CN116738316A - Electroencephalogram classification method for unilateral limb motor imagery tasks - Google Patents

Electroencephalogram classification method for unilateral limb motor imagery tasks Download PDF

Info

Publication number
CN116738316A
CN116738316A CN202310719853.XA CN202310719853A CN116738316A CN 116738316 A CN116738316 A CN 116738316A CN 202310719853 A CN202310719853 A CN 202310719853A CN 116738316 A CN116738316 A CN 116738316A
Authority
CN
China
Prior art keywords
motor imagery
probability distribution
electroencephalogram
matrix
eeg
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310719853.XA
Other languages
Chinese (zh)
Inventor
张锐
陈亚迪
张利朋
胡玉霞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhengzhou University
Original Assignee
Zhengzhou University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhengzhou University filed Critical Zhengzhou University
Publication of CN116738316A publication Critical patent/CN116738316A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2131—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on a transform domain processing, e.g. wavelet transform
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2123/00—Data types
    • G06F2123/02—Data types in the time domain, e.g. time-series data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

The invention discloses a unilateral limb motor imagery task electroencephalogram classification method, which sequentially comprises the following steps: a: constructing a single-side limb motor imagery data set and preprocessing to obtain motor imagery electroencephalogram signals; b: performing feature extraction and classification on the motor imagery electroencephalogram signals by using a deep neural network model to obtain a probability distribution matrix of the electroencephalogram signals; c: processing the motor imagery electroencephalogram signals by using a CSP method to obtain a time-frequency chart data set, and then carrying out feature extraction and classification on the time-frequency chart data set by using a deep neural network model to obtain a probability distribution matrix of the time-frequency chart signals; d: calculating an average cluster coefficient pair and optimizing two probability distribution matrixes; and then carrying out fusion decision by utilizing the two optimized probability distribution matrixes based on the D-S evidence theory to obtain a final classification result. The invention can effectively improve the classification accuracy of unilateral limb movement information tasks and provides a data basis for the development of brain-computer interface technology.

Description

Electroencephalogram classification method for unilateral limb motor imagery tasks
Technical Field
The invention relates to the technical field of biological signal recognition, in particular to a single-side limb motor imagery task electroencephalogram classification method based on multi-mode information fusion.
Background
The brain-computer interface (BCI) constructs an information interaction channel between the human brain and the external environment, and can directly read and analyze the electrical signals generated by the brain to identify the intention of a user, and the brain electrical signals are used for realizing the interaction between the human and the external equipment, and the core technology is the identification of the brain electrical signals. Motor imagery tasks (MI) are psychological processes that simulate movement without actual movement, i.e. the brain imagines the whole movement without actually contracting the muscles. In the neurophysiologic field, there are many similarities between the true actions of peripheral autonomic nerves and cortical potentials and moving images. MI, an important model of spontaneous BCI, has been widely used in the field of nerve rehabilitation to provide rehabilitation and motor assistance for disabled patients suffering from dyskinesia, such as patients suffering from brain stem injury, stroke, muscle atrophy, who possess a complete brain but cannot communicate with the outside because of peripheral nerve injury. Although the MI-BCI system cannot repair the physiological nerve transmission channel of the system, the brain consciousness can be transmitted to a controlled device for control. In the research of the current motor imagery tasks, the movements of different parts of the body are often involved, and the brain electrical signals generated by motor imagery of different body parts are identified, however, compared with the motor imagery tasks among different body parts, the combination of the motor imagery tasks of limbs at the same part and the movements of the corresponding body parts can be more natural and visual, for example, in the arm rehabilitation training combined with a rehabilitation robot, the robot executes the same movements to perform the rehabilitation training after imagining the different movements of the arm, thereby being more beneficial to training and rehabilitation and being more suitable for clinical application scenes. However, classification and identification of individual limbs is more difficult and complex than identification between different parts of the body, because the brain motor cortex areas activated when the same body part performs a motor task are very similar, which brings greater difficulty to research of decoding and identification, and therefore an effective feature extraction method is very important in the classification process.
The process of classifying and identifying the electroencephalogram signals comprises preprocessing of the electroencephalogram signals, feature extraction and feature classification, and in the process, effective feature extraction and a proper feature classifier are key for determining the identity recognition performance. In the process of extracting features of an electroencephalogram signal, representative feature extraction methods can be divided into time domain analysis (including amplitude, mean value and the like), frequency domain analysis (including power spectrum analysis, coherent analysis and the like), time-frequency domain analysis (including wavelet transformation, empirical mode decomposition and the like) and airspace analysis (including co-airspace mode method, independent component analysis and the like), and in order to improve accuracy of biological feature recognition, some researches combine different feature extraction methods to extract multidimensional features of an EEG signal so as to characterize the EEG from multiple domains. In the process of classifying the characteristics of the electroencephalogram signals, the deep learning method directly takes the original signals as the input of the model to carry out end-to-end training, and the characteristics of the signals do not need to be extracted any more, so that the current application is very wide. Convolutional neural networks and recurrent neural networks are commonly used as methods for biometric identification due to the temporal, frequency and spatial properties of EEG signals.
At present, a deep learning algorithm is increasingly applied to a classification task of an electroencephalogram signal, in the existing motor imagery electroencephalogram signal identification, an original electroencephalogram signal or a transformed electroencephalogram frequency domain feature and the like are generally used for classification identification, and a simple deep learning model can be utilized to obtain better performance, so that the deep learning algorithm is paid more attention to by more researchers. However, the existing motor imagery electroencephalogram signal identification often ignores the inherent relation existing between signals in different domains, does not pay attention to feature diversity and feature correlation between signals in different domains and different modes, and lacks more comprehensive feature information. Therefore, the method has important significance for improving the classification accuracy of the electroencephalogram signals by fusion application of the information of different characteristics of the electroencephalogram signals.
Disclosure of Invention
The invention aims to provide a unilateral limb movement imagination task electroencephalogram classification method, which utilizes a convolutional neural network to extract and classify characteristics of preprocessed original electroencephalogram signals and a two-dimensional time-frequency diagram obtained by Continuous Wavelet Transform (CWT), and performs fusion decision on results obtained by classification of two modal signals, so that the classification accuracy of unilateral limb movement information tasks can be effectively improved, and a data base is provided for development of brain-computer interface technology.
The invention adopts the following technical scheme:
the electroencephalogram classification method for the unilateral limb motor imagery task sequentially comprises the following steps:
a: constructing a single-side limb motor imagery data set, and performing data preprocessing on motor imagery electroencephalogram signals in the single-side limb motor imagery data set to obtain preprocessed motor imagery electroencephalogram signals;
b: and C, performing feature extraction and classification on the preprocessed motor imagery electroencephalogram signals obtained in the step A by using a deep neural network model to obtain a probability distribution matrix P of the electroencephalogram signals EEG The method comprises the steps of carrying out a first treatment on the surface of the Probability distribution matrix P EEG The probability that the preprocessed motor imagery electroencephalogram signal belongs to the corresponding action category is included;
c: c, performing spatial filtering on the preprocessed motor imagery electroencephalogram signal obtained in the step A by using a CSP method to obtain a feature matrix of the electroencephalogram signal; then, a time-frequency diagram data set is obtained through selection of a feature matrix and continuous wavelet transformation, and finally, the time-frequency diagram data set is subjected to feature extraction and classification by using a deep neural network model to obtain a probability distribution matrix P of a time-frequency diagram signal TF The method comprises the steps of carrying out a first treatment on the surface of the Probability distribution matrix P TF The time-frequency diagram signal comprises the probability of belonging to the corresponding action category;
d: using the probability distribution matrix P of the EEG signals obtained in the step B EEG And the probability distribution matrix P of the time-frequency diagram signals obtained in the step C TF Respectively calculating the clustering coefficient of each action category, and according to the calculated average clustering coefficient, performing a probability distribution matrix P EEG and PTF Optimizing to obtain an optimized probability distribution matrix P EEG* and PTF* The method comprises the steps of carrying out a first treatment on the surface of the Based on D-S evidence theory and utilizing probability distribution matrix P EEG* and PTF* And carrying out fusion decision to obtain a final classification result.
In the step A, the data preprocessing comprises band-pass filtering, downsampling, channel selection, independent component analysis and artifact removal and data selection of motor imagery electroencephalogram signals.
In the step A, in a single-side limb motor imagery data set, only motor imagery electroencephalogram signals under three actions of forward stretching arms, left rotating wrists and grabbing cups are selected;
in the step B of the process, wherein ,/> and />The probability that the sample results belong to the corresponding action categories in the electroencephalogram signal three-classification experiment is respectively;
in the step C, the step of, in the step C, wherein ,/> and />Respectively, sample results in the time-frequency diagram three-classification experiment belong to corresponding actionsProbability of category.
The step C comprises the following specific steps:
c1: performing spatial filtering on the preprocessed motor imagery electroencephalogram signals by using a CSP algorithm;
when the spatial filtering is carried out, firstly, a CSP projection matrix W is obtained by calculation, and then a feature matrix Z is obtained by calculation by utilizing the CSP projection matrix W and the preprocessed motor imagery electroencephalogram signals obtained in the step A; the calculation formula is as follows: z is Z M×T ＝W M×M E M×T The method comprises the steps of carrying out a first treatment on the surface of the Wherein W is CSP projection matrix, M is the channel number of the preprocessed brain electrical data, T is the data length, E is brain electrical data matrix with M x T formed by converting preprocessed motor imagery brain electrical signals;
c2: selecting a feature matrix Z' of CSP feature extraction;
and C3: c2, according to the feature matrix Z 'obtained in the step, performing time-frequency feature extraction by using a CWT algorithm, converting the feature matrix Z' into a two-dimensional time-frequency diagram, selecting Morlet wavelets as a basis function to perform wavelet transformation, and finally obtaining a motor imagery time-frequency diagram data set;
the formula of the wavelet transform is as follows:
the Morlet is a basis function commonly used in wavelet transformation and is used for carrying out decomposition operation on signals to be processed in the wavelet transformation process; phi (t) represents a basis function, and omega represents the wavelet center frequency; t represents time; i represents a time-varying sequence; phi α，β (t) represents a base function subjected to scale transformation and translation transformation, and α represents the scale transformationFactor, β represents a time shift factor, CWT (α, β) represents a result after performing wavelet transform on the signal, and f (t) represents an electroencephalogram signal;
and C4: according to the acquired motor imagery time-frequency diagram data set, performing feature extraction and classification in an image processing mode by using a convolutional neural network to obtain a probability distribution matrix P of time-frequency diagram signals TF ，
In the step C2, a feature matrix Z is selected M×T The first m rows and the last m rows of the data are taken as a feature matrix Z' for CSP feature extraction; wherein, 2m<M。
The step D comprises the following specific steps:
d1: according to the probability distribution matrix P EEG and PTF The probability of each action category contained in the model is calculated respectively, the clustering coefficient CC of each action category data in the two probability distribution matrixes is calculated, and the average clustering coefficient is obtained by averaging the clustering coefficients of all nodes of each action category data
wherein ,h represents the number of neighbors of the node; n represents the number of interconnected edges between all adjacent nodes of a node;
d2: the probability distribution matrix P is obtained EEG and PTF Taking the average cluster coefficient of (2) as a weight value, re-optimizing the two probability distribution matrixes to obtain an optimized probability distribution matrix P EEG* and PTF *；
D3: the D-S fusion rule is utilized to optimize the probability distribution matrix P EEG* and PTF * Performing fusion decision to obtain a new probability distribution matrix P, and accordingly obtaining a final classification result;
where k is an evidence conflict factor, k=Σ B∩C＝φ P EEG (B)P TF (C) A, B and C represent probability distribution matrices P, P, respectively EEG and PTF Is a three-way action category.
When the band-pass filtering is carried out, a 4-order Butterworth band-pass filter is selected to carry out 8-30Hz filtering treatment on the original electroencephalogram data, and a frequency band of electroencephalogram related to movement is obtained; when downsampling is performed, the electroencephalogram signal is downsampled to 250Hz; when the channel selection is carried out, 20 EEG channels of the sensory and motor cortex of the brain are selected for classification; when performing independent component analysis for artifact removal, EEGLAB is used for ICA calculation and artifact removal of the EEG signals.
In the step B, the convolutional neural network structure used is EEG-CNN, and comprises a convolutional layer, a pooling layer and a full-connection layer; the two convolution layers and the maximum pooling layer are combined into a feature extraction module, and a one-dimensional convolution kernel in the convolution layers is used for extracting features of each channel to serve as a feature map output by the layer.
In the step C, the convolutional neural network structure is TF-CNN, and VGG16 is adopted as a basic network frame to extract time-frequency diagram characteristic information.
The convolutional neural network in the step B and the step C uses beta in training 1 ＝0.9，β 2 Adam optimizer=0.999, updates trainable parameters for each network layer, and the initial learning rate is set to 0.01.
The invention has the beneficial effects and advantages that:
1. the invention completes the classification and identification of the motor imagery electroencephalogram signals by carrying out fusion decision on the information of the electroencephalogram and the time-frequency diagram. The invention converts the time-sequence electroencephalogram signals into the two-dimensional time-frequency diagram containing rich time-frequency domain and space information, and supplements the time-sequence electroencephalogram signals as the electroencephalogram signals so as to realize more complete decision making and achieve the aim of improving the classification accuracy.
2. In the conversion of the time-frequency diagram, the generation of the image is formed by stacking all channels, and the invention adopts CSP algorithm to carry out spatial filtering on the preprocessed brain electrical signals to obtain a feature matrix, and replaces the brain electrical signals with the feature matrix to complete the conversion of the time-frequency diagram. The invention mainly concentrates the characteristic information contained in the generated characteristic matrix on the head and the tail of the characteristic matrix through the CSP algorithm, and selects the head and the tail of the characteristic matrix as the signal for generating the time-frequency diagram, thereby not only fully utilizing the characteristic information of multiple channels, but also facilitating the stacking and the acquisition of the time-frequency diagram and obtaining better classification effect.
Drawings
FIG. 1 is a schematic flow chart of the present invention;
FIG. 2 is a graph showing the classification effect of the D-S fusion method according to the present invention and EEG-CNN, TF-CNN and CSP methods.
Detailed Description
The invention is described in detail below with reference to the attached drawings and examples:
as shown in FIG. 1, the unilateral limb motor imagery task electroencephalogram classification method provided by the invention sequentially comprises the following steps:
a: constructing a single-side limb motor imagery data set, and performing data preprocessing on motor imagery electroencephalogram signals in the single-side limb motor imagery data set to obtain preprocessed motor imagery electroencephalogram signals;
in the step A, a single-side limb motor imagery data set adopts a multi-mode signal data set of a single upper limb in a Giga DB dataset during 11 visual movement tasks in a multi-time recording process, wherein the multi-mode signal data set comprises arm stretching movements in 6 directions including upward, downward, leftward, rightward, forward and backward, the grasping movements of 3 objects including a cup, a card and a ball, and motor imagery electroencephalogram signals under 2 wrist twisting movements including leftward rotation and rightward rotation; each type of motion is randomly executed for 50 times, and the motor imagery electroencephalogram signal is sampled at the sampling rate of 2500Hz, the time for executing the motor imagery motion is 4s, and each tested person carries out three identical recording processes.
In the step A, data preprocessing comprises bandpass filtering, downsampling, channel selection, independent component analysis and artifact removal and data selection of motor imagery electroencephalogram signals;
in the embodiment, band-pass filtering is used to limit the frequency range of the electroencephalogram signals in the used single-side limb motor imagery data set to 8-30Hz, and downsampling is carried out to 250Hz; simultaneously, channel selection is carried out, only the electroencephalogram data of 20 electroencephalogram channels of the cerebral sensory and motor cortex are selected for research, and EEGLAB tool boxes for electroencephalogram analysis and processing in MATLAB are used for Independent Component Analysis (ICA) of the electroencephalogram, so that interference signals generated by eye movement, muscle movement and the like in the electroencephalogram are removed. An Independent Component Analysis (ICA) algorithm is used as a blind source separation algorithm, and has wide application in the fields of biomedical signal processing, image denoising and the like. In the single-side limb motor imagery data set, only motor imagery electroencephalogram signals under three actions of extending arms forwards and rotating wrists leftwards and grabbing a water cup are selected for classification;
b: and C, performing feature extraction and classification on the preprocessed motor imagery electroencephalogram signals obtained in the step A by using a deep neural network model to obtain a probability distribution matrix P of the electroencephalogram signals EEG The method comprises the steps of carrying out a first treatment on the surface of the Probability distribution matrix P EEG The probability that the preprocessed motor imagery electroencephalogram signal belongs to the corresponding action category is included;
in the embodiment, in the step a, only the motor imagery electroencephalogram signals under the three actions of stretching the arm in the forward direction, rotating the wrist in the left direction and grasping the cup are selected for classification, so wherein ,/> and />Is divided intoThe probability that the sample result belongs to the corresponding action category in the electroencephalogram signal three-category experiment is distinguished by ++>The superscript EEG is English abbreviation of brain wave;
in the step B, the preprocessed motor imagery electroencephalogram is converted into an M multiplied by T matrix form, and then a convolutional neural network is used for extracting and classifying characteristics of the motor imagery electroencephalogram in the matrix form to obtain a probability distribution matrix of the electroencephalogram; wherein M is the number of channels, in this embodiment, the value of M is 20, and t is the data length; the motor imagery electroencephalogram signals used in the method are generated by three action stimuli of forward stretching arms, left rotating wrists and grabbing cups, the electroencephalogram signals belong to different types of actions, the main purpose of classification tasks is to identify action categories to which the electroencephalogram signals belong, three-classification experiments refer to classification and identification of the electroencephalogram signals corresponding to the three actions, the category with the largest probability value in a probability distribution matrix is generally used as a classification result, and the three-classification experiments belong to conventional technologies in the field and are not repeated.
C: c, performing spatial filtering on the preprocessed motor imagery electroencephalogram signal obtained in the step A by using a CSP method to obtain a feature matrix of the electroencephalogram signal; then, a time-frequency diagram data set is obtained through selection of a feature matrix and continuous wavelet transformation, and finally, the time-frequency diagram data set is subjected to feature extraction and classification by using a deep neural network model to obtain a probability distribution matrix P of a time-frequency diagram signal TF The method comprises the steps of carrying out a first treatment on the surface of the Probability distribution matrix P TF The time-frequency diagram signal comprises the probability of belonging to the corresponding action category;
in the embodiment, in the step a, only the motor imagery electroencephalogram signals under the three actions of stretching the arm in the forward direction, rotating the wrist in the left direction and grasping the cup are selected for classification, so wherein ,/> and />Respectively representing the probability of the sample result belonging to the corresponding action category in the time-frequency diagram three-classification experiment, and ++>The superscript TF is English abbreviation of time frequency;
in the invention, the obtained characteristic matrix of the electroencephalogram is used for replacing the electroencephalogram matrix to carry out subsequent processing, and each line of data of the characteristic matrix is used for replacing each channel electroencephalogram data of the electroencephalogram matrix to carry out continuous wavelet transformation so as to generate a time-frequency diagram data set, so that each line of data of the characteristic matrix is called a virtual channel.
The step C comprises the following specific steps:
c1: because the time-frequency diagram is formed by stacking all channels, in order to fully utilize the characteristic information of all 20 channels in the time-frequency diagram, the invention uses CSP algorithm to carry out airspace filtering on the preprocessed motor imagery electroencephalogram; when the spatial filtering is carried out, firstly, a CSP projection matrix W is obtained by calculation, and then a feature matrix Z is obtained by calculation by utilizing the CSP projection matrix W and the preprocessed motor imagery electroencephalogram signals obtained in the step A; the calculation formula is as follows:
Z M×T ＝W M×M E M×T ；
wherein W is CSP projection matrix, M is the channel number of the preprocessed brain electrical data, T is the data length, E is brain electrical data matrix with M x T formed by converting preprocessed motor imagery brain electrical signals;
c2: selecting a feature matrix extracted from CSP features;
because the characteristic matrix Z obtained in the step C1 is not equivalent in characteristic information distribution, the characteristic information is mainly concentrated at the head and the tail of the characteristic matrix Z, and the middle characteristic information is not obviously negligible, so the characteristic matrix Z is selected in the invention M×T The first m rows and the last m rows of the data are taken as a feature matrix Z' for CSP feature extraction;wherein, 2m<M；
And C3: c2, according to the feature matrix Z 'obtained in the step, performing time-frequency feature extraction by using a CWT algorithm, converting the feature matrix Z' into a two-dimensional time-frequency diagram with the resolution of 64 multiplied by 64, selecting Morlet wavelet as a basis function to perform wavelet transformation, and finally obtaining a motor imagery time-frequency diagram data set;
the formula of the wavelet transform is as follows:
the Morlet is a basis function commonly used in wavelet transformation and is used for carrying out decomposition operation on signals to be processed in the wavelet transformation process; phi (t) represents a basis function, and omega represents the wavelet center frequency; t represents time; i represents a time-varying sequence; phi α，β (t) represents a base function subjected to scale transformation and translation transformation, α represents a scale transformation factor, β represents a time translation factor, CWT (α, β) represents a result after wavelet transformation is performed on the signal, and f (t) represents an electroencephalogram signal.
And C4: according to the acquired motor imagery time-frequency diagram data set, performing feature extraction and classification in an image processing mode by using a convolutional neural network to obtain a probability distribution matrix P of time-frequency diagram signals TF ，
D: using the probability distribution matrix P of the EEG signals obtained in the step B EEG And the probability distribution matrix P of the time-frequency diagram signals obtained in the step C TF Calculating the clustering coefficient of each action category respectively, and according to the calculationThe calculated average cluster coefficient pair probability distribution matrix P EEG and PTF Optimizing to obtain an optimized probability distribution matrix P EEG* and PTF* The method comprises the steps of carrying out a first treatment on the surface of the Based on D-S evidence theory and utilizing probability distribution matrix P EEG* and PTF* Performing fusion decision to obtain a final classification result;
in the invention, the probability distribution matrix P of the obtained brain electrical signal EEG And probability distribution matrix P of time-frequency diagram signal TF Known as "evidence" in classification problems, can be used to determine that the classification result belongs to a certain category; and (3) completing uncertain reasoning based on the D-S evidence theory, and fusing a plurality of pieces of evidence to form comprehensive evidence, namely obtaining new basic probability distribution to realize the fusion of different information.
The synthesis rules of the D-S evidence theory are as follows:
if two mutually independent probability distribution matrixes are respectively P obtained by electroencephalogram signal classification EEG And P obtained by classifying time-frequency diagrams TF P is a new probability distribution matrix generated by fusion, and the synthesis rule is defined as:
where k is an evidence conflict factor, k= Σ B∩C＝φ P EEG (B)P TF (C) A, B and C represent probability distribution matrices P, P, respectively EEG and PTF Three action categories in (a); when k=1, the evidence is completely conflicted, the synthesis cannot be performed, and the synthesis rule is invalid; when k is more than 0 and less than 1, the probability distribution matrixes of different events are fused by utilizing a synthesis rule, so that a fusion decision process is realized.
Evidence of individual anomalies often results in evidence conflicts, making the fusion result inaccurate and even erroneous. In addition, two probability distribution matrixes obtained by extracting and classifying the characteristics of the electroencephalogram signal and the time-frequency chart signal are different and still can be optimized. Aiming at the problem, the invention introduces a clustering coefficient before the fusion decision, readjusts the basic probability assignment of each evidence body, and then carries out the fusion decision by using a fusion rule.
The clustering coefficient represents the aggregation level of nodes of one class, and in a specific network, the nodes of the same class always tend to be clustered. For example node v 1 Connected to node v 2 Node v 2 Connection and node v 3 Then node v 3 Most likely with v 1 The connection or the close relation is realized, and the phenomenon shows the dense connection property among partial nodes. The cluster coefficients of any node can be expressed as:
wherein h represents the number of all adjacent nodes of the node, namely the number of neighbors of the node; n represents the number of interconnected edges between all adjacent nodes of a node; the average clustering coefficient can be obtained by averaging the clustering coefficients of all nodes in the networkIn the present invention, the average cluster coefficient of each class obtained by calculation is +.>Optimizing the probability distribution matrix obtained by each classifier, and finally obtaining the optimized probability distribution matrix as follows:
the larger the clustering coefficient is, the larger the node aggregation degree is, and the larger the function in comprehensive evaluation is, so that the clustering coefficient can effectively represent the compactness of each group of data connection, the larger the numerical value is, the more similar the representing data has, and the more accurate the classification of the categories is. Therefore, the reliability of the clustering coefficient is represented by each class of clustering coefficient, the data is optimized by the average clustering coefficient, and the optimized P is reused EEG* and PTF* Instead of the original probability distribution matrix P EEG P TF And the fusion decision is carried out, so that a final classification effect is obtained, and the classification accuracy of the unilateral limb movement information task can be effectively improved.
Examples:
a: firstly, a multi-mode signal data set of 11 visual movement tasks of a single upper limb in a multi-time recording process in the Giga DB dataset is selected to construct a single-side limb movement imagination data set.
The dataset was completed by Ji-Hoon Jeong et al in 2020, and the dataset included not only EEG data, but also Electromyogram (EMG) and Electrooculogram (EOG) data, all collected synchronously in the same experimental environment without interference from each other. The electroencephalogram signal acquisition equipment used by the data set selects electrode positions according to the 10-20 international configuration, uses 60 EEG channels, 6 EMG channels and 4 EOG channels, samples at a sampling rate of 2500Hz, performs motor imagery for 4s, and respectively records actual motion data and imagery motion data by three identical recording processes of each tested object. The invention uses the original brain electricity data of the imagined movement in the data set.
Then, carrying out data preprocessing on the motor imagery electroencephalogram signals in the single-side limb motor imagery data set to obtain preprocessed motor imagery electroencephalogram signals;
in the embodiment, the data preprocessing includes bandpass filtering, downsampling, channel selection, independent component analysis and artifact removal, and data selection of motor imagery electroencephalogram signals, and the used tool is an electroencephalogram signal analysis processing tool box EEGLAB in MATLAB;
band-pass filtering: the noise in the signals can be effectively reduced by considering that the electroencephalogram data is filtered, and the signal-to-noise ratio of the signals of the interested frequency band is improved; a number of studies have shown that the frequency band associated with exercise is 8-30 Hz. Therefore, the invention selects 4-order Butterworth band-pass filter to carry out 8-30Hz filtering treatment on the original electroencephalogram data, thereby obtaining the frequency band of the electroencephalogram related to the movement.
Downsampling: because the sampling frequency of the original electroencephalogram signal is 2500Hz, the corresponding electroencephalogram data is huge, and the speed of subsequent preprocessing and characteristic analysis is greatly reduced, so the electroencephalogram signal is downsampled to 250Hz.
Channel selection: since the constructed single-sided limb motor imagery data set, the original electroencephalogram data contains 60 EEG channels, 4 EOG channels and 7 EMG channels, in order to obtain electroencephalogram data, we need to remove 7 EMG channels and 4 EOG channels, and then obtain 60 EEG channels. In the present invention, these 60 EEG channels were selected again and 20 channels of the sensory motor cortex of the brain were selected for classification.
Independent Component Analysis (ICA) de-artifacting: because the amplitude of the brain electrical signal is very weak and is interfered by eye movement, the data is distorted, the simplest method is to directly remove the distorted parts, but the effective components of the data are lost, and the subsequent research and analysis is greatly error. Therefore, the Independent Component Analysis (ICA) method is often selected to calculate and remove the interference during the preprocessing stage of the electroencephalogram signals. Independent component analysis is a method used to find implicit factors or components from multivariate (multidimensional) statistical data, and is the most common blind source separation technique in electroencephalogram artifact removal: the method decomposes multi-channel EEG data from different sources into Independent Components (ICs), judges whether the independent components are ocular electricity, myoelectricity and other artifacts, removes the components if the independent components are the artifacts, and reconstructs EEG data without the artifacts such as ocular electricity by using the rest independent components. Thereby realizing the effect of removing myoelectricity, electrooculogram and other artifacts. In the invention, EEGLAB is used for ICA calculation and artifact removal of the EEGLAB.
In the single-side limb motor imagery data set, only motor imagery electroencephalogram signals under three actions of extending arms forwards and rotating wrists leftwards and grabbing a water cup are selected for classification;
b: and C, performing feature extraction and classification on the preprocessed motor imagery electroencephalogram signals obtained in the step A by using a deep neural network model to obtain a probability distribution matrix P of the electroencephalogram signals EEG ，
The electroencephalogram data is typically in the form of a two-dimensional matrix of signal channels and sampling points, each row of the matrix representing the sampled data for each channel, the electroencephalogram data in this form including its spatial and temporal characteristics. The electroencephalogram signal data used in the invention is continuously executed for 4s each time, the sampling frequency is 250Hz, the data length is 1000, and finally the electroencephalogram signal data is processed into a 20×T matrix form, wherein 20 is the number of channels, and T=1000 is the data length.
In the invention, the convolutional neural network structure used is EEG-CNN, and comprises a convolutional layer, a pooling layer and a full-connection layer; wherein the two convolution layers and the maximum pooling layer are combined into a feature extraction module; electroencephalogram signal E input into network M×T Is a matrix of M rows and T columns, where M represents the number of channels and T represents the length of the electroencephalogram signal. In the invention, the most commonly used two-dimensional convolution is not selected in the construction of the network, but one-dimensional convolution is used, the convolution kernel of the one-dimensional convolution only moves in one direction, the convolution kernel only calculates along the length of a signal, and the time sequence characteristics are extracted from the electroencephalogram data. Classical CNN relies on multilayer convolution pooling to extract different basic layer features to improve classification accuracy, but the increase of network layers can increase network parameters by times, which is unfavorable for rapid convergence of network and seriously affects network performance. Considering that our input is a simpler matrix vector, we use shallower network depth, reduce the number of convolution layers used, use only two sets of convolution layers and pooling layers to compose a feature extraction module for extracting feature information of input signals, and the application of two convolution layers can make the network possess more nonlinear changes and adapt to complex modes. In this embodiment, the size of the convolution kernel is set to 3×1, the step size is 1, and N can be obtained after convolution w ×N f A formal feature map; wherein N is w As vectors, N f The number of convolution kernels; maximum poolThe layer is used for downsampling the data output by the convolution layer, and in the embodiment, the pooling layer selects the kernel size as 2×1 and the step length as 2; the full connection layer is used for flattening the features extracted by the convolution layer.
C: c, performing spatial filtering on the preprocessed motor imagery electroencephalogram obtained in the step A by using a CSP method to obtain a feature matrix Z of the electroencephalogram, selecting Z to obtain a feature matrix Z ', performing continuous wavelet transformation on Z' to obtain a motor imagery time-frequency diagram data set, and finally performing feature extraction and classification on the time-frequency diagram data set by using a deep neural network model to obtain a probability distribution matrix P of the time-frequency diagram signal TF ，
In the invention, the conversion of the time-frequency diagram is divided into the following 4 steps:
c1: obtaining a feature matrix Z;
because the conversion generation of the time-frequency diagram is carried out by each channel independently, 20 electroencephalogram channels used in the invention can not be combined together to be used as a picture, and if only a few channels are selected, information loss can be caused, and a lot of useful channel information is lost. In order to effectively utilize the information of each channel, the invention carries out preprocessing on the data to extract the characteristics, namely, the CSP method is used for carrying out spatial filtering on the electroencephalogram signals of 20 electroencephalogram channels to obtain virtual channels, and then a time-frequency diagram is generated. The CSP is a spatial filtering feature extraction algorithm aiming at two classification tasks, and can extract spatial distribution components of each class from multichannel electroencephalogram data, the basic principle of the algorithm is to find a group of optimal spatial filters for projection by utilizing diagonalization of a matrix, so that variance value difference of two classes of signals is maximized, and thus, feature vectors with higher distinction degree are obtained. And for three classification tasks, the CSP can be expanded by adopting an OVR strategy (One vs. Rest) to realize multi-classification CSP feature extraction. In the invention, a feature matrix Z of the electroencephalogram is obtained according to the calculated CSP projection matrix W and the preprocessed electroencephalogram obtained in the step A, and a calculation formula is as follows;
Z M×T ＝W M×M E M×T ；
wherein, W is CSP projection matrix, M is the number of channels of the preprocessed electroencephalogram data, m=20, T is the data length, E is an electroencephalogram data matrix with size of mxt formed by converting the preprocessed motor imagery electroencephalogram signals in step a;
c2: selecting a feature matrix extracted from CSP features;
in this embodiment, a feature matrix Z is selected M×T Data of the first and last lines (i.e., m=1) as feature matrices for CSP feature extraction to obtain feature matrix Z' 2×T The time-frequency diagram conversion is carried out instead of the preprocessed motor imagery electroencephalogram data.
And C3: generating a time-frequency diagram data set;
and extracting characteristic information from the electroencephalogram signal through CWT, and finally converting the characteristic information into a two-dimensional time-frequency diagram with the resolution of 64 multiplied by 64. Processing all the brain electrical signals by the method to finally obtain a time-frequency diagram data set of motor imagery;
and C4: extracting and classifying characteristics of time-frequency diagram signals;
according to the acquired motor imagery time-frequency diagram data set, then using a convolutional neural network to perform feature extraction and classification in an image processing mode to obtain a probability distribution matrix P of time-frequency diagram signals TF ，
The convolutional neural network structure is TF-CNN, VGG16 can be adopted as a basic network framework, and the VGG16 network is a typical structure of the convolutional neural network and is mainly characterized by comprising convolutional kernel calculation and a feedforward structure. Compared with other CNN, VGG16 has the characteristics of high precision and good stability, so that the method is widely applied. Therefore, the invention adopts the VGG16 network model to extract the time-frequency diagram characteristic information, the VGG16 network uses a plurality of continuous convolutions, has fewer parameters than a larger convolution kernel used alone, and simultaneously has more nonlinear changes than a single convolution, thereby being suitable for more complex modes. And features are extracted in series and multiple times by the convolution kernels, so that the features extracted by the convolution kernels are finer than those extracted by a single convolution kernel. And feature extraction can be better performed on time-frequency chart pictures with smaller differences. VGG16 contains 16 hidden layers (13 convolutional layers and 3 fully-connected layers), the convolution kernels used in the convolutional parts are all 3×3 in size, the step sizes are all 1, the maximum pooling layer size is 2×2, and the step size is 2.
In the invention, beta is used in the training stage of the classification model used by the electroencephalogram and the time-frequency chart 1 ＝0.9，β 2 Adam optimizer=0.999, updates trainable parameters for each network layer, and the initial learning rate is set to 0.01.
D: using the probability distribution matrix P of the EEG signals obtained in the step B EEG And the probability distribution matrix P of the time-frequency diagram signals obtained in the step C TF The clustering coefficient of each action category is calculated respectively, and the probability distribution matrix P is based on the average clustering coefficient EEG and PTF Optimizing to obtain an optimized probability distribution matrix P EEG* and PTF* The method comprises the steps of carrying out a first treatment on the surface of the Based on D-S evidence theory and utilizing probability distribution matrix P EEG* and PTF* Performing fusion decision to obtain a final classification result;
d1: according to the probability distribution matrix P EEG and PTF The probability of each action category contained in the model is calculated respectively, the clustering coefficient CC of each action category data in the two probability distribution matrixes is calculated, and the average clustering coefficient is obtained by averaging the clustering coefficients of all nodes of each action category data
D2: the probability distribution matrix P is obtained EEG and PTF Taking the average cluster coefficient of (2) as a weight value, re-optimizing and correcting the two probability distribution matrixes to obtain an optimized probability distribution matrix P EEG* and PTF *；
D3: the D-S fusion rule is utilized to optimize the probability distribution matrix P EEG* and PTF * Making fusion decisionsAnd obtaining a new probability distribution matrix P, and obtaining a final classification result according to the new probability distribution matrix P.
In the invention, the average classification accuracy of 25 testees in 3 acquisition processes is tested, and the classification effect comparison graph of the CSP method using the classical algorithm most commonly used in the classification of the electroencephalogram is shown in FIG. 2, wherein the classification effect comparison graph is obviously higher than that of other methods by using the method (which is called as a D-S fusion result herein), the network model (which is called as EEG-CNN herein) for classifying the electroencephalogram signals, the network model (which is called as TF-CNN herein) for classifying the time-frequency chart, and the classification effect comparison graph of the CSP method using the classical algorithm most commonly used in the classification of the electroencephalogram signals.

Claims (10)

1. The electroencephalogram classification method for the unilateral limb motor imagery task is characterized by comprising the following steps in sequence:
a: constructing a single-side limb motor imagery data set, and performing data preprocessing on motor imagery electroencephalogram signals in the single-side limb motor imagery data set to obtain preprocessed motor imagery electroencephalogram signals;
b: and C, performing feature extraction and classification on the preprocessed motor imagery electroencephalogram signals obtained in the step A by using a deep neural network model to obtain a probability distribution matrix P of the electroencephalogram signals EEG The method comprises the steps of carrying out a first treatment on the surface of the Probability distribution matrix P EEG The probability that the preprocessed motor imagery electroencephalogram signal belongs to the corresponding action category is included;
c: c, performing spatial filtering on the preprocessed motor imagery electroencephalogram signal obtained in the step A by using a CSP method to obtain a feature matrix of the electroencephalogram signal; then, a time-frequency diagram data set is obtained through selection of a feature matrix and continuous wavelet transformation, and finally, the time-frequency diagram data set is subjected to feature extraction and classification by using a deep neural network model to obtain a probability distribution matrix P of a time-frequency diagram signal TF The method comprises the steps of carrying out a first treatment on the surface of the Probability distribution matrix P TF The time-frequency diagram signal comprises the probability of belonging to the corresponding action category;
d: using the probability distribution matrix P of the EEG signals obtained in the step B EEG And the summary of the time-frequency diagram signal obtained in the step CRate distribution matrix P TF Respectively calculating the clustering coefficient of each action category, and according to the calculated average clustering coefficient, performing a probability distribution matrix P EEG and PTF Optimizing to obtain an optimized probability distribution matrix P EEG* and PTF* The method comprises the steps of carrying out a first treatment on the surface of the Based on D-S evidence theory and utilizing probability distribution matrix P EEG* and PTF* And carrying out fusion decision to obtain a final classification result.
2. The method for classifying brain waves of a motor imagery task of a side limb according to claim 1, wherein: in the step A, the data preprocessing comprises band-pass filtering, downsampling, channel selection, independent component analysis and artifact removal and data selection of motor imagery electroencephalogram signals.
3. The method for classifying brain waves of a motor imagery task of a side limb according to claim 1, wherein: in the step A, in a single-side limb motor imagery data set, only motor imagery electroencephalogram signals under three actions of forward stretching arms, left rotating wrists and grabbing cups are selected;
in the step B of the process, wherein ,/> and />The probability that the sample results belong to the corresponding action categories in the electroencephalogram signal three-classification experiment is respectively;
in the step C, the step of, in the step C, wherein ,/> and />And the probabilities that the sample results belong to the corresponding action categories in the time-frequency diagram three-classification experiment are respectively.
4. The method for classifying brain waves of a motor imagery task of a side limb according to claim 1, wherein the step C includes the following steps:
c1: performing spatial filtering on the preprocessed motor imagery electroencephalogram signals by using a CSP algorithm;
when the spatial filtering is carried out, firstly, a CSP projection matrix W is obtained by calculation, and then a feature matrix Z is obtained by calculation by utilizing the CSP projection matrix W and the preprocessed motor imagery electroencephalogram signals obtained in the step A; the calculation formula is as follows: z is Z M×T ＝W M×M E M×T The method comprises the steps of carrying out a first treatment on the surface of the Wherein W is CSP projection matrix, M is the channel number of the preprocessed brain electrical data, T is the data length, E is brain electrical data matrix with M x T formed by converting preprocessed motor imagery brain electrical signals;
c2: selecting a feature matrix Z' of CSP feature extraction;
and C3: c2, according to the feature matrix Z 'obtained in the step, performing time-frequency feature extraction by using a CWT algorithm, converting the feature matrix Z' into a two-dimensional time-frequency diagram, selecting Morlet wavelets as a basis function to perform wavelet transformation, and finally obtaining a motor imagery time-frequency diagram data set;
the formula of the wavelet transform is as follows:
the Morlet is a basis function commonly used in wavelet transformation and is used for carrying out decomposition operation on signals to be processed in the wavelet transformation process; phi (t) represents a basis function, and omega represents the wavelet center frequency; t represents time; i represents a time-varying sequence; phi α,β (t) represents a base function subjected to scale transformation and translation transformation, α represents a scale transformation factor, β represents a time translation factor, CWT (α, β) represents a result after wavelet transformation is performed on the signal, and f (t) represents an electroencephalogram signal;
and C4: according to the acquired motor imagery time-frequency diagram data set, performing feature extraction and classification in an image processing mode by using a convolutional neural network to obtain a probability distribution matrix P of time-frequency diagram signals TF ，
5. The method for classifying brain waves of a motor imagery task of a side limb according to claim 4, wherein: in the step C2, a feature matrix Z is selected M×T The first m rows and the last m rows of the data are taken as a feature matrix Z' for CSP feature extraction; wherein, 2m<M。
6. The method for classifying brain waves of a motor imagery task of a side limb according to claim 1, wherein: the step D comprises the following specific steps:
d1: according to the probability distribution matrix P EEG and PTF The probability of each action category contained in the model is calculated respectively, the clustering coefficient CC of each action category data in the two probability distribution matrixes is calculated, and the average clustering coefficient is obtained by averaging the clustering coefficients of all nodes of each action category data
wherein ,h represents the number of neighbors of the node; n represents the number of interconnected edges between all adjacent nodes of a node;
d2: the probability distribution matrix P is obtained EEG and PTF Taking the average cluster coefficient of (2) as a weight value, re-optimizing the two probability distribution matrixes to obtain an optimized probability distribution matrix P EEG* and PTF *；
D3: the D-S fusion rule is utilized to optimize the probability distribution matrix P EEG* and PTF * Performing fusion decision to obtain a new probability distribution matrix P, and accordingly obtaining a final classification result;
where k is an evidence conflict factor, k= Σ B∩C＝φ P EEG (B)P TF (C) A, B and C represent probability distribution matrices P, P, respectively EEG and PTF Is a three-way action category.
7. The method for classifying brain waves of a motor imagery task of a side limb according to claim 2, wherein: when the band-pass filtering is carried out, a 4-order Butterworth band-pass filter is selected to carry out 8-30Hz filtering treatment on the original electroencephalogram data, and a frequency band of electroencephalogram related to movement is obtained; when downsampling is performed, the electroencephalogram signal is downsampled to 250Hz; when the channel selection is carried out, 20 EEG channels of the sensory and motor cortex of the brain are selected for classification; when performing independent component analysis for artifact removal, EEGLAB is used for ICA calculation and artifact removal of the EEG signals.
8. The method for classifying brain waves of a motor imagery task of a side limb according to claim 1, wherein: in the step B, the convolutional neural network structure used is EEG-CNN, and comprises a convolutional layer, a pooling layer and a full-connection layer; the two convolution layers and the maximum pooling layer are combined into a feature extraction module, and a one-dimensional convolution kernel in the convolution layers is used for extracting features of each channel to serve as a feature map output by the layer.
9. The method for classifying brain waves of a motor imagery task of a side limb according to claim 1, wherein: in the step C, the convolutional neural network structure is TF-CNN, and VGG16 is adopted as a basic network frame to extract time-frequency diagram characteristic information.
10. The method for classifying brain waves of a motor imagery task of a side limb according to claim 1, wherein: the convolutional neural network in the step B and the step C uses beta in training 1 ＝0.9，β 2 Adam optimizer=0.999, updates trainable parameters for each network layer, and the initial learning rate is set to 0.01.
CN202310719853.XA 2023-02-06 2023-06-17 Electroencephalogram classification method for unilateral limb motor imagery tasks Pending CN116738316A (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN2023100674056 2023-02-06
CN202310067405 2023-02-06

Publications (1)

Publication Number Publication Date
CN116738316A true CN116738316A (en) 2023-09-12

Family

ID=87909381

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310719853.XA Pending CN116738316A (en) 2023-02-06 2023-06-17 Electroencephalogram classification method for unilateral limb motor imagery tasks

Country Status (1)

Country Link
CN (1) CN116738316A (en)

Similar Documents

Publication Publication Date Title
Hermosilla et al. Shallow convolutional network excel for classifying motor imagery EEG in BCI applications
CN112890827B (en) Electroencephalogram identification method and system based on graph convolution and gate control circulation unit
CN114947883B (en) Deep learning electroencephalogram noise reduction method based on time-frequency domain information fusion
CN114533086A (en) Motor imagery electroencephalogram decoding method based on spatial domain characteristic time-frequency transformation
Elessawy et al. A long short-term memory autoencoder approach for EEG motor imagery classification
Feng et al. Feature extraction algorithm based on csp and wavelet packet for motor imagery eeg signals
CN114387668A (en) Classification method based on multi-level neuromuscular coupling characteristic information fusion
Geng et al. A fusion algorithm for EEG signal processing based on motor imagery brain-computer interface
Ghonchi et al. Spatio-temporal deep learning for EEG-fNIRS brain computer interface
Mirzabagherian et al. Temporal-spatial convolutional residual network for decoding attempted movement related EEG signals of subjects with spinal cord injury
Abougharbia et al. A novel BCI system based on hybrid features for classifying motor imagery tasks
CN116738316A (en) Electroencephalogram classification method for unilateral limb motor imagery tasks
Zhao et al. GTSception: a deep learning eeg emotion recognition model based on fusion of global, time domain and frequency domain feature extraction
Bhattacharyya et al. Detection of fast and slow hand movements from motor imagery EEG signals
Guo et al. EEG signal analysis based on fixed-value shift compression algorithm
Rabby et al. Time-frequency based EEG motor imagery signal classification with deep learning networks
Narayan Analysis of MLP and DSLVQ classifiers for EEG signals based movements identification
Zhou et al. MI-EEG temporal information learning based on one-dimensional convolutional neural network
CN114533083B (en) Motor imagery state identification method based on multi-fusion convolutional neural network
CN114082169B (en) Disabled hand soft body rehabilitation robot motor imagery identification method based on electroencephalogram signals
Ghanbari et al. Wavelet and Hilbert transform-based brain computer interface
Massar et al. DWT-BSS: Blind Source Separation applied to EEG signals by extracting wavelet transform’s approximation coefficients
Ahirwal et al. Classification of 2D hand movement with power spectrum estimation
Anjum et al. Effect of artifact removal on EEG based motor imagery BCI applications
Liu et al. Low quality samples detection in motor imagery EEG data by combining independent component analysis and confident learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination