Electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network

Abstract

The invention discloses an electroencephalogram analysis method based on a time-frequency complex network and a convolutional probabilistic neural network. The electroencephalogram analysis method comprises the following steps: dividing multichannel electroencephalogram signals acquired by an electroencephalogram signal acquisition device into a training set and a test set; time-frequency complex networks are constructed for the training set samples and the test set samples respectively; inputting the time-frequency complex network obtained by the training set into a convolutional neural network for model training; building a mode layer of a paralleling probabilistic neural network based on the advanced feature vector output by the trained convolutional neural network, wherein the convolutional neural network and the paralleling probabilistic neural network jointly form the convolutional probabilistic neural network; and inputting the time-frequency complex network obtained by the test set into a convolutional probabilistic neural network for feature extraction and classification. According to the method, the brain topology dynamic representation capability of the time-frequency complex network, the deep feature extraction capability of the convolutional neural network and the excellent classification capability of the probabilistic neural network are fused, and high-precision classification of electroencephalogram signals under tasks such as epilepsy prediction, sleep detection and motor imagery identification can be realized.

Description

Technical Field

[0001] This invention relates to a method for analyzing electroencephalogram (EEG) signals. In particular, it relates to an EEG analysis method based on time-frequency complex networks and convolutional probabilistic neural networks. Background Technology

[0002] Brain-computer interfaces (BCIs) establish an external communication pathway between the human brain and external devices by analyzing electroencephalogram (EEG) signals, enabling direct interaction between humans and the external environment. This represents a novel and revolutionary human-computer interaction technology. BCIs involve multiple disciplines, including neuroscience, signal detection, signal processing, and pattern recognition, and play a crucial role in fields such as rehabilitation training for people with disabilities, biomedicine, gaming, aerospace, and military applications. Electroencephalogram (EEG) signals are among the most widely used BCI signals due to their safe, non-invasive, convenient, and cost-effective acquisition process. Effective extraction and classification of key features from EEG signals are crucial for building high-performance BCI systems. However, EEG signals are characterized by weak amplitude, strong background noise, and nonlinearity, making high-precision decoding a significant challenge.

[0003] The brain is the most complex known structure in the human body and even the universe. The topological interactions between brain neurons, different regions of the cerebral cortex, and other brain components ensure the brain's control over the human body. Complex networks, as abstract and simplified representations of complex systems, can effectively characterize the topological interactions between their components and are important tools for studying complex systems. Analyzing EEG signals under different brain states using complex networks can effectively uncover hidden features in the signals and improve EEG signal decoding performance. In recent years, convolutional neural networks (CNNs), as one of the most popular deep learning frameworks, have gained favor in EEG signal analysis. CNNs, with their local connectivity and weight sharing, can automatically extract discriminative features from EEG signals, demonstrating superior feature extraction capabilities. Probabilistic neural networks (PNs), a neural network model based on statistical principles, are commonly used for classification tasks and have advantages such as simple and fast learning processes, good fault tolerance, and strong interpretability. Combining the topological representation capabilities of complex networks, the deep feature extraction capabilities of CNNs, and the excellent classification capabilities of PNs can provide a new framework for decoding EEG signals. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an EEG analysis method based on time-frequency complex networks and convolutional probabilistic neural networks that can effectively analyze EEG signals and achieve high-precision classification of EEG signals.

[0005] The technical solution adopted in this invention is: an electroencephalogram (EEG) analysis method based on time-frequency complex networks and convolutional probabilistic neural networks, comprising the following steps:

[0006] 1) The multi-channel EEG signals acquired through the EEG signal acquisition device are divided into a training set and a test set. The training set contains... The test set contains [number] samples. One sample;

[0007] 2) For any sample in the training set and the test set Perform frequency band division to obtain frequency band set Multichannel EEG signals corresponding to four frequency bands;

[0008] 3) Regarding the sample In frequency band set Multichannel EEG signals from different frequency bands were used to construct time-frequency subnetworks, resulting in a four-layer time-frequency subnetwork. And four network adjacency matrices corresponding one-to-one with the four-layer time-frequency sub-networks. ;

[0009] 4) The network adjacency matrix corresponding to the four-layer time-frequency sub-network The adjacency matrix of the fused single-layer network is obtained by adding the two matrices. And by the single-layer network adjacency matrix Characterizes the corresponding single-layer time-frequency complex network ;

[0010] 5) Repeat steps 2) to 4) for each sample in both the training and test sets to obtain all samples in the training set. The set of single-layer adjacency matrices corresponding to each sample and all samples in the test set. The set of single-layer adjacency matrices corresponding to each sample;

[0011] 6) Add all training data to the training set. The single-layer adjacency matrix corresponding to each sample is input into a convolutional neural network for model training. The main structure of the convolutional neural network includes two convolutional layers, two pooling layers, and two fully connected layers. The parameters are iteratively updated, and the training process of the convolutional neural network is stopped when the accuracy of the model meets the preset threshold requirement.

[0012] 7) Input the high-level feature vector output by the first fully connected layer of the convolutional neural network described in step 6) into the parallel probabilistic neural network to build the mode layer of the parallel probabilistic neural network; the parallel probabilistic neural network includes at least a mode layer, a summation layer and an output layer, and the convolutional neural network and the parallel probabilistic neural network together constitute the convolutional probabilistic neural network.

[0013] 8) Add all test cases to the test set. The single-layer adjacency matrix corresponding to each sample is input into the convolutional probabilistic neural network for feature extraction and classification output.

[0014] This invention presents an EEG analysis method based on time-frequency complex networks and convolutional probabilistic neural networks, enabling effective analysis of EEG signals and achieving high-precision classification. This invention integrates the topological dynamics representation capabilities of time-frequency complex networks with the deep feature extraction capabilities of convolutional neural networks, reducing the influence of redundant features in EEG signals and effectively extracting key distinguishable features. Furthermore, it leverages the efficient and rapid classification capabilities of probabilistic neural networks, improving the accuracy and speed of identifying different types of EEG signals. This invention can be applied to various multi-channel EEG signal analysis tasks, such as epilepsy prediction, sleep monitoring, and motor imagery recognition. Attached Figure Description

[0015] Figure 1 This is an overall flowchart of the EEG analysis method based on time-frequency complex networks and convolutional probabilistic neural networks of the present invention;

[0016] Figure 2 This is a structural diagram of the convolutional probabilistic neural network of this invention. Detailed Implementation

[0017] The following detailed description of the electroencephalogram (EEG) analysis method based on time-frequency complex networks and convolutional probabilistic neural networks of the present invention, with reference to embodiments and accompanying drawings, is provided in conjunction with the embodiments and accompanying drawings.

[0018] This invention presents an EEG analysis method based on time-frequency complex networks and convolutional probabilistic neural networks. It provides a method for analyzing EEG signals that integrates complex network theory, convolutional neural networks, and probabilistic neural networks, and can be used for tasks such as epilepsy prediction, sleep monitoring, and motor imagery recognition. EEG signals are characterized by weak amplitude, strong background noise, and nonlinearity, making it difficult to extract key task-related features and posing significant challenges to tasks such as classifying different brain states. The interactions between different brain regions within the complex brain system ensure the normal functioning of the human body. Complex networks are abstract and simplified representations of complex systems, effectively characterizing the topological interactions between different brain regions. Convolutional neural networks can automatically extract discriminative deep features from EEG signals; probabilistic neural networks offer advantages such as simple and fast classification processes, good fault tolerance, and strong interpretability. To achieve high-precision classification of EEG signals, this invention proposes an EEG signal analysis method that integrates the advantages of complex network theory, convolutional neural networks, and probabilistic neural networks. This invention first constructs a time-frequency complex network from multi-channel EEG signals acquired by an EEG signal acquisition device to uncover hidden brain functional activity features within the EEG signals. Then, the time-frequency complex network is input into a convolutional neural network (CNN) to extract discriminative high-level feature vectors. Finally, these high-level feature vectors are input into a probabilistic neural network (PN) to complete a classification task. The pattern layer of the PN is constructed from the high-level feature vectors extracted from the training set by the CNN, and the CNN and PN together constitute a convolutional probabilistic neural network. This invention integrates the brain topological dynamics representation capabilities of the time-frequency complex network, the deep feature extraction capabilities of the CNN, and the excellent classification capabilities of the PN, enabling effective classification of different EEG signals.

[0019] like Figure 1 As shown, the EEG analysis method based on time-frequency complex networks and convolutional probabilistic neural networks of the present invention includes the following steps:

[0020] 1) The multi-channel EEG signals acquired through the EEG signal acquisition device are divided into a training set and a test set. The training set contains... The test set contains [number] samples. One sample;

[0021] The training set is represented as The test set is represented as Any sample in the training set and the test set can be represented as ,in This represents the number of channels in a single sample. This indicates the number of sampling points per channel for a single sample. The sample No. The first in the channel The signal value at each sampling point.

[0022] 2) EEG signals contain rich information about brain states and exhibit obvious frequency-dependent characteristics for any sample in the training and test sets. Perform frequency band division to obtain frequency band set Multichannel EEG signals corresponding to four frequency bands;

[0023] The frequency band set The frequency ranges of each band are 0.5-3Hz, 4-7Hz, 8-13Hz, and 14-30Hz, respectively; for any sample in the training and test sets... After frequency band division, for any Obtain the frequency band The multichannel EEG signals below are represented as ,in This represents the number of channels in a single sample. This indicates the number of sampling points per channel for a single sample. Indicates sample Filtered frequency band The The first in the channel The signal value at each sampling point.

[0024] 3) Regarding the sample In frequency band set Multichannel EEG signals from different frequency bands were used to construct time-frequency subnetworks, resulting in a four-layer time-frequency subnetwork. And four network adjacency matrices corresponding one-to-one with the four-layer time-frequency sub-networks. ;include:

[0025] (3.1) For any sample In any frequency band Multichannel EEG signals Calculate every two channels and Pearson correlation coefficient between :

[0026] ;

[0027] in, and Representing the first and second samples respectively The and the first EEG signals from each channel, and Representing the first and second samples respectively The and the first The first in the channel The signal value at each sampling point Represents the first sample The and the first Covariance between EEG signals of each channel and Representing the first and second samples respectively The and the first The standard deviation of the EEG signals in each channel, This represents the number of channels in a single sample. This indicates the number of sampling points per channel for a single sample. and Representing the first and second samples respectively The and the first Mean of EEG signals from each channel, Pearson correlation coefficient The larger the value, the more likely the sample channel is to be affected. and channels The stronger the correlation between the EEG signals;

[0028] (3.2) Frequency band Each channel of the multi-channel EEG signal is defined as a node, and the Pearson correlation coefficient between any two channels of a node determines the connection, thus constructing a connection corresponding to the frequency band. Single-layer time-frequency subnetwork The corresponding network adjacency matrix Represented as:

[0029] ;

[0030] (3.3) For frequency band sets Steps (3.1) and (3.2) were repeated for multi-channel EEG signals of each frequency band to obtain a four-layer time-frequency subnetwork. and the corresponding four network adjacency matrices .

[0031] 4) The network adjacency matrix corresponding to the four-layer time-frequency sub-network The adjacency matrix of the fused single-layer network is obtained by adding the two matrices. And by the single-layer network adjacency matrix Characterizes the corresponding single-layer time-frequency complex network ;include:

[0032] (4.1) For any sample The four-layer time-frequency subnetwork Corresponding network adjacency matrix The adjacency matrix of the fused single-layer network is obtained by adding the two matrices. The Characterizes the corresponding single-layer time-frequency complex network ,in:

[0033] ;

[0034] (4.2) For the network adjacency matrix Sparsification: The set of edge weights after removing the main diagonal elements. Select the top 25% of the largest edge weights, and denote the corresponding threshold as . and will satisfy The edges with the highest weights are retained, while the weights of the remaining edges are set to 0, resulting in the sparsed adjacency matrix of the single-layer network. , For channel index, This indicates the number of channels in a single sample.

[0035] 5) Repeat steps 2) to 4) for each sample in both the training and test sets to obtain all samples in the training set. The set of single-layer adjacency matrices corresponding to each sample and all samples in the test set. The set of single-layer adjacency matrices corresponding to each sample;

[0036] 6) Add all training data to the training set. The single-layer adjacency matrix corresponding to each sample is input into a convolutional neural network (CNN) for model training. The main structure of the CNN includes two convolutional layers, two pooling layers, and two fully connected layers. Parameters are iteratively updated, and the training process stops when the model's accuracy meets a preset accuracy threshold. The preset accuracy threshold for stopping CNN training is... The convolutional neural network specifically includes:

[0037] (6.1) The first convolutional layer has a kernel size of... The activation function is the ReLU function;

[0038] (6.2) The first pooling layer is the max pooling layer;

[0039] (6.3) The second convolutional layer has a kernel size of... The activation function is the ReLU function;

[0040] (6.4) The second pooling layer is an average pooling layer;

[0041] (6.5) Flattened layer;

[0042] (6.6) The first fully connected layer uses the ReLU activation function and outputs a high-level feature vector;

[0043] (6.7) The second fully connected layer has the Softmax function as its activation function and is used to output the classification probability.

[0044] 7) Input the high-level feature vector output from the first fully connected layer of the convolutional neural network described in step 6) into the parallel probabilistic neural network, using the single-layer adjacency matrix of each sample in the training set, to construct the pattern layer of the parallel probabilistic neural network; the parallel probabilistic neural network includes at least a pattern layer, a summation layer, and an output layer, and the convolutional neural network and the parallel probabilistic neural network together constitute as follows: Figure 2 The convolutional probabilistic neural network shown; the parallel probabilistic neural network sequentially includes:

[0045] (7.1) Pattern layer: The input to the pattern layer is the high-level feature vector output by the first fully connected layer in the convolutional neural network. The number of neurons in the pattern layer is the same as the total number of samples in the training set. Each neuron stores a high-level feature vector of a training sample; for the , High-level feature vectors of each test set sample With the High-level feature vectors of training set samples The output of the mode layer is:

[0046] ;

[0047] in, Indicates the sample index of the test set. This represents the index of the training set samples. The dimension of the high-level feature vector. The smoothing parameters directly affect the classification performance and generalization ability of the concatenated probabilistic neural network. It is a Euclidean norm. Representing an exponential function, it maps the Euclidean distance to a... Within the range;

[0048] (7.2) Summation layer: The number of neurons in the summation layer is the same as the total number of categories, denoted as... ;No. The summing layer neuron corresponds to the first... kind The pattern layer output is aggregated by category to obtain the test sample belonging to the first category. Probability density estimate of the class:

[0049] ;

[0050] in, Indicates that it belongs to the first in the training set. The number of samples in each category Indicates belonging to the first The first category High-level feature vectors of training set samples;

[0051] (7.3) Output layer, number of neurons in the output layer and total number of categories In accordance with Bayesian classification rules, the output layer selects the class with the highest probability density estimate as the output class for the test sample.

[0052] ;

[0053] in, For the first The output categories of each test set sample.

[0054] 8) Add all test cases to the test set. The single-layer adjacency matrix corresponding to each sample is input into the convolutional probabilistic neural network for feature extraction and classification output.

[0055] The EEG analysis method based on time-frequency complex networks and convolutional probabilistic neural networks of the present invention can be applied to various multi-channel EEG signal analysis tasks, such as epilepsy prediction, sleep detection, and motor imagery recognition.

[0056] The above description of the present invention and its embodiments is not limited thereto. The description in the embodiments is only one of the implementation methods of the present invention. Without departing from the spirit of the present invention, any structure or embodiment similar to the technical solution designed without creativity shall be within the protection scope of the present invention.

Claims

1. A method for electroencephalogram analysis based on time-frequency complex network and convolutional probabilistic neural network, characterized in that, Comprising the following steps: 1) The multi-channel electroencephalogram signals obtained by the brain electrical signal acquisition device are divided into a training set and a test set, the training set contains samples, and the test set contains samples; 2) for any sample in the training set and the test set performing band division to obtain a band set multi-channel electroencephalogram signals corresponding to the four bands 3) for the sample In the frequency band set The multi-channel electroencephalogram under each frequency band is respectively constructed into a time-frequency sub-network, and four layers of time-frequency sub-networks are obtained And four network adjacency matrices corresponding to the four layers of time-frequency sub-networks ; 4) adding the network adjacency matrices corresponding to the four-layer time-frequency sub-networks to obtain the fused single-layer network adjacency matrix , and the single-layer network adjacency matrix characterizes the corresponding single-layer time-frequency complex network ; 5) repeating steps 2) to 4) for each sample in the training set and the test set respectively to obtain a set of single-layer adjacency matrices corresponding to all samples in the training set and a set of single-layer adjacency matrices corresponding to all samples in the test set; 5) repeating steps 2) to 4) for each sample in the training set and the test set respectively to obtain a set of single-layer adjacency matrices corresponding to all samples in the training set and a set of single-layer adjacency matrices corresponding to all samples in the test set;​ 6) Add all training data to the training set. The single-layer adjacency matrix corresponding to each sample is input into a convolutional neural network for model training. The main structure of the convolutional neural network includes two convolutional layers, two pooling layers, and two fully connected layers. The parameters are iteratively updated, and the training process of the convolutional neural network is stopped when the accuracy of the model meets the preset threshold requirement. 7) inputting the single-layer adjacency matrix of each sample in the training set into the high-level feature vector output by the first fully connected layer in the convolutional neural network in step 6), and inputting the high-level feature vector into the joint probabilistic neural network to build the pattern layer of the joint probabilistic neural network; the joint probabilistic neural network at least comprises a pattern layer, a summation layer and an output layer, and the convolutional neural network and the joint probabilistic neural network jointly constitute a convolutional probabilistic neural network; 8) inputting the single-layer adjacency matrix corresponding to all samples in the test set into the convolutional probabilistic neural network for feature extraction and classification output. 8) inputting the single-layer adjacency matrix corresponding to all samples in the test set into the convolutional probabilistic neural network for feature extraction and classification output.

2. The electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network according to claim 1, characterized in that, The training set mentioned in step 1) is represented as follows: The test set is represented as Any sample in the training set and the test set can be represented as ,in This represents the number of channels in a single sample. This indicates the number of sampling points per channel for a single sample. The sample No. The first in the channel The signal value at each sampling point.

3. The electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network according to claim 1, characterized in that, Step 2) the frequency band set The frequency ranges of the bands are 0.5-3 Hz, 4-7 Hz, 8-13 Hz and 14-30 Hz in turn; for any sample in the training set and the test set After the band division, for any sample The frequency band The multi-channel electroencephalogram signal of the sample is represented as , wherein represents the number of channels of a single sample, represents the number of sampling points of each channel of a single sample, represents the signal value of the i-th sampling point in the j-th channel of the sample After filtering, the signal value of the i-th sampling point in the j-th channel of the frequency band .​​ 4. The electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network according to claim 1, characterized in that, Step 3) comprises: (3.1) for each sample in any frequency band of the multichannel electroencephalogram , a Pearson correlation coefficient and between each two channels is calculated ； in, and Representing the first and second samples respectively The and the first EEG signals from each channel, and Representing the first and second samples respectively The and the first The first in the channel The signal value at each sampling point Represents the first sample The and the first Covariance between EEG signals of each channel and Representing the first and second samples respectively The and the first The standard deviation of the EEG signals in each channel This represents the number of channels in a single sample. This indicates the number of sampling points per channel for a single sample. and Representing the first and second samples respectively The and the first Mean of EEG signals from each channel, Pearson correlation coefficient The larger the value, the more likely the sample channel is to be affected. and channels The stronger the correlation between the EEG signals; (3.2) defining each channel of the multi-channel electroencephalogram signal under the frequency band as a node, determining an edge between each two nodes according to the Pearson correlation coefficient between the corresponding channels, and constructing a single-layer time-frequency sub-network corresponding to the frequency band , and a corresponding network adjacency matrix is represented as:​ ； (3.3) To the frequency band set The multi-channel electroencephalogram signals of each frequency band in the middle are repeated step (3.1) and step (3.2), and four-layer time-frequency sub-networks are obtained And the corresponding four network adjacency matrices .

5. The electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network according to claim 1, characterized in that, Step 4) comprises: (4.1) For any sample Four layers of time-frequency sub-networks The corresponding network adjacency matrix Add to get the fused single-layer network adjacency matrix , the Characterize the corresponding single-layer time-frequency complex network Wherein: ； (4.2) The network adjacency matrix Sparse processing: select the top 25% of edge weights in the edge weight set after removing the main diagonal elements , and the corresponding threshold is , and the edges that satisfy are retained, and the rest of the edge weights are set to 0, obtaining the sparse single-layer network adjacency matrix , is the channel index, indicates the number of channels of a single sample.

6. The electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network according to claim 1, characterized in that, Step 6) the convolutional neural network comprises: (6.1) a first convolutional layer with a convolution kernel size of and a ReLU function as the activation function; (6.2) the first pooling layer is a maximum pooling layer; (6.3) a second convolutional layer with a convolution kernel size of and an activation function of a ReLU function; (6.4) the second pooling layer is an average pooling layer; (6.5) the flattening layer; (6.6) the first fully connected layer, the activation function is ReLU function, and the output high-level feature vector; (6.7) the second fully connected layer, the activation function is Softmax function, and is used for outputting classification probability.

7. The electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network according to claim 1, characterized in that, Step 6) the preset accuracy threshold for stopping training of the convolutional neural network is %.

8. The electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network according to claim 1, characterized in that, Step 7) the joint probabilistic neural network comprises in sequence: (7.1) Pattern layer: The input to the pattern layer is the high-level feature vector output by the first fully connected layer in the convolutional neural network. The number of neurons in the pattern layer is the same as the total number of samples in the training set. Each neuron stores a high-level feature vector of a training sample; for the , High-level feature vectors of each test set sample With the High-level feature vectors of training set samples The output of the mode layer is: ； wherein, denotes the test set sample index, denotes the training set sample index, denotes the dimension of the high-level feature vector, is a smoothing parameter, directly affecting the classification performance and generalization ability of the neural network, is the Euclidean norm, denotes the exponential function, mapping the Euclidean distance into an interval; (7.2) a summation layer, the number of neurons of the summation layer is consistent with the total number of classes, denoted as ; the th summation layer neuron corresponds to the th class , the mode layer output is aggregated by class to obtain the probability density estimation value of the test sample belonging to the th class. ； in, Indicates that it belongs to the first in the training set. The number of samples in each category Indicates belonging to the first The first category High-level feature vectors of training set samples; (7.3) the output layer, the number of neurons of the output layer is equal to the total number of classes In accordance with the Bayesian classification rule, the output layer selects the class with the maximum probability density estimate as the output class of the test sample: ； wherein, is the output class for the th test set sample.

9. The electroencephalogram analysis method based on time-frequency complex network and convolutional probabilistic neural network according to claim 1, characterized in that: It can be applied to various multi-channel electroencephalogram signal analysis tasks, including epilepsy prediction, sleep detection, motor imagination recognition.

Images