EEG Data Decoding Method Based on the Hjorth Parameter Width Learning System Model

Abstract

The present invention relates to a method for decoding electroencephalogram (EEG) data based on a width learning system model of Hjorth parameters, which comprises the following steps: Step 1: Obtain an EEG data set and preprocess the EEG data set by using a line impulse response band-pass filter; Step 2: Calculate the Hjorth parameters of the preprocessed EEG data set and construct an EEG data set feature matrix; Step 3: Establish a width learning system model according to the feature matrix of the Hjorth parameters; Step 4: Decode the EEG data according to the width learning system model. This method combines the characteristics of Hjorth parameters and the time-frequency information of EEG data, and establishes a width learning system model to achieve EEG data decoding. This model has the characteristics of simple structure, few input parameters, and fast model training; this method realizes the rapid fitting of EEG data features, effectively improves the accuracy of EEG data classification, improves the utilization rate of memory resources, and the model has strong applicability and good actual application effects.

Description

Technical Field

[0001] This application relates to the fields of bio-signal processing and pattern recognition, and particularly to an electroencephalogram data decoding method based on a width learning system model with Hjorth parameters. Background Art

[0002] With the development of BCI technology, in the process of the controllable range of patients with movement disorders and the popularization of BCI technology in the healthy population, the above research methods are insufficient because an efficient BCI system should allow users to freely interact with external devices through various senses such as vision and hearing, and provide additional functions and degrees of freedom for more complex tasks. Imagined speech signal is a new trend in BCI paradigms. When using imagined speech as a research paradigm, it is required that the user perform phoneme, word, and sentence-level pronunciations within the brain network without the output of movement behavior and sound information, and record the neural activities corresponding to the generation of imagined speech, and use signal processing and classification methods to decode the speech activity function in the cerebral cortex.

[0003] In the decoding based on electroencephalogram signal data, the commonly used extraction method is CSP (Common Spatial Pattern), an algorithm for extracting spatial filtering features under two-class tasks, which can extract the spatial distribution components of each class from multi-channel brain-computer interface data, and methods such as support vector machines, bag of features, and Riemannian manifolds for feature extraction. However, it also has limitations. For example, CSP is widely used in binary classification problems but rarely in multi-class problems. The commonly used methods for electroencephalogram data decoding are convolutional neural networks (CNNs), hybrid deep neural networks, deep convolutional neural networks (Deep ConvNet), shallow convolutional neural networks (Shallow ConvNet), and siamese neural networks (SNN), but their decoding performance does not reach a highly reliable performance ability.

[0004] To address this problem, the present invention proposes a width learning system with Hjorth parameters (hjorth parameters) to decode electroencephalogram data with high accuracy and fast decoding rate. The structure of the width learning system is more concise and tends to be flattened. In addition, there are fewer parameters to be learned during the training process, and the optimal fitting effect of the original data in the feature space can be achieved more quickly, effectively solving the problems of low training speed and excessive consumption of memory resources during the deep learning training process. Summary of the Invention

[0005] To overcome the deficiencies of the prior art, the present invention combines the characteristics of Hjorth parameters and the time-frequency information of electroencephalogram (EEG) data to establish a width learning system model, which optimally fits the imagined speech features in the feature space, effectively improves the accuracy of imagined speech classification, and realizes the decoding process of imagined speech. This model features a reasonable structure design, few input parameters, fast model training, and strong applicability.

[0006] To achieve the above object, the solution adopted by the present invention is as follows:

[0007] An EEG data decoding method based on a width learning system model of Hjorth parameters, comprising the following steps:

[0008] Step 1: Obtain an EEG data set and preprocess the EEG data set using a linear impulse response band-pass filter;

[0009] Perform band-pass filtering on the original EEG data set using a linear impulse response band-pass filter to obtain a filtered EEG data set; segment the filtered EEG data set and divide it into an EEG data training set and an EEG data test set;

[0010] Step 2: Calculate the Hjorth parameters of the preprocessed EEG data set and construct a feature matrix of the EEG data set;

[0011] The Hjorth parameters of the EEG data set include an activity feature f1, a mobility feature f2, and a complexity feature f3; determine the Hjorth parameters based on the EEG data training set and form a feature matrix F as shown in the following formula:

[0012]

[0013] In the formula: F represents the feature matrix of Hjorth parameters; f 11 ···f 61 respectively represent the activity features of the 1st... 6th channels; f 12 ···f 62 respectively represent the mobility features of the 1st... 6th channels; f 13 ···f 63 respectively represent the complexity features of the 1st... 6th channels;

[0014] Step 3: Establish a width learning system model based on the feature matrix of Hjorth parameters;

[0015] Step 31: Obtain the feature matrix of Hjorth parameters determined in Step 2, use the features of the input data mapping as feature nodes, and for n feature mapping nodes, each feature mapping generates k nodes. The feature mapping of the feature nodes is expressed as:

[0016] Zi = Φ(FW ei + β ei ), i = 1, ..., n;

[0017] Where: Z i represents the feature mapping of the i-th feature node, i = 1, …, n; W ei and β ei respectively represent the first and second weights generated by the i-th feature node; Φ represents the feature mapping transfer function of the feature node; n represents the number of feature mapping nodes;

[0018] The n feature mappings of the original EEG data obtain the feature mapping vectors Z n of a set of feature nodes, as follows:

[0019] Z n = [Z1, ..., Z n ;

[0020] Where: Z n represents the feature mapping vector of the feature node;

[0021] Step 32: Obtain the feature mapping of the enhanced nodes, and use the weights generated by the feature nodes to obtain the feature mapping H j of the j-th group of enhanced nodes through further feature mapping, as follows:

[0022] H j = ξ(Z n W hj ), j = 1, ..., m;

[0023] Where: H j represents the feature mapping of the j-th enhanced node, j = 1, …, m; W hj represents the weight generated by the j-th feature node through feature mapping, j = 1, …, m; ξ represents the feature mapping transfer function of the enhanced node; m represents the number of enhanced nodes;

[0024] Map the features of the first m groups of enhanced nodes to obtain the feature mapping vector H m of the enhanced nodes, as follows:

[0025] H m = [H1, ..., H m ;

[0026] Where: H m represents the feature mapping vector of the enhanced node;

[0027] Step 33: Combine the feature mapping vector Z n of the feature nodes with the feature mapping vector H mSpliced together as the input layer of the width learning system model, as shown below;

[0028] A=[Z n |H m ];

[0029] Where: A represents the input layer of the width learning system model;

[0030] Step 34: Obtain the weight matrix of the width learning system model on the EEG data decoding task, through the model's output vector Y and the input layer [Z n |H m ] between the weight model matrix The output expression is constructed as follows:

[0031]

[0032] Where: Y represents the output vector of the width learning system model; The weight model matrix representing the width learning system;

[0033] Weight model matrix The calculation results are as follows:

[0034]

[0035] Where: [Z n |H m ] + Denotes the input layer [Z n |H m ] is the pseudo-inverse; λ represents the ridge parameter of the weight model matrix; I represents the identity matrix of the weight model matrix; A T represents the transpose of the input layer of the width learning system model;

[0036] Step 4: Decode the EEG data based on the width learning system model;

[0037] Extract the Hjorth parameter features of the EEG data test set in step 1, and use the width learning system model established in step 3 to complete the decoding of the EEG data.

[0038] Preferably, in step 1, the original EEG data set is subjected to a 2-40 Hz bandpass filter.

[0039] Preferably, the method for obtaining the activity feature f1 in step 2 is as follows:

[0040] f1=Activity(S (t) )=var(S (t) );

[0041] where: f1 = Activity(S (t) ) represents the activity characteristic of the Hjorth parameter; S (t) represents the EEG dataset; var(S (t) ) represents the variance of the EEG dataset; t represents the EEG data acquisition time.

[0042] Preferably, the method for obtaining the mobility feature f2 in step 2 is as follows:

[0043]

[0044] where: f2 = Mobility(S (t) ) represents the mobility characteristic of the Hjorth parameter; dS (t) / dt represents the derivative of the EEG dataset with respect to time.

[0045] Preferably, the method for obtaining the complexity feature f3 in step 2 is as follows;

[0046]

[0047] where: f3 = Compllexity(S (t) ) represents the complexity characteristic of the Hjorth parameter.

[0048] Preferably, the input layer of the width learning system model in step 33 is specifically:

[0049] The input EEG feature set is mapped according to the set weight and bias values to obtain a set of feature nodes Z n , and when generating enhanced nodes, Z n is also mapped according to the set weight and bias values to obtain enhanced nodes H m , and the feature nodes Z n and the enhanced nodes H m are concatenated together as the input layer [Z n |H m of the width learning system model.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] (1) The present invention combines the characteristics of Hjorth parameters and the time-frequency information of EEG data, and establishes a width learning system model to achieve imagined speech decoding. This model has the characteristics of simple structure, few input parameters, fast model training, and high classification accuracy for EEG data.

[0052] (2) The present invention realizes the rapid fitting of imagined speech features, effectively improves the accuracy of imagined speech classification, and enhances the utilization rate of memory resources. The model features strong applicability and good practical application effects. Description of the Drawings

[0053] Figure 1 It is a control block diagram of the imagined speech decoding method of the width learning system based on Hjorth parameters in an embodiment of the present invention;

[0054] Figure 2 It is the overall flowchart of the imagined speech decoding method of the width learning system based on Hjorth parameters in an embodiment of the present invention;

[0055] Figure 3 It is the activity feature distribution diagram after Hjorth parameter extraction in an embodiment of the present invention;

[0056] Figure 4 It is the average classification confusion matrix of six imagined speech words of the optimal subject in an embodiment of the present invention;

[0057] Figure 5 It is a comparison chart of the recognition performance of different methods in decoding electroencephalogram imagined speech in an embodiment of the present invention. Detailed Embodiment

[0058] Hereinafter, the embodiments of the present invention will be described with reference to the drawings.

[0059] In an embodiment of the present invention, a width learning system model is established by combining the characteristics of Hjorth parameters and the time-frequency information of electroencephalogram data, further realizing imagined speech decoding; the method proposed in an embodiment of the present invention realizes the optimal fitting of imagined speech features in the feature space, effectively improves the accuracy of imagined speech classification, and enhances the utilization rate of computer memory resources; the width learning system model established by the present invention features a simple structure, few input parameters, and fast model training, and has good practical application effects. As Figure 1 shown is a control block diagram of the imagined speech decoding method of the width learning system based on Hjorth parameters in an embodiment of the present invention;

[0060] An embodiment of the present invention provides an imagined speech decoding method based on a width learning system of Hjorth parameters. As Figure 2 shown is the overall flowchart of the imagined speech decoding method of the width learning system based on Hjorth parameters in an embodiment of the present invention; to prove the applicability of the present invention, it is applied to an example, which specifically includes the following steps:

[0061] S1: Obtain an electroencephalogram data set, and preprocess the electroencephalogram data set by using linear impulse response band-pass filtering;

[0062] The original EEG dataset is band-pass filtered using a line impulse response band-pass filter to obtain the filtered EEG dataset;

[0063] The EEG data acquisition electrodes are selected at the positions of F3, F4, C3, C4, P3 and P4, while the reference and ground correspond to the electrodes on the left and right mastoids respectively.

[0064] The original EEG dataset is band-pass filtered at 2 - 40 Hz using a line impulse response band-pass filter to obtain the filtered EEG dataset; band-pass filtering the original EEG dataset at 2 - 40 Hz can eliminate the interference of external noise on the original EEG data.

[0065] The preprocessed EEG dataset is divided into an EEG data training set and an EEG data test set according to the ratio of 80% and 20%.

[0066] S2: Calculate the Hjorth parameters of the preprocessed EEG dataset and construct the EEG dataset feature matrix;

[0067] The Hjorth parameters of the EEG dataset include the activity feature f1, the mobility feature f2 and the complexity feature f3;

[0068] The method for obtaining the activity feature f1 is as follows:

[0069] f1 = Activity(S (t) ) = var(S (t) );

[0070] In the formula: f1 = Activity(S (t) ) represents the activity feature of the Hjorth parameter; S (t) represents the EEG dataset; var(S (t) ) represents the variance of the EEG dataset; t represents the EEG data acquisition time.

[0071] The method for obtaining the mobility feature f2 is as follows:

[0072]

[0073] In the formula: f2 = Mobility(S (t) ) represents the mobility feature of the Hjorth parameter; dS (t) / dt represents the derivative of the EEG dataset with respect to time.

[0074] The method for obtaining the complexity feature f3 is as follows;

[0075]

[0076] In the formula: f3 = Compllexity(S(t) ) represents the complexity characteristics of Hjorth parameters.

[0077] As Figure 3 shown, the activity feature in the Hjorth parameters obtained by processing the EEG (signal waveform diagram of event response) signals of the imagined speech of different words by the subjects is presented. The horizontal axis represents different imagined speech words, and the vertical axis indicates the number of experiments for each imagined speech word by the subjects. The distribution of the Hjorth parameters of the EEG signals reflecting different imagined speech words in the feature space is different through the figure.

[0078] Determine the Hjorth parameters according to the EEG data training set, and construct the feature matrix F as shown in the following formula:

[0079]

[0080] In the formula: F represents the feature matrix of Hjorth parameters; f 11 ···f 61 respectively represent the activity characteristics of the 1st... 6th channels; f 12 ···f 62 respectively represent the mobility characteristics of the 1st... 6th channels; f 13 ···f 63 respectively represent the complexity characteristics of the 1st... 6th channels;

[0081] S3: Establish a width learning system model according to the feature matrix of Hjorth parameters;

[0082] S31: Obtain the feature matrix of Hjorth parameters determined in S2, and use the features mapped by the input data as feature nodes. For n feature mapping nodes, each feature mapping generates k nodes, and the feature mapping of the feature nodes is expressed as:

[0083] Z i = Φ(FW ei +β ei ), i = 1,..., n;

[0084] In the formula: Z i represents the feature mapping of the i-th feature node, i = 1,..., n; W ei and β ei respectively represent the first and second weights generated by the i-th feature node; Φ represents the feature mapping transfer function of the feature node; n represents the number of feature mapping nodes;

[0085] The feature mapping of n feature mappings of the original EEG data obtains a set of feature mapping vectors Z n of feature nodes, as shown below:

[0086] Zn = [Z1,..., Z n ;

[0087] In the formula: Z n represents the feature mapping vector of the feature node;

[0088] S32: Obtain the feature mapping of the enhanced node, and use the weights generated by the feature nodes to obtain the feature mapping H j of the j-th group of enhanced nodes through further feature mapping, as follows:

[0089] H j = ξ(Z n W hj ), j = 1,..., m;

[0090] In the formula: H j represents the feature mapping of the j-th enhanced node, j = 1,..., m; W hj represents the weight generated by the j-th feature node through feature mapping, j = 1,..., m; ξ represents the feature mapping transfer function of the enhanced node; m represents the number of enhanced nodes;

[0091] Map the features of the first m groups of enhanced nodes to obtain the feature mapping vector H m of the enhanced nodes, as follows:

[0092] H m = [H1,..., H m ;

[0093] In the formula: H m represents the feature mapping vector of the enhanced node;

[0094] S33: Map the input EEG feature set according to the set weights and bias values to obtain a set of feature nodes Z n , and when generating enhanced nodes, also map Z n according to the set weights and bias values to obtain the enhanced node H m , and splice the feature node Z n with the enhanced node H m together as the input layer [Z n |H m of the width learning system model, as follows;

[0095] A = [Z n |H m ;

[0096] In the formula: A represents the input layer of the width learning system model;

[0097] S34: Obtain the weight matrix of the width learning system model in the EEG data decoding task, through the output vector Y of the model and the input layer [Z n |H m and the weight model matrix between them Construct the output expression as follows:

[0098]

[0099] In the formula: Y represents the output vector of the width learning system model; represents the weight model matrix of the width learning system;

[0100] Weight model matrix The calculation result of is as follows:

[0101]

[0102] In the formula: [Z n |H m + represents the pseudo-inverse of the input layer [Z n |H m ; λ represents the ridge parameter of the weight model matrix; I represents the identity matrix of the weight model matrix; A T represents the transpose of the input layer of the width learning system model;

[0103] S4: Decode the EEG data according to the width learning system model;

[0104] Extract the Hjorth parameter features of the EEG data test set in S1, and use the width learning system model established in S3 to complete the decoding of the EEG data.

[0105] Such as Figure 4 is the average classification confusion matrix of six imagined speech words of the optimal subject in the embodiment of the present invention, which is used to judge the quality of the classifier. The calculation method of calculating the classifier accuracy is as follows:

[0106]

[0107] In the embodiment, it can be seen from the confusion matrix that the calculation result of the classifier accuracy is as follows. The accuracy rate reaches more than 96%, proving that the calculation result of this method is good.

[0108]

[0109] Such as Figure 5 ​The figure shows a comparison chart of the recognition performance of different methods in the embodiment of the present invention when decoding electroencephalogram imagined speech; in the figure: (A) represents Hjorth-BLS (this method), (B) represents BLS (width learning system), (C) represents CSP-CNN (common spatial pattern + CNN convolutional neural network), and (D) represents CNN (convolutional neural network). Some existing methods for decoding electroencephalogram imagined speech, such as CNN, CSP-CNN, and BLS, are studied. The comparison between the proposed width learning system classification method based on Hjorth parameters and the existing methods gives the receiver operating characteristic curves and accuracies of different methods. It can be seen from the figure that the recognition performance of the BLS (width learning system) network (Hjorth-BLS) based on Hjorth parameters is better than that of CNN, CSP-CNN, and BLS.

[0110] In summary, the calculation results of the imagined speech decoding method of the width learning system based on Hjorth parameters in this case prove to have good effects.

[0111] (1) In the embodiment of the present invention, by combining the characteristics of Hjorth parameters and the time-frequency information of electroencephalogram data, a width learning system model is established, and further the imagined speech decoding is realized; it is proved by the attached drawings that the width learning system model established by the present invention has the characteristics of simple structure, few input parameters, and fast model training; it can maintain a high accuracy for complex classification processes such as six-class classification.

[0112] (2) The method proposed in the embodiment of the present invention realizes the rapid fitting of imagined speech features, effectively improves the accuracy of imagined speech classification, and improves the utilization rate of computer memory resources. It is proved by the embodiment that this method has the characteristics of strong applicability and good actual application effect.

[0113] The above-described embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solution of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for decoding EEG data based on a Hjorth parameter width learning system model, characterized in that: It includes the following steps: Step 1: Obtain the EEG dataset and preprocess it using line impulse response bandpass filtering; Using a linear impulse response bandpass filter to perform bandpass filtering on the original EEG dataset to obtain a filtered EEG dataset; Segment the filtered EEG data set into an EEG data training set and an EEG data test set; Step 2: Calculate the Hjorth parameter of the preprocessed EEG dataset and construct the EEG dataset feature matrix; The Hjorth parameters of the EEG dataset include activity feature f1, mobility feature f2, and complexity feature f3. The Hjorth parameters are determined based on the EEG data training set, and the feature matrix F is constructed as shown in the following formula: Where: F represents the characteristic matrix of Hjorth parameters; f 11 …f 61 Represent the activity characteristics of the 1st…6th channels respectively; f 12 …f 62 They represent the mobility characteristics of the 1st…6th channels respectively; f 13 …f 63 Represent the complexity characteristics of the 1st…6th channels respectively; Step 3: Establish a width learning system model based on the characteristic matrix of Hjorth parameters; Step 31: Obtain the feature matrix of the Hjorth parameter determined in step 2, use the features of the input data mapping as feature nodes, for n feature mapping nodes, each feature mapping generates k nodes, and the feature map of the feature node is expressed as: Z i =Φ(FW ei +β ei ),i=1,...,n; Where: Z i represents the feature map of the i-th feature node, i = 1, ..., n; W ei and β ei They represent the first and second weights generated by the i-th feature node respectively; Φ represents the feature mapping transfer function of the feature node; n represents the number of feature mapping nodes; The n feature maps of the original EEG data are used to obtain a set of feature node feature mapping vectors Z n , as shown below: WITH n =[Z1,...,Z n ]; Where: Z n The feature map vector representing the feature node; Step 32: Obtain the feature map of the enhanced node, and use the weights generated by the feature node to obtain the feature map H of the jth group of enhanced nodes through further feature mapping j , as shown below: H j =ξ(Z n W hj ),j=1,...,m; Where: H j represents the feature map of the j-th enhancement node, j = 1, ..., m; W hj represents the weight of the jth feature node generated by feature mapping, j = 1, ..., m; ξ represents the feature mapping transfer function of the enhanced node; m represents the number of enhanced nodes; Map the features of the first m groups of enhanced nodes to obtain the feature mapping vector H of the enhanced nodes m , as shown below: H m =[H1,...,H m ]; Where: H m The feature map vector representing the enhanced node; Step 33: Map the feature node's feature vector Z n and the feature map vector H of the enhanced node m Spliced together as the input layer of the width learning system model, as shown below; A=[Z n |H m ]; Where: A represents the input layer of the width learning system model; Step 34: Obtain the weight matrix of the width learning system model on the EEG data decoding task, through the model's output vector Y and the input layer [Z n |H m ] between the weight model matrix The output expression is constructed as follows: Where: Y represents the output vector of the width learning system model; The weight model matrix representing the width learning system; Weight model matrix The calculation results are as follows: Where: [Z n |H m ] + Denotes the input layer [Z n |H m ] is the pseudo-inverse; λ represents the ridge parameter of the weight model matrix; I represents the identity matrix of the weight model matrix; A T represents the transpose of the input layer of the width learning system model; Step 4: Decode the EEG data based on the width learning system model; Extract the Hjorth parameter features of the EEG data test set in step 1, and use the width learning system model established in step 3 to complete the decoding of the EEG data.

2. The EEG data decoding method based on the Hjorth parameter width learning system model according to claim 1, characterized in that: In step 1, the original EEG data set is subjected to a 2-40 Hz band-pass filter.

3. The EEG data decoding method based on the Hjorth parameter width learning system model according to claim 1, characterized in that: The method for obtaining the activity feature f1 in step 2 is as follows: f1=Activity(S (t) )=var(S (t) ); Where: f1=Activity(S (t) ) represents the activity characteristics of the Hjorth parameter; S (t) Represents the EEG dataset; var(S (t) ) represents the variance of the EEG data set; t represents the EEG data acquisition time.

4. The EEG data decoding method based on the Hjorth parameter width learning system model according to claim 1, characterized in that: The method for obtaining the mobility feature f2 in step 2 is as follows: Where: f2=Mobility(S (t) ) represents the mobility characteristic of Hjorth parameter; dS (t) / dt represents the time derivative of the EEG dataset.

5. The EEG data decoding method based on the Hjorth parameter width learning system model according to claim 1, characterized in that: The method for obtaining the complexity feature f3 in step 2 is as follows: Where: f3 = Completion (S (t) ) represents the complexity characteristic of the Hjorth parameter.

6. The EEG data decoding method based on the Hjorth parameter width learning system model according to claim 1, characterized in that: The input layer of the width learning system model in step 33 is specifically: The input EEG feature set is mapped according to the set weights and bias values to obtain a set of feature nodes Z n , when generating the enhanced node, also set Z n Mapping is performed according to the set weight and bias value to obtain the enhanced node H m , the feature node Z n With enhanced node H m Spliced together as the input layer of the width learning system model [Z n |H m ].

Images