An electroencephalogram signal dictionary learning method based on optimization of label consistency

Abstract

The application discloses a brain electrical signal dictionary learning method based on label consistency optimization, which comprises the following steps: preprocessing training set sample data of brain electrical signals; decomposing the preprocessed brain electrical signals into intrinsic mode functions by using empirical wavelet transform (EWT); initializing control relative contribution coefficients alpha, control relative contribution coefficients beta, initial dictionary D 0 , transformation matrix A 0 and classifier parameters W 0 ; obtaining dictionary model parameters X new and D new by using the initialized parameters; updating the dictionary model parameters D new by using a golden eagle algorithm label consistency optimization method; obtaining a dictionary D, a transformation matrix A and classifier parameters W that meet classification requirements from the updated dictionary; inputting test set brain electrical signal samples y i ; performing sparse representation on the test set brain electrical signal samples y i by using the dictionary D, so as to obtain coefficients s i ; and reconstructing the test set brain electrical signal by using a reconstruction algorithm OMP, and simultaneously performing corresponding discriminant classification on the test set brain electrical signal by using a linear classifier. The application improves the classification accuracy and reconstruction quality of brain electrical signals.

Description

Technical Field

[0001] The present invention relates to the field of signal processing technology, and in particular to an electroencephalogram (EEG) signal dictionary learning method based on optimized label consistency. Background Art

[0002] EEG signals reflect the synaptic potential activity generated by brain neurons, as recorded on the scalp or its surface. EEG data contains a wealth of information and is a key area of ​​research. In-depth analysis of EEG signals can help us understand the brain's information processing processes and accelerate the exploration of its appropriate applications. Furthermore, EEG signals contain a wide range of information about the human body. Through in-depth research, we can provide scientific diagnosis and treatment for patients with brain disorders, and also provide data support for the prevention of certain brain diseases.

[0003] To obtain high-quality EEG signal data, a data acquisition system must consist of electrodes, amplifiers, analog-to-digital converters, and data recorders. The Nyquist theorem is a typical signal acquisition method, where signals are sampled, compressed, and encoded at the acquisition end, and finally decompressed to obtain the original signal at the receiving end. As the number of channels increases, this method generates massive amounts of signal data, posing significant challenges to its transmission and storage. Compressed sensing technology overcomes the drawbacks of traditional compressed sampling methods, conserving storage space and reducing the power consumption required for signal acquisition and transmission, alleviating the pressure on the acquisition system. It also improves system hardware, reducing size and enhancing portability.

[0004] The quality of dictionary learning directly impacts the performance of sparse representation, which in turn affects the reconstruction results of compressed sensing technology. With the deepening of research in the field of dictionary learning, scholars have proposed numerous effective dictionary learning methods. Unsupervised dictionary learning methods mainly include multi-component dictionaries, K-singular value decomposition (K-SVD) dictionaries, and optimal direction algorithms, with K-SVD being the most widely used. Supervised dictionaries utilize the signal's label data to effectively reconstruct the original signal and can also be applied to classification problems. Avnessh S et al. directly used EEG signal training samples as dictionaries, calculated the sparse representation vector based on the L1 norm, and finally calculated the reconstruction error to determine the type of sample data. This classification method does not require dictionary training and is therefore computationally efficient, but it loses a lot of classification information hidden in the training sample set. Separating dictionary learning from classifier learning may reduce classification accuracy. Jiang et al. proposed the Label Consistent K-SVD (LC-KSVD) algorithm for face recognition. They incorporated two constraints into the K-SVD objective function: a classification error term, used to optimize the linear classifier; and a label regularization term, a sparse representation error term, used to constrain the sparse coding coefficients. This allows the dictionary to discriminate between signals while learning the dictionary. Yan Chunman et al. used principal component analysis to improve dictionary updating, thereby increasing face recognition efficiency. Ma, Yang Xiaoling et al. extracted image or signal features and used the LC-KSVD algorithm to learn dictionaries, achieving relatively good results.

[0005] In summary, this supervised dictionary learning approach is currently still primarily in the application phase. Most literature combines feature extraction or preprocessing with LC-KSVD for specific signals or images to learn a dictionary that combines EEG signal reconstruction and classification. However, dictionary learning has limited ability to classify and reconstruct unevenly distributed signals, and the speed of dictionary learning needs to be improved. Therefore, further research is needed on supervised dictionary learning methods. Summary of the Invention

[0006] The purpose of the present invention is to provide an EEG signal dictionary learning method based on optimized label consistency (Empirical Wavelet Transform, Golden Eagle Optimizer and LCKSVD, EWT+GEOLCKSVD), which improves the EEG signal classification accuracy and reconstruction quality.

[0007] To achieve the above objectives, this application proposes an EEG signal dictionary learning method based on optimized label consistency, including:

[0008] Step 1: Preprocess the training set sample data of EEG signals, such as segmentation, filtering, denoising, etc.

[0009] Step 2: The preprocessed EEG signal is decomposed into intrinsic mode functions using empirical wavelet transform (EWT);

[0010] Step 3: Initialize the control relative contribution coefficient α, the control relative contribution coefficient β, and the initial dictionary D 0 , transformation matrix A 0 and the classifier parameters W 0 ;

[0011] Step 4: Use the initialization parameters to obtain the dictionary model parameters X new and D new ;

[0012] Step 5: Use the Golden Eagle algorithm to optimize the label consistency method LCKSVD to update the dictionary model parameters D new ;

[0013] Step 6: Obtain the dictionary D, transformation matrix A and classifier parameters W that meet the classification requirements from the updated dictionary;

[0014] Step 7: Input the test set EEG signal sample y i ;

[0015] Step 8: Use dictionary D to analyze the EEG signal samples y in the test set i Perform sparse representation and obtain its coefficient s i ;

[0016] Step 9: Use the reconstruction algorithm OMP to reconstruct the EEG signals of the test set, and use the linear classifier to perform corresponding discrimination classification on them, that is, the label l corresponding to the largest element is the category corresponding to this EEG signal.

[0017] Furthermore, based on the prior knowledge and the specific real EEG signal data, the optimal values ​​of the control relative contribution coefficients α and β are found; at the same time, the Golden Eagle algorithm is used to optimize the label consistency method LCKSVD to obtain the dictionary items of each type of EEG signal intrinsic mode function, and the dictionary items of each type of EEG signal are merged to obtain the initial dictionary D 0 .

[0018] Furthermore, according to formula (1), the initial dictionary D 0 , obtain the sparse coding coefficient S of sample X through iteration 0 :

[0019]

[0020] Where s is a column of the sparse coding coefficient matrix S; st is the constraint function;

[0021] Initialize the transformation matrix A using the multivariate ridge regression model shown in formula (2) 0 :

[0022]

[0023] A 0 =QS T (SS T +λ1I) -1 (3)

[0024] For the classifier parameter W 0 Initialization, again using the ridge regression model, we get:

[0025] W 0 =HS T (SS T +λ2I) -1 (4)

[0026] Among them, λ1, λ2 are positive numbers, I represents the unit matrix, S T represents the transpose of S; Q is the discriminant sparse coding; H is the class label matrix.

[0027] Furthermore, the dictionary model parameters X are obtained using the initialization parameters new and D new , specifically;

[0028]

[0029] Furthermore, the model of the Golden Eagle algorithm optimization label consistency method LCKSVD is:

[0030]

[0031] Among them, the parameters α and β are the control relative contribution coefficients, which are used to adjust the proportions of the three terms in the model. The matrix H is the class label matrix of the input sample X, and the matrix W is the classifier parameter. The first term represents the reconstruction error term of the sample data; the second term is the label regularization term, which represents the discriminant sparse coding error; and the third term is the classification error term.

[0032] Furthermore, when solving the sparse coding coefficient matrix S, equation (6) is expressed as:

[0033]

[0034] Then, using formula (5) to optimize formula (7) is equivalent to using the Golden Eagle algorithm to optimize the label consistency method LCKSVD, that is, solving formula (8):

[0035]

[0036] In the process of updating the dictionary atoms, first obtain the error matrix Perform singular value decomposition E on it k =U∑V T , where U, Σ, and V represent E k The left singular matrix, singular value matrix, and right singular matrix of d are then updated using the Golden Eagle algorithm. k , and use the first column of V multiplied by the first element of ∑(1,1) as the sparse coefficient s k , until the maximum number of iterations is reached, and the dictionary and sparse coding coefficient matrix are output.

[0037] Furthermore, the dictionary D, transformation matrix A and classifier parameters W that meet the classification requirements are obtained from the updated dictionary, specifically:

[0038]

[0039] Among them, d1 is the first column of dictionary D, d k The kth column of the dictionary D; a1 is the first column of the transformation matrix A, a k is the kth column of the transformation matrix A; w1 is the first column of the classifier parameter W, w k is the k-th column of the classifier parameters W.

[0040] Furthermore, the label l corresponding to the largest element is:

[0041]

[0042] Where Ws j Represents the classifier W and sparse coefficient s j The product of .

[0043] The above technical solution adopted by the present invention has the advantages compared with the existing technology: the present invention fully considers the characteristics of EEG signals and reduces the dependence of dictionary learning on the EEG signals of the training set. First, the EEG signal training set data is subjected to denoising, segmentation and other processing, and then it is decomposed into intrinsic mode functions using empirical wavelets. Next, the Golden Eagle optimized LCKSVD method is used to learn a dictionary and a linear classifier, and then the test EEG signal is sparsely represented. Finally, the existing orthogonal matching pursuit algorithm is used to reconstruct and classify the signal. The method of the present invention improves the sparse representation capability of the dictionary, and at the same time uses the Golden Eagle optimization algorithm to update atoms to speed up the dictionary learning, thereby improving the EEG signal classification accuracy and reconstruction quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Collect norm maps for BCIⅡ competition datasetⅢ;

[0045] Figure 2This is the bandpass filter diagram of the data in EEGLAB;

[0046] Figure 3 This is the ICA denoised image in EEGLAB;

[0047] Figure 4 This is the effect diagram of parameter T on classification results;

[0048] Figure 5 This is the effect of parameter alpha on the classification results;

[0049] Figure 6 This is a comparison chart of the sparse index of the dictionary. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application. That is, the embodiments described are only part of the embodiments of this application, not all of them.

[0051] Example 1

[0052] In the present invention, label-consistent dictionary learning utilizes the labels of EEG signals and dictionary atoms, and integrates reconstruction and classification error terms in the objective function, which can effectively reconstruct and classify the EEG signals to be tested. However, this method has low classification ability for unevenly distributed EEG signals and the dictionary learning speed needs to be improved. To address this problem, a EEG signal dictionary learning method based on optimized label consistency is proposed, which utilizes the intrinsic mode function decomposed by the empirical wavelet transform and a label-consistent dictionary optimized by the Golden Eagle algorithm to learn a sparse dictionary. The reconstruction error term, sparse coding error term and classification error term are combined in one objective function. When sparsely representing similar EEG signals, the corresponding dictionary atoms can be selected by label. EEG signals from the same class have similar sparse coding, while EEG signals from different classes have different sparse coding. This allows for sharing features between classes, shortens the time for updating dictionary atoms, and further improves the efficiency of compressed sensing in processing EEG signals.

[0053] In order to verify the performance of the method proposed in the present invention, the proposed simulation experiment uses the dataset III of the 2003 BCI II competition. This dataset contains a tester performing two motor imagery tasks for the right hand and the left hand. On the same day, the tester performed seven sets of experiments, with a break in between each set of experiments. Each set of experiments performed 40 motor imagery tasks, so a total of 280 sets of experimental data were obtained. Since the EEG signal movement is mainly concentrated in the three leads C3, C4 and Cz, this experiment only records the data of these three leads. The electrode settings and acquisition time norms of the experiment are as follows: Figure 1As shown, in each experiment, the EEG signal acquisition time was 9 seconds. The first two seconds were rest and relaxation time. At the second second, a prompt sound was issued and a "+" sign was displayed on the screen. At the third second, a black arrow appeared on the computer screen, prompting the tester to imagine the left hand or the right hand. The task ended after 6 seconds.

[0054] Therefore, when processing the data, we need to remove unnecessary signals and only need to intercept the motor imagery signal within 3-9 seconds. Since the sampling rate is 128Hz, EEGLAB uses the prior knowledge that EEG signal information is mainly in the range of 0.5-30Hz to filter this frequency band. Figure 2 As shown in the figure, the denoising process of principal component analysis is as follows: Figure 3 As shown, the length of each signal segment after truncation is 768 sampling points.

[0055] In the objective function of the dictionary learning algorithm proposed in this invention, the values ​​of the parameters T, α and β will affect the classification accuracy. In order to find the parameter values ​​that can achieve the best classification and reconstruction effects, different values ​​are taken for them and then tested. The results are as follows: Figure 4 and Figure 5 shown.

[0056] It can be seen that when the sparse threshold T0 is 3, the classification accuracy of the three methods is the highest. Therefore, T0 = 3 in this experiment. As the value of α increases, the classification accuracy also increases. When α = 3, the accuracy curve tends to be flat. Further increases in the value almost stop the accuracy increase and also consume time. Conversely, if the value of α is too small, the classification accuracy is low. Therefore, in this embodiment, α = 3, and similarly β = 2.

[0057] Table 1 Comparison of classification accuracy of different methods

[0058]

[0059]

[0060] As shown in Table 1, the accuracy of the EWT+GEOLCKSVD method proposed in this paper is 84.21%. Compared with other methods, the accuracy is improved. It is 26.68% higher than the LC-KSVD1 method, 15.25% higher than the LC-KSVD2 method, 14.84% higher than the EWT+LCKSVD1 method, 9.16% higher than the EWT+LCKSVD2 method, about 1.63% higher than the references

[11] and

[12] , and 0.05% higher than the reference

[13] . Therefore, it can be shown that EWT+GEOLCKSVD has a better classification effect on EEG signals. The aforementioned document

[11] is Liu Chong, Zhao Haibin, Li Chunsheng, and Wang Hong. Motor imagery EEG signal classification based on CSP and SVM algorithm [J]. Journal of Northeastern University (Natural Science Edition), 2010, 31(08): 1098-1101; document

[12] is Luo Fei, Liu Pengfei, Luo Yuan, and Zhu Simeng. Motor imagery EEG feature extraction method based on multi-feature fusion [J]. Computer Applications, 2020, 40(2): 616-620; document

[13] is Wu L, Wang T, Wang Q, et al. EEG Signal Processing Based on Multivariate Empirical Mode Decomposition and Common Spatial Pattern Hybrid Algorithm [J]. International Journal of Pattern Recognition and Artificial Intelligence, 2019, 33(9): 1-10.

[0061] The goal of dictionary learning is to perform sparse representation of EEG signals. The sparsity index is an important indicator of sparse representation performance. The smaller the value, the stronger the sparse representation ability of the dictionary. When the number of dictionary atoms is 256, the sparse index of different dictionaries is as follows: Figure 6 As shown. Figure 6 It can be seen that under the condition of the same number of dictionary atoms, the LC-KSVD1 dictionary has the highest sparsity index of 6.8690, the LC-KSVD2 dictionary has a sparsity index of 6.2761, the EWT+LCKSVD1 dictionary has a sparsity index of 6.2173, the EWT+LCKSVD2 dictionary has a sparsity index of 5.9436, and the EWT+GEOLCKSVD dictionary proposed in this invention has a sparsity index of 5.7706, which means that the EWT+GEOLCKSVD dictionary has a stronger sparse representation capability.

[0062] In order to further verify the effectiveness of the learning algorithm of the present invention in compressed sensing, the mean square error of different dictionaries under the OMP reconstruction algorithm is calculated under the conditions that the number of dictionary atoms is 256 and the compression ratio is 0.4, and the reconstruction time is recorded.

[0063] Table 2 Comparison of reconstruction error and time of different algorithms

[0064]

[0065] As can be seen from Table 2, the EWT+GEOLCKSVD method has the smallest mean square error and the shortest reconstruction time, which further verifies the effectiveness of this dictionary learning method.

[0066] In summary, on the dataset III of the BCI II competition, the learning dictionary proposed in this paper achieved higher classification and reconstruction accuracy, proving the effectiveness of the proposed method from two perspectives.

[0067] The foregoing descriptions of specific exemplary embodiments of the present invention are for purposes of illustration and description. These descriptions are not intended to limit the invention to the precise forms disclosed, and it is apparent that many variations and modifications are possible in light of the foregoing teachings. The exemplary embodiments have been selected and described for the purpose of explaining the specific principles of the invention and their practical application, thereby enabling those skilled in the art to realize and utilize a variety of exemplary embodiments of the invention and various options and modifications. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for learning an EEG signal dictionary based on optimized label consistency, characterized in that: include: Step 1: Preprocess the training set sample data of EEG signals; Step 2: The preprocessed EEG signal is decomposed into intrinsic mode functions using empirical wavelet transform (EWT); Step 3: Initialize the control relative contribution coefficient α, the control relative contribution coefficient β, and the initial dictionary D 0 , transformation matrix A 0 and the classifier parameters W 0 ; Step 4: Use the initialization parameters to obtain the dictionary model parameters X new and D new ; Step 5: Use the Golden Eagle algorithm to optimize the label consistency method LCKSVD to update the dictionary model parameters D new ; Step 6: Obtain the dictionary D, transformation matrix A and classifier parameters W that meet the classification requirements from the updated dictionary; Step 7: Input the test set EEG signal sample y i ; Step 8: Use dictionary D to analyze the EEG signal samples y in the test set i Perform sparse representation and obtain its coefficient s i ; Step 9: Use the reconstruction algorithm OMP to reconstruct the EEG signals of the test set, and use the linear classifier to perform corresponding discrimination classification on them, that is, the label l corresponding to the largest element is the category corresponding to this EEG signal.

2. The method for learning an EEG signal dictionary based on optimized label consistency according to claim 1, characterized in that: According to the prior knowledge and the specific real EEG signal data, the optimal values ​​of the control relative contribution coefficients α and β are found; at the same time, the label consistency method LCKSVD is optimized according to the Golden Eagle algorithm to obtain the dictionary items of the intrinsic mode function of each type of EEG signal, and the dictionary items of each type of EEG signal are merged to obtain the initial dictionary D 0 .

3. The method for learning an EEG signal dictionary based on optimized label consistency according to claim 2, characterized in that: According to formula (1), use the initial dictionary D 0 , obtain the sparse coding coefficient S of sample X through iteration 0 : Where s is a column of the sparse coding coefficient matrix S; st is the constraint function; Initialize the transformation matrix A using the multivariate ridge regression model shown in formula (2) 0 : A 0 =QS T (SS T +λ1I) -1 (3) For the classifier parameter W 0 Initialization, again using the ridge regression model, we get: W 0 =HS T (SS T +λ2I) -1 (4) Among them, λ1, λ2 are positive numbers, I represents the unit matrix, S T represents the transpose of S; Q is the discriminant sparse coding; H is the class label matrix.

4. The method for learning an EEG signal dictionary based on optimized label consistency according to claim 3, characterized in that: Use the initialization parameters to obtain the dictionary model parameters X new and D new , specifically; 5. The method for learning an EEG signal dictionary based on optimized label consistency according to claim 1, characterized in that: The model of the Golden Eagle algorithm optimization label consistency method LCKSVD is: Among them, the parameters α and β are the control relative contribution coefficients, which are used to adjust the proportions of the three terms in the model. The matrix H is the class label matrix of the input sample X, and the matrix W is the classifier parameter. The first term represents the reconstruction error term of the sample data; the second term is the label regularization term, which represents the discriminant sparse coding error; and the third term is the classification error term.

6. The method for learning an EEG signal dictionary based on optimized label consistency according to claim 5, characterized in that: When solving the sparse coding coefficient matrix S, equation (6) is expressed as: Then, using formula (5) to optimize formula (7) is equivalent to using the Golden Eagle algorithm to optimize the label consistency method LCKSVD, that is, solving formula (8): In the process of updating the dictionary atoms, first obtain the error matrix Perform singular value decomposition E on it k =U∑V T , where U, Σ, and V represent E k The left singular matrix, singular value matrix, and right singular matrix of d are then updated using the Golden Eagle algorithm. k , and use the first column of V multiplied by the first element of ∑(1,1) as the sparse coefficient s k , until the maximum number of iterations is reached, and the dictionary and sparse coding coefficient matrix are output.

7. The method for learning an EEG signal dictionary based on optimized label consistency according to claim 1, characterized in that: Obtain the dictionary D, transformation matrix A, and classifier parameters W that meet the classification requirements from the updated dictionary, specifically: Among them, d1 is the first column of dictionary D, d k The kth column of the dictionary D; a1 is the first column of the transformation matrix A, a k is the kth column of the transformation matrix A; w1 is the first column of the classifier parameter W, w k is the k-th column of the classifier parameters W.

8. The method for learning an EEG signal dictionary based on optimized label consistency according to claim 1, characterized in that: The label l corresponding to the largest element is: Where Ws j Represents the classifier W and sparse coefficient s j The product of .

Images