A hierarchical classification and error correction mechanism-based EEG signal classification method

Abstract

This invention discloses a hierarchical classification and error correction mechanism for EEG signal classification, comprising the following steps: acquiring multi-channel motor imagery EEG data from subjects and performing preliminary preprocessing; using RBF neural networks and DEMATEL to solve the characteristic representation degree of EEG channels, and determining the weight of each channel based on the characteristic representation degree; calculating the ordering entropy of each channel and using discrete Fréchet curve similarity judgment to pre-classify the EEG signals into two major categories (left and right hands, and fists and feet); using RCSP to extract the spatial features of samples in each major category, and using SVM for fine classification; judging and correcting error-prone samples according to the error correction mechanism; and finally determining the final classification category of each sample. This invention, through the design of a hierarchical classification method and the use of different classification methods, avoids the single model structure when a multi-task classification problem is transformed into multiple binary classification problems, thus optimizing the multi-task classification model structure.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal processing, and more specifically, relates to an EEG signal classification method based on hierarchical classification and error correction mechanisms. Background Technology

[0002] Brain-computer interface (BCI) is a system that allows interaction with computers or other devices directly by collecting signals from the brain, without relying on the peripheral nervous system and muscles. Research on BCI systems can provide new channels for communication for elderly or disabled individuals with normal cognitive function but impaired motor function, thereby improving their autonomy and quality of life. Feature extraction and classification of electroencephalogram (EEG) signals are the main research directions in BCI systems. Commonly used EEG signals include event-related P300 potentials, steady-state visual evoked potentials (SSVEPS), and [α] and [β] rhythms related to motor imagery (MI). The classification of motor imagery EEG signals is currently a research hotspot; however, EEG signals exhibit significant nonlinearity and non-stationarity, meaning their characteristics change over time. Furthermore, the collected EEG signals are often mixed with noise, posing challenges to their analysis. Therefore, more effective measures should be explored to better achieve the analysis and classification of EEG tasks.

[0003] While binary classification of motor imagery in EEG has achieved high accuracy, the limited number of categories restricts its application. Multi-class classification in EEG remains a challenging problem. Current multi-task EEG classification methods typically transform multi-class problems into binary classification problems using "one-to-many" or "one-to-one" approaches, but the overall recognition model is relatively simplistic. Summary of the Invention

[0004] This invention utilizes the results of EEG binary classification and achieves multi-task identification and classification through hierarchical classification combined with error correction mechanisms. It can effectively optimize the structure of multi-task EEG identification models and has very important practical significance.

[0005] To address at least one of the aforementioned technical problems, according to one aspect of the present invention, a method for classifying electroencephalogram (EEG) signals based on hierarchical classification and error correction mechanisms is provided, comprising the following steps:

[0006] S1. Acquire and preprocess multi-class, multi-channel motor imagery EEG signals from the subjects;

[0007] S2. For the preprocessed data obtained in S1, calculate the relevant feature matrix and input it into the RBF neural network. Combine it with DEMATEL to calculate the feature representation degree of each channel, and determine the feature weight of each channel based on the obtained feature representation degree.

[0008] S3. Calculate the relevant entropy value of each channel and weight the entropy value according to the feature weight of each channel obtained in S2. Use the discrete Fréchet distance curve similarity judgment to pre-classify the EEG signal into two categories (left and right hands and fists and feet).

[0009] S4. Based on the two major categories obtained in S3, feature extraction is performed using RCSP, and SVM is used to further refine the pre-classification results.

[0010] S5. Based on the pre-classification and fine-classification results, and combined with the error correction mechanism, conduct a comprehensive analysis and error correction on the two classification results.

[0011] Furthermore, in S3, the entropy values ​​of the EEG signals are weighted using channel feature weights, and pre-classification of the EEG signals is performed based on the similarity of discrete Fréchet distance curves. The 64-dimensional EEG signal entropy values ​​can form a one-dimensional entropy vector of size 1*64. The weighting rules for the entropy vector based on the feature correlation ranking of each channel of the EEG signal obtained by RBF and DEMATEL are as follows:

[0012]

[0013] Among them, C L For the Lth channel, x i Let η be the time series corresponding to the i-th channel. i The weights are the weights of the i-th time series; the entropy vector is compared with the standard vectors of each type using the discrete Fréchet rule to achieve sample pre-classification.

[0014] The specific method for the error correction mechanism of S5 for the two classifications (pre-classification and fine classification) is as follows:

[0015] S51. Regarding the determination of error-prone samples in the two classifications: For pre-classification, after statistically analyzing the similarity values ​​between all test samples and the standard sample, a judgment threshold is set. If the similarity value between the two types is greater than the threshold, it indicates that the accuracy of the pre-classification of the sample is not high, and the sample is classified as an error-prone sample. For fine-grained classification, the distance of the test sample from the hyperplane in the fine-grained SVM classifier is calculated. The closer the sample is to the hyperplane, the lower the accuracy and reliability of the classification, and the sample is classified as an error-prone sample. The distance of the test sample from the hyperplane is expressed as follows:

[0016]

[0017] Where: (x1, y2) are sample points, x1 is the first type of sample, y2 is the second type of sample, w1 and w2 are the parameters of the hyperplane boundary, and b is the displacement term;

[0018] S52. For the error-prone samples identified in S51, train four SVM classifiers: left fist and two feet, left hand and two fists, right hand and two feet, and right hand and two fists. Combine these with the two SVM classifiers already trained in the fine classification to form a classifier group. Design an error correction mechanism based on the classification performance of the error-prone samples in these six classifiers. Different classification performances correspond to different error correction methods. The main error correction methods are as follows:

[0019] If three classifiers in a classifier group produce the same result, then the final category of the error-prone sample is determined directly based on that result.

[0020] If two classifiers in the classifier group produce the same result, the final category of the error-prone sample is determined by comparing the similarity between samples in the pre-classification process with the classification results of the classifier group.

[0021] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the electroencephalogram (EEG) signal classification method based on hierarchical classification and error correction mechanism of the present invention.

[0022] According to another aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the electroencephalogram signal classification method based on hierarchical classification and error correction mechanism of the present invention.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects:

[0024] This invention, through the design of a hierarchical classification method and the use of different classification methods, avoids the problem of a single model structure when a multi-task classification problem is transformed into multiple binary classification problems, thus optimizing the multi-task classification model structure. Furthermore, through the design of an error correction mechanism, it can fully utilize the characteristics of different models to correct errors for error-prone samples, thereby improving the accuracy of multi-task EEG classification. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention and are not intended to limit the present invention.

[0026] Figure 1 This is a flowchart of the method of the present invention;

[0027] Figure 2 This is a temporal experimental diagram of a single motion imagination task database used in this invention under one embodiment;

[0028] Figure 3This is a schematic diagram of discrete Fréchet after trajectory discretization in one embodiment. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0030] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0031] like Figure 1-3 As shown,

[0032] Example 1:

[0033] This invention discloses a method for classifying electroencephalogram (EEG) signals based on hierarchical classification and error correction mechanisms. This method can simplify the structure of a four-class EEG model to a certain extent and improve the classification accuracy of motor imagery tasks. Figure 1 The proposed solution flowchart includes the following steps:

[0034] Step 1: Acquire and preprocess the multi-class, multi-channel motor imagery EEG signals of the subjects.

[0035] Step 2: For the preprocessed data obtained in Step 1, calculate the relevant feature matrix and input it into the RBF neural network in conjunction with DEMATEL to calculate the feature representation degree of each channel. Determine the feature weight of each channel based on the obtained feature representation degree.

[0036] Step 3: Calculate the relevant entropy value of each channel and weight the entropy value according to the feature weight of each channel obtained in Step 2. Use the discrete Fréchet distance curve similarity judgment to pre-classify the EEG signal into two major categories (left and right hands and fists and feet).

[0037] Step 4: Based on the two major categories obtained in Step 3, feature extraction is performed using RCSP, and SVM is used to further refine the pre-classification results.

[0038] Step 5: Based on the pre-classification and fine-classification results, and combined with the error correction mechanism, perform a comprehensive analysis and error correction on the two classification results.

[0039] In this embodiment: the international public dataset used is based on the BCI2000 system, with EEG signals collected at a sampling frequency of 160Hz, recording EEG signals from 64 channels of motor imagery. Subjects performed (or imagined) four different motor tasks, including opening and closing the left fist, the right fist, both fists, and both feet. During the classification process, the training set data and the test set data maintained a 7:3 ratio.

[0040] In this embodiment, the specific process for calculating the feature weights of each channel in step 2 is as follows:

[0041] This invention selects the channels that best characterize the aforementioned features through comprehensive analysis of EEG data in both the time and time-frequency domains. The time-domain features include sample entropy (SampEn), Shannon entropy (ShEn), and approximate entropy (ApEn). The time-frequency features include wavelet entropy (WE), power spectral density (PSD), and root mean square of sub-band energy, calculated based on wavelet transform decomposition into various frequency bands. These six features together constitute a feature matrix, which is then input into an RBF neural network to obtain the direct correlation matrix between the various feature factors.

[0042]

[0043] Where b 11 =0, This represents the influence of the i-th indicator relative to the j-th indicator. ω represents the weights of the RBF neural network.

[0044] Based on the obtained direct correlation matrix, the full correlation matrix among the influencing factors of each feature is obtained using the Decision Laboratory Analysis (DEMATEL) method. Then, the centrality of each channel with respect to the given feature is calculated based on the full correlation matrix, and the centrality is used to characterize the feature representation degree of the channel.

[0045] The full correlation matrix T is calculated as follows:

[0046]

[0047]

[0048]

[0049] Matrix B is the direct correlation matrix, representing the degree of mutual influence among system factors; matrix G is the normalized direct correlation matrix, with each coefficient ranging from 0 to 1; and Z is the identity matrix.

[0050] Define matrix T, and let the sum of its rows and the sum of its columns be D and M, respectively:

[0051] T=(t ij ) n×n (7)

[0052] D=(t i· ) n×n (8)

[0053] M=(t ·j ) n×n (9)

[0054] D represents the combined influence of the variable on other variables, and M represents the combined influence of the variable on other variables. D+M represents the centrality of the channel, indicating the role of the channel in the evaluation system and reflecting its importance. The centralities are then sorted in descending order to obtain the characteristic representation degree of the EEG channel.

[0055] In this embodiment, sorting entropy is used as the discrimination criterion for pre-classification. The sorting entropy values ​​of the 64-bit EEG signals are weighted according to the obtained feature representation degree of each channel. Specifically, this includes:

[0056] We use electroencephalogram (EEG) signal ordering entropy (PE) to study the local structure of dynamic time series. A given time series is transformed into a series of ordered patterns, each describing the sequential relationship between the current and equidistant past values ​​within a given time period.

[0057] Based on the obtained sorting entropy and channel feature representation degree, the sorting entropy of the 64-channel EEG signal is segmented and weighted. After calculating the sorting entropy of each channel for each EEG sample, the 64-channel sorting entropy can form a one-dimensional vector of size 1*64. The test sample and the comparison standard sample are segmented separately using four channels as the segmentation unit and a step size of 1, resulting in 61 sub-vectors of size 1*4. Then, the segmented vectors are weighted according to the correlation ranking of the time-frequency features of the EEG signal obtained by DEMATEL, as follows:

[0058]

[0059] Among them, C L For the Lth channel, x i Let η be the time series corresponding to the i-th channel. i The weights are the weights of the i-th time series.

[0060] In this embodiment: Based on the weighted sorting entropy obtained in step 3, the present invention utilizes channel-weighted discrete Fréchet distance curve similarity to perform pre-classification of EEG signals. Specific steps include:

[0061] To determine the standard sample for similarity judgment using the Fréchet distance curve, this invention employs a training sample inner loop to identify the standard sample that best represents the current type. Then, the samples are sequentially compared with the standard sample for similarity judgment. The definition of Fréchet distance is as follows: Figure 3 As given in, such as Figure 3 The diagram shows two lines, R and G, each with a moving point α and β. Assume the length of R is l and the length of G is s. The positions of the two moving points can be described by continuously increasing functions α(t) and β(t). For ease of discussion, this invention constrains the variable t to [0,1], α(0) = 0, α(1) = l, β(0) = 0, β(1) = s. The positions of the two moving points on their respective curves at the same time can be described as R(α(t)) and G(β(t)). The distance between the two moving points varies depending on the two continuously increasing functions α(t) and β(t) and the change in variable t. The Fréchet distance is used to find a pair of functions to minimize the maximum distance between the two moving points. The continuous Fréchet distance can be other forms of distance; this invention uses the Eulerian distance. Its representation is as follows:

[0062]

[0063] Based on the defined continuous Fréchet similarity judgment, the discretization of continuous Fréchet similarity judgment can be achieved. The discrete Fréchet definition is based on the continuous Fréchet, discretizing the trajectory. Let curve R consist of r trajectory points and curve G consist of g trajectory points, quantizing the two curves into two ordered sets:

[0064] σ(R)=(u1,...,u r ),σ(G)=(v1,...,v g (11)

[0065] By mapping the elements in the two sets above to the sequence of trajectory points, we can obtain:

[0066] (u1,v1),(u2,v2),...,(u r ,v g (12)

[0067] The length ||L|| between pairs of R and G trajectory sequences is defined as the maximum Euclidean distance between each pair, and the expression is as follows:

[0068] ||L||=max d(u i ,v j (13)

[0069] Where i∈(1,r), j∈(1,g). In this invention, R=G, so i=j.

[0070] Therefore, the distance expression for discrete Fréchet is as follows:

[0071] δ dF (R,G)=min||L|| (14)

[0072] Finally, the discrete Fréchet rule is used to compare the curve similarity of all sample subvectors with the standard samples of each type. Then, the weighted subvectors are added together to obtain the similarity value. After comparison, the sample pre-classification is achieved.

[0073]

[0074] Here, ζ represents the category of EEG, and F represents the sum of all weighted similarities. The sample is compared with all category standards; the lower the F value, the greater the likelihood that the sample belongs to that category.

[0075] In this embodiment: RCSP is used to extract regularized spatial features of EEG signals and SVM is used to further classify the two major categories that have been pre-classified, ultimately achieving four-class classification of motor imagery EEG signals.

[0076] In this embodiment: the two classification processes cannot guarantee that all samples are classified correctly. To further improve the classification accuracy of EEG signals, this invention designs an error correction mechanism based on the analysis of the sample classification results from the two classifications, as follows:

[0077] For the selection of error correction targets, pre-classification is based on the similarity value of curve similarity, so the similarity value is the key to the selection of error correction targets. In this experiment, four similarity values ​​were obtained for each test sample and the other four standard samples based on the discrete Fréchet criterion. The lower the value obtained by the Fréchet criterion, the more similar the two curves are. After counting all similarity values, a fixed threshold (0.2) was set. If the similarity value between two types is greater than 0.2, it indicates that the accuracy of the pre-classification is not high.

[0078] In fine-grained classification, this invention selects samples based on the distance between the two classes of samples in SVM and the hyperplane. The closer a sample is to the hyperplane, the lower its classification accuracy and reliability. The specific judgment method is as follows:

[0079] Distance of sample points from the hyperplane:

[0080]

[0081] Where (x1, y2) are sample points, x1 is a first-class sample and y2 is a second-class sample. w1 and w2 are parameters of the hyperplane boundary, and b is the displacement term.

[0082] For the selected error-prone samples, this invention trains four SVM classifiers: left fist and two feet, left hand and two fists, right hand and two feet, and right hand and two fists. These are combined with two SVM classifiers already trained in the fine-grained classification to form a classifier group. Based on the classification performance of the error-prone samples across these six classifiers, and combined with an error correction mechanism, the final category of the error-prone sample is determined. The error correction method is as follows:

[0083] If three classifiers in a classifier group produce the same classification result, then that result is used to directly determine the final category of the error-prone sample.

[0084] If two classifiers in the classifier group have the same classification result, the final category of the error-prone sample is determined by comparing the similarity between samples in the pre-classification with the classification result of the classifier group.

[0085] The error correction mechanism proposed in this invention effectively improves the accuracy of EEG signal classification by correcting error-prone samples. Ultimately, the model achieved a 93.46% accuracy rate in the four-class classification of motor imagery EEG. This represents an improvement compared to similar methods and databases.

[0086] Example 2:

[0087] The computer-readable storage medium of this embodiment stores a computer program that, when executed by a processor, implements the steps in the EEG signal classification method based on hierarchical classification and error correction mechanism of Embodiment 1.

[0088] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0089] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0090] Example 3:

[0091] The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the EEG signal classification method based on hierarchical classification and error correction mechanism of Embodiment 1.

[0092] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0093] Those skilled in the art will understand that the content disclosed in the embodiments can be provided as a method, system, or computer program product. Therefore, this solution can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this solution can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage) containing computer-usable program code.

[0094] This solution is described with reference to flowchart illustrations and / or block diagrams of methods and computer program products according to embodiments of this solution. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0098] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

Claims

1. A method for classifying electroencephalogram (EEG) signals based on hierarchical classification and error correction mechanisms, characterized in that, Includes the following steps: S1. Acquire and preprocess multi-class, multi-channel motor imagery EEG signals from the subjects; S2. For the preprocessed data obtained in S1, calculate the relevant feature matrix and input it into the RBF neural network. Combine it with DEMATEL to calculate the feature representation degree of each channel, and determine the feature weight of each channel based on the obtained feature representation degree. S3. Calculate the relevant entropy value of each channel and weight the entropy value according to the feature weight of each channel obtained in S2. Use the similarity judgment of discrete Fréchet distance curve to pre-classify the EEG signal into two major categories. S4. Based on the two major categories obtained in S3, feature extraction is performed using RCSP, and SVM is used to further refine the pre-classification results. S5. Based on the pre-classification and fine-classification results, and combined with the error correction mechanism, conduct a comprehensive analysis and error correction of the two classification results; S5 is detailed below: S51. Regarding the determination of error-prone samples in the two classifications: For pre-classification, after calculating the similarity values ​​of all test samples and the standard sample, a judgment threshold is set. If the similarity value between the two types is greater than the threshold, it indicates that the accuracy of the pre-classification of the sample is not high, and the sample is classified as an error-prone sample. For fine-grained classification, the distance of the test sample from the hyperplane in the fine-grained SVM classifier is calculated. The closer the sample is to the hyperplane, the lower the accuracy and reliability of the classification of the sample, and the sample is classified as an error-prone sample. The distance of the test sample from the hyperplane is expressed as follows: Where: (x1, y2) are sample points, x1 is a sample of the first class, y2 is a sample of the second class, and w1 and w2 are parameters of the hyperplane boundary; S52. For the error-prone samples identified in S51, train four SVM classifiers: left fist and two feet, left hand and two fists, right hand and two feet, and right hand and two fists. Combine the results of the two SVM classifiers in the fine classification to form a classifier group. Design an error correction mechanism based on the classification results of the error-prone samples in these six classifiers. Different classification results correspond to different error correction methods. The main error correction methods are as follows: If three classifiers in a classifier group produce the same result, then the final category of the error-prone sample is determined directly based on that result. If two classifiers in the classifier group produce the same result, the final category of the error-prone sample is determined by comparing the similarity between samples in the pre-classification with the classification result of the classifier group.

2. The EEG signal classification method based on hierarchical classification and error correction mechanism according to claim 1, characterized in that, In S3, EEG signals are pre-divided into two main categories: left and right hands, and fists and feet.

3. The EEG signal classification method based on hierarchical classification and error correction mechanism according to claim 2, characterized in that, In S3, the entropy values ​​of EEG signals are weighted using channel feature weights, and pre-classification of EEG signals is performed based on the similarity of discrete Fréchet distance curves. The 64-dimensional EEG signal entropy values ​​constitute a one-dimensional entropy vector of size 1*64. The weighting rules for the entropy vector are as follows, based on the feature correlation ranking of each channel of the EEG signal obtained by RBF and DEMATEL: Among them, C L For the Lth channel, x i Let η be the time series corresponding to the i-th channel. i The weights are the weights of the i-th time series; the entropy vector is compared with the standard vectors of each type using the discrete Fréchet rule to achieve sample pre-classification.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps in the EEG signal classification method based on hierarchical classification and error correction mechanism as described in any one of claims 1 to 3.

5. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the EEG signal classification method based on hierarchical classification and error correction mechanism as described in any one of claims 1 to 3.

Images