An electroencephalogram attention level monitoring method based on a dual-domain deep supervision feature fusion network

Abstract

The application discloses a kind of based on double-domain deep supervision feature fusion network's electroencephalogram attention level monitoring method, comprising: collection electroencephalogram (Electroencephalography, EEG) and Electrooculography (Electrooculography, EOG) signal, form attention level database;Double-domain deep supervision feature fusion network is constructed, network data processing module analysis EEG signal time series data and spectral data, and PERCLOS value is calculated based on EOG signal as attention level quantitative index, construct time series dataset and spectral dataset;Time attention module extracts EEG signal time domain feature, and outputs main function loss value;Spectral residual module extracts EEG signal spectral feature, and outputs main function loss value;Deep supervision module receives and processes the output value from TransformEncoder and the first residual block, provides two auxiliary loss values, integrates all main loss values and auxiliary loss values, enhances the capturing ability of network to feature details;Feature fusion module monitors and evaluates pilot's attention level by linear regression algorithm.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of attention level calculation and prediction, and relates to an electroencephalogram attention level monitoring method based on a dual-domain deep supervision feature fusion network. BACKGROUND

[0002] With the rapid development of the aviation field, the complexity of flight tasks has significantly increased, and the challenges to flight safety have also intensified. Human factors are the main cause of flight accidents, and improper attention allocation caused by fatigue accumulation and decreased vigilance of pilots is an important inducement to accidents. Therefore, effectively monitoring the attention level of pilots and timely discovering and correcting potential attention deviation are crucial to reducing human errors and improving flight safety. At present, physiological signal monitoring technologies such as EEG and EOG capture the physiological response changes of pilots during the execution of flight tasks, providing a scientific basis for quantitative assessment of attention states.

[0003] However, existing attention monitoring methods are generally limited to one-sided analysis of a single physiological signal source and a single feature. This single-dimensional analysis method not only limits a comprehensive and in-depth understanding of physiological responses, but also directly affects the accuracy and reliability of attention evaluation results due to the lack of key information. SUMMARY

[0004] The purpose of the application is to solve the problem of lack of objective and accurate monitoring means for monitoring the attention level of pilots, and to provide an electroencephalogram attention level monitoring method based on a dual-domain deep supervision feature fusion network.

[0005] Technical scheme: In order to achieve the above purpose, the application provides an electroencephalogram attention level monitoring method based on a dual-domain deep supervision feature fusion network, comprising the following steps:

[0006] S1: Real-time capture the physiological response of pilots during the execution of tasks through high-precision EEG and EOG devices, collect EEG data and EOG data, and form an attention level database;

[0007] S2: Preprocess the obtained EEG data;

[0008] S3: Align the EOG data and the preprocessed EEG data according to the time stamp, select 4s as the time window, segment the continuous EOG data, and calculate the PERCLOS value in each time window as the quantitative indicator of attention level;

[0009] S4: Construct a dual-domain deep supervision feature fusion network, including a data processing module, a time attention module, a frequency spectrum residual module, a deep supervision module, and a feature fusion module;

[0010] S5: Process the EEG signal time series data and the EEG signal spectrum data respectively through the data processing module to obtain time series data and spectrum data;

[0011] S6: According to the PERCLOS value obtained in step S3 and the time series data and spectrum data obtained in step S5, construct a time series data set and a spectrum data set; each piece of time series data and spectrum data is aligned and labeled with the corresponding PERCLOS value to form a data-label pair required for machine learning, and the PERCLOS value is taken as a label and input into a regression network model to guide the learning and training of the model;

[0012] S7: For the obtained time series data set, the multi-head attention mechanism of the time attention module captures the key information in the EEG signal time series data, and outputs the main function loss loss temporal and the time domain feature F temporal ;

[0013] S8: For the obtained spectrum data set, the spectrum residual module deep mines high-order attention related information in the spectrum data by stacking multiple residual blocks, and outputs the main function loss loss spectral and the spectrum feature F spectral ;

[0014] S9: The output of the multi-head attention mechanism in step S7 and the output of the first residual block obtained in step A8 are respectively input into a depth supervision module, and auxiliary losses loss aux1 and loss aux2 are generated through two depth supervision sub-modules, combined with the main function loss loss temporal , loss spectral , and all main loss values and auxiliary loss values are integrated;

[0015] S10: According to the time domain feature F temporal and the spectrum feature F spectral , the attention level of the pilot is predicted through a linear regression algorithm.

[0016] S11: Input the time series data set and the spectrum data set into the DBPNet model, output the error with the real PERCLOS value, and evaluate the performance of the model through the mean squared error (MSE), the mean absolute error (MAE), and the correlation coefficient (CORR) index.

[0017] Further, the preprocessing of the EEG data in step S2 includes re-reference, down-sampling, and band-pass filtering, specifically:

[0018] A1: re-reference the EEG data, the formula is:

[0019]

[0020] wherein N represents the total number of electrodes, EEG i (t) represents the signal of the i-th electrode;

[0021] A2: downsample the EEG data, the formula is:

[0022] x'[n] = x[Mn]

[0023] wherein x'[n] is the downsampled sequence, x[Mn] is to take one sample every M samples, and M is the downsampling factor;

[0024] A3: Butterworth band-pass filter the EEG data, the formula is:

[0025]

[0026] wherein ω l and ω h are the normalized high and low frequency cutoff frequencies, respectively, and Q is the quality factor.

[0027] Further, the calculation method of the attention level quantization index in the step S3 is:

[0028] B1: calculate the PERCLOS value in each time window, the formula is:

[0029]

[0030] wherein CLOS represents the eye closure time, blink represents the eye blink time, fixation represents the gaze time, and saccade represents the saccade time;

[0031] B2: obtain the attention level quantization label Labels = {PERCLOS1, PERCLOS2,..., PERCLOS n}.

[0032] Further, in the data processing module of the step S5, the EEG signal time series data is processed by common space mode analysis, and the EEG signal spectrum data is processed by applying multi-band DE feature extraction and 2D topological map mapping technology, and the processing flow is specifically:

[0033] C1: analyze the EEG time series data {X1, X2,..., X N} by regression common space mode, and calculate the covariance matrix of each sample:

[0034]

[0035] Compute the correlation covariance matrix:

[0036]

[0037] where y i is the target variable value associated with the i-th sample;

[0038] Find the projection matrix that maximizes the objective function:

[0039]

[0040] where R is the average of all sample covariance matrices;

[0041] Generate regression task features:

[0042] Z i = W T X i ;

[0043] C2: Decompose the EEG signal into five bands, namely the delta band, theta band, alpha band, beta band and gamma band;

[0044] C3: For each frequency band, convert the time-domain signal to the frequency-domain signal by Fast Fourier Transform (FFT):

[0045]

[0046] where n is the time index and N is the number of samples;

[0047] Compute the PSD:

[0048]

[0049] Calculate the DE value using the normalized PSD:

[0050] DE = -∑ f∈band p[f]log(p[f])

[0051] where f is the band.

[0052] Further, the time attention module of the step S7 is composed of a TransformerEncoder, a convolution layer, an average pooling layer and a fully connected layer, and the processing flow is specifically:

[0053] D1: Capture the EEG time series data features by the Transformer Encoder:

[0054] EncoderOutput = TransformerEncoder(EEG temporal );

[0055] D2: Time attention module outputs time domain feature F temporal :

[0056]

[0057] Wherein, W is the weight of the full connection layer, and b is the bias of the full connection layer.

[0058] D3: Calculate the loss value of the time attention module output by the mean absolute error (Mean Squared Error, MAE) loss function:

[0059]

[0060] Wherein, y i is the true value, is the predicted value, and N is the total number of samples.

[0061] Further, the step S8 spectrum residual module is composed of a convolution layer, three multi-band fusion residual blocks and a pooling layer, and the specific processing process of the spectrum residual module is:

[0062] E1: Capture the EEG spectrum data features by the multi-band fusion residual block:

[0063] ResBlockOutput = σ (F (x) + x)

[0064] Wherein, x is the input of the residual block, F(x) is the internal operation of the residual block, and σ is the activation function.

[0065] E2: The spectrum residual module outputs the spectrum feature F spectral :

[0066] F spectral = Avg (ResBlock (ResBlock (ResBlock (Conv (x) )))

[0067] E3: Calculate the loss value of the spectrum residual module output by the MAE loss function:

[0068]

[0069] Wherein, y i is the true value, is the predicted value, and N is the total number of samples.

[0070] Further, the two deep supervision sub-modules in the deep supervision module of step S9 are respectively a time domain deep supervision sub-module and a frequency domain deep supervision sub-module, the time domain deep supervision sub-module is composed of a convolution layer and two fully connected layers, and the frequency domain deep supervision sub-module is composed of two convolution layers, an average pooling layer and a fully connected layer.

[0071] Further, the processing flow of the deep supervision module in step S9 is:

[0072] F1: directly receiving the TransformerEncoder output value EncoderOutput through the time domain deep supervision sub-module, and calculating the auxiliary loss loss aux1 :

[0073]

[0074] F2: directly receiving the first residual block output value ResBlockOutput through the frequency domain deep supervision sub-module, and calculating the auxiliary loss loss aux2 :

[0075]

[0076] F3: integrating the main function loss value and the auxiliary loss value to obtain the total loss value:

[0077] L total = loss temporal + loss spectral + λ1loss aux1 + λ2loss aux2 .

[0078] Further, in step S10, the time domain feature F temporal and the frequency domain feature F spectral are obtained, the features are fused by connection, and the fused feature is expressed as: F = [F temporal , F spectral ].

[0079] The method of the present application is based on simulating an aircraft cockpit scene, designing flight task scenes with different difficulty gradients to induce changes in pilot attention levels, and synchronously collecting high-precision EEG and EOG signals to form an attention level database; a dual-domain deep supervision feature fusion network is constructed, the data processing module in the network analyzes EEG signal time series data and spectral data, and calculates the PERCLOS value based on the EOG signal as an attention level quantification index to construct a time series dataset and a spectral dataset; the time attention module extracts EEG signal time domain features and outputs the main function loss value; the spectral residual module extracts EEG signal spectral features and outputs the main function loss value; the deep supervision module receives and processes the output values from the TransformerEncoder and the first residual block, provides two auxiliary loss values, integrates all main loss values and auxiliary loss values, and enhances the network's ability to capture feature details; the feature fusion module monitors and evaluates the pilot's attention level through a linear regression algorithm.

[0080] In order to realize accurate monitoring and evaluation of pilot attention level, the present application proposes a brain electrical attention level monitoring method based on dual-domain deep supervision feature fusion network. The method forms a multi-modal attention level database by collecting EEG and eye movement signals; constructs a time attention module and a spectral residual module to extract time domain and frequency domain features of EEG data respectively, and calculates the total loss value; the deep supervision module further extracts local feature information and outputs auxiliary loss value; the feature fusion module fuses time domain features and frequency domain features through linear regression algorithm; the model is optimized based on total loss value and auxiliary loss value, and finally realizes monitoring and evaluation of pilot attention level.

[0081] Advantages: Compared with the prior art, the method proposed by the present application uses dual-domain deep supervision feature fusion network based on multi-modal data input of EEG and EOG, pays attention to global time series features through time attention module, and obtains more information from frequency domain features through spectral residual module, which improves the model's ability to capture details and the richness of feature expression. The deep supervision module optimizes the combination of main loss and auxiliary loss, enhances the model's feature learning ability and generalization ability, so that the model can effectively adapt to data changes in complex task environment, improves the accuracy of the model in predicting attention level, and is suitable for attention monitoring and evaluation of pilots, drivers and other complex operation environments. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 The flowchart of the method of the present application;

[0083] Figure 2 The architecture diagram of the dual-domain deep supervision feature fusion network proposed in the present application;

[0084] Figure 3Evaluation index values of different models in the cross-subject condition in the embodiment;

[0085] Figure 4 Evaluation index values of different models in the within-subject condition in the embodiment. DETAILED DESCRIPTION

[0086] The application will be further clarified by the following examples and drawings, which should not be interpreted as limiting the scope of the application. After reading the application, those skilled in the art will be able to modify various equivalent forms of the application, which are within the scope of the appended claims.

[0087] The application provides an electroencephalogram attention level monitoring method based on a dual-domain deep supervision feature fusion network, as shown in Figure 1 The method comprises the following steps:

[0088] S1: Based on the aircraft cockpit simulation platform, different flight task scenarios are designed, and the scenarios are arranged in ascending order of difficulty to induce the attention changes of pilots under different stress levels. Through high-precision EEG and EOG devices, the physiological responses of pilots during task execution are captured in real time, and EEG data and EOG data are collected to form an attention level database.

[0089] S2: The obtained EEG data is preprocessed, and the preprocessing steps include re-reference, downsampling and band-pass filtering, aiming to improve the quality of EEG data and reduce noise interference. The specific calculation formula is:

[0090] A1: Re-reference reduces noise caused by improper selection of reference electrode position by reselecting reference electrode, and the formula is:

[0091]

[0092] Wherein, N represents the total number of electrodes, EEG i (t) represents the signal of the i-th electrode;

[0093] A2: Downsampling reduces the amount of calculation and data storage requirements, and improves the effective information amount of the signal. The formula is:

[0094] x'[n] = x[Mn]

[0095] Wherein, x'[n] is the sequence after downsampling, x[Mn] is to take one sample every M samples, and M is the downsampling factor.

[0096] A3: Band-pass filtering can effectively remove irrelevant frequency components, and retain the brain wave frequency band related to attention level. The EEG data is subjected to Butterworth band-pass filtering, and the formula is:

[0097]

[0098] Where, ω l and ω h are the normalized high and low frequency cutoff frequencies respectively, and Q is the quality factor.

[0099] S3: Align the EOG data and the preprocessed EEG data according to the time stamp, select 4s as the time window, segment the continuous EOG data, and calculate the PERCLOS value in each time window as the quantitative indicator of attention level;

[0100] The calculation method of the quantitative indicator of attention level is:

[0101] B1: Calculate the PERCLOS value in each time window, and the formula is:

[0102]

[0103] Where, CLOS represents the closed eye time, blink represents the blink time, fixation represents the fixation time, and saccade represents the saccade time;

[0104] B2: Obtain the attention level quantitative label Labels={PERCLOS1, PERCLOS2,..., PERCLOS n}.

[0105] S4: Construct a dual-domain deep supervision feature fusion network, as shown in Figure 2 The dual-domain deep supervision feature fusion network includes a data processing module, a time attention module, a spectral residual module, a deep supervision module and a feature fusion module; refer to Figure 2 Steps S5-S10 are executed;

[0106] S5: Process the EEG signal time series data and the EEG signal spectrum data through the data processing module respectively, and obtain the time series data and the spectrum data;

[0107] In the present application, the EEG signal time series data is analyzed and processed by common space mode, and the EEG signal spectrum data is processed by applying multi-band DE feature extraction and 2D topological map mapping technology, and the processing flow is as follows:

[0108] C1: Analyze the EEG time series data {X1, X2,..., X N} by regression common space mode, and calculate the covariance matrix of each sample:

[0109]

[0110] Calculate the correlation covariance matrix:

[0111]

[0112] where y i is the target variable value associated with the i-th sample;

[0113] Find the projection matrix that maximizes the objective function:

[0114]

[0115] where R is the average of all sample covariance matrices;

[0116] Generate regression task features:

[0117] Z i = W T X i ;

[0118] C2: Decompose the EEG signal into five wave bands, namely delta wave band, theta wave band, alpha wave band, beta wave band and gamma wave band;

[0119] C3: For each frequency wave band, convert the time domain signal to the frequency domain signal by fast Fourier transform (FFT):

[0120]

[0121] where n is the time index and N is the number of samples;

[0122] Calculate the power spectral density PSD:

[0123]

[0124] Calculate the DE value using the normalized PSD:

[0125] DE = -∑ f∈band p[f]log(p[f])

[0126] where f is the wave band.

[0127] S6: According to the PERCLOS value obtained in step S3 and the time series data and spectral data obtained in step S5, construct a time series data set and a spectral data set, each piece of time series data and spectral data is aligned and labeled with the corresponding PERCLOS value to form a data-label pair required for machine learning, and the PERCLOS value is taken as the label and input into the regression network model to guide the learning and training of the model;

[0128] S7: For the acquired time series data set, the key information in the EEG signal time series data is captured by the multi-head attention mechanism of the time attention module, and the main function loss loss is output temporal and the time domain feature F temporal ;

[0129] The time attention module is composed of TransformerEncoder, convolution layer, average pooling layer and full connection layer, and the processing process is as follows:

[0130] D1: Capture the EEG time series data features by TransformerEncoder:

[0131] EncoderOutput = TransformerEncoder(EEG temporal )

[0132] D2: The time attention module outputs the time domain feature F temporal :

[0133]

[0134] Where W is the weight of the full connection layer, and b is the bias of the full connection layer.

[0135] D3: Calculate the loss value of the time attention module output by the MAE loss function:

[0136]

[0137] Where y i is the true value, is the predicted value, and N is the total number of samples.

[0138] S8: For the acquired frequency spectrum data set, the frequency spectrum residual module deep mines high-order attention related information in the frequency spectrum data by stacking multiple residual blocks, and outputs the main function loss loss spectral and the frequency spectrum feature F spectral ;

[0139] The frequency spectrum residual module is composed of convolution layer, three multi-frequency band fusion residual blocks and pooling layer, and the specific processing process of the frequency spectrum residual module is as follows:

[0140] E1: Capture the EEG frequency spectrum data features by multi-frequency band fusion residual block:

[0141] ResBlockOutput = σ(F(x) + x)

[0142] Where x is the input of the residual block, F(x) is the internal operation of the residual block, and σ is the activation function.

[0143] E2: the spectral residual module outputs a spectral feature F spectral :

[0144] F spectral = Avg(ResBlock(ResBlock(ResBlock(Conv(x)))));

[0145] E3: calculate the loss value of the spectral residual module output by the mean absolute error loss function:

[0146]

[0147] where y i is the true value, is the predicted value, and N is the total number of samples.

[0148] S9: input the output of the multi-head attention mechanism in step S7 and the output of the first residual block obtained in step A8 into the deep supervision module, respectively, to generate auxiliary losses loss aux1 and loss aux2 , combine the main function losses loss temporal , loss spectral , and integrate all main loss values and auxiliary loss values.

[0149] The two deep supervision sub-modules in the deep supervision module are a time domain deep supervision sub-module and a frequency domain deep supervision sub-module. The time domain deep supervision sub-module is composed of a convolution layer and two fully connected layers, and the frequency domain deep supervision sub-module is composed of two convolution layers, an average pooling layer, and a fully connected layer.

[0150] The processing flow of the deep supervision module is as follows:

[0151] F1: directly receive the Transformer Encoder output value EncoderOutput through the time domain deep supervision sub-module, and calculate the auxiliary loss loss aux1 :

[0152]

[0153] F2: directly receive the first residual block output value ResBlockOutput through the frequency domain deep supervision sub-module, and calculate the auxiliary loss loss aux2 :

[0154]

[0155] F3: integrate the main function loss value and the auxiliary loss value to obtain the total loss value L total, the model can improve the sensitivity to local details of EEG data while maintaining the understanding of global information, improve the accuracy and robustness of attention level prediction, and calculate the total loss value L total :

[0156] L total = loss temporal + loss spectral + lambda1 * loss aux1 + lambda2 * loss aux2

[0157] S10: Obtain time domain features F temporal and frequency domain features F spectral The feature fusion module fuses the features through connection, and the fused features are expressed as F = [F temporal , F spectral ], and the attention level of the pilot is predicted through a linear regression algorithm.

[0158] S11: Input the time series data set and the spectrum data set into the DBPNet model, output the error with the real PERCLOS value, and evaluate the performance of the model through the MSE, MAE and CORR indicators.

[0159] In order to verify the effectiveness and effect of the method of the application, the method of the application is compared with the existing attention monitoring method in the embodiment, which is as follows:

[0160] In the embodiment, the prediction performance of different models under cross-subject conditions is evaluated, and Table 1 lists the evaluation index values of different models under cross-subject conditions, Figure 3 which intuitively shows the index values of each model, wherein (a) is the MAE value of different models, (b) is the RMSE value of different models, and (c) is the CORR value of different models. In the embodiment, all models are repeatedly run five times under the same hyperparameter configuration, and the mean and standard deviation of each index are obtained. Compared with the optimal values of the prediction performance of the classic EEGNet model and the improved EEGNet model, the average values of MAE and RMSE of the DSFNet model are reduced by 0.0482 and 0.0425 respectively, and the average value of CORR is increased by 0.6154, which verifies that the DSFNet model provided by the application can effectively predict the attention level state under cross-subject conditions.

[0161] Table 1 Comparison of results of different models under cross-subject conditions

[0162] Method MAE RMSE CORR EEGNet 0.1380 0.1736 0.0261 SE_EEGNet 0.1356 0.1705 -0.0035 ECA_EEGNet 0.1344 0.1701 -0.0111 SK_EEGNet 0.1355 0.1703 0.0400 FC_SK_EEGNet 0.1336 0.1674 0.0279 DBPNet 0.0854 0.1249 0.6043

[0163] At the same time, in the embodiment, the data of a subject is selected to evaluate the prediction performance of different models under intra-subject conditions. Table 2 lists the evaluation index values of different models under intra-subject conditions,Figure 4 The values of the indicators of each model are directly shown, wherein (a) is the MAE value of different models, (b) is the RMSE value of different models, and (c) is the CORR value of different models. The average values of MAE and RMSE of the DSFNet model are reduced by 0.0182 and 0.0183 respectively compared with the optimal values in other models, and the average value of CORR is increased by 0.0676.

[0164] Table 2 Comparison of different model results in subjects

[0165] Method MAE RMSE CORR EEGNet 0.1257 0.1562 -0.0086 SE_EEGNet 0.1253 0.1557 -0.0127 ECA_EEGNet 0.1123 0.1430 -0.0145 SK_EEGNet 0.1046 0.1297 -0.0101 FC_SK_EEGNet 0.1029 0.1280 -0.0220 DBPNet 0.0847 0.1097 0.0590

Claims

1. A dual-domain deep supervision feature fusion network-based electroencephalogram attention level monitoring method, characterized in that, The method comprises the following steps: S1: Real-time capture physiological response of the pilot during task execution through EEG and EOG devices, collect EEG data and EOG data, and form an attention level database; S2: Preprocess the obtained EEG data; S3: Align the EOG data and the preprocessed EEG data according to timestamps, select a time window, segment the continuous EOG data, and calculate the PERCLOS value in each time window as a quantitative indicator of the attention level; S4: Construct a dual-domain deep supervision feature fusion network, including a data processing module, a time attention module, a frequency spectrum residual module, a deep supervision module and a feature fusion module; S5: Process the EEG signal time series data and the EEG signal frequency spectrum data through the data processing module to obtain time series data and frequency spectrum data; S6: According to the PERCLOS value obtained in step S3 and the time series data and frequency spectrum data obtained in step S5, construct a time series data set and a frequency spectrum data set; S7: For the acquired time series data set, the key information in the EEG signal time series data is captured through the multi-head attention mechanism of the time attention module, and the main function loss loss is output temporal and time domain features F temporal ; S8: For the obtained spectrum data set, the spectrum residual module deeply mines high-order attention related information in the spectrum data by stacking multiple residual blocks, and outputs a main function loss loss spectral and the spectrum feature F spectral ; S9: Input the output of the multi-head attention mechanism in step S7 and the output of the first residual block obtained in step A8 into the deep supervision module respectively, and generate auxiliary loss through the two deep supervision submodules. aux1 and loss aux2 , combined with the main function loss loss temporal 、loss spectral , integrate all main loss values ​​and auxiliary loss values; S10: predicting the attention level of the pilot based on the time domain feature F temporal and the frequency domain feature F spectral by a linear regression algorithm.

2. The electroencephalogram attention level monitoring method based on the dual-domain deep supervision feature fusion network according to claim 1, characterized in that, The preprocessing of the EEG data in step S2 includes re-reference, downsampling and band-pass filtering, specifically: A1: Re-reference the EEG data, the formula is: where N represents the total number of electrodes, EEG i (t) represents the signal of the i-th electrode; A2: Downsample the EEG data, the formula is: x'[n] = x[Mn] Where x'[n] is the downsampled sequence, x[Mn] is to take one sample every M samples, and M is the downsampling factor; A3: Butterworth band-pass filter the EEG data, the formula is: where ω l and ω h are the normalized high and low frequency cutoff frequencies, respectively, and Q is the quality factor.

3. The electroencephalogram attention level monitoring method based on the dual-domain deep supervision feature fusion network according to claim 1, characterized in that, The calculation method of the quantitative indicator of the attention level in step S3 is: B1: Calculate the PERCLOS value in each time window, the formula is: Where CLOS represents the closed eye time, blink represents the blink time, fixation represents the fixation time, and saccade represents the saccade time; B2: Obtain attention level quantization labels Labels = {PERCLOS1, PERCLOS2,..., PERCLOS n}.

4. The electroencephalogram attention level monitoring method based on the dual-domain deep supervision feature fusion network according to claim 1, characterized in that, In the data processing module of step S5, the EEG signal time series data is processed by common space pattern analysis, and the EEG signal frequency spectrum data is processed by multi-band differential entropy (DE) feature extraction and 2D topological map mapping technology, and the processing flow is specifically: C1 : Compute the covariance matrix for each sample by regressing the EEG time series data {X1, X2,..., Xn} using the regression co-space mode analysis N} Calculate the correlation covariance matrix: where y i is the target variable value associated with the ith sample; Find the projection matrix that maximizes the objective function: Where R is the average of all sample covariance matrices; Generate regression task features: Z i = W T X i ; C2: Decompose the EEG signal into five bands, namely delta band, theta band, alpha band, beta band and gamma band; C3: For each frequency band, convert the time domain signal to the frequency domain signal by fast Fourier transform: Where n is the time index, and N is the sample number; Calculate the power spectral density (PSD): Calculate the DE value using the normalized PSD: DE = -∑ f∈band p[f] log(p[f]) Where f is the band.

5. The electroencephalogram attention level monitoring method based on the dual-domain deep supervision feature fusion network according to claim 1, characterized in that, The time attention module of step S7 is composed of TransformerEncoder, convolution layer, average pooling layer and full connection layer, and the processing flow is specifically: D1: Capture the EEG time series data features through TransformerEncoder: EncoderOutput = TransformerEncoder(EEG temporal ); D2: temporal attention module output time domain feature F temporal : Wherein, W is the weight of the full connection layer, b is the bias of the full connection layer; D3: Calculate the loss value of the time attention module output by the mean absolute error loss function: where y i is the true value, is the predicted value, and N is the total number of samples.

6. The electroencephalogram attention level monitoring method based on the dual-domain deep supervision feature fusion network according to claim 5, characterized in that, The spectral residual module of step S8 is composed of a convolution layer, three multi-band fusion residual blocks and a pooling layer, and the specific processing procedure of the spectral residual module is: E1: Capture the EEG spectral data features by the multi-band fusion residual block: ResBlockOutput = σ (F (x) + x) Wherein, x is the input of the residual block, F (x) is the internal operation of the residual block, and σ is the activation function; E2: The spectral residual module outputs the spectral feature F spectral : F spectral = Avg(ResBlock(ResBlock(ResBlock(Conv(x))))) ; E3: Calculate the loss value of the spectral residual module output by the mean absolute error loss function: Among them, y i is the true value, is the predicted value, and N is the total number of samples.

7. The electroencephalogram attention level monitoring method based on the dual-domain deep supervision feature fusion network according to claim 1, characterized in that, The two deep supervision submodules in the deep supervision module of step S9 are respectively the time domain deep supervision submodule and the frequency domain deep supervision submodule, the time domain deep supervision submodule is composed of a convolution layer and two full connection layers, and the frequency domain deep supervision submodule is composed of two convolution layers, an average pooling layer and a full connection layer.

8. The electroencephalogram attention level monitoring method based on the dual-domain deep supervision feature fusion network according to claim 7, characterized in that, The processing procedure of the deep supervision module in step S9 is: F1: directly receive the output value EncoderOutput of the TransformEncoder by the time domain deep supervision submodule, calculate the auxiliary loss loss aux1 : F2: directly receiving the first residual block output value ResBlockOutput by the frequency domain deep supervision submodule, calculating the auxiliary loss loss aux2 : F3: Integrate the main function loss value and the auxiliary loss value to obtain the total loss value: L total = loss temporal + loss spectral + λ1loss aux1 + λ2loss aux2 .

9. The electroencephalogram attention level monitoring method based on the dual-domain deep supervision feature fusion network according to claim 1, characterized in that, The time domain feature F obtained in the step S10 temporal and the frequency domain feature F spectral are fused by connection, and the fused feature is expressed as F=[F temporal ,F spectral ].

Images