CN113128384A - Brain-computer interface software key technical method of stroke rehabilitation system based on deep learning - Google Patents

Brain-computer interface software key technical method of stroke rehabilitation system based on deep learning Download PDF

Info

Publication number
CN113128384A
CN113128384A CN202110376347.6A CN202110376347A CN113128384A CN 113128384 A CN113128384 A CN 113128384A CN 202110376347 A CN202110376347 A CN 202110376347A CN 113128384 A CN113128384 A CN 113128384A
Authority
CN
China
Prior art keywords
electroencephalogram
signal
sequence
layer
sample entropy
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.)
Granted
Application number
CN202110376347.6A
Other languages
Chinese (zh)
Other versions
CN113128384B (en
Inventor
王卓峥
宋霖涛
董雨萌
任博雯
丁熠辉
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.)
Beijing University of Technology
Original Assignee
Beijing University of Technology
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 Beijing University of Technology filed Critical Beijing University of Technology
Priority to CN202110376347.6A priority Critical patent/CN113128384B/en
Publication of CN113128384A publication Critical patent/CN113128384A/en
Application granted granted Critical
Publication of CN113128384B publication Critical patent/CN113128384B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • 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)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Theoretical Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Public Health (AREA)
  • Biophysics (AREA)
  • Medical Informatics (AREA)
  • Signal Processing (AREA)
  • Pathology (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Veterinary Medicine (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Physiology (AREA)
  • Psychiatry (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Databases & Information Systems (AREA)
  • Fuzzy Systems (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

A stroke rehabilitation system brain-computer interface software key technical method based on deep learning belongs to the technical field of deep learning. According to the invention, in the electroencephalogram signal feature extraction and classification stage, the autoregressive model and the sample entropy are adopted to extract features of the electroencephalogram signal, and the classification is carried out by utilizing the CNN algorithm, so that the accuracy of electroencephalogram signal classification can be improved, and the electroencephalogram signal classification method can be applied to the brain-computer interface software of a stroke rehabilitation system to realize rehabilitation treatment of stroke patients.

Description

Brain-computer interface software key technical method of stroke rehabilitation system based on deep learning
Technical Field
The invention belongs to the technical field of deep learning.
Background
At present, the Brain electrical signal correlation technology and Brain Computer Interface (BCI) technology are continuously developed, and the application of the Brain electrical signal in the recovery of stroke is more and more extensive. Relevant research results show that the motor imagery electroencephalogram signals of stroke patients can help them to carry out effective rehabilitation treatment. Researches find that the activity of the electroencephalogram signals can not change due to the damage of limbs and muscles, and the electroencephalogram signals can still accurately reflect the electrophysiological activity performed by the cerebral cortex of the human beings. Therefore, the related art of brain-computer interface based on motor imagery has become a key point in the rehabilitation research of stroke patients.
The deep learning is different from the traditional machine learning method, an end-to-end learning mode is realized, and the original data is used as model input and is directly mapped to generate the output of the model. The more classical deep learning models include a convolutional neural network and a cyclic neural network. However, the method has disadvantages that, for example, a deep neural network requires a large amount of data to train a model, and it is difficult to sufficiently train the model for most small-scale data sets; lack of sufficient label data to effectively train the deep learning model; there is a huge computational demand. Deep forest (gcForest) is a feature fusion model as a deep neural network model including a convolutional neural network and a cyclic neural network, and has been a result of attention in various fields such as vision, text, and speech. In deep forests, the strategy of multi-granularity scanning is used, so that the relevance between the features captured by the model can be helped, and the capability of model characterization learning is further improved.
The traditional electroencephalogram signal processing method mainly has the time filtering, the space filtering, the principal component analysis, the independent component analysis and the like. At present, the mainstream electroencephalogram feature extraction method is a power spectral density and public space mode algorithm. The traditional machine learning method is to apply a proper classification algorithm to classify samples on the basis of finishing data preprocessing and artificially constructing sample characteristics. In the subject, a deep learning mode is tried to be adopted, the algorithm is further optimized, and the processing speed and accuracy of the algorithm are improved.
The invention in this field of China is as follows: the data enhancement-based convolutional neural network motor imagery electroencephalogram classification method [1] classifies electroencephalogram signals through a convolutional neural network, and the average accuracy of the obtained second, third and fourth classification tasks respectively reaches 87.32%, 76.26% and 64.72%. The method [2] for recognizing the characteristics of the motor imagery electroencephalogram signals based on the LFFCNN-GRU algorithm model extracts the frequency domain characteristics of the electroencephalogram signals by using a convolutional neural network based on interlayer characteristic fusion, and further extracts the time domain characteristics of the electroencephalogram signals by using a gated cyclic network. The electroencephalogram identification method [3] based on CWT and MLMSFFCNN fully extracts time, frequency and space domain feature information of signals by utilizing MLMSFFCNN feature fusion capability and multi-resolution computing capability, and improves classification accuracy.
[1] Liuyue, Dubin, Yuekang, Tiankuang. convolutional neural network motor imagery electroencephalogram classification method based on data enhancement [ P ]. Beijing City: CN111950366A,2020-11-17.
[2] Mao Xuefeng, Shexiangrong, Zhang Yi, Rouyan, a motor imagery electroencephalogram signal feature identification method based on LFFCNN-GRU algorithm model [ P ]. Chongqing city: CN111950455A,2020-11-17.
[3] Li Minai, Korea Jianfu, Yangjin Fu, Sun Yan jade, CWT and MLMSFFCNN-based electroencephalogram identification method [ P ]. Beijing City: CN111582041A,2020-08-25.
Disclosure of Invention
According to the invention, in the electroencephalogram signal feature extraction and classification stage, the autoregressive model and the sample entropy are adopted to extract features of the electroencephalogram signal, and the classification is carried out by utilizing the CNN algorithm, so that the accuracy of electroencephalogram signal classification can be improved, and the electroencephalogram signal classification method can be applied to the brain-computer interface software of a stroke rehabilitation system to realize rehabilitation treatment of stroke patients.
The technical scheme of the invention comprises the following stages: in the electroencephalogram signal acquisition and preprocessing stage, an electroencephalogram signal acquisition system is used for acquiring an EEG signal, then the acquired EEG signal is preprocessed by EEGLAB to remove electrooculogram, artifacts and the like, the electroencephalogram signal with noise filtered is obtained, and the dimensionality reduction is carried out through principal component analysis; in the EEG signal feature extraction stage, an Autoregressive Model (Autoregressive Model) and sample entropy are respectively utilized to extract EEG signal features of motor imagery, two types of linear and nonlinear signal features can be respectively obtained from a frequency domain and a space domain, wherein AR Model parameters are used as feature vectors; and a characteristic classification stage, namely classifying the electroencephalogram signals based on a CNN (convolutional neural network) design classifier, respectively adding two types of signal characteristics obtained by characteristic extraction into the CNN for classification to obtain different classification results, and integrating the results according to a certain weight to obtain a final classification result.
The invention improves the classification accuracy of the motor imagery electroencephalogram signals by simultaneously extracting the characteristics of frequency domain electroencephalogram signals, spatial domain electroencephalogram signals, linear electroencephalogram signals and nonlinear electroencephalogram signals.
Description of the drawings:
FIG. 1 is a flow chart of the present invention.
The specific implementation mode is as follows:
1. electroencephalogram signal acquisition stage
The invention uses an EEG signal acquisition instrument to acquire signals, selects a sampling frequency of 250Hz, arranges electrodes according to an international 10-20 standard electrode arrangement method, selects a plurality of healthy subjects, respectively acquires left-hand and right-hand motor imagery EEG signals, uses a band-stop filter to filter, and selects 1Hz-50Hz EEG signals.
2. Preprocessing stage of EEG signal
The electroencephalogram signal of the human body is very weak, and the background noise is very large. Due to the influence of external factors such as instruments and the muscle movement of the human body, various noise interferences such as baseline drift interference, power frequency interference, myoelectric interference, electrode contact noise and the like can be generated when the electroencephalogram signals are collected. The above interference will affect both the time domain analysis and the frequency domain analysis of the EEG.
The main purposes of preprocessing are to improve the signal-to-noise ratio of the brain electrical signal, stabilize the baseline, reduce artifacts, and remove glitches. The traditional preprocessing method generally adopts Fourier transform, but because the electroencephalogram signals have the characteristic of random instability, the effect of denoising the electroencephalogram signals by adopting the Fourier transform is not ideal. In recent years, with the deep research on electroencephalogram signals, methods such as wavelet transformation and independent component analysis gradually play an important role in signal preprocessing. And the fast fixed point algorithm (FastICA) in the independent component analysis has the characteristics of parallelism, distribution, less occupied memory, fast convergence and the like.
EEGLAB is used for preprocessing, FastICA is used for denoising experimental data, and principal component analysis is used for dimensionality reduction of the data.
3. Feature extraction stage
After EEG signal preprocessing, the processed EEG signal is respectively subjected to feature extraction by using an Autoregressive Model (Autoregressive Model) and sample entropy to obtain two types of signal features.
3.1 Autoregressive Model (Autoregensive Model)
An Autoregressive Model (Autoregressive Model) is a common form in a time series and utilizes a linear combination of random variables at a plurality of moments in the early stage to describe a linear regression Model of random variables at a certain moment in the later stage. Suppose the EEG signal sequence is y1,y2,…,ynThe autoregressive model of order P (abbreviated AR (P)) indicates y in the sequencetIs a function of the linear combination of the first P sequences and the error term, and the general form of the mathematical model is:
Figure BDA0003004020900000041
wherein the content of the first and second substances,
Figure BDA0003004020900000042
is a constant term that is used to determine,
Figure BDA0003004020900000043
is a model parameter, etIs white noise with a mean value of 0 and a variance σ. In the invention
Figure BDA0003004020900000044
As EEG signal feature vectors.
3.2 sample entropy
The sample entropy is a nonlinear analysis method, and can reflect the nonlinear characteristics of the electroencephalogram signal by measuring the complexity of the electroencephalogram signal.
Let us say the electroencephalogram one-dimensional time sequence is { y (i) }, i ═ 1,2, …, n, where y (i) denotes the electroencephalogram signal sequence of the ith second, and n denotes the total time length. The sample entropy can be calculated as follows:
(1) forming m-dimensional vectors by the sequence { y (i) } in sequence, and selecting m as 2 in the invention, namely
Ym(i)＝[y(i)y(i+1)…y(i+m-1)]
i＝1,2,…,n-m+1
(2) For each time i, a vector Y is calculatedm(i) And vector Ym(j) Is a distance therebetween, i.e.
d[Ym(i),Ym(j)]＝max|y(i+k)-y(j+k)|
k＝0,1,…,m-1；i＝1,2,…,n-m+1；i≠j
(3) Given threshold r (r)>0) In the invention, r is 0.2 times of the standard deviation of the original time sequence, and d [ Y ] is counted for each time im(i),Ym(j)]A number less than r, is recorded
Figure BDA0003004020900000045
And the ratio of the number to the total number of distances n-m is recorded as
Figure BDA0003004020900000046
Namely, it is
Figure BDA0003004020900000047
(4) To find
Figure BDA0003004020900000048
Average of all values, denoted Bm(r) that
Figure BDA0003004020900000049
(5) Forming m + 1-dimensional vectors by the sequence { y (i) } in sequence, and repeating the steps (1) to (4) to obtain
Figure BDA00030040209000000410
And Bm+1(r)
(6) The sample entropy of the sequence y (i) } is
Figure BDA00030040209000000411
In practical calculation, because the sequence length is limited, the finally obtained sample entropy estimated value when the sequence length is n is as follows
sampEn(m,r,n)＝-ln[Bm+1(r)/Bm(r)]}
The specific steps of extracting the nonlinear characteristics of the electroencephalogram signal by utilizing the sample entropy are as follows: firstly, adding a sliding time window with the length of 1s to the electroencephalogram signal, calculating the sample entropy of the electroencephalogram signal, moving the window by one sampling point each time, calculating the sample entropy of the electroencephalogram signal of the next 1s time window until the sample entropy of the electroencephalogram signal of the last 1s time window is calculated, thereby obtaining the time sequence of the sample entropy of the group of electroencephalogram signals, and then superposing and averaging the group of sample entropy sequences to obtain the sample entropy of the group of sample electroencephalogram signal data.
4. Feature classification phase
The invention uses a Convolutional Neural Network (CNN) as a classifier to classify the two extracted types of feature vectors respectively, and then the classification results of the two types of feature vectors are weighted to finally obtain the classification results.
4.1CNN construction and principles thereof
The CNN used by the invention consists of six convolutional layers, two pooling layers and two full-connecting layers, and the sequence is as follows: the three-layer rolling layer, the one-layer pooling layer, the three-layer rolling layer, the one-layer pooling layer and the two full-connection layers are all realized by using Tensorflow.
Dividing the data set into training set and test set according to the ratio of 7:3, and extracting two types of feature vectors obtained by feature extraction
Figure BDA0003004020900000051
(where i ∈ [1, n ]]N represents the total time length) and SEi＝{i∈[1,n]| sampEn (m, r, i) } (whichWhere n represents the total time length) as input, three prediction possibilities are output (T0, T1, T2, respectively representing resting state, imaginary left hand, imaginary right hand). An Adam optimizer was used, and the learning rate was 1 x 10-5.
(1) Convolutional layer (convolutional layer): the size of each convolution kernel of each convolution layer in the convolution neural network is 3 x 3, and the parameter of each convolution unit is obtained by optimization through a back propagation algorithm, so that different input features can be extracted.
(2) Pooling layer (Pooling layer): after the convolution, features of large dimensions (size [28 x 20 x 64]) were obtained, and the features were cut into regions of [2,2], respectively, the maximum of which was taken to obtain new features of smaller dimensions (size [14 x 10 x 64 ]).
(3) Supervised learning
The model uses supervised learning, using a back propagation algorithm to calculate the parameters of each convolution unit.
The main process is as follows:
calculating a loss function using Euclidean distances
E＝(x-pred)2
Where x is the value of the actual feature vector and pred is the value of the predicted feature vector.
Propagating the loss function from the output layer back to the hidden layer until propagating to the input layer; in the back propagation process, adjusting the value of the parameter of each convolution unit according to a loss function; and continuously iterating the process until convergence.
4.2 weighted integration
In the last step, the results of the classification of the two types of feature vectors are summed by respectively weighting 50% to obtain the final prediction result. As can be seen from the above, the present invention,
Figure BDA0003004020900000061
and SEpredThe prediction result vectors are respectively two kinds of feature vectors in the form of
Figure BDA0003004020900000062
And
Figure BDA0003004020900000064
Figure BDA0003004020900000065
wherein predi(i＝T0,T1,T2) The predicted probability that the group of brain electrical signals in the model is of the second class is shown. The final prediction result vector is
Figure BDA0003004020900000063
The final prediction result can be obtained by final.

Claims (2)

1. A stroke rehabilitation system brain-computer interface software key technical method based on deep learning is characterized by comprising the following stages: in the electroencephalogram signal acquisition and preprocessing stage, an electroencephalogram signal acquisition system is used for acquiring an EEG signal, then the acquired EEG signal is preprocessed by EEGLAB to remove electrooculogram and artifacts, the electroencephalogram signal with noise filtered is obtained, and the dimensionality is reduced through principal component analysis;
in the electroencephalogram signal feature extraction stage, an autoregressive model and sample entropy are respectively utilized to extract the features of a motor imagery EEG signal, and linear and nonlinear signal features can be respectively obtained from a frequency domain and a space domain, wherein AR model parameters are used as feature vectors; and a characteristic classification stage, namely classifying the electroencephalogram signals based on a CNN (convolutional neural network) design classifier, respectively adding two types of signal characteristics obtained by characteristic extraction into the CNN for classification to obtain different classification results, and integrating the results to obtain a final classification result.
2. The method of claim 1, wherein:
1) electroencephalogram signal acquisition stage
Collecting signals by using an electroencephalogram signal collector, selecting a sampling frequency of 250Hz, arranging electrodes according to an international 10-20 standard electrode arrangement method, selecting a plurality of healthy subjects, respectively collecting left-hand and right-hand motor imagery electroencephalogram signals, filtering by using a band-stop filter, and selecting 1Hz-50Hz EEG signals;
2) preprocessing stage of electroencephalogram signals
EEGLAB is used for preprocessing, FastICA is used for denoising experimental data, and principal component analysis is used for reducing the dimension of the data;
3) a feature extraction stage
After preprocessing, respectively performing feature extraction on the processed EEG signal by using an autoregressive model and sample entropy to obtain two types of signal features;
3.1 autoregressive model
The autoregressive model describes a linear regression model of random variables at a later moment by utilizing linear combinations of the random variables at a plurality of earlier moments; suppose the EEG signal sequence is y1，y2，...，ynThe autoregressive model of order P, denoted AR (P), indicates y in the sequencetIs a function of the linear combination of the first P sequences and the error term, and the mathematical model is:
Figure FDA0003004020890000021
wherein the content of the first and second substances,
Figure FDA0003004020890000022
is a constant term that is used to determine,
Figure FDA0003004020890000023
is a model parameter, etWhite noise with mean 0 and variance σ; will be provided with
Figure FDA0003004020890000024
As EEG signal feature vectors;
3.2 sample entropy
Setting the electroencephalogram signal one-dimensional time sequence as { y (i) }, i ═ 1,2,. and n, wherein y (i) represents the electroencephalogram signal sequence of the ith second, and n represents the total time length; the sample entropy is calculated as follows:
(1) forming m-dimensional vectors by the sequence { y (i) } in sequence, and selecting m as 2, namely
Ym(i)＝[y(i)y(i+1)...y(i+m-1)]
i＝1，2，...，n-m+1
(2) For each time i, a vector Y is calculatedm(i) And vector Ym(j) Is a distance therebetween, i.e.
d[Ym(i)，Ym(j)]＝max|y(i+k)-y(j+k)|
k＝0，1，...，m-1；i＝1，2，...，n-m+1；i≠j
(3) Given a threshold r, r > 0, where r is 0.2 times the standard deviation of the original time series, d [ Y ] is counted for each i timem(i)，Ym(j)]A number less than r, is recorded
Figure FDA0003004020890000025
And the ratio of the number to the total number of distances n-m is recorded as
Figure FDA0003004020890000026
Namely, it is
Figure FDA0003004020890000027
(4) To find
Figure FDA0003004020890000028
The average value of all i values is recorded as Bm(r) that
Figure FDA0003004020890000029
(5) Forming m + 1-dimensional vectors by the sequence { y (i) } in sequence, and repeating the steps (1) to (4) to obtain
Figure FDA00030040208900000211
And Bm+1(r)
(6) The sample entropy of the sequence y (i) } is
Figure FDA00030040208900000210
Obtaining a sample entropy estimated value when the sequence length is n as
sampEn(m，r，n)＝-ln[Bm+1(r)/Bm(r)]}
The specific steps of extracting the nonlinear characteristics of the electroencephalogram signal by utilizing the sample entropy are as follows: firstly, adding a sliding time window with the length of 1s to the electroencephalogram signal, calculating the sample entropy of the electroencephalogram signal, moving the window by one sampling point each time, calculating the sample entropy of the electroencephalogram signal of the next 1s time window until the sample entropy of the electroencephalogram signal of the last 1s time window is calculated, thereby obtaining the time sequence of the sample entropy of the group of electroencephalogram signals, and then superposing and averaging the group of sample entropy sequences to obtain the sample entropy of a group of sample electroencephalogram signal data;
4) stage of feature classification
Using a convolutional neural network CNN as a classifier to classify the two extracted types of feature vectors respectively, and then performing weighting processing on classification results of the two types of feature vectors respectively to finally obtain a classification result;
4.1CNN construction and principles thereof
The CNN used consists of six convolutional layers, two pooling layers and two full-link layers in the following sequence: the three-layer rolling layer, the one-layer pooling layer, the three-layer rolling layer, the one-layer pooling layer and the two full-connection layers are all realized by using Tensorflow;
dividing the data set into training set and test set, and extracting the two kinds of feature vectors
Figure FDA0003004020890000031
Where i ∈ [1, n ]]N represents the total time length; and SEi＝{i∈[1，n]I sampEn (m, r, i) } is used as input, three kinds of prediction possibility are available for output, T0, T1 and T2 respectively represent a rest state, a left imagination hand and a right imagination hand; an Adam optimizer is used, and the learning rate is 1 x 10^ -5;
A. and (3) rolling layers: the size of each convolution kernel of each convolution layer in the convolution neural network is 3 x 3, and the parameter of each convolution unit is obtained by optimization through a back propagation algorithm, so that different input features are extracted;
B. a pooling layer: obtaining a feature with the dimension of [28 × 20 × 64] after the convolution layer, cutting the feature into regions with the dimensions of [2,2], and taking the maximum value to obtain a new feature with the dimension of [14 × 10 × 64 ];
C. supervised learning
Calculating parameters of each convolution unit by using a back propagation algorithm;
calculating a loss function using Euclidean distances
E＝(x-pred)2
Wherein, x is the value of the actual feature vector, and pred is the value of the predicted feature vector;
propagating the loss function from the output layer back to the hidden layer until propagating to the input layer; in the back propagation process, adjusting the value of the parameter of each convolution unit according to a loss function; continuously iterating the above process until convergence;
4.2 weighted integration
In the last step, the results after the two types of feature vectors are classified adopt a method of respectively weighting 50% to sum to obtain a final prediction result; as can be seen from the above, the present invention,
Figure FDA0003004020890000041
and SEpredThe prediction result vectors are respectively two kinds of feature vectors in the form of
Figure FDA0003004020890000042
And
Figure FDA0003004020890000043
Figure FDA0003004020890000044
wherein predi(i＝T0，T1，T2) Is shown in this modelThe prediction probability of the group of electroencephalogram signals in the model as the ith class; the final prediction result vector is
Figure FDA0003004020890000045
Final prediction result is final.
CN202110376347.6A 2021-04-01 2021-04-01 Brain-computer interface software key technical method of cerebral apoplexy rehabilitation system based on deep learning Active CN113128384B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110376347.6A CN113128384B (en) 2021-04-01 2021-04-01 Brain-computer interface software key technical method of cerebral apoplexy rehabilitation system based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110376347.6A CN113128384B (en) 2021-04-01 2021-04-01 Brain-computer interface software key technical method of cerebral apoplexy rehabilitation system based on deep learning

Publications (2)

Publication Number Publication Date
CN113128384A true CN113128384A (en) 2021-07-16
CN113128384B CN113128384B (en) 2024-04-05

Family

ID=76775280

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110376347.6A Active CN113128384B (en) 2021-04-01 2021-04-01 Brain-computer interface software key technical method of cerebral apoplexy rehabilitation system based on deep learning

Country Status (1)

Country Link
CN (1) CN113128384B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114082169A (en) * 2021-11-22 2022-02-25 江苏科技大学 Disabled hand soft body rehabilitation robot motor imagery identification method based on electroencephalogram signals
CN114664434A (en) * 2022-03-28 2022-06-24 上海韶脑传感技术有限公司 Cerebral apoplexy rehabilitation training system for different medical institutions and training method thereof

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107958213A (en) * 2017-11-20 2018-04-24 北京工业大学 A kind of cospace pattern based on the medical treatment of brain-computer interface recovering aid and deep learning method
CN109620223A (en) * 2018-12-07 2019-04-16 北京工业大学 A kind of rehabilitation of stroke patients system brain-computer interface key technology method
CN112370066A (en) * 2020-09-30 2021-02-19 北京工业大学 Brain-computer interface method of stroke rehabilitation system based on generation of countermeasure network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107958213A (en) * 2017-11-20 2018-04-24 北京工业大学 A kind of cospace pattern based on the medical treatment of brain-computer interface recovering aid and deep learning method
CN109620223A (en) * 2018-12-07 2019-04-16 北京工业大学 A kind of rehabilitation of stroke patients system brain-computer interface key technology method
CN112370066A (en) * 2020-09-30 2021-02-19 北京工业大学 Brain-computer interface method of stroke rehabilitation system based on generation of countermeasure network

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114082169A (en) * 2021-11-22 2022-02-25 江苏科技大学 Disabled hand soft body rehabilitation robot motor imagery identification method based on electroencephalogram signals
CN114664434A (en) * 2022-03-28 2022-06-24 上海韶脑传感技术有限公司 Cerebral apoplexy rehabilitation training system for different medical institutions and training method thereof

Also Published As

Publication number Publication date
CN113128384B (en) 2024-04-05

Similar Documents

Publication Publication Date Title
CN111012336B (en) Parallel convolutional network motor imagery electroencephalogram classification method based on spatio-temporal feature fusion
Zhao et al. ECG authentication system design incorporating a convolutional neural network and generalized S-Transformation
CN110969108B (en) Limb action recognition method based on autonomic motor imagery electroencephalogram
CN110353702A (en) A kind of emotion identification method and system based on shallow-layer convolutional neural networks
CN110367967B (en) Portable lightweight human brain state detection method based on data fusion
CN114052735B (en) Deep field self-adaption-based electroencephalogram emotion recognition method and system
CN104771163A (en) Electroencephalogram feature extraction method based on CSP and R-CSP algorithms
CN114533086B (en) Motor imagery brain electrolysis code method based on airspace characteristic time-frequency transformation
CN113158964B (en) Sleep stage method based on residual error learning and multi-granularity feature fusion
Zhang et al. Human identification driven by deep CNN and transfer learning based on multiview feature representations of ECG
Deepa et al. Epileptic seizure detection using deep learning through min max scaler normalization
CN113128552A (en) Electroencephalogram emotion recognition method based on depth separable causal graph convolution network
CN113128384B (en) Brain-computer interface software key technical method of cerebral apoplexy rehabilitation system based on deep learning
CN111797674A (en) MI electroencephalogram signal identification method based on feature fusion and particle swarm optimization algorithm
Aly et al. Bio-signal based motion control system using deep learning models: a deep learning approach for motion classification using EEG and EMG signal fusion
CN116340824A (en) Electromyographic signal action recognition method based on convolutional neural network
CN113116361A (en) Sleep staging method based on single-lead electroencephalogram
CN113705398A (en) Music electroencephalogram space-time characteristic classification method based on convolution-long and short term memory network
CN115414051A (en) Emotion classification and recognition method of electroencephalogram signal self-adaptive window
CN115795346A (en) Classification and identification method of human electroencephalogram signals
CN115238796A (en) Motor imagery electroencephalogram signal classification method based on parallel DAMSCN-LSTM
CN113476056B (en) Motor imagery electroencephalogram signal classification method based on frequency domain graph convolution neural network
Samal et al. Ensemble median empirical mode decomposition for emotion recognition using EEG signal
CN112259228B (en) Depression screening method by dynamic attention network non-negative matrix factorization
Azami et al. Automatic signal segmentation based on singular spectrum analysis and imperialist competitive algorithm

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
GR01 Patent grant
GR01 Patent grant