the structural information between different brain regions, thereby improving the recognition performance of EEG signals.

[0004] The present invention adopts the following technical solutions to solve the above problems:

[0005] An electroencephalogram (EEG) signal recognition method that fuses structural information between brain regions includes the following steps:

[0006] Step 1: Collect EEG signals.

[0007] Step 2: Preprocess the EEG signals to remove noise and select a specific time period to obtain the original EEG signals where represents the nth EEG signal, m and d respectively represent the number of electrode channels and the number of time sampling points of the EEG signal, and yn ∈{-1, +1} is the label corresponding to the nth EEG signal.

[0008] Step 3: Divide the EEG signals into corresponding local brain region EEG signals according to channels based on different regions of the brain.

[0009] Step 4: Input the local and global brain region EEG signals into a deep neural network to learn their deep features respectively, and unfold and stack the deep features of the local and global brain regions into an EEG feature matrix.

[0010] Step 5: Classify the EEG feature matrix, and use the nuclear norm regularization term to learn the structural information between local brain regions and between local and global brain regions.

[0011] Step 6: Use the backpropagation method to learn the parameters of the classifier and the deep neural network.

[0012] Preferably, the method for fusing the structural information between brain regions uses a deep learning method to learn the deep features of the local and global brain region EEG signals, abandoning the disadvantages of time-consuming, laborious and error-prone manual extraction of EEG features, and having a powerful adaptive feature learning ability.

[0013] Preferably, the method for fusing the structural information between brain regions stacks the deep feature vectors of the local and global brain regions into a feature matrix, and uses the nuclear norm regularization term to learn the structural information between brain regions. The final classifier objective function is:

[0014]

[0015] where W and b represent the regression matrix and bias respectively, R(W) represents the square Frobenius norm of W, G(W) represents the nuclear norm of W, and H(·) represents the loss function.

[0016] Beneficial effects:

[0017] 1. Divide the EEG signals into corresponding local brain region EEG signals according to channels based on different regions of the brain, and use a deep neural network to learn the deep features of the local and global brain regions, which can learn higher-level and more abstract deep feature representations of EEG signals in different brain regions.

[0018] 2. This paper introduces the nuclear norm regularization term to learn the structural information between different brain regions based on the EEG feature matrix, which can further assist in brain decoding. Traditional deep learning methods ignore the relevant information existing between different brain regions, resulting in limited representation ability and poor performance in dealing with complex EEG signal classification problems. Description of the Drawings

[0019] Figure 1It is the network framework diagram of the EEG signal recognition method that fuses the structural information between brain regions in the present invention. Detailed implementation manners

[0020] Please refer to Figure 1 as shown below:

[0021] The following further explains the present invention in combination with examples.

[0022] The main implementation process of the present invention is as follows. For the relevant network framework, see Figure 1 .

[0023] Step 1: Collect EEG signals.

[0024] Step 2: Preprocess the EEG signals, including band-pass filtering and intercepting corresponding time periods to remove noise and select signals in specific time periods, so as to obtain the original EEG signals where represents the nth EEG signal, m represents the number of EEG signal channels, and d represents the number of time sampling points of the EEG signal. y n ∈{-1, +1} is the label corresponding to the nth EEG signal.

[0025] Step 3: Divide the EEG signal X n into J corresponding local regions according to different regions of the brain. The EEG signal corresponding to the jth local brain region can be expressed as where, m j represents the number of channels of the EEG signal corresponding to the jth local brain region, and m 1 +m 2 +…+m J =m. In addition, use to represent the EEG signal of the global brain region.

[0026] Step 4: Learn the EEG deep features of local and global brain regions, and stack the EEG deep features of all regions into a feature matrix. The specific method is as follows:

[0027] 1) Input the EEG signals of local and global brain regions into the deep neural network respectively, and use the deep neural network to learn their deep features. Then, the deep neural network for learning the EEG signal of the jth local brain region can be expressed as: where θ j represents the network parameters of this deep neural network, represents the learned EEG deep feature of the jth brain region. Among them, d 1 and d 2 represent the dimensions of the EEG deep feature. At the same time, use the deep neural network to learn the EEG deep feature of the global brain region, which is expressed as: where θg denote the network parameters of the deep neural network, denote the EEG depth features of the j-th learned brain region.

[0028] 2) Use to represent the depth feature vector after unfolding the EEG depth features of the j-th local brain region, and use to represent the depth feature vector after unfolding the EEG depth features of the global brain region. Stack the EEG depth feature vectors of the local and global brain regions to obtain the feature matrix

[0029] Step 5: Classify the EEG feature matrix obtained in Step 4. The objective function of the classifier is:

[0030]

[0031] where W and b represent the regression matrix and the bias respectively. denote the squared Frobenius norm of the regression matrix W, which is mainly used to control the complexity of the network model and prevent overfitting during the training process. G(W) = τ||W|| * denote the nuclear norm of the regression matrix W, and the parameter τ > 0 is the penalty coefficient. The nuclear norm is used to learn the structural information between local brain regions and between local and global brain regions to improve the recognition performance of EEG signals.

[0032] In addition, use the Hinge function as the loss function, and C is the balance parameter used to constrain the loss term. H(·) can be expressed as:

[0033]

[0034] Step 6: Take the partial derivatives of the regression matrix W and the bias b, and we can get:

[0035]

[0036]

[0037] Finally, use the backpropagation method to learn the parameters of the classifier and the deep neural network.

[0038] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the content of the specification and drawings of the present invention, or directly or indirectly applied in the relevant technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for electroencephalogram (EEG) signal recognition that fuses structural information between brain regions, characterized in that, the steps are as follows: Step 1: Collect EEG signals; Step 2: Preprocess the EEG signals to remove noise and select a specific time period to obtain the original EEG signals where X n represents the nth EEG signal, m represents the number of EEG signal channels, d represents the number of time sampling points of the EEG signals, and y n ∈{-1, +1} is the label corresponding to the nth EEG signal; Step 3: According to different regions of the brain, divide the global brain region EEG signals into corresponding local brain region EEG signals according to channels; Step 4: Input the local brain region EEG signals and the global brain region EEG signals into a deep neural network to learn their deep features respectively, and unfold and stack the EEG deep features of the local brain region EEG signals and the global brain region EEG signals into an EEG feature matrix; Step 5: Classify the EEG feature matrix, and use the nuclear norm regularization term to learn the structural information between local brain regions and between local and global brain regions; Step 6: Use the backpropagation method to learn the parameters of the classifier and the deep neural network; Step 7: Input the EEG signal to be classified into the final network model obtained by the above steps to obtain its class label.

2. The method for EEG signal recognition that fuses structural information between brain regions according to claim 1, characterized in that, In step 3, the electroencephalogram signal X is divided according to different regions of the brain n into J corresponding local regions according to channels, and the EEG signal corresponding to the j-th local brain region is expressed as where m j represents the number of channels of the EEG signal corresponding to the j-th local brain region, and m 1 + m 2 + … + m J = m.

3. The method for EEG signal recognition that fuses structural information between brain regions according to claim 2, characterized in that, In step 4, the EEG signals of the local brain regions and the global brain region are respectively input into the deep neural network, and the deep neural network is used to learn their deep features. The EEG deep feature of the j-th local brain region learned is expressed as: Meanwhile, the EEG deep feature of the global brain region learned is expressed as: where d 1 and d 2 represent the dimensions of the EEG deep features; is used to represent the deep feature vector after expanding the EEG deep feature of the j-th local brain region, and is used to represent the deep feature vector after expanding the EEG deep feature of the global brain region. The EEG deep feature vectors of the local and global brain regions are stacked to obtain the feature matrix 4. The method for EEG signal recognition that fuses structural information between brain regions according to claim 3, characterized in that, In step 5, when classifying the EEG feature matrix, the objective function of the classifier is: where \(W\) and \(b\) represent the regression matrix and the bias, respectively; denotes the squared Frobenius norm of \(W\), \(G(W)=\tau||W||\) * denotes the nuclear norm of \(W\), \(H(\cdot)\) is the loss function, expressed as:

5. The method for EEG signal recognition that fuses structural information between brain regions according to claim 4, characterized in that, In step 6, when using the backpropagation method to learn the parameters of the classifier and the deep neural network, taking partial derivatives of the regression matrix W and the bias b, we can obtain:

Images