WO2020143300A1

WO2020143300A1 - Auditory attention state arousal level recognition method and apparatus, and storage medium

Info

Publication number: WO2020143300A1
Application number: PCT/CN2019/117074
Authority: WO
Inventors: 陆云; 王明江; 韩宇菲; 张啟权
Original assignee: 哈尔滨工业大学（深圳）
Priority date: 2019-01-07
Filing date: 2019-11-11
Publication date: 2020-07-16
Also published as: CN109740544B; CN109740544A

Abstract

An electroencephalogram signal-based auditory attention state arousal level recognition method and apparatus, and a storage medium. Said method comprises: acquiring required electroencephalogram signals; on the basis of the acquired electroencephalogram signals and a first-level main component filter constructed in the training process, performing first-level feature extraction based on ensemble empirical mode decomposition and main component filtering; on the basis of a first-level feature extraction signal and a second-level main component filter constructed in the training process, performing second-level feature extraction based on ensemble empirical mode decomposition and main component filtering; on the basis of a second-level feature extraction signal, performing variance statistics-based feature vector calculation on the extraction feature signals; and on the basis of a feature vector calculation result and a machine learning classifier constructed in the training process, extracting an electroencephalogram signal-based auditory attention state arousal level in a test process. The invention realizes arousal level recognition of an auditory attention state, and facilitates the improvement of the accuracy and effectiveness of the auditory attention state arousal level recognition.

Description

Method, device and storage medium for recognition of arousal state of auditory attention state

Technical field

The invention relates to the technical field of EEG signal feature extraction and pattern recognition, in particular to a method, device and storage medium for recognition of arousal attention state awakening degree based on EEG signals.

Background technique

Emotion is an attitude experience produced by whether a person satisfies his needs for objective things. It combines the state of human feelings, thoughts and behaviors. At present, researchers have not yet given a unified definition of emotion, but it is generally agreed that emotion is a subjective state that is generated under strong nerve impulses and is inseparable from the cerebral cortex. It can cause people to produce positive or negative psychological reactions. So that the corresponding body organization can act. Emotions are complex, a conscious experience or experience of special significance to humans, and contain a coordinated set of responses, which may include oral, physiological, behavioral, and neurological mechanisms. At present, the three widely accepted emotion models are discrete emotion model, dimensional emotion model and emotion model based on cognitive evaluation. However, with the in-depth study of emotion problems, scholars have found that discrete emotions cannot reflect the complexity and richness of emotions in the process of expression and transmission. For example, people often show more complex and delicate emotions in daily communication, such as Thinking, disappointment, embarrassment, appreciation, etc. Therefore, the dimensional sentiment model has gradually attracted the attention of researchers. At present, the most commonly used dimensional emotion model is: Awakening-valence model (Arousal-Valence model). With the development of emotion models, although only a few emotion models are still accepted, the valence-wake model still occupies the main position. Most dimensional models recognize the existence of the two dimensions of potency and arousal.

Human hearing, sound perception and cognition have certain rules, which can be divided into auditory awareness, auditory attention, auditory orientation, auditory discrimination, auditory memory, auditory selection and auditory feedback, and finally form the auditory concept and make sound information For the correct response, these stages are related to each other and promote each other. Auditory attention is a kind of psychological activity related to hearing. It is the activity that people pour into the sound in order to meet certain psychological needs. It is based on the awareness of hearing, and this sound still has a certain effect on the listener. Only a certain degree of meaning will produce auditory attention. Auditory attention helps humans quickly and accurately extract interesting or important sounds (cocktail party effect) from a noisy sound environment, and make further responses accordingly. In the state of auditory attention, the physiological emotion of the listener and the awakening of the central nervous system are closely related to the degree of awakening of specific emotions. For people to listen to specific auditory affairs, the success of their auditory behavior requires the listener to have a high awakening degree. If the level of awakening is low, the auditory response will be slow and the judgment will be inaccurate, and it will be easy for the auditory behavior to fail. Therefore, how to recognize the degree of awakening in the state of auditory attention is a very practical research.

At present, the detection methods for the degree of awakening mainly include subjective evaluation, biological response test, physiological signal detection, and biochemical methods. Among them, physiological signals are signals that directly reflect changes in the human body, and are more and more widely used in the detection of arousal. Because it can more accurately reflect changes in brain arousal, the study of EEG signals has attracted more and more attention from scholars. However, because EEG signals are generally weak and easily affected by the environment, research on EEG signals is still in the laboratory research stage. The existing awakening recognition algorithm based on EEG signals has the following limitations in feature representation and extraction: the recognition accuracy is not enough to effectively extract useful feature information; the EEG signals cannot be effectively extracted Non-linear features in the computer; there are certain requirements for the stability of the EEG signal. The more unstable the EEG signal, the more limited the effective feature mode it extracts.

Summary of the invention

The invention provides a method, device and storage medium for auditory attention state awakening recognition based on electroencephalogram signals, and realizes awakening degree recognition of auditory attention state by using cognitive electroencephalogram signals, which can effectively improve recognition accuracy and recognition effectiveness.

To achieve the above objective, the present invention provides a method for recognizing the arousal state of auditory attention state based on EEG signals, including the following steps:

Obtain the EEG signal to be tested;

Based on the tested EEG signals and the first-level principal component filter constructed during the training process, perform the first-stage feature extraction based on set empirical mode decomposition and principal component filtering;

Based on the first-level feature extraction signal signal and the second-level principal component filter constructed during the training process, the second-level feature extraction based on set empirical mode decomposition and principal component filtering is performed;

Based on the second-level feature extraction signal, the feature vector calculation based on variance statistics is performed on the extracted feature signal;

Based on the calculation result of the feature vector and the machine learning classifier constructed during the training process, the arousal degree of the auditory attention state based on the EEG signal during the testing process is extracted.

Wherein, before the step of obtaining the EEG signal to be tested further includes:

Obtain EEG signals during training;

Based on the EEG signals in the training process, perform the first-level set-based empirical mode decomposition, construct the first-level principal component filter and feature extraction, and perform the second-level set-based empirical mode decomposition, Construct the second-level principal component filter and feature extraction of principal component filter;

Perform feature vector calculation based on variance statistics for the extracted feature signals;

Construct a machine learning classifier based on the calculation results of feature vectors.

Among them, the steps of feature extraction based on set empirical mode decomposition, constructing principal component filters and principal component filters include:

Obtain time series signals of EEG signals during training;

Integrate empirical mode decomposition of time series signals to obtain the first four (first, second, third and fourth) eigenmode function components of time series signals under high awakening degree and low awakening degree;

Construct the principal component filter by using the first four order (first, second, third and fourth) eigenmode function components of the time series signal under high awakening degree and low awakening degree;

Using the constructed principal component filter, the first four-order eigenmode function components are subjected to principal component filter processing to obtain four time-series signals with feature extraction.

Wherein, the step of performing aggregate empirical mode decomposition on the time series signal includes:

Assuming the initial value of the window length w of the phase space reconstruction of the EEG signal is 1, for the one-dimensional time series signal t(n) of length N, n=1, 2, 3, ···,N, add white noise and zero Averaging processing to get the signal x(n);

Determine all local maxima and minima of signal x(n);

Use the cubic spline curve to fit all local maxima of the signal x(n) to form the upper envelope env_max(n); fit all local minima of the signal x(n) to form Lower envelope env_min(n);

Calculate the average value of the upper and lower envelope m(n)=(env_max(n)+env_min(n))/2;

Extract the detail signal h(n)=t(n)–m(n);

Check whether h(n) satisfies the iteration termination condition of the eigenmode function;

After the screening iteration termination condition is satisfied, w=w+1; the first eigenmode function IMF1(n)=h1,k(n) is obtained, and the remaining signal r(n)=x(n)-IMF1(n) ;

Determine whether the remaining signal r(n) meets the stop condition;

If the remaining signal r(n) is a constant or the change satisfies the preset condition, then all the iterative processes are terminated, otherwise, based on r(n), repeat the second step to the seventh step of the above process to enter the next round Iterate until the conditions for stopping iteration are met;

After the iteration stop condition is satisfied, the set of empirical mode decomposition of the time series signal is completed to obtain the eigenmode function components of each order.

Wherein, the step of constructing the principal component filter by using the first fourth-order eigenmode function components of the time series signal under high awakening degree and low awakening degree includes:

Obtain the first-order fourth-order eigenmode function components of the time series signal under high awakening degree and low awakening degree;

According to the first-order fourth-order eigenmode function components of the time series signal under high awakening degree and low awakening degree, the mixed spatial covariance matrix is obtained;

Perform matrix feature decomposition on the mixed space covariance matrix to obtain a whitened eigenvalue matrix;

A principal component filter is constructed based on the whitening eigenvalue matrix.

Wherein, the step of performing feature vector calculation based on variance statistics on the extracted features includes:

The input of the second level feature extraction has four time series signals. After each time series signal undergoes set empirical mode decomposition, the first four-order eigenmode function components are extracted and subjected to principal component filter dimensionality reduction filtering, which becomes two time Serial signal

Calculate the variance Z of these two time series signals separately;

According to the mathematical formula F=log ₁₀ (1+Var(Z)), eight feature values are calculated, which form a feature vector, which is sent to the machine learning classifier at the next level.

Among them, the model used by the machine learning classifier includes: support vector machine, linear decision, and neural network model.

The present invention also proposes an apparatus for recognizing the state of auditory attention based on electroencephalogram signals, including: a memory, a processor, and a computer program stored on the memory, the computer program is implemented as described above when the processor runs The steps of the method described.

The present invention also provides a computer-readable storage medium having a computer program stored on the computer-readable storage medium. The computer program is executed by a processor to implement the steps of the method described above.

Compared with the prior art, the present invention proposes a method, device and storage medium for the recognition of auditory attention state based on electroencephalogram signals, and its feature representation and extraction is a feature representation and extraction based on its own data, which is very suitable for Non-linear and non-stationary EEG signal feature extraction realizes how to recognize the awakening degree of auditory attention state, and improves the recognition accuracy and recognition effectiveness.

BRIEF DESCRIPTION

FIG. 1 is a system block diagram of an auditory attention state arousal recognition algorithm based on EEG signals of the present invention;

2 is a flowchart of feature extraction based on set empirical mode decomposition and principal component filtering according to the present invention;

FIG. 3 is a flow chart of aggregate empirical mode decomposition of time series signals involved in the present invention;

4 is a flowchart of the construction of the principal component filter of the eigenmode function component of the time series signal involved in the present invention;

5 is a flowchart of a method for calculating a feature vector based on variance statistics according to the present invention.

The implementation, functional characteristics and advantages of the present invention will be further described in conjunction with the embodiments and with reference to the drawings.

detailed description

It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

Specifically, please refer to FIG. 1, which is a schematic flowchart of an embodiment of a method for recognizing arousal state of auditory attention state based on EEG signals provided by the present invention.

An embodiment of the present invention provides a method for recognizing arousal state of auditory attention based on EEG signals, including the following steps:

S1, obtaining the EEG signal to be tested;

S2, based on the tested EEG signals, perform the first stage of EEG signals to be tested based on set empirical mode decomposition to obtain the eigenmode function components of the EEG signals, and use the first-level principal component filter constructed by the training process To realize the feature extraction of the principal component filtering of the eigenmode function components of the EEG signal to be tested;

S3. Extract the signal based on the first-level feature, perform the second-level set-based empirical mode decomposition, and then use the second-level principal component filter constructed by the training process to perform the second-level set-based empirical mode-based eigen Modular function components, feature extraction for principal component filtering;

S4. Perform feature vector calculation based on variance statistics for the extracted feature signals;

S5. Based on the calculated feature vector and the machine learning classifier constructed during the training process, the recognition of the arousal degree of the auditory attention state based on the EEG signal is realized.

Further, the step S1 before acquiring the EEG signal to be tested further includes:

S01: Obtain the trained EEG signal;

S02, based on the EEG signals in the training process, performing the first stage of EEG signals based on set empirical mode decomposition to obtain the eigenmode function components of the trained EEG signals; using the trained EEG signals Eigenmode function components, construct the first-level principal component filter, perform feature extraction of the eigenmode function components of the EEG signal, and perform the second stage based on set empirical mode decomposition to construct the second-level principal component Feature extraction of component filter and principal component filter;

S03: Perform feature vector calculation based on variance statistics on the extracted feature signals;

S04. Based on the calculated feature vectors, train a machine learning classifier to implement the construction of the machine learning classifier.

Obtain EEG signals during training;

Perform collective empirical mode decomposition on the EEG signals to obtain the first four (first, second, third, and fourth) eigenmode function components of the EEG signals at high and low awakening levels;

Construct the principal component filter using the first four orders (the first, second, third, and fourth order) eigenmode function components of the EEG signals at high and low awakening levels;

Using the constructed principal component filter, the first four-order eigenmode function components are subjected to principal component filter processing to obtain four time series signals with feature extraction.

Wherein, the step of performing collective empirical mode decomposition on EEG signals includes:

Determine all local maxima and minima of signal x(n);

Extract the detail signal h(n)=t(n)–m(n);

Determine whether the remaining signal r(n) meets the stop condition;

After the iteration stop condition is satisfied, the collective empirical mode decomposition of the EEG signals is completed to obtain the eigenmode function components of each order.

Wherein, the step of constructing the principal component filter by using the first four-order eigenmode function components of the EEG signals under high awakening degree and low awakening degree includes:

Obtain the first four-order eigenmode function components of EEG signals at high and low awakening levels;

According to the first four-order eigenmode function components of the EEG signals at high and low awakening degree, the mixed spatial covariance matrix is obtained;

Wherein, the step of performing feature vector calculation based on variance statistics on the extracted feature signals includes:

The input of the second-level feature extraction has four time series signals, and each time series signal undergoes set empirical mode decomposition to extract the first four (first, second, third, fourth) eigenmode function components for principal component filtering Dimensionality reduction filter to obtain two time series signals;

Calculate the variance Z of these two time series signals separately;

The present invention proposes an awakening degree recognition algorithm for auditory attention state based on electroencephalogram signals. The awakening degree recognition algorithm based on electroencephalogram signals is characterized and extracted based on its own data-driven feature representation and extraction. The advantages are as follows:

Directly use EEG signals to realize the high-awakening degree and low-awakening degree of auditory attention state;

The pattern features of the auditory attention state arousal based on EEG signals are not only feature representation and extraction based on set empirical mode decomposition and principal component filtering; a deep feature representation and The extraction framework is an efficient feature representation and extraction based on its own data;

The recognition algorithm of auditory attention state awakening degree based on EEG signal proposed by the present invention is a pattern extraction algorithm based on its own data drive, which is very suitable for non-linear and non-stationary EEG signal feature extraction.

The technical solutions and embodiments of the present invention will be described in detail below with reference to the drawings.

Referring to FIG. 1, FIG. 1 is a system block diagram of an arousal recognition state arousal recognition algorithm based on EEG signals of the present invention. The system block diagram of the auditory attention state awakening degree recognition algorithm is mainly composed of training and testing. The training process mainly involves four modules, which are: first level feature extraction based on set empirical mode decomposition and principal component filtering; second level feature extraction based on set empirical mode decomposition and principal component filtering; based on variance statistics Feature vector calculation; machine learning classifier. The main function of the training process is to realize the construction of the first and second principal component filters and the training of the machine learning classifier. In the testing process, the designed module process is the same as the training process as a whole, the main difference is that the first and second principal component filters and machine learning classifiers constructed by the training process are directly used by the testing process; 1. Construction of the second principal component filter and construction of the machine learning classifier. Therefore, based on the model parameters of the training process, the recognition of the high arousal degree and the low arousal degree of the auditory attention state based on the EEG signal can be realized.

In the system block diagram of the auditory attention state arousal recognition algorithm based on EEG signals described in FIG. 1, the first level feature extraction module based on collective empirical mode decomposition and principal component filtering is a combination of collective empirical mode decomposition and The feature extraction algorithm based on its own data drive realized by the main component filtering is shown in Figure 2. The first level of feature extraction module based on set empirical mode decomposition and principal component filtering mainly involves three sub-processes: first sub-process, set empirical mode decomposition of time series signals; the present invention, the second sub-process, before use Fourth-order (1st, 2nd, 3rd, and 4th) eigenmode function components, constructing the principal component filter; the third sub-process, using the constructed principal component filter, processes the first fourth-order eigenmode function components In order to achieve the first level of feature extraction based on set empirical mode decomposition and principal component filtering.

In the system block diagram of the auditory attention state arousal recognition algorithm based on the EEG signal described in FIG. 1, the EEG signal undergoes the first level of feature extraction based on set empirical mode decomposition and principal component filtering, and the dimension of the time series signal is The original one-dimensional time series signal becomes a four-dimensional time series signal. Then, the four time series signals after the first level feature extraction processing are processed, and then the second level feature extraction based on set empirical mode decomposition and principal component filtering is performed. The second level feature extraction module based on set empirical mode decomposition and principal component filtering has the same implementation method as the first level feature extraction algorithm flow based on set empirical mode decomposition and principal component filtering. The first-level and second-level feature extraction methods are effectively combined into a deep feature extraction model by cascading. This deep feature extraction model is an automatic feature extraction method based on its own data drive, which is realized by combining empirical mode decomposition and principal component filtering.

Next, the technical implementation schemes of the three sub-processes of the first (second) level feature extraction module based on set empirical mode decomposition and principal component filtering are described in detail.

Further, FIG. 2 is a detailed scheme of sub-process technology implementation of set empirical mode decomposition, as shown in FIG. 3. Mainly consists of the following processes:

Assuming that the initial value of the window length w of the phase space reconstruction of the EEG signal is 1, for a one-dimensional time series signal t(n) of length N, n=1, 2, 3, ..., N;

Add white noise and zero averaging to get the signal x(n);

Determine all local maxima and minima of signal x(n);

Extract the detail signal h(n)=t(n)–m(n);

Check whether h(n) satisfies the iteration termination condition of the eigenmode function:

h _1,k (i) represents the value of the kth iteration of the first detail signal, SD is the iterative filtering threshold (generally 0.2-0.3), m _1,1 is the average of the upper and lower envelopes, the detail signal h _1, The initial value of _k (i) is obtained by subtracting the average of the upper and lower envelopes from x (n). In the embodiment of the present invention, the value is 0.2. When SD is less than 0.2, the screening iteration of the eigenmode function component of the current round is terminated.

After the screening iteration termination condition is satisfied, w=w+1; the first eigenmode function IMF1(n)=h1,k(n) is obtained, and the remaining signal r(n)=x(n)-IMF1(n).

Determine whether the remaining signal r(n) meets the stop condition.

If the residual signal r(n) is finally a constant or the change is small enough, terminate all the iterative processes, otherwise, based on r(n), repeat the second to seventh steps of the above process to enter the next round of iterations, Until the conditions for stopping iterations are met.

After the iteration stop condition is satisfied, the set empirical mode decomposition of the time series signal is completed to obtain the eigenmode function components of each order.

Further, the detailed solution implemented by the sub-process technology for constructing the principal component filter in FIG. 2 is shown in FIG. 4. Mainly involves the following steps:

In the first step, the mixed spatial covariance matrix is obtained according to the first-order eigenmode function components of the time series signal under high awakening degree and low awakening degree. Assuming that the matrix composed of the first four order eigenmode function components of the time series signal under high awakening degree and low awakening degree is: IMF4 ₁ and IMF4 _{2 respectively} , and the physiological time series with length N, for IMF4 ₁ and IMF4 ₂ The expressed matrix dimensions are all 4×N.

The second step is to solve the normalized covariance matrix of IMF4 ₁ and IMF4 ₂ , respectively R ₁ and R ₂ , the specific mathematical expression is as follows,

Among them, trace (·) represents the sum of the elements on the diagonal of the matrix.

In the third step, the mixed space covariance matrix R is obtained as

with

The mean covariance matrix of the singular spectral components IMF4 ₁ and IMF4 ₂ of physiological time series in two physiological states respectively.

The fourth step is to perform feature decomposition on the mixed spatial covariance matrix R to obtain the whitened eigenvalue matrix. First, the eigenvalue decomposition of the mixed space covariance matrix R, U and λ are: the eigenvector matrix and its corresponding eigenvalue matrix (the eigenvalues of the eigenvalue matrix, arranged in descending order).

^{R = U × λ × U T} (5)

Thus, the whitening value matrix P can be expressed as follows:

The fifth step is to construct the principal component filter. Based on the whitening value matrix, the matrix R ₁ and R ₂ are transformed as follows:

S ₁ = P × R ₁ × P ^T (7)

S ₂ = P × R ₂ × P ^T (8)

Then, feature decomposition of the matrices S ₁ and S ₂ ,

It can be proved that the eigenvector of matrix S ₁ and the eigenvector matrix of matrix S ₂ are equal, that is,

B ₁ = B ₂ = B (11)

At the same time, the sum of the diagonal matrix λ1 and λ2 of the two eigenvalues is the identity matrix, namely:

λ1+λ2＝I (12)

Since the sum of the eigenvalues of the two types of matrices is always 1, the eigenvector corresponding to the largest eigenvalue of S ₁ makes S ₂ have the smallest eigenvalue, and vice versa. The transformation of the whitening physiological time series to the eigenvector corresponding to the largest eigenvalue in λ1 and λ2 is optimal for separating the variance in the two signal matrices.

Therefore, the optimal principal component filter W can be constructed at this time, and its mathematical form is,

W＝B ^T ×P (13)

Further, in FIG. 2, the sub-process of processing the components of the first fourth-order eigenmode function using the constructed principal component filter, the main technical implementation scheme, such as the following mathematical expression:

For the training process, for the time series signals with high awakening degree and low awakening degree, after processing by the constructed principal component filter W, the extracted features Z1 and Z2 are:

Z1＝W×IMF4 ₁ (14)

Z2＝W×IMF4 ₂ (15)

For the testing process, for the tested time series signal, after processing by the constructed principal component filter W, the extracted feature Z_test is obtained as:

Z_test＝W×IMF4_test (16)

In the system block diagram of the auditory attention state arousal recognition algorithm based on EEG signals described in FIG. 1, the technical implementation method of feature vector calculation based on variance statistics is shown in FIG. 5. The input of the second level feature extraction has 4 time series signals. After each time series undergoes set empirical mode decomposition, the first four order eigenmode function components are extracted and subjected to principal component filter dimensionality reduction filtering to obtain two time series Signal, and then calculate the variance Z of these two time series signals separately. According to the mathematical formula F=log ₁₀ (1+Var(Z)), 8 feature values can be obtained to form a feature vector, which is sent to the next level of machine learning classifier.

In the system block diagram of the auditory attention state arousal recognition algorithm based on EEG signals described in FIG. 1, the technical implementation scheme of the machine learning classifier can use classic classifier models, such as support vector machines, linear decisionrs, and neural networks Wait.

In addition, the present invention also proposes an apparatus for recognizing the state of auditory attention based on EEG signals, including: a memory, a processor, and a computer program stored on the memory, the computer program is implemented when the processor runs The steps of the method as described above.

In addition, the present invention also provides a computer-readable storage medium having a computer program stored on the computer-readable storage medium. The computer program is executed by a processor to implement the steps of the method described above.

Compared with the prior art, the present invention provides a method, device and storage medium for auditory attention state arousal recognition based on electroencephalogram signals, and its feature representation and extraction is a feature representation and extraction based on its own data to achieve How to recognize the awakening degree of the auditory attention state, and improve the recognition accuracy and recognition effectiveness.

The above are only the preferred embodiments of the present invention, and therefore do not limit the patent scope of the present invention. Any equivalent structure or process transformation made by using the description and drawings of the present invention, or directly or indirectly used in other related technical fields The same reason is included in the patent protection scope of the present invention.

Claims

A method for recognizing awakening degree of auditory attention state based on EEG signals, which is characterized by comprising the following steps:

Obtain the EEG signal to be tested;

Based on the EEG signals to be tested, the first-level set-based empirical mode decomposition is performed, and then the first-level principal component filter constructed by the training process is used to perform the first-level set of empirical mode decomposition eigenmode function components. Feature extraction of principal component filtering;

Based on the first-level feature extraction signal, the second-level set-based empirical mode decomposition is performed, and then the second-level principal component filter constructed by the training process is used to perform the principal component on the second-level set of empirical mode decomposition eigenmode function components Filtered feature extraction;

Based on the second-level feature extraction signal, the feature vector calculation based on variance statistics is performed on the extracted feature signal;

Based on the calculation result of the feature vector and the machine learning classifier constructed during the training process, the arousal degree of the auditory attention state based on the EEG signal during the testing process is extracted.
The method according to claim 1, wherein before the step of acquiring the EEG signal to be tested further comprises:

Obtain EEG signals during training;

Based on the EEG signals in the training process, perform the first-level set-based empirical mode decomposition, construct the first-level principal component filter and feature extraction, and perform the second-level set-based empirical mode decomposition, Construct the second-level principal component filter and feature extraction of principal component filter;

Perform feature vector calculation based on variance statistics for the extracted feature signals;

Construct a machine learning classifier based on the calculation results of feature vectors.
The method according to claim 2, wherein the steps of feature extraction based on set empirical mode decomposition, constructing principal component filters and principal component filtering include:

Obtain time series signals of EEG signals during training;

Integrate empirical mode decomposition of time series signals to obtain the first four (first, second, third and fourth) eigenmode function components of time series signals under high awakening degree and low awakening degree;

Construct the principal component filter by using the first four order (first, second, third and fourth) eigenmode function components of the time series signal under high awakening degree and low awakening degree;

Using the constructed principal component filter, the first four-order eigenmode function components are subjected to principal component filter processing to obtain four time-series signals with feature extraction.
The method according to claim 3, wherein the step of performing collective empirical mode decomposition on the time series signal comprises:

Assuming the initial value of the window length w of the phase space reconstruction of the EEG signal is 1, for the one-dimensional time series signal t(n) of length N, n=1, 2, 3, ···,N, add white noise and zero Averaging processing to get the signal x(n);

Determine all local maxima and minima of signal x(n);

Use the cubic spline curve to fit all local maxima of the signal x(n) to form the upper envelope env_max(n); fit all local minima of the signal x(n) to form Lower envelope env_min(n);

Calculate the average value of the upper and lower envelope m(n)=(env_max(n)+env_min(n))/2;

Extract the detail signal h(n)=t(n)–m(n);

Check whether h(n) satisfies the iteration termination condition of the eigenmode function;

After the screening iteration termination condition is satisfied, w=w+1; the first eigenmode function IMF1(n)=h1,k(n) is obtained, and the remaining signal r(n)=x(n)-IMF1(n) ;

Determine whether the remaining signal r(n) meets the stop condition;

If the remaining signal r(n) is a constant or the change satisfies the preset condition, then all the iterative processes are terminated, otherwise, based on r(n), repeat the second step to the seventh step of the above process to enter the next round Iterate until the conditions for stopping iteration are met;

After the iteration stop condition is satisfied, the set of empirical mode decomposition of the time series signal is completed to obtain the eigenmode function components of each order.
The method according to claim 4, wherein the step of constructing a principal component filter using the first fourth-order eigenmode function components of the time series signal at high awakening and low awakening includes:

Obtain the first-order fourth-order eigenmode function components of the time series signal under high awakening degree and low awakening degree;

According to the first-order fourth-order eigenmode function components of the time series signal under high awakening degree and low awakening degree, the mixed spatial covariance matrix is obtained;

Perform matrix feature decomposition on the mixed space covariance matrix to obtain a whitened eigenvalue matrix;

A principal component filter is constructed based on the whitening eigenvalue matrix.
The method according to claim 4, wherein the step of performing feature vector calculation based on variance statistics on the extracted feature signal includes:

The input of the second level feature extraction has four time series signals. After each time series signal undergoes set empirical mode decomposition, the first four-order eigenmode function components are extracted and subjected to principal component filter dimensionality reduction filtering to obtain two time series signal;

Calculate the variance Z of these two time series signals separately;

According to the mathematical formula F=log 10 (1+Var(Z)), eight feature values are calculated, which form a feature vector, which is sent to the machine learning classifier at the next level.
The method according to claim 6, wherein the model adopted by the machine learning classifier includes: a support vector machine, a linear decider, and a neural network model.
A device for recognizing the state of auditory attention based on electroencephalogram signals, comprising: a memory, a processor, and a computer program stored on the memory, the computer program is implemented as a right when the processor runs The method of any one of claims 1-7 is required.
A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and the computer program is executed by a processor to implement the method according to any one of claims 1-7 step.