Electroencephalogram emotion signal recognition method based on graph regularized non-negative matrix factorization

Abstract

The application discloses a kind of electroencephalogram emotion signal recognition method based on graph regular non-negative matrix decomposition, comprising the following steps: step one, using nearest neighbor method for electroencephalogram emotion signal constructs an adjacent matrix;Step two, establish a graph regular non-negative matrix decomposition model, and the non-negative constraint of matrix after decomposition is carried out;Step three, introduce a projection matrix, further to matrix implement three decomposition;Step four, the model is optimized;Step five, using the new representation PX obtained after the multiplication of trained model parameter P and sample X replaces the representation matrix V of original model, joins corresponding label matrix and trains a classifier in linear SVM classifier;Step six, the class of test sample is predicted by inputting into classifier.The application can effectively identify electroencephalogram emotion signal, compared with other classic electroencephalogram emotion recognition method, the application effectively improves recognition rate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electroencephalogram signal processing and pattern recognition, and relates to an electroencephalogram emotional signal recognition method based on graph regular non-negative matrix factorization. BACKGROUND

[0002] The method of graph regular non-negative matrix factorization is a classical method in machine learning and pattern recognition. The electroencephalogram emotional signal is a data type of graph network structure, and therefore the graph method can be used to well utilize the graph structure of the electroencephalogram signal for electroencephalogram emotion recognition.

[0003] At present, there are many methods for electroencephalogram emotional signal emotion recognition, including SVM and DBN methods. The SVM method is a classical classification method, can process high-dimensional data, has good generalization ability, and is beneficial to electroencephalogram signal classification, but the method is sensitive to noise, and the performance of the model may be affected when there is noise in the data, and noise is easily generated during electroencephalogram signal acquisition. The DBN method is a classical deep learning model, and the DBN can automatically extract high-order features of the electroencephalogram signal, but the training process of the network is relatively complex, and the model needs to be carefully designed and the parameters need to be set, which is easy to fall into a local optimal solution and is also more likely to have the phenomenon of overfitting. SUMMARY

[0004] The purpose of the application is to provide an electroencephalogram emotional signal recognition method based on graph regular non-negative matrix factorization, which can effectively recognize electroencephalogram emotional signals, and effectively improve the recognition rate compared with other classical electroencephalogram emotional recognition methods.

[0005] To achieve the above purpose, the application provides the following technical scheme: an electroencephalogram emotional signal recognition method based on graph regular non-negative matrix factorization, comprising the following steps:

[0006] Step one, using the k-nearest neighbor algorithm to calculate the distance or similarity between each channel of the electroencephalogram signal, and constructing an adjacency matrix according to the distance or similarity between them;

[0007] Step two, constructing a graph regular non-negative matrix factorization model, and performing non-negative constraint on the decomposed matrix;

[0008] Step three, introducing a projection matrix to further implement three decompositions on the matrix;

[0009] Step four, optimizing the model;

[0010] Step five, multiplying the trained model parameters P with the sample X to obtain a new representation PX, and inputting the new representation PX and the corresponding label matrix into the SVM classifier to train a classifier;

[0011] Step six, input the test sample into the classifier to predict its category.

[0012] Further, the specific method of step one is to use the k-nearest neighbor algorithm to calculate the distance or similarity between each channel of the electroencephalogram signal, which will generate a 62x62 similarity matrix, and further convert the similarity matrix into an adjacency matrix.

[0013] Further, the specific method of step three is to introduce a projection matrix P to transform the representation matrix V into PX, X being the sample, thereby transforming the original two-factorization into a three-factorization form, and additionally, a regularization term is added to make the model smoother

[0014] Further, the specific method of step four is to use the KKT condition to optimize and solve the model based on the non-negative constraint of the model.

[0015] Advantages of the present application: the method of the present application constructs a graph regular non-negative matrix factorization method to perform emotion recognition on electroencephalogram emotional signals. Compared with the prior art such as SVM and DBN, the method in the present application has the following advantages:

[0016] (1) The introduction of the graph regular method can well establish the effective connection between each channel of the electroencephalogram emotional signal, and the method can obtain feature information more effectively by calculating the distance or similarity between each channel and other channels to obtain a distance or similarity matrix.

[0017] (2) In the construction of the model, we improve the traditional matrix factorization model, change the traditional two-factorization into a three-factorization model, and represent the decomposition matrix V as V=PX, so that the process of solving the generalized inverse of the matrix V is omitted during testing and application, and only the sample needs to be multiplied by the matrix P obtained in the three-factorization, that is, PX. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flowchart of the present application;

[0019] Figure 2 The layout diagram of the 62-channel electroencephalogram cap. DETAILED DESCRIPTION

[0020] The present application will be described in detail below in combination with the drawings and examples.

[0021] As Figure 1 shown, a graph regular non-negative matrix factorization based electroencephalogram emotional signal recognition method includes the following steps:

[0022] Step one, using k-neighborhood algorithm to calculate the similarity between each channel of electroencephalogram signal, and constructing an adjacent matrix according to the similarity between them.

[0023] Step two, based on electroencephalogram emotional signal, a traditional graph regular non-negative matrix model is constructed to learn the parameter matrix V, in order to avoid solving the pseudo-inverse of the parameter matrix V in the test process, a projection matrix P is introduced, and V is expressed as V=PX.

[0024] Step three, the model is optimized and calculated, and an effective parameter matrix is learned;

[0025] Step four, based on the trained P and label information, an LSVM is used to train a classifier, and the real data and test set are predicted.

[0026] In the preferred scheme of the application, the electroencephalogram emotional data comes from the database of Shanghai Jiaotong University (SJTU Emotion EEG Dateset, SEED).

[0027] The application is based on the matrix decomposition of graph regularization, so the construction of the adjacent matrix is introduced first, and then the model of the projection matrix decomposition is introduced on this basis.

[0028] 1. Based on the EEG electroencephalogram emotional signal related signal channel, the graph regularization method is used to establish the adjacent matrix L of the EEG signal V ,L V =D V -W V .

[0029] Wherein, we define

[0030]

[0031] W V Is a symmetric weight matrix, N(x) represents the neighbor of sample x, matrix D V Is a diagonal matrix, and the diagonal elements are the row sum or column sum of matrix W V , that is

[0032] 2. Construct a model based on graph regular non-negative matrix decomposition

[0033] A standard graph regular non-negative matrix decomposition model is as follows:

[0034]

[0035] Wherein, X=[x1,x2,...,x N ]∈R M×N Is the given EEG emotional signal data, wherein N is the number of samples, xi (i = 1, 2, ..., N) represents the i-th sample, which is the result of vectorizing the EEG signal. It is an M×N matrix, and each element of it is non-negative. This model represents the input data as the product of two non-negative encoding matrices U and V, thus representing the sample well. However, when testing the validity of the sample, we may need to solve the pseudo-inverse of U. This process cannot guarantee the non-negativity of the new test sample, thus violating the basic assumption of factorization of non-negative matrices. To avoid this trouble, we further modify each test sample by introducing a projection matrix in model (2) to maintain a low-dimensional non-negative representation. Let V = [v1, v2, ..., v N We assume that each component v i It is from sample x i It is obtained through linear transformation, i.e.

[0036] v i =Px i Or V = PX.

[0037] Here P∈R K×M It is the projection matrix. Therefore, the objective function (2) can be rewritten as:

[0038]

[0039] After obtaining the projection matrix P, we can easily obtain a new representation for the test sample.

[0040]

[0041] It is worth noting It is non-negative and computationally more efficient.

[0042] From model (3), we can see that the sample matrix X can be represented as UPX, thus we can approximate UP as an identity matrix I. To ensure a smoother objective function, we add an extra penalty function term to formula (3). Therefore, we obtain a new objective function:

[0043]

[0044] Where λ1, λ2 ≥ 0, λ1, λ2 are adjustable hyperparameters used to balance the three regularization terms in the objective function, I ∈ R. M ×M It is an identity matrix with the same dimension as UP. During training, we set λ1 = 10 and λ2 = 1.

[0045] 3. Optimize the model

[0046] We will use the Lagrange multiplier method to deal with the constraints in the model, assuming and ψ ij are the Lagrange multipliers, respectively, with u ij ≥0, p ij ≥0, and Ψ=[ψ ij ], then the Lagrange function is:

[0047]

[0048] The partial derivatives of the model L with respect to the basis matrix U and the projection matrix P are respectively:

[0049]

[0050]

[0051] From the KKT conditions φ ij u ij =0 and ψ ij p ij =0, we get the following equations:

[0052]

[0053]

[0054] From the previous L V =D V -W V , so equation (9) can be written as:

[0055]

[0056] According to equations (7) and (9), we further obtain the update rules of u ij and p ij :

[0057]

[0058]

[0059] 4、Test

[0060] After obtaining the projection matrix P, we use the LSVM classifier to train a linear classifier with PX and the corresponding label matrix, and then use this classifier to predict the test samples or actual data. When predicting, only the projection matrix P is multiplied by the test sample or actual data and input into the classifier, and the sample emotion category can be predicted.

[0061] The action recognition experiment process of the method of the application is as follows:

[0062] The graph construction process used in this embodiment is to use the k- nearest neighbor algorithm to calculate the distance or similarity between each channel, and construct an adjacency matrix according to the distance or similarity between them.

[0063] EEG signals are brainwave signals collected by an electrode cap with many electrode positions touching the scalp. A public electrode cap includes 64 or 128 electrode positions in accordance with the international 10-20 system, and the electrode positions are also referred to as channels. When the participants are stimulated by a given visual and auditory stimulus, the brain can generate brainwave signals corresponding to the emotions, and these signals are collected by the electrode cap worn. Scientists divide the brainwave signals into different frequency bands based on frequency differences, including Alpha (a), Beta (b), gamma (g), Theta (0) and Delta (d) bands. Table 1 shows the range of the five frequency bands.

[0064] Table 1 Frequency band range of EEG signals

[0065] Waveband Beta Delta Theta Alpha Gamma Frequency 0-3Hz 4-7Hz 8-13Hz 14-30Hz 31-50Hz

[0066] We use the SJTU Emotion EEG Dataset (SEED) and the Multimodal Psychophysiological Emotional Database (MPED) to evaluate the proposed method. SEED selects 15 participants (7 males and 8 females), and each group has three separate sessions. In the signal processing process, the subjects are assigned to watch movie clips corresponding to three emotions (negative, neutral and positive), so that the brain can generate corresponding emotional signals. The SEED emotional signals are taken from 62 channels of electrodes (as shown in Figure 2 Table 1) using the ESI NeuroScan system with a sampling rate of 1000 Hz. We use five single frequency bands and the full frequency band to comprehensively evaluate the proposed method.

[0067] Based on the experimental protocol provided by the SEE database, in Table 2 we use the EEG emotion recognition results to compare with our method, which includes two baseline methods (linear support vector machine (SVM) and deep belief network (DBN)). In addition, we empirically set the hyperparameters in the model to λ1=10 and λ2=1.

[0068] Classification and evaluation:

[0069] Table 2 shows the experimental results of the three methods in terms of recognition accuracy (%). From the results in Table 2, we can clearly see that in the full band (a + b + g + 0 + d) and most of the single bands, the performance of the proposed method is much higher than that of the SVM and DBN methods. The experimental results of each band show that in the single band, the recognition rate of a and b bands is significantly better than that of 0, d and g bands. The reason for this phenomenon may be that the low-frequency signals of EEG features (such as 0, d and g bands) are related to the brain in sleep or drowsy state, while the high-frequency signals (such as a and b bands) contain more useful information related to the cognitive activities of the human brain. In addition, the three methods in Table 2 all achieve the best recognition rate in the full band, and the recognition accuracy is more than 80%. From these results, we believe that the full band of EEG signals may contain more comprehensive and useful emotional information of the human brain in cognitive activities.

[0070] Table 2 Experimental results on SEED database

[0071] Method Delta Theta Alpha Beta Gamma Complete SVM 60.50 60.95 66.64 80:76 79.56 83.99 DBN 64.32 60.77 64.01 78.92 79.19 86.08 GPNMF 66:78 69:54 71:65 80.22 81:02 86:26

[0072] The above description only expresses the preferred embodiments of the present application, which is described in more detail and in more detail, but it cannot be understood as limiting the scope of the patent of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, several modifications, improvements and substitutions can be made, which belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for recognizing EEG emotion signals based on graph regular nonnegative matrix factorization, characterized in that: Includes the following steps: Step 1: Use the k-nearest neighbor algorithm to calculate the distance or similarity between each channel of each EEG signal, and construct an adjacency matrix based on the distance or similarity between them; Step 2: Construct a graph regular nonnegative matrix decomposition model and impose nonnegativity constraints on the decomposed matrix; Step 3: Introduce a projection matrix , will represent matrix Transformation ,in Using samples, the original binary decomposition is transformed into a ternary decomposition. Additionally, a regularization term is added to make the model smoother. ; Step 4: Optimize the model; Step 5: Use the trained model parameters With sample The new representation obtained after multiplication The corresponding label matrix is ​​then input into an SVM classifier to train a classifier. Step 6: Input the test sample into the classifier to predict its category.

2. The EEG emotion signal recognition method based on graph regular nonnegative matrix factorization according to claim 1, characterized in that: The specific method of step one is as follows: use the k-nearest neighbor algorithm to calculate the distance or similarity between each channel of each EEG signal, which will generate a 62x62 similarity matrix, and further convert the similarity matrix into an adjacency matrix.

3. The EEG emotion signal recognition method based on graph regular nonnegative matrix factorization according to claim 1, characterized in that: The specific method for step four is to optimize and solve the model using KKT conditions based on the model's non-negative constraints.

Images