CN115374812A - A signal feature extraction method for multiple feature fusion extraction of EEG signals - Google Patents

Abstract

The invention discloses a signal feature extraction method for multi-feature fusion extraction of EEG signals. The method comprises: acquiring EEG signal data and performing signal preprocessing on the EEG signal data to obtain preprocessed signals; The final signal is subjected to multi-dimensional feature extraction and matrix construction to obtain the original feature matrix; the original feature matrix is fused and dimensionally reduced to obtain the final feature matrix; the final feature matrix is input to the pre-trained classification model for classification, and the classification result is output. The invention overcomes the problem of incomplete EEG signal information in the single-domain feature extraction algorithm in the traditional algorithm, and effectively improves the classification performance. The invention can be widely used in the field of signal processing.

Description

A signal feature extraction method for multiple feature fusion extraction of EEG signals

technical field

The invention relates to the field of signal processing, in particular to a signal feature extraction method for multi-feature fusion extraction of electroencephalogram signals.

Background technique

EEG signals are non-linear, non-stationary time-series signals that can be detected by sensors on electrodes on the scalp, and these signals are the external manifestations of neuron membrane potential. Research based on EEG signals can be used for emotion recognition. At present, most of the emotion recognition is performed by extracting the features of a single field, but the feature extraction of a single field can only contain part of the information of the EEG signal, which will lead to unsatisfactory classification performance. .

Contents of the invention

The purpose of the present invention is to provide a signal feature extraction method for multi-feature fusion extraction of EEG signals, which overcomes the problem of incomplete EEG signal information in single-domain feature extraction algorithms in traditional algorithms.

The first technical solution adopted in the present invention is: a signal feature extraction method for multi-element feature fusion extraction of EEG signals, comprising the following steps:

Obtain the EEG signal data and perform signal preprocessing on the EEG signal data to obtain the preprocessed signal;

Perform multi-dimensional feature extraction and matrix construction on the preprocessed signal to obtain the original feature matrix;

Perform fusion and dimensionality reduction processing on the original feature matrix to obtain the final feature matrix;

Input the final feature matrix to the pre-trained classification model for classification, and output the classification result.

Further, the step of obtaining the EEG signal data set and performing signal preprocessing on the EEG signal data set to obtain the preprocessed signal specifically includes:

Obtain EEG signal data;

Based on the independent component analysis method, the EEG signal data is subjected to noise separation processing to obtain the separated signal;

Frequency filtering is performed on the separated signal based on the Butterworth bandpass filter to obtain the filtered signal;

Normalize the filtered signal to obtain the preprocessed signal.

Further, the step of performing multi-dimensional feature extraction and matrix construction on the preprocessed signal to obtain the original feature matrix specifically includes:

Perform feature extraction on the preprocessed signal in the time domain to obtain time domain features;

Based on the AR model power spectrum estimation method, the preprocessed signal is extracted in the frequency domain to obtain the frequency domain features;

Based on the Hilbert-Huang transform method, the preprocessed signal is extracted in the time-frequency domain to obtain the time-frequency domain features;

Based on the nonlinear dynamic analysis, the preprocessed signal is extracted in the nonlinear domain to obtain the nonlinear domain features;

Based on the public space pattern analysis, the preprocessed signal is extracted in the airspace to obtain the airspace features;

According to the time domain feature, frequency domain feature, time-frequency domain feature, nonlinear domain feature and space domain feature, the matrix is constructed to obtain the original feature matrix.

Further, the time-domain features include standard deviation, root mean square, and the mean value of the absolute value of the first-order difference, and the nonlinear features include approximate entropy, fuzzy entropy, and sample entropy.

Further, the step of performing fusion and dimensionality reduction processing on the original feature matrix to obtain the final feature matrix specifically includes:

Perform de-meaning processing on the original feature matrix to obtain the centered matrix;

Calculate the covariance matrix of the centered matrix and perform eigenvalue decomposition to obtain the eigenvalues and corresponding eigenvectors;

The eigenvalues are sorted and the eigenvectors corresponding to the preset number of eigenvalues are taken in order to construct a matrix to obtain the final eigenmatrix.

Further, the step of inputting the final feature matrix to the pre-trained classification model for classification, and outputting the classification result specifically includes:

Optimize the parameters of the SVM-KNN classifier based on the particle swarm optimization algorithm and the pre-built training set to obtain the pre-trained classification model;

Taking the final feature matrix as input, calculate the distance between the sample to be tested and the optimal hyperplane based on the pre-trained classification model, and obtain the sample distance;

Compare the sample distance with a preset threshold;

It is judged that the absolute value of the sample distance is less than the preset threshold, the KNN algorithm is used for classification, and the classification result is output;

Judging that the absolute value of the sample distance is greater than or equal to the preset threshold, the SVM algorithm is used for classification, and the classification result is output.

Further, the calculation formula of the sample distance is as follows:

In the above formula, a _i is the Lagrangian multiplier, y _i a _i is the coefficient of the support vector in the decision function, k is the system constant, and b is the constant term of the decision function in SVM.

The second technical solution adopted in the present invention is: a signal feature extraction system for the fusion and extraction of multiple features of EEG signals, including:

The preprocessing module is used to obtain the EEG signal data and perform signal preprocessing on the EEG signal data to obtain the preprocessed signal;

The feature extraction module is used to perform multi-dimensional feature extraction and matrix construction on the preprocessed signal to obtain the original feature matrix;

A dimensionality reduction module is used to perform fusion and dimensionality reduction processing on the original feature matrix to obtain the final feature matrix;

The classification module is used to input the final feature matrix to the pre-trained classifier for classification, and output the classification result.

The third technical solution adopted by the present invention is: a signal feature extraction device for the fusion and extraction of multiple features of EEG signals, including:

at least one processor;

at least one memory for storing at least one program;

When the at least one program is executed by the at least one processor, the at least one processor implements the signal feature extraction method for multi-element feature fusion extraction of EEG signals as described above.

The beneficial effect of the method of the present invention is: the present invention retains the information contained in the EEG signal as completely as possible through multi-dimensional features, can effectively improve the classification performance of the classifier, and makes the calculation of the classifier more convenient and convenient through dimensionality reduction processing. Further improve the performance of the classifier.

Description of drawings

Fig. 1 is the step flow chart of the signal feature extraction method of the multiple feature fusion extraction of a kind of EEG signal of the present invention;

FIG. 2 is a structural block diagram of a signal feature extraction system for multi-element feature fusion extraction of EEG signals according to the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. For the step numbers in the following embodiments, it is only set for the convenience of illustration and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art sexual adjustment.

As shown in Figure 1, the present invention provides a kind of signal feature extraction method of the multiple feature fusion extraction of EEG signal, and this method comprises the following steps:

S1. Obtain the EEG signal data and perform signal preprocessing on the EEG signal data to obtain a preprocessed signal;

Specifically, since there are many noises in the EEG signal, and most of the emotion-related information is concentrated in the 0-50 Hz frequency band, it is necessary to preprocess the EEG signal data.

S1.1. Obtain EEG signal data;

S1.2. Perform noise separation processing on the EEG signal data based on the independent component analysis method to obtain separated signals;

Specifically, we need to separate the oculoelectricity, myoelectricity and other bioelectricity in the human body from the EEG signal. At present, the removal of artifacts can use Independent Component Analysis (ICA) to separate the EEG signal from other noises. separated.

S1.3. Perform frequency screening processing on the separated signal based on the Butterworth bandpass filter to obtain the filtered signal;

Specifically, the content of 0-50 Hz in the EEG signal is preserved through a Butterworth band-pass filter.

S1.4. Normalize the filtered signals to obtain preprocessed signals.

Specifically, normalization is finally performed through the mapminmax function that comes with matlab.

S2. Perform multi-dimensional feature extraction and matrix construction on the preprocessed signal to obtain the original feature matrix;

S2.1. Perform feature extraction on the preprocessed signal in the time domain to obtain time domain features;

Specifically, the time-domain feature is to use the signal statistical result of the EEG signal as a feature in the time domain, and the time-domain feature includes standard deviation, root mean square, and the mean value of the absolute value of the first-order difference.

The formula for calculating the standard deviation is as follows:

Among them, x _i is the value of the i-th sampling point in the time series, μ _x the mean value of the current time series, and N is the length of the time series.

The root mean square calculation formula is as follows:

The formula for calculating the mean of the absolute value of the first-order difference is as follows:

S2.2, based on the AR model power spectrum estimation method, perform feature extraction on the preprocessed signal in the frequency domain to obtain frequency domain features;

Specifically, when extracting EEG frequency domain features, the signals are first mapped to corresponding frequency bands, and then the feature quantities of each frequency band are obtained. Since the EEG signal is a random signal, this paper uses the AR model power spectrum estimation method for frequency domain analysis and feature extraction. The AR model is also called the autoregressive moving average model, and the Y-W equation needs to be solved to estimate the model parameters. For general random signals, the AR model can be expressed as follows:

Among them, y(n) is the nth sampling value of the signal, p is the model order, a _i is the model coefficient, and r(n) is the residual of zero-mean white noise.

System function of AR model:

Model output power spectrum:

Use the AR model method to obtain the power spectrum of the EEG signal, and then calculate the spectral energy or power corresponding to each rhythm of the EEG, as the frequency domain characteristics of the EEG θ, α, β, and γ rhythms

S2.3, based on the Hilbert-Huang transform method, perform feature extraction on the preprocessed signal in the time-frequency domain to obtain time-frequency domain features;

Specifically, the Hilbert-Huang transform method mainly includes Empirical Mode Decomposition (EMD) and Hilbert Spectral Analysis (HSA).

Empirical mode decomposition is to obtain the intrinsic mode function (Intrinsic mode function, IMF), which has adaptive, orthogonality, completeness, and modulation characteristics of IMF components. EMD satisfies the following two conditions:

1) The number of signal extreme points is equal to or differs by 1 from the number of zero points;

2) The local mean of the upper envelope defined by the maximum value and the lower envelope defined by the minimum value of the signal is 0.

The EMD process is as follows:

1) Find all the maximum and minimum points of the input signal.

2) Use cubic splines to fit the maximum and minimum points, find the curves of the upper and lower envelopes, calculate the mean function, and then calculate the difference h between the signal to be analyzed and the mean.

3) Investigate whether h satisfies the IMF condition, and if so, take h as the first IMF; otherwise, perform the first two steps until the kth step satisfies the IMF condition, then obtain the first IMF, and obtain the original The difference r between the signal and the IMF.

4) Take the difference r as the signal to be decomposed until the remaining r is a monotonic signal or there is only one pole, the obtained expression is as follows:

Among them, x(t) is the original signal, C _i (t) represents the IMF component obtained by the i-th screening, N is the number of screening times, and R _n (t) is the final residual component.

After the EMD process, carry out Hilbert spectrum analysis (HSA), the following formula carries out Hilbert spectrum transformation to each IMF component:

Parse the signal as:

Among them, A _i (t) is the instantaneous amplitude, Pi (t) is the instantaneous phase, and the instantaneous frequency ω _i (t) can be obtained further, namely:

Then the distribution of the signal amplitude in the time-frequency domain can be described, that is, the Hilbert spectrum:

Among them, Re is the real part.

According to the above formula, the instantaneous energy spectrum (IES) and marginal energy spectrum (MES) can be further obtained:

Where: [ω ₁ , ω ₂ ] is the frequency range of the signal, and [t ₂ , t ₂ ] is the time range of the signal.

S2.4. Based on nonlinear dynamic analysis, perform feature extraction on the preprocessed signal in the nonlinear domain to obtain nonlinear domain features;

Specifically, nonlinear features include approximate entropy, fuzzy entropy, and sample entropy.

The use of approximate entropy as one of the nonlinear dynamic features is based on the fact that approximate entropy has the advantage of quantifying the regularity and unpredictability of EEG signals. It can represent the complexity of EEG signals and reflect the possibility of new information in the signals.

The processing steps of approximate entropy on the preprocessed EEG signal x(t) are as follows:

(1) The N-dimensional original signal time series is sampled at equal time intervals, and the m-dimensional vectors X(1), X(2),...,X(N-m+1) are reconstructed, where Xi=[u(i), u(i+1),...,u(i+m-1)].

(2) For 1,2,...,N-m+1, count the number of vectors that satisfy the following conditions

Among them, the condition to be satisfied by X(i) is d[X(i), X(j)]≤r, d[X,X ^* ]=max|u(a)-u*(a)|, where u( a) is an element of vector x, d represents the distance between two vectors X, the value range of j is l≤j≤N-m+1, and j can be equal to i.

(3) Definition:

(4) The approximate entropy can be defined as:

ApEn＝Φ ^m (r)-Φ ^m+1 (r)

In the formula, the parameter m=2 or m=3 is usually set, and m=3 can reconstruct the dynamic evolution process of the system in more detail; the value of r mainly depends on the application, and usually choose r=0.2*std (std is the standard of time series Difference).

Using fuzzy entropy as one of the nonlinear dynamic features, while inheriting the advantages of sample entropy, reduces the dependence on the length of the time series. Due to its good continuity and robustness, it can be effectively used for EEG time series. Analyzing.

The processing steps of fuzzy entropy on the preprocessed EEG signal x(t) are as follows:

(1) The given N-dimensional signal time series is the same as the approximate entropy, define the dimensions of the phase space as m (m<N-2) and r, and reconstruct the phase space: X(i)=[u(i),u (i+1),...,u(i+m-1)]-u ₀ (i)

(2) Introduce the fuzzy relationship function A(x), and calculate:

并定义：(3) get

and define:

(4) The fuzzy entropy can be defined as:

FuzzyEn＝InΦ ^m (r)-InΦ ^m+1 (r)

The sample entropy is used as one of the nonlinear dynamic characteristics. The sample entropy reflects the complexity of the time series and can measure the probability of generating new patterns in the time series. The larger the sample entropy value, the greater the probability of generating new patterns, and the higher the sequence is. complex. Under different emotional states, the degree of activation of the corresponding cerebral cortex is different, and the complexity of brain electrical activity is also different. Therefore, the sample entropy feature can reflect the change of emotional EEG.

The processing steps of the sample entropy on the preprocessed brain signal x(t) are as follows:

改为

将

的分母N-m+1，改为N-m，且j≠i。(1) The sample entropy and approximate entropy in the first three steps are the same, where the approximate entropy

changed to

Will

The denominator of N-m+1 is changed to Nm, and j≠i.

对所有i的平均值，为：(2) seeking

The average value for all i is:

(3) Make k=m+1, repeat the first and second steps of sample entropy, and get:

(4) The sample entropy can be defined as:

S2.5. Based on the public space pattern analysis, perform feature extraction on the preprocessed signal in the airspace to obtain airspace features;

Specifically, the "one-to-one" (OVO) method is used to extend the multi-classification of the CSP algorithm. The OVO-CSP method is used to extract spatial features of EEG signals. This method splits multi-classification into several binary classification problems, so the specific implementation process of the traditional CSP algorithm for binary classification is explained.

(1) Find the spatial covariance matrix of two types of data:

表示矩阵对角线上元素之和。Where E _i is the data matrix converted from two types of signals,

Represents the sum of elements on the diagonal of a matrix.

(2) Calculate the average covariance matrix C _i of each category, and then sum:

(3) Perform eigenvalue decomposition and whitening processing on the mixed space covariance matrix according to the formula to obtain S ₁ and S ₂ with the same eigenvector, and then perform eigenvalue decomposition on the eigenvectors S ₁ and S ₂ respectively.

S ₁ =Bλ ₁ B ^T S ₂ =Bλ ₂ B ^T

B is the common eigenvector of S ₁ and S ₂ , and the sum of eigenvalues is 1.

(4) After constructing the spatial filter, the EEG signal matrix E is filtered to obtain z, and _Zi is used as the eigenvalue after performing the following operations:

where p=1,2,...,2m (2m<n).

S2.6. Perform matrix construction according to time domain features, frequency domain features, time-frequency domain features, nonlinear domain features and space domain features to obtain an original feature matrix.

Specifically, all f _p constitute the final feature moment F={f ₁ , f ₂ , . . . f _2m }, and a set of EEG features is obtained.

S3. Perform fusion and dimensionality reduction processing on the original feature matrix to obtain the final feature matrix;

Specifically, the dimensionality of the combination of feature vectors is high, which is not conducive to classifier recognition, making the algorithm not accurate and fast enough, so the PCA algorithm is selected to achieve dimensionality reduction. According to the principle of variance maximization, PCA uses a set of linearly independent and mutually orthogonal new vectors to represent the original vector. This set of new vectors is the principal component, which is the linear combination of the original vectors. The implementation steps are as follows:

S3.1. Perform demeaning processing on the original feature matrix {x ₁ , x ₂ ,…,x _n } to obtain a centered matrix;

Specifically, the decentralized matrix is expressed as follows:

S3.2. Calculate the covariance matrix C of the centered matrix X and perform eigenvalue decomposition to obtain the eigenvalue λ and the corresponding eigenvector U;

Specifically, λ is a diagonal matrix, representing the size of the variance of the principal components.

S3.3. Sorting the eigenvalues and taking the eigenvectors corresponding to the preset number of eigenvalues in sequence to construct a matrix to obtain the final eigenmatrix.

Specifically, λ is sorted: λ ₁ ≥λ ₂ ≥...≥λ _d , and the eigenvectors corresponding to the first d' eigenvalues are taken to form a projection matrix W=(u ₁ ,u ₂ ,...,u _d′ ). Among them, d' represents the dimension after dimension reduction, which is determined by the cumulative contribution rate of each component, namely:

Generally, the C _rate is set to be greater than 85%.

According to W, the new feature quantity F=X*W of the low-dimensional space is obtained. The new features processed by PCA represent the useful information to the greatest extent, and those eigenvectors corresponding to the d-d' eigenvalues are eliminated to increase the sample sampling density and play a denoising role to a certain extent. As the next step of the classifier Type better.

S4. Input the final feature matrix to the pre-trained classification model for classification, and output the classification result.

Specifically, an SVM-KNN classifier optimized by the particle swarm optimization algorithm is selected in this embodiment.

S4.1. Optimizing the parameters of the SVM-KNN classifier based on the particle swarm optimization algorithm and the pre-built training set to obtain a pre-trained classification model;

Specifically, in the PSO algorithm, the particles in the particle swarm continuously modify the position and velocity of the particles in the n-dimensional space by combining their own flight experience with the flight experience of the particle swarm, where n is the number of parameters that need to be optimized the number of parameters. The particle swarm updates its position and velocity by continuously iteratively updating and tracking the two values of the individual extremum (Pbest) and the overall extremum (Gbest). The update formulas of speed and position are as follows:

V _id (t+1)＝ω·V _id (t)+c ₁ r ₁ (Pbest(t)-x _id (t))+c ₂ r ₂ (Gbest(t)-x _id (t))

X _id (t+1)＝X _id (t)+V _id (t+1)

In the formula, 1≤i≤n, 1≤d≤m, m is the number of particles in the example group; ω is the inertia weight, c ₁ and c ₂ are acceleration constants, generally 2, r ₁ and r ₂ are in ( 0, 1), t is the iteration algebra.

Before the PSO algorithm optimizes the parameters, it is necessary to set the range of some parameters: where ω is the inertia weight set to 1, c ₁ and c ₂ are set to 2, the maximum flight speed upper limit V _max is set to 200, and the total number of population particles is 50 , and the maximum number of iterations is set to 10000. In this article, the parameters that need to be tuned should also be set within the range. The value range of the penalty factor C in the SVM algorithm and the parameter γ of the RBF kernel function are both set to (-100, 100), and the selected value range of the parameter K value in KNN is (0, 50), and the range of another threshold ω for judging distance is set to (0.3, 0.8).

After the parameters are set, the parameter optimization steps are as follows:

(1) Initialize the position and velocity of all particles.

(2) Calculate the classification accuracy of the SVM-KNN classification model under the parameters, and use it as a fitness function to calculate the fitness of each particle.

(3) Update Pbest by comparing the current individual extremum with the searched fitness, compare all Pbests and select the best overall extremum Gbest from them, and update Gbest.

(4) For each update, the SVM penalty factor C is reset to create a larger research space for particles and avoid falling into the local region of the current optimal value.

(5) Update the velocity and position of the particle according to the update formula of velocity and position mentioned above.

(6) When the maximum number of iterations is reached, output the optimal parameters; otherwise, return to step (2).

After tuning the parameters of the SVM-KNN classification model through the PSO algorithm, a set of optimal parameters is obtained, and the classification model is set according to the parameters.

Rely on the optimal parameters found by the PSO algorithm, set the penalty factor C of the SVM algorithm and the parameter γ of the kernel function, the K value in the KNN algorithm, and the threshold ω in the SVM-KNN algorithm.

S4.2. Calculate the distance between the sample to be tested and the optimal hyperplane based on the pre-trained classification model to obtain the sample distance;

Specifically, the distance in the classification model in this paper is calculated by the following formula:

In the above formula, a _i is the Lagrangian multiplier, y _i a _i is the coefficient of the support vector in the decision function, k is the system constant, and b is the constant term of the decision function in SVM.

The distance of judging the distance is determined by the set threshold. The smaller the threshold, the larger the proportion of the SVM algorithm in the fusion algorithm, and the overall algorithm is more biased towards the SVM algorithm. When the threshold is 0, the fusion algorithm is equivalent to SVM Algorithm; on the contrary, the larger the threshold, the greater the proportion of the improved KNN algorithm in the fusion algorithm. When the threshold is as large as infinity, this algorithm is equivalent to the KNN classification algorithm.

S4.3. Comparing the sample distance with a preset threshold;

S4.4. Determine that the absolute value of the distance to the sample is less than the preset threshold, use the KNN algorithm to classify, and output the classification result;

Specifically, g(x) is compared with the threshold ω, if |g(x)|<ω, this sample point can be regarded as a sample point close to the optimal hyperplane, and the KNN algorithm is used for classification, and the calculation is to be Measure the distance between the sample point and all the feature vectors, select the sample points with the minimum K distances, and count their corresponding categories. The category of the sample to be tested is the same as the category with a high probability of occurrence.

S4.5. Determine that the absolute value of the distance to the sample is greater than or equal to the preset threshold, use the SVM algorithm to classify, and output the classification result.

Specifically, if |g(x)|》ω, this sample point can be regarded as a sample point far away from the optimal hyperplane, and the SVM algorithm is used for classification. In the SVM algorithm, calculate F(x)=sgn( g(x)) yields the categorical categories.

The present invention proposes a novel signal classification method—based on the PSO improved SVM-KNN classification algorithm. The classification algorithm solves the problem of the SVM algorithm in the face of samples that are relatively close to the hyperplane by integrating the SVM and the improved KNN algorithm. Point-time classification is not accurate. Utilize PSO algorithm to optimize traditional SVM-KNN algorithm at the same time, seek the optimum parameter of SVM-KNN classifier, have improved the classification performance of system, have had a substantial promotion to the efficiency of sentiment classification; Adopted time domain in the present invention , frequency domain, time-frequency domain, non-linear, and spatial domain features to extract the features of the EEG signal, and retain the information contained in the EEG signal as completely as possible. Then calculate the contribution rate of each feature through PCA, remove all features with a contribution rate lower than 85%, and construct a new feature matrix, which can not only retain the information in the signal to the greatest extent, but also improve the computational efficiency of the classifier.

As shown in Figure 2, a signal feature extraction system for multi-element feature fusion extraction of EEG signals, including:

The preprocessing module is used to obtain the EEG signal data and perform signal preprocessing on the EEG signal data to obtain the preprocessed signal;

The feature extraction module is used to perform multi-dimensional feature extraction and matrix construction on the preprocessed signal to obtain the original feature matrix;

A dimensionality reduction module is used to perform fusion and dimensionality reduction processing on the original feature matrix to obtain the final feature matrix;

The classification module is used to input the final feature matrix to the pre-trained classifier for classification, and output the classification result.

The content in the above-mentioned method embodiments is applicable to this system embodiment. The specific functions realized by this system embodiment are the same as those of the above-mentioned method embodiments, and the beneficial effects achieved are also the same as those achieved by the above-mentioned method embodiments.

A signal feature extraction device for multiple feature fusion extraction of EEG signals:

at least one processor;

at least one memory for storing at least one program;

When the at least one program is executed by the at least one processor, the at least one processor implements the signal feature extraction method for multi-element feature fusion extraction of EEG signals as described above.

The content in the above-mentioned method embodiment is applicable to this device embodiment, and the specific functions realized by this device embodiment are the same as those of the above-mentioned method embodiment, and the beneficial effects achieved are also the same as those achieved by the above-mentioned method embodiment.

A storage medium, which stores processor-executable instructions, is characterized in that: the processor-executable instructions are used to realize the above-mentioned multi-element feature fusion extraction of EEG signals when executed by the processor Signal feature extraction method.

The content in the above-mentioned method embodiments is applicable to this storage medium embodiment. The functions realized by this storage medium embodiment are the same as those of the above-mentioned method embodiments, and the beneficial effects achieved are also the same as those achieved by the above-mentioned method embodiments. same.

The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the described embodiments, and those skilled in the art can also make various equivalent deformations or replacements without violating the spirit of the present invention. , these equivalent modifications or replacements are all within the scope defined by the claims of the present application.

Claims (7)

1. A signal feature extraction method for multi-element feature fusion extraction of an electroencephalogram signal is characterized by comprising the following steps:

acquiring electroencephalogram signal data and performing signal preprocessing on the electroencephalogram signal data to obtain a preprocessed signal;

performing multi-dimensional feature extraction and matrix construction on the preprocessed signals to obtain an original feature matrix;

performing fusion dimensionality reduction processing on the original feature matrix to obtain a final feature matrix;

and inputting the final characteristic matrix into a pre-trained classification model for classification, and outputting a classification result.

2. The method for extracting signal features of multi-feature fusion extraction of electroencephalogram signals according to claim 1, wherein the step of obtaining an electroencephalogram signal data set and performing signal preprocessing on the electroencephalogram signal data set to obtain a preprocessed signal specifically comprises:

acquiring electroencephalogram signal data;

carrying out noise separation processing on the electroencephalogram signal data based on an independent component analysis method to obtain a separated signal;

performing frequency screening processing on the separated signals based on a Butterworth band-pass filter to obtain screened signals;

and normalizing the screened signals to obtain preprocessed signals.

3. The method for extracting signal features of multi-feature fusion extraction of electroencephalogram signals according to claim 2, wherein the step of performing multi-dimensional feature extraction and matrix construction on the preprocessed signals to obtain an original feature matrix specifically comprises:

performing feature extraction on the preprocessed signals on a time domain to obtain time domain features;

performing feature extraction on the preprocessed signal on a frequency domain based on an AR model power spectrum estimation method to obtain frequency domain features;

performing feature extraction on the preprocessed signal on a time-frequency domain based on a Hilbert-Huang transform method to obtain time-frequency domain features;

performing feature extraction on the preprocessed signals on a nonlinear domain based on nonlinear dynamics analysis to obtain nonlinear domain features;

performing feature extraction on the preprocessed signals on a space domain based on public space mode analysis to obtain space domain features;

and constructing a matrix according to the time domain characteristics, the frequency domain characteristics, the time-frequency domain characteristics, the nonlinear domain characteristics and the space domain characteristics to obtain an original characteristic matrix.

4. The method for extracting signal features of multi-feature fusion extraction of electroencephalogram signals according to claim 3, wherein the time-domain features comprise standard deviation, root mean square, mean of first-order difference absolute values, and the non-linear features comprise approximate entropy, fuzzy entropy, and sample entropy.

5. The method for extracting signal features through multi-feature fusion extraction of electroencephalogram signals according to claim 3, wherein the step of performing fusion dimensionality reduction processing on the original feature matrix to obtain a final feature matrix specifically comprises the following steps:

carrying out mean value removing processing on the original characteristic matrix to obtain a centralized matrix;

calculating a covariance matrix of the centralized matrix and decomposing an eigenvalue to obtain an eigenvalue and a corresponding eigenvector;

and sequencing the characteristic values and sequentially taking the characteristic vectors corresponding to the characteristic values of the preset number to construct a matrix to obtain a final characteristic matrix.

6. The method for extracting signal features through multi-feature fusion extraction of electroencephalogram signals according to claim 5, wherein the step of inputting the final feature matrix into a pre-trained classification model for classification and outputting a classification result specifically comprises:

performing parameter optimization on the SVM-KNN classifier based on a particle swarm algorithm and a pre-constructed training set to obtain a pre-trained classification model;

calculating the distance between the sample to be tested and the optimal hyperplane based on a pre-trained classification model by taking the final characteristic matrix as input to obtain a sample distance;

comparing the sample distance with a preset threshold value;

judging that the absolute value of the distance to the sample is smaller than a preset threshold value, classifying by adopting a KNN algorithm, and outputting a classification result;

and judging whether the absolute value of the distance to the sample is larger than or equal to a preset threshold value, classifying by adopting an SVM algorithm, and outputting a classification result.

7. The method for extracting signal features of multi-feature fusion extraction of electroencephalogram signals according to claim 6, wherein the calculation formula of the sample distance is as follows:

in the above formula, a _i Is a Lagrangian multiplier, y _i a _i And k is a system constant and b is a constant term of the decision function in the SVM for the coefficient of the support vector in the decision function.

Images