A driving fatigue state detection method based on a core brain network and tensor decomposition

Abstract

The application discloses a driving fatigue state detection method based on a core brain network and tensor decomposition. The application innovatively combines correlation analysis and graph theory centrality principle as the basis for judging the importance of nodes, considers the change of the functional network with the mental state, and determines the core brain network affecting the mental state by retaining key nodes and corresponding edges. In view of the problem that the correlation between the brain network characteristics of different frequency bands of the multi-layer brain network cannot be effectively mined, the application uses the brain network data as a high-order tensor, uses a continuous low-rank non-negative Tucker decomposition algorithm, uses a tensor decomposition method to retain the mutual relationship between the brain networks of different frequency bands, and quickly and efficiently extracts the classification features of the multi-layer core brain network.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electroencephalogram signal processing, and particularly relates to a driving fatigue state detection method based on a core brain network and tensor decomposition. BACKGROUND

[0002] A driver needs to maintain high vigilance during driving, and the brain in a high vigilance state for a long time will cause mental fatigue. Mental fatigue will make the driver's reaction time longer, which may cause traffic accidents and casualties. Therefore, fatigue driving monitoring is a very important research topic and has high application value. Electroencephalogram (EEG) is a direct manifestation of brain activity, has good time resolution and the advantage of non-invasive signal acquisition, and therefore is considered an important basis for fatigue assessment.

[0003] Traditional fatigue assessment based on electroencephalogram directly extracts time-frequency features in electroencephalogram signals, such as power spectrum, cross entropy, chaos entropy and sample entropy. However, the above method can only analyze the signals collected by the local brain area of the brain, and cannot reflect the functional connection between different brain areas and describe the overall functional state of the brain. The brain often needs to complete a task in coordination with multiple regions, and there is dynamic interaction between different regions. Researchers set the electrodes of EEG as nodes and the functional connections between electrode pairs as edges of the network. These functional connections can be represented by different estimation techniques, such as correlation coefficient, phase synchronization, etc. This functional connection-based brain network analysis method is widely used in detecting diseases and mental states. At present, brain function network graph analysis is widely used in the field of fatigue detection. Researchers analyze the brain network through graph analysis and find that there are significant differences in the topological characteristics of the brain network between the fatigue state and the wakeful state.

[0004] The large amount of redundant information and noise in the nodes and edges of the brain function network will cause the detection performance to decrease, and therefore it is necessary to extract key regions to simplify the functional brain network. In EEG brain network analysis, selecting a suitable core network can make signal analysis simple and efficient, and is conducive to finding regions that have an important impact on the overall brain network, exploring key regions that have a greater impact on the overall brain network, and selecting a core brain network to simplify the brain function network and quickly and efficiently obtain the mental state of the brain, which helps to build a high-precision, high-speed electroencephalogram-based driving fatigue detection system.

[0005] There are two points worth discussing in the classification using the core brain network, the first point is how to extract the effective core network from the complete brain network, and the second point is how to extract the appropriate features from the core network for classification. For how to extract the effective core network, researchers based on the centrality theory in network science find the key electrodes with high centrality, and based on these electrodes, the sub-network composed of key electrodes and corresponding edges is retained for classification. However, these methods are for all samples of a certain state or all samples of two states, and the selection of key electrodes lacks research on electrodes that affect the change of mental state. Therefore, the present application focuses on the key changes of network evolution channels, and proposes a new driving fatigue detection method based on core brain function network by combining correlation analysis and graph centrality principle. For the problem of selecting core network and extracting appropriate features for classification, the method of using graph theory features of brain network to detect mental state of brain is more common at present. However, when combining the graph theory features of different frequency bands, the multi-layer brain network features are linearly spliced, and a method should be selected to link the brain network features of different frequency bands and mine the potential information between different frequency bands. The present application proposes a tensor decomposition feature extraction method based on multi-layer brain network to overcome the shortcoming of linear splicing of graph theory features of different frequency band brain networks. The brain network features are high-order tensors composed of channel x channel x frequency band, which cover multiple structural information. Through tensor decomposition, not only the channel and frequency band information can be retained, but also the potential relationship between different frequency band brain networks can be further mined. SUMMARY

[0006] The purpose of the present application is to solve two problems, one is the problem of information redundancy and a large amount of noise in multi-layer brain network, and the other is the problem that the correlation between brain network features of different frequency bands in multi-layer brain network cannot be effectively mined. A driving fatigue state detection method based on core brain network and tensor decomposition is proposed. For the problem of information redundancy and a large amount of noise in multi-layer brain network, the present application obtains the core network for rapid and efficient classification by retaining the key nodes and corresponding edges of the brain network. The present application innovatively combines correlation analysis and graph centrality principle as the basis for judging the importance of nodes, considers the change of functional network with mental state, and determines the core brain network affecting mental state by retaining key nodes and corresponding edges. For the problem that the correlation between brain network features of different frequency bands in multi-layer brain network cannot be effectively mined, the present application uses continuous low-rank non-negative Tucker decomposition algorithm to retain the mutual relationship between different frequency band brain networks by using tensor decomposition method, and quickly and efficiently extracts the classification features of multi-layer core brain network.

[0007] The present application comprises the following steps:

[0008] Step (1), electroencephalogram data set acquisition and preprocessing:

[0009] Step 1-1, obtaining electroencephalogram signals with labels of fatigue state and non-fatigue state;

[0010] As a preferred, the sampling frequency is selected as 500Hz.

[0011] Step 1-2, pre-processing the electroencephalogram signals with different labels;

[0012] As a preferred, the pre-processing includes filtering and removing artifacts.

[0013] Step 1-3, decomposing the pre-processed electroencephalogram signals into four frequency bands of δ (1-4Hz), θ (4-8Hz), α (8-13Hz) and β (13-30Hz).

[0014] Step 1-4, integrating all the decomposed electroencephalogram signal data into a three-dimensional matrix of channel number × sample point number × frequency band number, and labeling the task category, setting the label of alert state (i.e. non-fatigue state) as 0 and the label of fatigue state as 1; then taking a part of the data as a training set and the other part as a test set.

[0015] Step (2), brain network data set construction:

[0016] There are functional connections between different electrodes, and the Spearman correlation coefficient is used to calculate the correlation coefficient between different electrode signals, which corresponds to the functional connection strength between different electrodes; at the same time, the three-dimensional matrix of electroencephalogram signals processed in step 1-4 is converted into a four-dimensional brain network matrix of channel number × channel number × frequency band number × total sample amount.

[0017] Step (3), core brain network construction

[0018] Step 3-1, building a backbone brain network

[0019] The backbone brain network includes Nn single-layer brain networks, Nn = total sample amount × frequency band number; each single-layer brain network includes Mn nodes Node and edges Edge connecting adjacent nodes, Mn = channel number, the node represents the electrode, the edge represents the functional connection between adjacent nodes, and the size of the edge is the functional connection strength;

[0020] Step 3-2, selecting key nodes in the training set backbone brain network, and finally retaining the key nodes and corresponding edges in the whole data set backbone brain network to form a core network.

[0021] Step 3-2-1, obtaining the first key degree Nv1 of the node according to the correlation between degree centrality and fatigue degree; specifically:

[0022] 1) Calculate the node degree centrality of each sample in each frequency band of the training set backbone brain network, the degree centrality DC(i) of node i:

[0023] DC(i) = å j∈G aij (1)

[0024] Where aij represents the strength of the edge between node i and adjacent node j, and G represents the single-layer brain network in which node i is located. The degree centrality DC(i) of node Ni in the same frequency band of all samples is expanded into a vector Vn1 with a length of the total number of samples, and the fatigue degree corresponding to all samples is expanded into a vector Vn2 with a length of the total number of samples. The Spearman correlation coefficient of the two vectors Vn1 and Vn2 is calculated, and the absolute value of the correlation coefficient represents the importance of node Ni in the current frequency band. The importance of the node in the four frequency bands is added as the first key degree Nv1 of the node. After calculating the first key degree Nv1 of all nodes, the Nv1 of each node is updated to the proportion of the first key degree of the node in the total sum of the first key degrees of all nodes.

[0025] Step 3-2-2, according to the correlation between functional connection and fatigue degree, the second key degree Nv2 of the node is obtained; specifically:

[0026] 1) In the same frequency band of the training set, in the same single-layer brain network, the strength of the functional connection Eij between node Ni and adjacent node Nj of all samples is expanded into a vector Ve1 with a length of the total number of samples, and the fatigue degree corresponding to all samples is expanded into a vector Ve2 with a length of the total number of samples. The Spearman correlation coefficient of the two vectors Ve1 and Ve2 is calculated, and the absolute value of the correlation coefficient represents the importance of the functional connection Eij in the current frequency band. Repeat the above operation to obtain the correlation coefficient of all functional connection strengths and fatigue degrees.

[0027] 2) In the same frequency band and in the same single-layer brain network, the electrode is taken as the node, and the absolute value of the correlation coefficient of the functional connection strength and the fatigue degree is taken as the edge to construct the corresponding correlation degree graph Gc of the functional connection strength and the fatigue degree.

[0028] 3) Calculate the degree centrality DCC(i) of the correlation degree graph according to formula (2):

[0029] DCC(i) = å j∈Gc bij (2)

[0030] where bij represents the absolute value of the correlation coefficient between the functional connection strength and the fatigue degree between node i and adjacent node j, the degree centrality directly reflects the change degree of the importance of the node with the fatigue degree, and the degree centrality of the correlation degree graph is taken as the importance of the node in the frequency band, and the importance of the node in the four frequency bands is added as the second key degree Nv2 of the node; after the second key degree Nv2 of all nodes is calculated, the Nv2 of each node is updated as the proportion of the second key degree of the node in the total sum of the second key degrees of all nodes;

[0031] Step 3-2-3, combining the updated node Nv1 and Nv2 of step 3-2-1 and step 3-2-2, obtaining the final key degree of the node, and then selecting the core network, specifically:

[0032] 1) The first key degree Nv1 and the second key degree Nv2 of the node are combined according to formula (3) to obtain the final key degree Nv of the node:

[0033] Nv=aNv1+(1-a)Nv2 (3)

[0034] Wherein a represents the weight, 0≤a≤1.

[0035] 2) The key degrees Nv of all nodes are arranged in descending order, and the first k key degrees corresponding nodes are selected and put into the key node set Nim; then all nodes in the key node set Nim in the training set and the test set and the edges connecting these nodes are retained to form the core brain network Nc;

[0036] Step (4), feature extraction of the core brain network Nc 4-1 The edges of the core brain network Nc in the training set are taken as the original tensor data Y1, and the format of Y1 is I1×I2×I3×I4, that is, the number of channels×the number of channels×the number of frequency bands×the total number of samples. According to formula (4), the continuous low-rank non-negative Tucker decomposition algorithm is used to obtain the approximate factor matrices 1A(1), 2A(2), 3A(3), and 4A(4), with sizes of I1×R1, I2×R2, I3×R3, and I4×R4, respectively;

[0037] Y1≈CT1×1A(1)×2A(2)×3A(3)×4A(4) (4)

[0038] Wherein CT1 represents the core tensor obtained by decomposing Y1;

[0039]

[0040] The mathematical expression of the non-negative Tucker decomposition algorithm is as formula 5, Y represents the original tensor data, 1A(1), 2A(2), N A(T) represents the factor matrix of the corresponding dimension.

[0041] 4-2 Apply the approximate factor matrices 1A(1), 2A(2), 3A(3) obtained from the training set to the core brain network in the training set and the test set, and the edges of the core brain network of each sample are taken as the original tensor data Y2, the format of Y2 is I1xI2xI3, i.e. the number of channels x the number of channels x the number of frequency bands, Y2 is directly projected into the feature subspace composed of the three approximate factor matrices 1A(1), 2A(2), 3A(3), and the core tensor CT2 of the current sample core brain network is updated in reverse according to formula (6), and the size is R1xR2xR3, which is taken as the tensor feature of the core brain network of each sample.

[0042] Y2≈CT2x1A(1)x2A(2)x3A(3) (6)

[0043] Step (5). The tensor feature CT2 of the core network is input into the classifier for fatigue degree classification. The classifier is a support vector machine classifier based on a Gaussian kernel function, and the core tensor obtained by decomposing the core brain network of each sample in step (4) is taken as the feature input into the classifier, and the corresponding accuracy is obtained, and the experimental results are based on five-fold cross-validation.

[0044] Another object of the present application is to provide an electronic device comprising a processor and a memory, the memory storing machine executable instructions executable by the processor, and the processor executes the machine executable instructions to implement the above method.

[0045] The beneficial effects of the present application are:

[0046] 1) The present application proposes a fatigue driving detection method based on core brain network, which simplifies brain function network, quickly and efficiently obtains mental state of brain, and helps to construct high-precision, high-speed fatigue driving detection system based on electroencephalogram.

[0047] 2) The present application obtains the core network by retaining the key nodes and the corresponding edges, and selecting the core brain network can help us explore the influence of mental state change on the main area of the brain.

[0048] 3) The electroencephalogram signal is converted into multi-layer brain network data by using the Spearman correlation coefficient, the tensor decomposition is carried out by using the continuous low-rank non-negative Tucker decomposition algorithm, the core tensor is obtained for classification, and the classification accuracy of all nodes is 91.21%, which has excellent classification performance.

[0049] 4) The present application converts the electroencephalogram signal into multi-layer brain network data by using the Spearman correlation coefficient, extracts the core tensor by using the core network for classification, and when the number of nodes is 24, the accuracy reaches 91.83%, which is improved by 0.62% compared with using all nodes, and the calculation resources are less and the calculation rate is faster. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 Flow chart of the method of the present application;

[0051] Figure 2 Spearman correlation test diagram of the node centrality of the multi-layer brain network and the fatigue degree in the present application;

[0052] Figure 3 Spearman correlation test diagram of the functional connection strength of the multi-layer brain network and the fatigue degree in the present application;

[0053] Figure 4 Flow chart of the tensor decomposition experiment;

[0054] Figure 5 Tensor decomposition feature classification accuracy when using the top k nodes with the highest key degree to form a subnetwork, high first key degree represents the classification accuracy of the core network formed by the top k nodes with the highest first key degree, high second key degree represents the classification accuracy of the core network formed by the top k nodes with the highest second key degree, the abscissa is the number of selected nodes, and the ordinate is the classification accuracy;

[0055] Figure 6 Tensor decomposition feature classification accuracy when using the top 24 nodes with the highest key degree to form a subnetwork according to the final key degree, the abscissa is the proportion of the first key degree, and the ordinate is the classification accuracy. DETAILED DESCRIPTION

[0056] The driving fatigue state detection method based on the core brain network and the tensor decomposition of the present application will be described in detail below with reference to the accompanying drawings, see Figure 1 .

[0057] Step (1). EEG signal dataset acquisition and preprocessing.

[0058] Step 1.1 Acquire task category labels and build a dataset:

[0059] The present application adopts the open electroencephalogram dataset published by Cao Z et al. in 2019. The experiment adopts 30 electroencephalogram physiological signal channels and 2 reference electrodes. The subjects realize the sustained attention driving task in the simulated VR driving simulator. The participants are required to drive and keep the car in the center of the lane. Lane deviation events are randomly introduced, which make the car drift from the central lane to the left or right. The degree of drowsiness of the subject every time is measured by the local reaction time, which is defined by the interval length between the start of the car drift and the reaction of the participant. In addition, another index called global reaction time is introduced to measure the degree of drowsiness in a relatively long period - it is calculated by averaging the reaction time of all tests within a 90-second window before the deviation starts. This method considers individual performance by defining the baseline "alert reaction time" as the 5th percentile of the local reaction time throughout the session. When both the local and global reaction times are less than 1.5 times the alert reaction time, the test is marked as "alert", and when both the local and global reaction times are longer than 2.5 times the alert reaction time, the test is marked as "fatigue". All the above sample data are labeled with task category labels, and the label of the alert state is set to 0, and the label of the fatigue state is set to 1. The greater the value of the label value, the higher the degree of reaction fatigue. There are 11 subjects participating in the experiment in the dataset, and 1011 alert state and fatigue state samples are selected for each of the 11 subjects, ensuring that the number of samples selected for each subject in the two states is the same, and the total number of samples for each subject is greater than 100.

[0060] Step 1-2. Data preprocessing: The present study uses a pre-processed version of the data. The original electroencephalogram signal is filtered by a 1Hz high-pass and 50Hz low-pass finite impulse response (FIR) filter. For artifact rejection, obvious blink contamination is removed manually. Eye and muscle artifacts are removed by the automatic artifact rejection plugin provided in the EEGLAB software. The original data is down-sampled from 500Hz to 128Hz, and 3s long EEG samples are extracted before the deviation starts on each path.

[0061] Step 1-3. The signal of each channel is divided into theta (4-8Hz), alpha (8-14Hz), beta (14-31Hz), gamma (31-50Hz), and finally each electroencephalogram sample data is an electroencephalogram signal three-dimensional matrix with a size of 30x384x4 (number of electrodes x number of sample points x number of frequency bands).

[0062] Step 1-4. All the obtained electroencephalogram signal sample data are merged to obtain a series of electroencephalogram signal four-dimensional matrices with a size of 30x384x4x2022 (number of channels x number of sample points x number of frequency bands x total number of samples).

[0063] Step 1-5. Using 80% of the data as the training set and 20% of the data as the test set, the experimental results were obtained using five-fold cross-validation.

[0064] Step (2). Construction of brain network dataset:

[0065] Step 2-1. The functional connection strength between different electrode pairs was calculated using the Spearman correlation coefficient. The format of the electroencephalogram data in the sample was converted from a 30x384x4x2022 four-dimensional electroencephalogram signal matrix to a 30x30x4x2022 brain network four-dimensional matrix.

[0066] Step (3). Construction of the core network of the brain network

[0067] Step 3-1 Building the backbone brain network

[0068] The backbone brain network includes 2022x4 single-layer brain networks; each single-layer brain network includes 30 nodes Node and edges Edge connecting adjacent nodes, the node represents the electrode, the edge represents the functional connection between adjacent nodes, and the size of the edge is the functional connection strength;

[0069] Step 3-2 Selecting key nodes in the training set backbone network and retaining key nodes and corresponding edges in the entire dataset backbone brain network to form the core network.

[0070] Step 3-2-1 According to the correlation between the degree centrality and the fatigue degree of the training set, the first key degree Nv1 of the node is obtained by formula (1), Figure 2 To select the key node set Nim1 according to the correlation between the degree centrality and the fatigue degree of the training set. The degree centrality DC(i) of the node Ni in the same frequency band of all samples is expanded into a vector Vn1 with a length of the total number of samples, and the fatigue degree of the corresponding sample is expanded into a vector Vn2 with a length of the total number of samples. The Spearman correlation coefficient of the two vectors Vn1 and Vn2 is calculated, and the correlation between the node centrality and the fatigue degree of all frequency bands is calculated. The absolute value of the correlation coefficient represents the importance of the node Ni in the current frequency band. The importance of the node in the four frequency bands is added as the first key degree Nv1 of the node. After calculating the first key degree Nv1 of all nodes, the Nv1 of the node is updated to the proportion of the first key degree of the node to the total sum of the first key degrees of all nodes;

[0071] Step 3-2-2 According to the correlation between the functional connection and the fatigue degree of the training set, the second key degree Nv2 of the node is obtained.

[0072] Figure 3Spearman correlation test diagram for functional connection strength and fatigue degree of multi-layer brain network. In the training set, under the same single-layer brain network in the same frequency band, the strength of the functional connection Eij between all sample nodes Ni and adjacent nodes Nj is unfolded into a sample total vector Ve1, and the corresponding sample fatigue degree is unfolded into a sample total vector Ve2. The Spearman correlation coefficient of the two vectors Ve1 and Ve2 is calculated, and the absolute value of the correlation coefficient represents the importance of the functional connection Eij under the current frequency band.

[0073] In the same frequency band and under the same single-layer brain network, the electrodes are taken as nodes, and the correlation between the functional connection strength and the fatigue degree is taken as an edge to construct a functional connection strength and fatigue degree corresponding correlation degree graph.

[0074] The absolute value of the size of the edge in the correlation degree graph is taken, and the degree centrality of the correlation degree graph is calculated. The degree centrality directly reflects the change degree of the importance of the fatigue degree to the node. The degree centrality of the correlation degree graph is taken as the importance of the node in the frequency band. The importance of the node in the four frequency bands is added to obtain the second key degree Nv2 of the node. After the second key degree Nv2 of all nodes is calculated, the Nv2 of the node is updated to the proportion of the second key degree of the node in the total sum of the second key degrees of all nodes.

[0075] Step 3-2-3, combining the updated nodes Nv1 and Nv2 of steps 3-2-1 and 3-2-2 to obtain the final key degree of the node, and then selecting the core network.

[0076] Step 4. Feature extraction of the core brain network Nc of the whole data set

[0077] Figure 4 For the tensor decomposition flowchart, the edges of the core brain network Nc in the training set are taken as the original tensor data Y1, and the format of Y1 is 30x30x4x1616 (channel numberxchannel numberxfrequency band numberxsample total amount). The continuous low-rank non-negative Tucker decomposition algorithm is used for Y1 to obtain the corresponding approximate factor matrices 1A(1), 2A(2), 3A(3), and 4A(4), which have sizes of 30xR1, 30xR2, 4xR3, and 1616xR4, respectively.

[0078] Then the factor matrices obtained from the training set are applied to the edges of the core brain network of the whole data set. For the edges Y2 of the core brain network of each sample, the format of Y2 is 30x30x4 (channel numberxchannel numberxfrequency band number). Y2 is directly projected into the feature subspace composed of the three factor matrices 1A(1), 2A(2), and 3A(3) obtained above, and the core tensor CT2 (with a size of R1xR2xR3) of the core brain network of the sample is updated in reverse to serve as the tensor feature of the core brain network of each sample.

[0079] The core tensor size R1xR2xR3 needs to be determined, R1 and R2 are both derived from the number of channels, so let R1 equal to R2, the core size of the present application needs to be determined is R1xR1xR3. R1 takes 5-30, interval is 5, R3 takes 1-4, interval is 1. When the core size is 20x20x4, the classification accuracy of using full node network for tensor decomposition is the highest, which is 91.21%. Therefore, R1 takes 20, R2 takes 20, and R3 takes 4.

[0080] Step 5. Fatigue degree classification using core network

[0081] The classifier is a support vector machine classifier based on Gaussian kernel function, the core tensor obtained by decomposing the core brain network of each sample in step (4) is input into the classifier as a feature, and the corresponding accuracy is obtained. The experimental results are based on five-fold cross-validation.

[0082] From Figure 5 It can be seen that when the number of selected nodes is small, both the first key degree selected core network and the second key degree selected core network have good effect. When the number of nodes is more than 7, the accuracy is greater than 86%. It is proved that selecting appropriate key sub-networks helps to simplify brain function network and quickly and efficiently obtain brain mental state. Figure 6 It can be seen that when the number of nodes is 24, the combination ratio of the first key degree and the second key degree is 2:8, the accuracy is 91.83%, compared with 91.21% using all nodes, the accuracy is improved by 0.62%, and the number of nodes is reduced by 6.

Claims

1. A method for detecting driver fatigue state based on core brain networks and tensor decomposition, characterized in that... The method includes the following steps: Step (1), acquisition and preprocessing of EEG signal dataset: Step 1-1: Obtain EEG signals labeled as fatigued and non-fatigue states; Steps 1-2: Preprocess the EEG signals of different labels; Steps 1-3: Decompose the preprocessed EEG signal into four frequency bands: δ, θ, α, and β. Steps 1-4: Integrate all the decomposed EEG signal data into a three-dimensional EEG signal matrix with a size of number of channels × number of sample points × number of frequency bands, and label it with task category; then use a part of the data as the training set and the other part as the test set. Step (2), Brain network dataset construction: Functional connections exist between different electrodes. The Spearman correlation coefficient is used to calculate the correlation coefficient between the signals of different electrodes, which corresponds to the strength of the functional connection between different electrodes. At the same time, the three-dimensional matrix of EEG signals processed in steps 1-4 is transformed into a four-dimensional brain network matrix of EEG signals with channel number × channel number × frequency band number × total number of samples. Step (3): Constructing the core brain network Step 3-1: Build the backbone brain network The backbone brain network consists of Nn single-layer brain networks, where Nn = total number of samples × number of frequency bands; each single-layer brain network consists of Mn nodes and edges connecting adjacent nodes, where Mn = number of channels, nodes represent electrodes, edges represent functional connections between adjacent nodes, and the size of the edge is the functional connection strength. Step 3-2: Select key nodes in the backbone brain network of the training set, and finally retain key nodes and corresponding edges in the backbone brain network of the entire dataset to form the core network. Step 3-2-1: Based on the correlation between degree centrality and fatigue level, obtain the first criticality Nv1 of the node; specifically: 1) Calculate the node degree centrality of each sample in each frequency band of the backbone brain network of the training set, and the degree centrality DC(i) of node i: DC(i)=∑ j∈G aij (1) Where aij represents the strength of the edge between node i and its neighboring node j, and G represents the single-layer brain network in which node i is located; Expand the degree centrality DC(i) of node Ni in the same frequency band of all samples into a vector Vn1 of length equal to the total number of samples. Expand the fatigue degree corresponding to all samples into a vector Vn2 of the total number of samples. Calculate the Spearman correlation coefficient between these two vectors Vn1 and Vn2. The absolute value of this correlation coefficient represents the importance of node Ni in the current frequency band. Sum the importance of node Ni in the four frequency bands to obtain the first criticality Nv1 of node Ni. After calculating the first criticality Nv1 of all nodes, update the Nv1 of each node to the proportion of the first criticality of the node to the sum of the first criticalities of all nodes. Step 3-2-2: Based on the correlation between functional connectivity and fatigue level, obtain the second criticality level Nv2 of the node; specifically: 1) Within the same frequency band and the same single-layer brain network in the training set, the strength of the functional connection Eij between nodes Ni and their adjacent nodes Nj of all samples is expanded into a vector Ve1 representing the total number of samples. The fatigue level corresponding to all samples is expanded into a vector Ve2 representing the total number of samples. The Spearman correlation coefficient between these two vectors Ve1 and Ve2 is calculated. The absolute value of this correlation coefficient represents the importance of the functional connection Eij in the current frequency band. Repeat the above operation to obtain the correlation coefficient between the strength of all functional connections and the fatigue level. 2) Within the same frequency band and under the same single-layer brain network, electrodes are used as nodes, and the absolute values ​​of the correlation coefficients between functional connectivity strength and fatigue level are used as edges to construct a correlation graph Gc between functional connectivity strength and fatigue level. 3) Calculate the degree centrality DCC(i) of the association degree graph according to formula (2): DCC(i)=∑ j∈Gc at (2) Where bij represents the absolute value of the correlation coefficient between the functional connection strength and fatigue level between node i and its neighboring node j. This degree centrality directly reflects the degree to which fatigue level changes the importance of a node. The degree centrality of the correlation graph is used as the importance of the node in the current frequency band. The importance of the current node in the four frequency bands is added together as the second criticality Nv2 of the current node. After calculating the second criticality Nv2 of all nodes, the Nv2 of each node is updated to the proportion of the node's second criticality to the sum of the second criticalities of all nodes. Step 3-2-3: Combine the updated nodes Nv1 and Nv2 from steps 3-2-1 and 3-2-2 to obtain the final criticality of the nodes, and then select the core network. Specifically: 1) Combine the first criticality Nv1 and the second criticality Nv2 of the node according to formula (3) to obtain the final criticality Nv of the node: Nv=aNv1+(1-a)Nv2 (3) Where a represents the weight, 0≤a≤1; 2) Sort all nodes by their criticality level Nv in descending order, select the nodes corresponding to the top k criticality levels, and put them into the critical node set Nim; then retain all nodes in the critical node set Nim in the training set and the edges connecting these nodes to form the core brain network Nc. Step (4): Extract features from the core brain network Nc 4-1 The edges of the core brain network Nc in the training set are used as the original tensor data Y1. The format of Y1 is I1×I2×I3×I4, that is, the number of channels × the number of channels × the number of frequency bands × the total number of samples. According to the formula (4), the continuous low-rank non-negative Tucker decomposition algorithm is used on Y1 to obtain the approximate factor matrix 1A(1), 2A(2), 3A(3), 4A(4), with sizes of I1×R1, I2×R2, I3×R3, I4×R4 respectively. Y1≈CT1×1A(1)×2A(2)×3A(3)×4A(4) (4) Where CT1 represents the core tensor obtained from Y1 decomposition; 4-2 The approximate factor matrices 1A(1), 2A(2), and 3A(3) obtained from the training set are applied to the core brain network in the training and test sets. The edge of the core brain network of each sample is used as the original tensor data Y2. The format of Y2 is I1×I2×I3, that is, the number of channels × the number of channels × the number of frequency bands. Y2 is directly projected into the feature subspace composed of the three approximate factor matrices 1A(1), 2A(2), and 3A(3). The core tensor CT2 of the current sample core brain network is updated in reverse according to formula (5). The size is R1×R2×R3, which is used as the tensor feature of the core brain network of each sample. Y2≈CT2×1A(1)×2A(2)×3A(3) (5) Step (5): Input the tensor feature CT2 of the core network into the classifier to achieve fatigue level classification.

2. The method according to claim 1, characterized in that... Step 1-1: The sampling frequency is selected as 500Hz.

3. The method according to claim 1, characterized in that... The preprocessing described in steps 1-2 includes filtering and artifact removal.

4. The method according to claim 1, characterized in that... The classifier in step (5) is a support vector machine classifier based on the Gaussian kernel function.

5. An electronic device comprising a processor and a memory, the memory storing machine-executable instructions executable by the processor, the processor executing the machine-executable instructions to implement the method of any one of claims 1-4.

Images