Four-dimensional attention recognition method based on electroencephalogram feature fusion selection

Abstract

The invention provides a four-dimensional attention recognition method based on electroencephalogram feature fusion selection, and belongs to the technical field of electroencephalogram signal processing. Attention specifically comprises four specific qualities of distribution, breadth, stability and transfer, electroencephalogram signals and attention quality are combined, multi-domain electroencephalogram features are extracted to construct a fusion vector for four-dimensional attention recognition, and a ReliefF-SBE-L1 three-layer mixed feature selection model is provided to improve recognition accuracy and calculation efficiency. According to the method, efficient four-dimensional attention state recognition can be realized while the calculation complexity is remarkably reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of electroencephalogram signal processing, in particular to the field of attention recognition based on electroencephalogram signals, and specifically is a four-dimensional attention recognition method based on electroencephalogram feature fusion selection. Background Art

[0002] Attention refers to the ability of a person's mental activities to be directed and focused on something. It is real and one of the most thoroughly studied psychological phenomena. Most studies focus on attention in a narrow sense, using only a single dimension of attention quality to determine an individual's comprehensive attention level. In fact, attention is composed of a multi-dimensional structure and cannot be defined by a single concept. Its connotation should include the development of four specific qualities: 1) breadth of attention (referring to the number of objects that an individual can clearly perceive at the same time); 2) stability of attention (referring to the ability to maintain attention on a certain activity for a long time); 3) allocation of attention (referring to the ability of an individual to direct attention to different objects at the same time when engaged in two or more activities); 4) attention transfer (referring to the individual's active transfer of attention from one activity to another according to a certain purpose).

[0003] Attention recognition is often based on subjective scales, questionnaires, etc., but these methods cannot be objective and accurate. More and more scholars are focusing on brain science research, and EEG signals have also been proven to be one of the most accurate and efficient physiological bases for studying attention. By preprocessing, extracting and selecting features of EEG signals, and using machine learning classification algorithms, it is possible to identify and evaluate the attention level, thereby accurately grasping the state of attention.

[0004] Attention recognition based on EEG signals mainly considers two aspects: recognition accuracy and recognition complexity, that is, precision and efficiency. The effective information contained in a single feature is not rich enough to achieve a good classification and recognition effect. Therefore, it is necessary to extract multiple electrode channel parameters of multiple EEG features to obtain a higher recognition accuracy. However, too high a feature dimension will lead to a large redundancy in the model, greatly increasing the computational complexity and affecting the classification and recognition accuracy to a certain extent. It is easy to produce overfitting, resulting in poor generalization performance of the model. Summary of the invention

[0005] In order to address the shortcomings of existing research, the present invention provides a four-dimensional attention recognition method based on EEG feature fusion selection. In the EEG-based attention recognition task, a multi-domain feature fusion vector is constructed to improve the traditional single feature recognition method, and the "ReliefF-SBE-L1" three-layer hybrid feature selection model is introduced to overcome the problem of complex feature redundancy calculation, and for the four-dimensional attention states of allocation, breadth, stability and transfer, efficient recognition effect can be achieved.

[0006] The present invention provides a four-dimensional attention recognition method based on EEG feature fusion selection, comprising the following steps:

[0007] Step 1. Experimental paradigm design: Based on the development of attention quality in four dimensions, namely, attention allocation, attention breadth, attention stability, and attention shifting, design an experiment to induce four-dimensional attention;

[0008] Step 2, EEG signal acquisition: using EEG acquisition equipment to collect EEG signals while inducing four-dimensional attention;

[0009] Step 3: Signal preprocessing: Preprocess the original mixed EEG signal collected in step 2 based on the EEGLAB toolbox and Matlab software;

[0010] Step 4: Feature extraction and fusion: From the EEG signal processed in step 3, 16 features are extracted based on time domain, frequency domain and time-frequency domain comparison. Each feature contains the parameters of 17 electrode channels. The extracted feature parameters are fused to construct a multi-domain feature fusion vector.

[0011] Step 5: Feature selection and dimensionality reduction: Use the "ReliefF-SBE-L1" three-layer hybrid feature selection model to perform feature selection and dimensionality reduction on the multi-domain fusion feature vector obtained in step 4.

[0012] Step 6: SVM recognition: Import the multi-domain feature fusion vector obtained in step 5 into the SVM classifier to realize four-dimensional attention recognition.

[0013] Furthermore, in step 2, when collecting EEG signals, the experimental subjects, experimental environment and equipment requirements are as follows:

[0014] (1) Experimental subjects and environment: Graduate students or undergraduates aged between 19 and 25 years old, with normal or corrected vision and no history of major diseases were selected as subjects in a quiet and undisturbed environment;

[0015] (2) Experimental equipment: Select a multi-channel wet electrode EEG acquisition instrument (17 channels), equipped with conductive paste, syringe and 19-conductivity electrode cap.

[0016] Furthermore, the step 3 comprises the following steps:

[0017] Step 3.1, EEGLAB toolbox processing;

[0018] The EEGLAB toolbox was used for electrode positioning, 50 Hz power frequency interference was removed, and 1-40 Hz bandpass filtering was performed. The sampling rate was reduced to 256 Hz and bad segments were removed. The signal was re-referenced with bilateral mastoid electrodes at T3 and T4, and independent component analysis (ICA) was performed to remove artifact components such as electrooculography, electromyography, and electrocardiography.

[0019] Step 3.2, wavelet threshold denoising;

[0020] Wavelet threshold denoising includes wavelet decomposition, threshold processing, and wavelet reconstruction; db4 wavelet is used for denoising, and a fixed threshold calculation formula is used Where N is the signal length.

[0021] Step 3.3, sample set generation;

[0022] Samples are generated for the signal processed in step 3.2, and a time window with a time length of 5s and a step length of 2.5s is selected, that is, the EEG signal is segmented with a sliding window of 5s, with an overlap rate of 50%, to obtain a sample set.

[0023] Furthermore, step 4 includes the following steps:

[0024] Step 4.1: Obtain time domain features.

[0025] After analysis and comparison, the present invention selects the following statistical features and nonlinear features, a total of 8 time domain features:

[0026] (1) Standard deviation (Std), which is the arithmetic square root of the variance:

[0027] x(i) is the N-point signal sequence, is the mean of the signal sequence, and F is the standard deviation;

[0028] (2) Hjorth activity parameter (HA) refers to the variance of the EEG signal, i.e., the signal power, which indicates the width:

[0029] σ0 is the standard deviation of the signal, and HA is the Hjorth activity parameter.

[0030] (3) Hjorth mobility parameter (HM), used to estimate the average frequency of the EEG signal:

[0031] σ1 is the standard deviation of the first-order derivative of the signal, and HM is the Hjorth mobility parameter.

[0032] (4) Hjorth complexity parameter (HC), which refers to the mobility of the first-order derivative relative to the EEG itself:

[0033] σ2 is the standard deviation of the second-order derivative of the signal, and HC is the Hjorth complexity parameter.

[0034] (5) Lempel-Ziv complexity (LZC) refers to the direct measurement of data complexity without data decomposition or transformation, which is used to characterize the corresponding time series:

[0035] C N is the complexity, N is the number of cycles, and C is the normalized Lempel-Ziv complexity.

[0036] (6) Higuchi fractal dimension (HFD), which can be used to describe the complexity and self-similarity of signals;

[0037] The larger the value of Higuchi fractal dimension, the stronger the self-similarity of the signal and the more complex the signal. The present invention uses Higuchi algorithm to calculate fractal dimension, divides the original time series into k sub-segments and calculates the average value L(k) of the curve length of the sub-segments. According to ln L(k) With ln 1 / k The linear relationship between them is fitted into a straight line, and its slope is the Higuchi fractal dimension D.

[0038] (7) Permutation entropy (Pec), a nonlinear method used to detect the complexity (randomness) of time series or dynamic mutations, can quantitatively evaluate the random noise contained in the signal.

[0039] The smaller the value of permutation entropy, the simpler and more regular the signal is; the larger the value, the more complex and random the signal is. The basic principle algorithm flow is shown in Table 1 below:

[0040] Table 1 Principle of permutation entropy algorithm

[0041]

[0042]

[0043] (8) Sample entropy, which is improved from approximate entropy, is used to describe the complexity of time series and the probability of a new pattern when the dimension changes. The higher the complexity and the greater the probability, the greater the sample entropy value. Its algorithm is shown in Table 2 below:

[0044] Table 2 Sample entropy algorithm principle

[0045]

[0046]

[0047] The present invention selects parameter values ​​of m=2 and r=0.15σ to calculate sample entropy characteristics of 17 electrode channels.

[0048] Step 4.2: Obtain frequency domain features.

[0049] Frequency domain feature extraction requires separating the delta wave, theta wave, alpha wave, and beta wave from the original signal for analysis. Since the EEG signal is a non-stationary signal, the present invention uses wavelet packet decomposition to analyze the signal, taking into account both the low-frequency part and the high-frequency part for decomposition, and extracting the following frequency domain features:

[0050] (1) Energy and ratio;

[0051] The present invention calculates the square sum of the wavelet packet coefficients of different frequency bands after decomposition to obtain the energy of the corresponding frequency bands of δ wave, θ wave, α wave and β wave, and then calculates the energy ratios of the corresponding frequency bands of θ wave and α wave, θ wave and β wave, and α wave and β wave. The frequency band energy calculation formula is as follows:

[0052]

[0053] Among them, i represents the corresponding frequency band, n represents the number of coefficients, D i,j represents the jth wavelet packet coefficient of frequency band i, E i is the frequency band energy.

[0054] (2) Differential entropy (DE);

[0055] Differential entropy is Shannon information entropy The generalized form for continuous variables is:

[0056]

[0057] Among them, p(x) represents the probability density function of continuous information, and [a,b] represents the interval of information value. For a certain length, it is approximately Gaussian distributed. The differential entropy of the EEG signal is:

[0058]

[0059] The differential entropy is equal to the logarithm of its energy spectrum in a specific frequency band. For the convenience of calculation, the power spectrum can be used to approximate the differential entropy. The specific formula of differential entropy is:

[0060]

[0061] Among them, PSD is the power spectral density of the signal, which can be obtained by the periodogram method or the autocorrelation method.

[0062] Step 4.3, multi-domain fusion feature vector;

[0063] In order to achieve better recognition effect, the present invention fuses the extracted features, constructs a multi-domain fusion feature vector, and obtains a 272-dimensional (16 features * 17 channels) feature vector.

[0064] Furthermore, in step 5, the "ReliefF-SBE-L1" three-layer hybrid feature selection model includes three layers of selection: first, based on the Filter method, the ReliefF algorithm is selected for the first-layer channel screening, and the weights of each channel under a single feature are calculated according to the correlation between the channel parameters and the category labels, and redundant channels are eliminated; secondly, based on the Wrapper method, the backward search SBE algorithm is selected for the second-layer feature selection, and the built-in SVM classifier continuously removes features to select the feature subset with the best performance indicators; finally, based on the Embedded method, Lasso regression, i.e., L1 regularization, is used for the third-layer global optimization to further eliminate redundant optimization models.

[0065] Beneficial effects: The four-dimensional attention recognition method based on EEG feature fusion selection proposed in the present invention improves the recognition accuracy and computational efficiency; four-dimensional attention recognition is performed by extracting multi-domain EEG features to construct a fusion vector, thereby improving the traditional single feature recognition method; by proposing the "ReliefF-SBE-L1" three-layer hybrid feature selection model, the computational complexity is significantly reduced while achieving efficient four-dimensional attention state recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a framework diagram of the four-dimensional attention recognition method based on EEG feature fusion selection proposed by the present invention;

[0067] Figure 2 It is a four-dimensional attention induction experimental paradigm designed by the present invention;

[0068] Figure 3 It is a schematic diagram of the electrode channels of the EEG acquisition device of the present invention;

[0069] Figure 4 It is a schematic diagram of the process of preprocessing EEG signals in the present invention;

[0070] Figure 5 It is a partial processing schematic diagram of the EEGLAB toolbox of the present invention;

[0071] Figure 6 It is a comparison diagram of the signal before and after the wavelet threshold denoising is adopted in the present invention;

[0072] Figure 7 is a schematic diagram of sample data generation of the present invention;

[0073] Figure 8 It is a schematic diagram of wavelet packet decomposition selected when extracting frequency domain features in the present invention;

[0074] Fig. 9 It is a schematic diagram of the combination of multi-domain fusion feature vectors constructed by the present invention;

[0075] Fig.10It is a schematic diagram of the "ReliefF-SBE-L1" three-layer hybrid feature selection model proposed in the present invention;

[0076] Fig.11 This is a schematic diagram of the present invention using an SVM classifier to perform four-dimensional attention recognition. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0078] The four-dimensional attention recognition method based on EEG feature fusion selection provided by the present invention has a framework as follows: Figure 1 As shown; the present invention combines EEG signals with four-dimensional attention recognition, designs different experiments to accurately induce the four-dimensional attention of the subjects while collecting EEG signals; preprocesses the collected EEG signals, extracts multi-domain features from the preprocessed signals, extracts statistical feature standard deviation, nonlinear feature sample entropy, differential entropy, permutation entropy, Hjorth parameter, Lempel-Ziv complexity, Higuchi fractal dimension in the time domain, extracts 4 energy features and 3 energy ratio features based on wavelet packet decomposition in the frequency domain, and 16 features of the extracted 17 channels are combined into a multi-domain feature fusion vector for classification. At the same time, a "ReliefF-SBE-L1" three-layer hybrid feature selection model is proposed to achieve complementary advantages and disadvantages, and the ReliefF algorithm is used to screen channels based on filtering method, packaging method and embedding method, and the backward search SBE algorithm is used to select features and the L1 regularization method is used for global optimization. The multi-domain feature fusion vector obtained in this way can more efficiently identify the four-dimensional attention state while greatly reducing the computational complexity.

[0079] This embodiment provides a four-dimensional attention recognition method based on EEG feature fusion selection, and the specific steps are as follows:

[0080] Step 1: Design an experimental paradigm to induce four-dimensional attention;

[0081] Different from the previous general attention induction experiments, the present invention combines the operational definition of attention quality to focus on the development of attention quality in four dimensions, namely, attention allocation, attention breadth, attention stability and attention shifting, and designs an experiment to induce four-dimensional attention.

[0082] The four-dimensional attention induction experimental paradigm designed in this embodiment is as follows: Figure 2As shown; four tests are used to induce the four corresponding dimensions of attention allocation, breadth, stability, and transfer. Test 1, Graphic Identification Test: The test consists of 2 circular rings of different sizes with gaps. Due to the different directions of the gaps in the circular rings, many similar different figures will be formed. The total number of circular rings is 15*20=300. The subjects are required to find the specified 2 figures and mark them. Test 2, Select 4 Circles Test: The test consists of small squares with different numbers of circles; the total number of squares is 26*50=650. The subjects are required to find the squares with 4 circles and mark them. Test 3, Visual Tracking Test: The test consists of multiple curves starting from the left and ending on the right. The total number of curves is 25. The subjects are required to use their eyes to track a curve from the left and write the serial number at the beginning of the curve with a pen in the square where the curve ends on the right. Test 4, addition and subtraction test: The test consists of natural numbers. The subjects are required to perform "addition" and "subtraction" operations alternately and write the result between the two numbers. The total number of results is 22*12=264.

[0083] Step 2: EEG signal acquisition;

[0084] (1) Experimental subjects and environment:

[0085] Since the EEG signals of different subjects are different, their age, gender, and medical history will have different degrees of influence. In order to ensure the authenticity and effectiveness of the experimental data and the final effect, the subjects participating in the recruitment should be carefully screened. Select subjects who meet the conditions of being a graduate student or undergraduate, aged between 19 and 25, with normal vision or corrected vision, and no history of major diseases. Since EEG signals are easily interfered by external signals, the experiment requires a comfortable, quiet and less disturbing environment.

[0086] (2) Experimental equipment: A multi-channel wet electrode EEG acquisition instrument produced by Nanjing Weishi Medical was used, which was equipped with conductive paste, syringe and 19-conductor electrode cap.

[0087] The distribution diagram of the corresponding electrode channels of the EEG acquisition device designed in this embodiment is as follows: Figure 3 As shown, the distribution of 19-lead electrodes conforms to the international 10-20 system standard electrode placement method, including forehead electrodes Fp1, Fp2, frontal lobe electrodes F3, F4, F7, F8, Fz, central area electrodes C3, C4, CZ, parietal area electrodes P3, P4, PZ, occipital area electrodes O1, O2, temporal lobe electrodes T3, T4, T5, T6 and reference electrodes A1, A2.

[0088] Step 3: EEG signal preprocessing;

[0089] EEG signals are easily affected by factors such as environmental noise and physical magnetic fields. In order to obtain clean and effective EEG signals with high signal-to-noise ratio, it is necessary to preprocess the original mixed EEG signals based on the EEGLAB toolbox and Matlab software.

[0090] The schematic diagram of the process of EEG signal preprocessing in this embodiment is as follows Figure 4 The preprocessing is based on EEGLAB toolbox and Matlab software, and is divided into three parts: EEGLAB toolbox processing, wavelet threshold denoising, and sample set generation.

[0091] Step 3.1, the EEGLAB toolbox processing includes using the EEGLAB toolbox to locate electrodes, remove 50 Hz power frequency interference and 1-40 Hz bandpass filtering, reduce the sampling rate to 256 Hz and then remove bad segments, re-reference the signal with the T3 and T4 bilateral mastoid electrodes of the EEG acquisition device, perform independent component analysis (ICA), and remove electrooculogram, electromyography, and electrocardiogram artifact components. Figure 5 This is a partial processing diagram of the EEGLAB toolbox, showing the comparison of the effects of removing bad segments and removing artifacts such as electrooculogram in the EEGLAB toolbox series of processing.

[0092] Step 3.2, wavelet threshold denoising includes wavelet decomposition, threshold processing, and wavelet reconstruction; wavelet filtering can reduce noise on the basis of maximally ensuring the authenticity of the original signal. Its algorithm flow mainly includes three parts: selecting wavelet basis, threshold, and threshold function. The present invention uses db4 wavelet for denoising and adopts a fixed threshold calculation formula Where N is the signal length, and the hard threshold function is used to prevent errors caused by feature loss after soft threshold denoising.

[0093] Figure 6 The following is a comparison of the signal before and after wavelet threshold denoising. The original signal and the signal after wavelet threshold denoising are compared in the time domain waveform and the signal-to-noise ratio of each channel. It can be clearly seen that the signal after denoising and reconstruction has a larger signal-to-noise ratio while effectively retaining the complete information.

[0094] Step 3.3, sample generation refers to segmenting the signal obtained after wavelet threshold denoising to obtain a sample set. Figure 7 The above processed signal is used for sample generation, and a time window with a time length of 5s and a step length of 2.5s is selected, that is, the EEG signal is segmented with a sliding window of 5s, and the overlap rate is 50%, and a sample set with a single sample length of 5s is obtained.

[0095] Step 4: Feature extraction and fusion: The purpose of feature extraction is to identify features from EEG signals that can effectively reflect different attention states.

[0096] From the EEG signal processed in step 3, 16 features are extracted based on the comparison of time domain, frequency domain and time-frequency domain. Each feature contains the parameters of 17 electrode channels. Figure 8 This is a schematic diagram of the wavelet packet decomposition used when extracting frequency domain features. Since the required brain wave frequency band is concentrated in 0-32 Hz, this embodiment decomposes and displays it from the second layer of the wavelet tree, where node (5,0) mainly corresponds to δ waves, node (5,1) mainly corresponds to θ waves, node (5,2) mainly corresponds to α waves, and nodes (5,3) to (5,7) mainly correspond to β waves. The extracted feature parameters are fused to construct a multi-domain feature fusion vector.

[0097] Fig. 9 It is a combined schematic diagram of multi-domain feature fusion vectors. In order to effectively identify four-dimensional attention, the present invention fuses the extracted features, constructs a multi-domain feature fusion vector, and obtains a 272-dimensional (17 channels * 16 feature parameters) feature vector F. The feature parameters represented by F1 to F16 in the figure are standard deviation (Std), Hjorth activity parameter (HA), Hjorth mobility parameter (HM), Hjorth complexity parameter (HC), Lempel-Ziv complexity (LZC), fractal dimension (HFD), permutation entropy (Pec), sample entropy (SampEn), delta wave energy (E_Delta), theta wave energy (E_Theta), alpha wave energy (E_Alpha), beta wave energy (E_Beta), theta wave to alpha wave energy ratio (E_Theta / Alpha), theta wave to beta wave energy ratio (E_Theta / Beta), alpha wave to beta wave energy ratio (E_Alpha / Beta), differential entropy (DE), and all of them are feature vectors containing 17 electrode channels.

[0098] Step 5: Feature selection and dimensionality reduction: Use the "ReliefF-SBE-L1" three-layer hybrid feature selection model to perform feature selection on the multi-domain fusion feature vector obtained in step 4.

[0099] Feature selection is performed on high-dimensional multi-domain feature fusion vectors to reduce redundancy, and a three-layer hybrid feature selection model "ReliefF-SBE-L1" is proposed to achieve complementary advantages among different feature selection methods and screen out the best feature subset for classifier performance.

[0100] Fig.10This is a schematic diagram of the "ReliefF-SBE-L1" three-layer hybrid feature selection model proposed in the present invention. The model mainly includes three layers of selection: first, based on the Filter method, the ReliefF algorithm is selected for the first layer of channel screening, and the weights of each channel under a single feature are calculated according to the correlation between the channel parameters and the category label, and the redundant channels are eliminated; secondly, based on the Wrapper method, the backward search SBE algorithm is selected for the second layer of feature selection, and the built-in SVM classifier continuously removes features to select the feature subset with the best performance index; finally, based on the Embedded method, Lasso regression, i.e., L1 regularization, is used for the third layer of global optimization to further eliminate redundant optimization models. The principles of the backward search SBE and L1 regularization algorithms are shown in Tables 3 and 4 below, respectively.

[0101] Table 3 SBE algorithm principle

[0102]

[0103] Table 4 Feature selection algorithm process based on L1 regularization

[0104]

[0105]

[0106] Step 6: Import the multi-domain feature fusion vector obtained in step 5 into the SVM classifier to realize four-dimensional attention recognition.

[0107] Fig.11 The schematic diagram of using SVM classifier for four-dimensional attention recognition is shown in the figure. It can be clearly seen from the results in the figure that after the recognition method proposed by the present invention, the computational complexity is significantly reduced while achieving more efficient four-dimensional attention state recognition.

Claims

1. A four-dimensional attention recognition method based on EEG feature fusion selection, characterized in that: The steps include: Step 1: Based on the development of attention quality in four dimensions, namely, attention allocation, attention breadth, attention stability, and attention shifting, design an experiment to induce four-dimensional attention; Step 2: Using an EEG acquisition device to collect EEG signals while inducing four-dimensional attention; Step 3: preprocessing the original mixed EEG signal collected in step 2 based on EEGLAB toolbox and Matlab software; Step 4: extract multiple features from the EEG signal processed in step 3 based on time domain, frequency domain and time-frequency domain comparison, each feature includes parameters of multiple electrode channels, and fuse the extracted feature parameters to construct a multi-domain feature fusion vector; Step 5: Use the ReliefF-SBE-L1 three-layer hybrid feature selection model to perform feature selection and dimensionality reduction on the multi-domain fusion feature vector obtained in step 4; Step 6: Import the multi-domain feature fusion vector obtained in step 5 into the SVM classifier to realize four-dimensional attention recognition.

2. The four-dimensional attention recognition method according to claim 1, characterized in that: When collecting EEG signals in step 2, the experimental subjects, experimental environment and equipment requirements are as follows: (1) Experimental subjects and environment: Graduate students or undergraduates aged between 19 and 25 years old with normal or corrected vision and no history of major diseases were selected; the environment was quiet and undisturbed; (2) Experimental equipment: A wet electrode EEG acquisition instrument with 17 channels was selected.

3. The four-dimensional attention recognition method according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1, EEGLAB toolbox processing: Use EEGLAB toolbox to locate electrodes, remove 50Hz power frequency interference and 1-40Hz bandpass filtering, reduce the sampling rate to 256Hz and remove bad segments, re-reference the signal with T3 and T4 bilateral mastoid electrodes of the EEG acquisition device, perform independent component analysis (ICA), and remove electrooculography, electromyography, and electrocardiogram artifacts; Step 3.2, wavelet threshold denoising: including wavelet decomposition, threshold processing, wavelet reconstruction; db4 wavelet is used for denoising, and a fixed threshold calculation formula is used Where N is the signal length; Step 3.3, sample set generation: Generate samples for the signal processed in step 3.2, and segment it with a sliding window of 5s with an overlap rate of 50% to obtain a sample set.

4. The four-dimensional attention recognition method according to claim 1, characterized in that: From the EEG signal processed in step 3, 16 features are extracted based on the comparison of time domain, frequency domain and time-frequency domain, each of which contains the parameters of 17 electrode channels; The step 4 comprises the following steps: Step 4.1, obtain 8 time domain features; (1) Standard deviation: x(i) is the N-point signal sequence, is the mean of the signal sequence, and F is the standard deviation; (2) Hjorth activity parameter refers to the variance of the EEG signal, i.e., the signal power, which indicates the width: σ0 is the standard deviation of the signal, HA is the Hjorth activity parameter; (3) Hjorth mobility parameter, used to estimate the average frequency of the EEG signal: σ1 is the standard deviation of the first-order derivative of the signal, HM is the Hjorth mobility parameter; (4) Hjorth complexity parameter, which refers to the mobility of the first-order derivative relative to the EEG itself: σ2 is the standard deviation of the second-order derivative of the signal, HC is the Hjorth complexity parameter; (5) Lempel-Ziv complexity refers to the complexity of directly measuring data without decomposing or transforming the data, and is used to characterize the corresponding time series: C N is the complexity, N is the number of cycles, and C is the normalized Lempel-Ziv complexity; (6) Higuchi fractal dimension, used to describe the complexity and self-similarity of signals; The fractal dimension is calculated by using the Higuchi algorithm. The original time series is divided into k sub-segments and the average length of the curve of the sub-segments is calculated. L(k) With ln 1 / k The linear relationship between them is fitted to obtain a straight line, whose slope is the Higuchi fractal dimension D; (7) Permutation entropy, a nonlinear method used to detect time series complexity or dynamic mutations, can quantitatively evaluate the random noise contained in the signal; (8) Sample entropy, which is used to describe the complexity of the time series and the probability of a new pattern in the series when the dimension changes; Step 4.2, obtain 8 frequency domain features; (1) Delta wave energy, theta wave energy, alpha wave energy, beta wave energy, theta wave to alpha wave energy ratio, theta wave to beta wave energy ratio, alpha wave to beta wave energy ratio; From the EEG signal processed in step 3, the δ wave, θ wave, α wave, and β wave are separated for analysis. The wavelet packet decomposition is used to analyze the signal, and the low-frequency part and the high-frequency part are decomposed. The square sum of the wavelet packet coefficients of different frequency bands after decomposition is calculated to obtain the energy of the corresponding frequency bands of the δ wave, θ wave, α wave, and β wave, and then the energy ratio of the corresponding frequency bands of the θ wave and the α wave, the θ wave and the β wave, and the α wave and the β wave is calculated. The frequency band energy calculation formula is as follows: Among them, i represents the corresponding frequency band, n represents the number of coefficients, D i,j represents the jth wavelet packet coefficient of frequency band i, E i is the frequency band energy; (2) Differential entropy; The differential entropy is equal to the logarithm of its energy spectrum in a specific frequency band. For the convenience of calculation, the power spectrum can be used to approximate the differential entropy. The specific formula of differential entropy is: Among them, PSD is the power spectrum density of the signal, which can be obtained by the periodogram method or the autocorrelation method; Step 4.3, multi-domain fusion feature vector; The features extracted in step 4.1 and step 4.2 are fused to construct a multi-domain fusion feature vector and obtain a 272-dimensional feature vector.

5. The four-dimensional attention recognition method according to claim 1, characterized in that: In step 5, the ReliefF-SBE-L1 three-layer hybrid feature selection model includes three layers of selection: first, based on the Filter method, the ReliefF algorithm is selected for the first-layer channel screening, and the weights of each channel under a single feature are calculated according to the correlation between the channel parameters and the category labels, and redundant channels are eliminated; secondly, based on the Wrapper method, the backward search SBE algorithm is selected for the second-layer feature selection, and the built-in SVM classifier continuously removes features to select the feature subset with the best performance index; finally, based on the Embedded method, Lasso regression, i.e., L1 regularization, is used for the third-layer global optimization to further eliminate redundant optimization models.

Images