CN113919387A

CN113919387A - Electroencephalogram signal emotion recognition based on GBDT-LR model

Info

Publication number: CN113919387A
Application number: CN202110946383.1A
Authority: CN
Inventors: 陈宇; 吴华健
Original assignee: Northeast Forestry University
Current assignee: Northeast Forestry University
Priority date: 2021-08-18
Filing date: 2021-08-18
Publication date: 2022-01-11

Abstract

The invention relates to an electroencephalogram signal emotion recognition method based on a GBDT-LR model, wherein the electroencephalogram signal emotion recognition method based on the GBDT-LR model comprises the following steps: firstly, acquiring a data set of related electroencephalogram signals, preprocessing the data set, removing artifacts in the data set by using a filtering method, reducing the dimension of the data set by using principal component analysis to obtain clean electroencephalogram signals, then extracting wavelet segment characteristics in the signals by using a wavelet packet decomposition algorithm, extracting permutation entropy, approximate entropy and sample entropy characteristics in the signals by using an information theory algorithm, and combining the permutation entropy, the approximate entropy and the sample entropy characteristics into a characteristic matrix. And sending the features into a Gradient Boosting Decision Tree (GBDT) model for automatic feature re-screening and selection, carrying out One-hot coding on the GBDT results to form new training data, and finally training and testing the new features in a classifier, wherein the classifier is a Logistic Regression (LR) classifier.

Description

Electroencephalogram signal emotion recognition based on GBDT-LR model

The technical field is as follows:

the invention relates to the field of brain-computer interfaces, in particular to an electroencephalogram signal emotion recognition method based on a GBDT-LR model.

Background art:

at present, the electroencephalogram has a plurality of important discoveries in the research field of emotion recognition. Papez in 1937 linked human physiological activity to emotion production, and he proposed a marginal system mechanism of emotion production, the Papez loop. He believes that the thalamus, upon receiving sensory information related to emotional stimuli, spreads to the sensory cortex and hypothalamus. The connection from the hypothalamus to the anterior nucleus of the thalamus and then to the cingulum cortex is proposed, when the signal from the hypothalamus and the information from the sensation are integrated by the cingulum cortex, the emotion is generated, and the output from the cingulum cortex to the hippocampus and then to the hypothalamus produces the emotional response cortex control from top to bottom. Subsequently, the psychologist Maclean has added the concept of visceral brain to the Papez theory, who thought that all organs related to emotion were regulated by the visceral brain and that the corresponding visceral and skeletal responses were mediated by the visceral brain through the hypothalamus, and Rizon et al extracted statistical features of brain electrical signals using three different wavelet basis functions of "db8", "sym8" and "coif5" in wavelet transform, and found that the three features were highly related to emotion. Koelstra and Patras recorded EEG signals evoked by video stimulation from several volunteers according to a valence-wake model, finding the feasibility of a combination of facial expression video and EEG signals for emotion recognition.

The invention content is as follows:

the invention provides electroencephalogram emotion recognition based on a GBDT-LR model, aiming at overcoming the defects and shortcomings of the existing method, and particularly relates to a combination of the GBDT-LR model and the LR model to solve the problem of electroencephalogram emotion recognition.

The electroencephalogram signal emotion recognition method based on the GBDT-LR model comprises the following steps:

step 1: acquiring a data set disclosed by university such as Mary empress university, performing noise reduction on the acquired data, eliminating artifacts such as electro-ocular signals and electromyographic signals in original signals by using a filtering method, performing down-sampling to 128Hz, and performing dimensionality reduction on the signals by using a Principal Component Analysis (PCA) algorithm to obtain clean electroencephalographic signals;

step 2: extraction of features using wavelet packet decomposition and information theory algorithms

And step 3: the extracted features are combined into a feature matrix, and then the extracted features are re-screened and combined using a Gradient Boosting Decision Tree (GBDT) model, so that the most significant features are extracted and further used in the training and testing of the final classification model, i.e., the Logistic Regression (LR) classifier. And finally, taking the characteristics of the GBDT model screening combination as new training data, and training and testing the LR classifier.

The implementation of step 1 comprises:

step 1.1: data published by university such as mary queen university are acquired, a data set collects physiological signals and corresponding emotion data of 32 volunteers, each volunteer watches 40 music videos containing different emotions, and the physiological signals of the volunteers are recorded into data files s01-s32. dat. When recording physiological signals, the system comprises 40 leads (front 32 lead brain electrical signals + back 8 lead peripheral physiological signals;

step 1.2: removing artifacts in the original signal by using a filtering method;

step 1.3: using a band pass filter, the sampling frequency was brought to 125 Hz;

step 1.4: using Principal Component Analysis (PCA) algorithm to reduce the dimension of the brain electrical signal, and when a sample is input, calculating:

wherein r is_ijCalculating the eigenvalue and the eigenvector for the element value of the sample characteristic matrix, and finally obtaining the principal component load:

finally, obtaining a principal component score;

the method for recognizing emotion of brain electrical signal based on GBDT-LR model as claimed in claim 1, wherein said step 2 includes the following steps:

step 2.1: extracting four wavelet segment characteristics from the preprocessed data by using a wavelet packet decomposition algorithm;

step 2.2: extracting three characteristics of approximate entropy, sample entropy and permutation entropy by using an information theory algorithm;

step 2.3: the extracted features are combined into a feature matrix.

The EEG signal emotion recognition method based on GBDT-LR model as claimed in claim 1, wherein said step 3 includes the steps of:

step 3.1: sending the combined feature matrix into a Gradient Boosting Decision Tree (GBDT) model for automatic re-screening and combination of features;

step 3.2: performing one-hot coding on the result obtained by the GBDT model to form new training data;

step 3.3: and sending the new training data into a Logistic Regression (LR) classifier for final training to obtain a final classifier model and a final classifier result.

Description of the drawings:

FIG. 1 is a flow chart of an electroencephalogram signal emotion recognition method based on a GBDT-LR model. .

Fig. 2 is a diagram comparing wavelet decomposition with wavelet packet decomposition principle.

FIG. 3 is a diagram of the GBDT-LR model architecture.

The specific implementation mode is as follows:

the technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. It is to be understood that the described embodiments are merely exemplary of the invention, and not restrictive of the full scope of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Fig. 1 is a schematic overall flow chart of the implementation of the present invention, and as shown in fig. 1, the method includes:

1. extracting wavelet packet characteristics in the original signal by using a wavelet packet decomposition algorithm:

wavelet decomposition overcomes the limitation of Fourier transform, has good local characteristics in time domain and frequency domain, can focus on any detail of an object, can decompose a signal into signals of multiple frequency channels, is an analysis method with multiple resolutions, and is a more detailed analysis and reconstruction method for the signal, which is extended from the wavelet decomposition. The wavelet decomposition is to decompose the signal into two parts of low frequency and high frequency, and in the next layer of decomposition, the decomposed low frequency part is decomposed into two parts of low frequency and high frequency, and so on, the frequency resolution of the wavelet transform is reduced along with the increase of the frequency. Wavelet packet decomposition not only decomposes the low frequency part of the signal, but also decomposes the high frequency part which is not subdivided by the wavelet decomposition, and can adaptively select a corresponding frequency band to be matched with the frequency spectrum of the signal according to the characteristics of the analyzed signal, and the formula is as follows:

let φ (t) be a scale function:

as a function of wavelets

Wherein h (n) and g (n) ═ 1^1-nh (1-n) is a pair of quadrature mirror filters. Wavelet function

Additional "admissible conditions" must be fulfilled, i.e.

In wavelet packet decomposition, for uniform function representation, let phi 0(t) be phi (t),

the following wavelet basis can be constructed from the two-scale equation:

in the formula, i is a node number; j is the number of decomposition stages. The wavelet packet decomposition coefficient of the signal f (t) ═ d00 at the j-th stage and k point can be expressed by the following recursion formula:

assuming that the original signal length is m × 2N points, the complete reconstruction of the f (t) signal can be expressed as:

in the formula:

and

is a wavelet packet basis function constructed according to a two-scale equation;

and

is a signal

At j-th order, k pointWavelet packet decomposition coefficients.

2. Extracting permutation entropy, sample entropy and approximate entropy in the electroencephalogram signals by using an information theory method:

the permutation entropy calculation method is as follows:

a time sequence { x (i) ═ 1,2, …, n } is set, and phase space reconstruction is performed on the time sequence to obtain a matrix:

in the formula: m and r are respectively embedding dimension and delay time; k — n — 1 (m — r). Each row in the matrix can be considered as one reconstruction component, for a total of K reconstruction components. Rearranging the J-th reconstruction component (x (J), x (J + r), …, x [ J + (m-1) r ]) in the X (i) reconstruction matrix according to the numerical value and ascending order, wherein J1, J2, … and jm represent the index of the column where each element in the reconstruction component is positioned, namely

x[i+(j₁-1)r]≤x[i+(j₂-1)r]≤…≤x[i+(j_m-1)r

If there are equal values in the reconstructed components, i.e.

x[i+(j₁-1)r]＝x[i+(j₂-1)r]

At this time according to j₁、j₂Is ordered by the value of j, i.e. when j is₁<j₂When there is

x[i+(j₁-1)r]≤x[i+(j₂-1)r]

Therefore, for each row in any time sequence X (i) reconstructed matrix, a set of symbol sequences can be obtained

S(l)＝(j₁，j₂，…，j_m)

In the formula: z is 1,2 …, k, and k is equal to m! M-dimensional phase space maps different symbol sequences (j)₁，j₂，…，j_m) Total of m! The symbol sequence S (l) is one permutation thereof. If the probability of each symbol sequence occurrence is calculated as (p)₁，p₂，…，p_k) Then according to the form of Shannon entropy, timeThe permutation entropy of the k different symbol sequences of the sequence X (i) can be defined as

When in use

When H is present_p(m) reaches a maximum ln (m!). For convenience, H is usually represented by ln (m!)_p(m) normalization, i.e.

0≤H_p＝H_p/ln(m！)≤1

The approximate entropy calculation is as follows:

let there be an N-dimensional time series u (1), u (2), … …, u (N) sampled at equal time intervals

Defining algorithm related parameters m, r, wherein m is an integer and represents the length of comparison vector, r is a real number and represents a metric value of' similarity

Reconstructing an m-dimensional vector X (1), X (2),.., X (N-m +1), where X (i) ═ u (i), u (i +1),.., u (i + m-1) ]

For i is more than or equal to 1 and less than or equal to N-m +1, counting the number of vectors meeting the following conditions

Where d [ X, X ] is defined as d [ X, X ] ═ max | u (a) -u (a) |, u (a) is an element of the vector X, d represents a distance between the vector X (i) and X (j), and is determined by a maximum difference between the corresponding elements, and j has a value range of [1, N-m +1], including j ═ i.

Defining:

the approximate entropy (ApEn) is defined as

The sample entropy is calculated as follows:

a time sequence X with a length N is provided, where X is { X (1), X (2),.., X (N) }, and the sample entropy is calculated as follows:

the time series X is constructed as an m-dimensional vector, i.e.

X(i)＝{x(i)，x(i+1)，…，x(i+m-1)} (1)

Wherein i is 1,2, …, N-m +1.

Defining the distance between X (i) and X (j) as d [ X (i), X (j) ] (i is not equal to j), and the distance is the largest one of the two corresponding elements, namely

Given a threshold r (r > 0), counting the number d [ X (i), X (j) ] < r and the ratio of the number of the total vectors N-m, i.e.

Averaging all the results obtained from equation (3), i.e.:

adding 1 to the dimension m, and repeating the above process, so that the sample entropy of the sequence is theoretically:

but in practice N cannot be infinite but a finite value, the estimated value of the entropy of the sample is

3. Classified training by adopting GBDT-LR model

The classification adopts two model combinations, firstly, the Gradient Boosting Decision Tree (GBDT) is used for automatic feature re-screening and combination, then, the result of the GBDT is subjected to one-hot coding to obtain new training data, and then, a Logistic Regression (LR) model is used for training. The model structure diagram is as follows in principle:

GBDT (gradient boosting decision tree) is a decision tree model based on integrated thought, which is essentially based on residual learning, can process various types of data, has high accuracy, and has strong robustness to abnormal values, but can not train data in parallel. GBDT uses an additive model to regress or classify data by continuously reducing the residual error generated during the training process. The GBDT performs multiple iterations, each iteration generates a weak classifier CART regression tree, and the classifier is obtained by training on the basis of the residual error result of the last classifier. The requirements for weak classifiers are low variance, high bias (low variance ensures that the model will not be over-fitted + high bias will be reduced during training, thus improving accuracy). In order to reduce the loss function as fast as possible, the negative gradient of the loss function is used as an approximation of the residual, and then the CART regression tree is fitted. The algorithm principle is as follows:

(1) initialization weak learning device

(2) For M1, 2, M has:

(a) for each

sample i

1,2

(b) Taking the residual error obtained in the previous step as a new true value of the sample, and taking the data (x)_i,r_im) N (x is used as training data for the next tree, resulting in a new regression tree f_m(x) Corresponding leaf thereofThe nodal region is R_jmJ is 1,2, …, J. Wherein J is the number of leaf nodes of the regression tree t.

(c) Calculating a best fit value for leaf region J ═ 1, 2.. J

(d) Updating strong learning device

(3) Get the final learner

LR (logistic regression) is a two-class model, and the logistic regression assumes that data are distributed according to the Bernoulli principle, and solves parameters by using gradient descent through a method of maximizing a likelihood function, so as to achieve the purpose of two-class data. Logistic regression has two assumptions:

assume one: assume that the data obeys a bernoulli distribution (0-1 distribution, coin toss). It assumes that h θ (x) is the probability that a sample is in the positive class and 1-h θ (x) is the probability that a sample is in the negative class. The model can be expressed as:

h_θ(x；θ)＝p

assume two: the probability p for a sample to be positive is assumed to be:

values are mapped between 0-1 by the sigmod function, so the final LR model is:

in logistic regression, the most common is that the cost function is cross entropy:

Claims

1. the electroencephalogram signal emotion recognition method based on the GBDT-LR model is characterized by comprising the following steps:

And 3, combining the extracted features into a feature matrix, and then using a Gradient Boosting Decision Tree (GBDT) model to re-screen and combine the extracted features, so that the most significant features are extracted and further used in the training and testing of the final classification model, namely a Logistic Regression (LR) classifier. And finally, taking the characteristics of the GBDT model screening combination as new training data, and training and testing the LR classifier.

2. The method for recognizing emotion of electroencephalogram signal based on GBDT-LR model according to claim 1, wherein the step 1 comprises the following steps:

finally, obtaining a principal component score;

3. the method for recognizing emotion of brain electrical signal based on GBDT-LR model as claimed in claim 1, wherein said step 2 includes the following steps:

step 2.3: the extracted features are combined into a feature matrix.

4. The EEG signal emotion recognition method based on GBDT-LR model as claimed in claim 1, wherein said step 3 includes the steps of: