CN110516711A - The training set method for evaluating quality of MI-BCI system and the optimization method of single training sample - Google Patents
The training set method for evaluating quality of MI-BCI system and the optimization method of single training sample Download PDFInfo
- Publication number
- CN110516711A CN110516711A CN201910677300.6A CN201910677300A CN110516711A CN 110516711 A CN110516711 A CN 110516711A CN 201910677300 A CN201910677300 A CN 201910677300A CN 110516711 A CN110516711 A CN 110516711A
- Authority
- CN
- China
- Prior art keywords
- training set
- single test
- acc
- quality
- 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.)
- Granted
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/369—Electroencephalography [EEG]
-
- 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
- A61B5/7221—Determining signal validity, reliability or quality
-
- 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
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Veterinary Medicine (AREA)
- Public Health (AREA)
- Animal Behavior & Ethology (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- Biomedical Technology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Surgery (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Psychiatry (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Signal Processing (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physiology (AREA)
- Psychology (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention discloses the optimum choice methods of a kind of training set method for evaluating quality of MI-BCI system and single training sample, acquire single test MI-EEG data;Design corresponding ICA spatial filter;Calculate single discrimination Acc (k) and overall discrimination Acc;Determine the quantizating index for being used for single test data quality accessment;Calculate the comprehensive score Cs of training sample quality;It is assessed according to quality of the numerical value of Cs to training set.Further according to gained Acc, quantizating index and comprehensive score Cs, the single test sample in training set is in optimized selection.The optimization method of training set method for evaluating quality and single training sample of the invention has the advantages that preferable stability, accuracy and computational complexity are low etc..
Description
Technical field
The present invention relates to a kind of processing method of data, especially a kind of training set quality evaluation for MI-BCI system
The optimization method of method and single training sample.
Background technique
Brain-computer interface (brain computer interface, BCI) is a kind of with brain electricity
(electroencephalography, EEG) is the novel man-machine interface of information carrier, is received in recent years extensive
Concern.The realization process of brain-computer interface technology is extracted in EEG and appoints by the way that multi-lead EEG signal is analyzed and handled
It is engaged in relevant feature mode and being converted into order, and then realizes that brain directly controls external equipment.The ultimate mesh of BCI technology
Mark is that the direct information channel of a control external equipment is provided for physical disabilities (or dyskinesia) group.Meanwhile BCI
Technology can also be applied to motion function rehabilitation training, brain-machine control game exploitation, spirit and psychological condition are assessed and special
Man-machine interaction etc. under environment.Mental imagery BCI (Motor Imagery BCI, MI-BCI) is that a kind of Endogenous Type brain-machine is logical
Implementation pattern is believed, the basic principle is that the event-related design that Mental imagery induces and the phenomenon (event-related that desynchronizes
Synchronization/desynchronization, ERS/ERD).The advantages of MI-BCI system is user not by outside
Stimulation, but external equipment is directly controlled by the realization of itself thinking activities.The design of MI-BCI system and training this two
During a, a large amount of Mental imagery EEG data is required.But in Mental imagery EEG (Motor Imagery EEG, MI-
EEG) in the collection process of data, subject's state of mind (such as fatigue, energy do not collect neutralization and lack experience) and it is random go out
Existing non-nervous activity artifacts, will affect the quality of data of MI-EEG.
According to existing research data, the automatic evaluation method research for the MI-EEG quality of data is had not been reported.In
During current MI-BCI system design and implementation, generallys use manual inspection mode and reject some EEG numbers by severe jamming
According to.But the low efficiency of manual type, error is big, does not have practical value.
Application No. is 201610066356.4 Chinese invention patent " a kind of EEG signals quality real-time judgment method ", packets
It includes following steps: obtaining EEG signals segment corresponding with the time window generated;Calculate the brain in actual time window
The average value of electric signal segment goes out the standard deviation of the EEG signals segment in actual time window using the mean value calculation;It is logical
Cross the fluctuation sex determination result for the EEG signals segment that calculated standard deviation obtains in actual time window;Utilize current brain
The fluctuation sex determination result of electric signal segment and the fluctuation sex determination of EEG signals segment before as a result, determine current jointly
The quality of EEG signals.The present invention quickly and efficiently can make real-time judgment analysis, effective guarantee to EEG signals quality
The accuracy of subsequent brain electricity applied analysis.But since this method uses the Time-domain Statistics feature (mean value and variance) of EEG signal
Its quality is assessed, and temporal signatures can only substantially describe EEG degree polluted by noise, and can not be to task correlation
The quality of EEG (such as MI-EEG) carries out thoroughly evaluating.Because the MI-EEG quality of data not only will receive the shadow of noise artefact
It rings, it is also related with the factors such as the state of mind of subject and experiment experience.Therefore this method is not suitable for the quality evaluation of MI-EEG.
Other research fields for making a general survey of processing of biomedical signals, in relation to electrocardio and pulse signal quality evaluation research report
It is relatively more.However, due to the biological signal datas such as electrocardio and pulse and EEG signal when-three characteristic of field of frequency-sky and generation
Mechanism etc. there were significant differences, therefore the signal quality evaluating methods such as existing electrocardio and pulse not can be used directly in
EEG signal.
Therefore, it is necessary to establish the Data Quality Assessment Methodology of effectively practical MI-EEG a kind of, to be MI-BCI system
Design provides the training sample of high quality.
Summary of the invention
The present invention be to avoid shortcoming present in above-mentioned prior art, provide it is a kind of have preferable stability,
The optimization method of the training set method for evaluating quality and single training sample of accuracy and the low MI-BCI system of computational complexity.
The present invention uses following technical scheme in order to solve the technical problem.
The training set method for evaluating quality of MI-BCI system comprising following steps:
Step 1: on-test, the single test data acquisition of Mental imagery EEG obtain the MI-EEG number of single test
According to;
Step 2: using the MI-EEG data of the single test, designing ICA spatial filter;
Step 3: single discrimination Acc (k) and overall discrimination Acc are calculated according to ICA spatial filter;
Step 4: determining the quantizating index of the quality evaluation of single test data;
Step 5: the comprehensive score Cs of training sample quality is calculated according to the quantizating index of the step 4;
Step 6: according to the numerical value of the comprehensive score Cs of the step 5, the quality of training set being assessed.
The characteristics of optimization method of single training sample of the invention, lies also in:
In the step 1, single test data acquisition are as follows: in each on-test, computer issues prompt tone,
Remind subject's single test that will start, computer display will appear the type prompts of arrow shaped Mental imagery after 1 second kind
Symbol, left/right direction arrow prompt subject carry out left/right hand Mental imagery, and down arrow then prompts double-legged Mental imagery;MI class
Type prompt arrow shows time T on the computer screenMIt is 5-6 seconds, then computer display blank screen 3-4 seconds, indicates single
Off-test;Complete single test time T is 10 seconds, the time interval of adjacent single test 2-3 seconds.
In the step 2, the design process of the ICA spatial filter are as follows: selection single test data carry out independent point
Amount analysis, obtains the source signal ingredient and ICA mixed model A and separation matrix W for constituting scalp EEG;Utilize ICA mixed model A
The isolated component spatial feature for being included determines from N number of output isolated component and moves relevant isolated component ingredient;Selection
The row vector of separation matrix W corresponding with the isolated component ingredient is denoted as: w respectively as MI spatial filterl, wr, wf, and protect
It deposits.
In the step 3, using MI spatial filter to training set D={ [xi,yi], i=1, L, I } in whole I it is single
Secondary test data carries out airspace filter, then carries out feature extraction and classification;And combine single test Mental imagery type and
The corresponding MI type label of single test calculates the discrimination Acc (k) of single test, and then obtains overall discrimination Acc.
In the step 4, the quantizating index includes Med, Max and Min;Wherein, Med=median (Acc), Max=
Max (Acc), Min=Min (Acc).
In the step 5, the calculation formula of comprehensive score Cs are as follows:
Wherein w1, w2, w3For weight coefficient.
The invention also discloses a kind of single training sample optimization methods according to above-mentioned appraisal procedure, including walk as follows
It is rapid:
Step 01: on-test, the single test data acquisition of Mental imagery EEG obtain single MI-EEG data;
Step 02: utilizing the single MI-EEG data, design ICA spatial filter;
Step 03: overall discrimination Acc is calculated according to ICA spatial filter;
Step 04: determining the quantizating index of the quality evaluation of single test data;
Step 05: the comprehensive score Cs of training sample quality is calculated according to the quantizating index of the step 4;
Step 06: according to the numerical value of the comprehensive score Cs of the step 5, the quality of training set being assessed;
Step 07: being commented according to gained Acc={ Acc (k), k=1 ..., I }, quantizating index { Med, Max, Min } and synthesis
Divide Cs, the single test sample in training set is in optimized selection.
In the step 07 of the optimum choice, according to Acc={ Acc (k), k=1 ..., I }, quantizating index Med, Max,
Min } and comprehensive score Cs optimization method include:
Step 07-1: the single test sample that Acc (k) is significantly lower than intermediate value Med is rejected from training set D;
Step 07-2: the single test sample architecture for selecting from training set D Acc (k) value to be apparently higher than intermediate value Med is new
Training set;
Step 07-3: if the comprehensive score C of training set DSValue is obvious relatively low, then it is integrally invalid can to mark this training set.
Compared with the prior art, the invention has the advantages that:
A kind of training set method for evaluating quality of MI-BCI system of the invention, includes the following steps:
Step 1: on-test, the single test data acquisition of Mental imagery EEG obtain single MI-EEG data;
Step 2: utilizing the single MI-EEG data, design ICA spatial filter;
Step 3: overall discrimination Acc is calculated according to ICA spatial filter;
Step 4: determining the quantizating index of the quality evaluation of single test data;
Step 5: the comprehensive score Cs of training sample quality is calculated according to the quantizating index of the step 4;
Step 6: according to the numerical value of the comprehensive score Cs of the step 5, the quality of training set being assessed.
The optimization method of single training sample of the invention is carried out on the basis of above-mentioned training set method for evaluating quality
's.Further include following steps 07 other than above-mentioned 6 steps: being referred to according to gained Acc={ Acc (k), k=1 ..., I }, quantization
{ Med, Max, Min } and comprehensive score Cs are marked, the single test sample in training set is in optimized selection.
The training set method for evaluating quality of MI-BCI system of the invention and the optimization method of single training sample, acquisition are single
Secondary test MI-EEG data;Design corresponding ICA spatial filter;Calculate single discrimination Acc (k) and overall discrimination Acc;
Determine the quantizating index for being used for single test data quality accessment;Calculate the comprehensive score Cs of training sample quality;According to Cs's
Numerical value assesses the quality of training set.It can determine description MI-EEG according to gained Acc, quantizating index and comprehensive score Cs
The quantizating index of the quality of data, and on this basis, the single test sample in training set is screened, is MI-BCI system
The design and training of key modules (such as spatial filter and classifier) provide good training sample, have preferable stablize
Property, accuracy and the technical effects such as computational complexity is low.
A kind of training set method for evaluating quality of MI-BCI system of the invention and the optimization method of single training sample, tool
User is aloowed to exchange, have a wide range of application by way of limbs voice non-verbal language, scalability is strong, uses
Comfortably, the advantages that interactivity is good, strong robustness.
Detailed description of the invention
Fig. 1 a is MI single test time normal form of the invention.
Fig. 1 b is the EEG distribution of electrodes location drawing of 26 leads of the invention.
Fig. 1 c is 26 lead single MI-EEG data of the invention.
Fig. 2 is the MI-EEG data quality accessment schematic diagram of the invention based on single test.
Fig. 3 a is data set D of the invention1Corresponding discrimination.
Fig. 3 b is data set D of the invention2Corresponding discrimination.
Fig. 3 c is data set D of the invention3Corresponding discrimination.
Fig. 3 d is data set D of the invention4Corresponding discrimination.
Fig. 4 a is data set D1In single MI-EEG x43Corresponding waveform diagram.
Fig. 4 b is data set D1In single MI-EEG x46Corresponding waveform diagram.
Fig. 5 a is data set D of the invention1Corresponding difference matrix YD(j) visual presentation (bianry image).
Fig. 5 b is data set D of the invention2Corresponding difference matrix YD(j) visual presentation (bianry image).
Fig. 5 c is data set D of the invention3Corresponding difference matrix YD(j) visual presentation (bianry image).
Fig. 5 d is data set D of the invention4Corresponding difference matrix YD(j) visual presentation (bianry image).
Below by way of specific embodiment, and in conjunction with attached drawing, the invention will be further described.
Specific embodiment
Referring to Fig. 1~5d, the training set method for evaluating quality of MI-BCI system, characterized in that include the following steps:
Step 1: on-test, the single test data acquisition of Mental imagery EEG obtain the MI-EEG number of single test
According to;
Step 2: using the MI-EEG data of the single test, designing ICA spatial filter;
Step 3: single discrimination Acc (k) and overall discrimination Acc are calculated according to ICA spatial filter;
Step 4: determining the quantizating index of the quality evaluation of single test data;
Step 5: the comprehensive score Cs of training sample quality is calculated according to the quantizating index of the step 4;
Step 6: according to the numerical value of the comprehensive score Cs of the step 5, the quality of training set being assessed.
In the step 1, single test data acquisition are as follows: in each on-test, computer issues prompt tone,
Remind subject's single test that will start, computer display will appear the type prompts of arrow shaped Mental imagery after 1 second kind
Symbol, left/right direction arrow prompt subject carry out left/right hand Mental imagery, and down arrow then prompts double-legged Mental imagery;MI class
Type prompt arrow shows time T on the computer screenMIt is 5-6 seconds, then computer display blank screen 3-4 seconds, indicates single
Off-test;Complete single test time T is 10 seconds, the time interval of adjacent single test 2-3 seconds.
MI-EEG training set is usually to be made of several single tests (a trial), the time of single test data acquisition
Normal form is as shown in Figure 1.In each on-test, computer issues prompt tone " Beep ", reminds subject's single test will
Start, after 1 second kind, computer display will appear the type prompts arrow shaped Mental imagery (MI) symbol, and left/right direction arrow mentions
Show that subject carries out left/right hand Mental imagery, down arrow then prompts double-legged Mental imagery.The type prompts MI arrow is in computer screen
Time (T is shown on curtainM) it is about 5-6 seconds, then computer display blank screen 3-4 seconds, indicate that single test terminates.Completely
Single test time T is about 10 seconds, the time interval of adjacent single test 2-3 seconds.It is adopted during single motion imagination test
The N lead EEG data of collection is referred to as single MI-EEG data.Training set D is made of I single MI-EEG data, it may be assumed that D=
{[xi,yi], i=1, L, I }.Wherein Xi=[x1,…,xN]TIt is the EEG signal x by N leadj, the MI- of j=1 ..., N composition
EEG signal matrix, size are N × L, L=T × FsIndicate single lead EEG sample length, FsFor the sample frequency of signal, T mono-
Single test total duration.yi∈ { ' l ', ' r ', ' f ' } it is the corresponding MI type label of single test, ' l ', ' r ' and ' f ' right respectively
Answer left hand, the right hand and foot Mental imagery.Fig. 1 (b) is the single MI-EEG data of one 26 leads.
In the step 2, the design process of the ICA spatial filter are as follows: selection single test data carry out independent point
Amount analysis, obtains the source signal ingredient and ICA mixed model A and separation matrix W for constituting scalp EEG;Utilize ICA mixed model A
The isolated component spatial feature for being included determines from N number of output isolated component and moves relevant isolated component ingredient;Selection
The row vector of separation matrix W corresponding with the isolated component ingredient is denoted as: w respectively as MI spatial filterl, wr, wf, and protect
It deposits.
As shown in Fig. 2, utilizing training set D={ [xi,yi], i=1, L, I } in single MI-EEG data x design ICA it is empty
Then domain filter tests its classifying quality (discrimination) for being applied to training set D.It is designed in view of ICA performance of filter with it
The quality of sample x is closely related, and therefore, gained discrimination can be used for the quality evaluation to single sample x.The specific implementation of method
Process is as shown in Figure 2.
Variable k=1 is set, from the one single test data x ∈ { x of selection in training set Di, i=1 ..., I } and it carries out independently
Component Analysis (independent component analysis, ICA) obtains the source signal ingredient s=for constituting scalp EEG
[s1,…,sN]TAnd ICA mixed model A and separation matrix W=A-1.ICA mixed model A, separation matrix W, selection one single examination
The relationship between the source signal ingredient s of data x, EEG is tested as shown in (1) and (2) formula, respectively indicates the mixed model for more leading EEG
And disjunctive model, aj=[a1.j,…,aN,j]TIt is arranged for the jth of hybrid matrix A, wj=[wj,1,…,wj,N]TFor separation matrix W's
Jth row.
Isolated component (IC) spatial feature for being included using ICA mixed model A, from N number of output isolated component s=
[s1,…,sN]TIn (i.e. the source signal ingredient s) of EEG, it is determining to move relevant isolated component ingredient (with right-hand man and foot three classes
For MI), it is denoted as sl, sr, sf, s thereinl、sr、sfRespectively indicate the isolated component ingredient of left hand, the right hand and foot.Selection with
sl、sr、sfCorresponding separation matrix row vector is denoted as: w respectively as MI spatial filterl、wr、wf, and save, wl、wr、wfPoint
It Biao Shi not be with sl、sr、sfCorresponding separation matrix row vector.
It is currently, there are the ICA algorithm of diversified forms, such as Infomax, FastICA, Sobi and Jade etc..The present invention is set
A kind of information maximum ICA algorithm of simplification is counted, as shown in table 1 below, the initial value of algorithm is set as W=eye/100, learning rate
Lrate=0.02, the number of iterations Num=300;Bandpass filtering range 7-40Hz.Based on the analysis and survey to a large amount of measured datas
It tries, ICA algorithm designed by the present invention, which compares classical ICA algorithm, has better stability and applicability.For different tested
The setting of the MI-EEG data set of person and different time acquisition, aforementioned initialization parameter has preferable versatility.
In the step 3, using MI spatial filter to training set D={ [xi,yi], i=1, L, I } in whole I it is single
Secondary test data carries out airspace filter, then carries out feature extraction and classification;And combine single test Mental imagery type and
The corresponding MI type label of single test calculates the discrimination Acc (k) of single test, and then obtains overall discrimination Acc.
Utilize gained wl, wr, wf, to training set D={ [xi,yi], i=1, L, I } in whole I single test data into
Row airspace filter, then carries out feature extraction and classification, and I is the total degree of single test.Wherein, feature sl, sr, sfIn MI
Execute the variance of period (referring to Fig. 1 (a)).Based on the ERD phenomenon of movement correlation mu rhythm and pace of moving things ingredient, by comparing the big of variance
It is small, the Mental imagery type of single test can be estimated:For the valuation of MI type label;Combined training
The true tag value y provided is providedi, i=1, L, I calculate discrimination Acc (k) and save, simultaneously willI=1, L, I, which are stored in, to be estimated
Count label matrixRow k, it may be assumed that
K ← k+1 repetition step 1,2.That is, next single test data is selected to carry out ICA airspace filter from training set D
Device design and its corresponding discrimination test, until k=N, saves full income discrimination Acc={ Acc (k), k=1 ..., I }
With estimation label matrix
In the step 4, the quantizating index includes Med, Max and Min;Wherein, Med=median (Acc), Max=
Max (Acc), Min=Min (Acc).
In the step 5, the calculation formula of comprehensive score Cs are as follows:
Wherein w1, w2, w3For weight coefficient.
According to Acc, quality evaluation is carried out to I single test data in MI-EEG training set.Table 2 is that the present invention proposes
Every quantizating index, they respectively from different angles reflect training sample quality.Formula (5) is to training sample quality
Comprehensive score, value range: 0-100.
Quality evaluation quantizating index of the table 2 based on ACC
Index name | Discrimination intermediate value | Maximum discrimination score | Minimum discrimination |
Calculation formula | Med=median (Acc) | Max=max (Acc) | Min=min (Acc) |
Ideal value | 1 | 1 | 1 |
In formula (5), weight coefficient value is required to meet: w1+w2+w3=3, and w1>=1,0<w2≤ 1,0 < w3≤1;Typical case takes
Value;w1=2, w2=0.5, w3=0.5.It is proved by multiple test, value w1=2, w2=0.5, w3When=0.5,
The accuracy highest of assessment.
A kind of single training sample optimization method according to above-mentioned appraisal procedure, includes the following steps:
Step 01: on-test, the single test data acquisition of Mental imagery EEG obtain single MI-EEG data;
Step 02: utilizing the single MI-EEG data, design ICA spatial filter;
Step 03: overall discrimination Acc is calculated according to ICA spatial filter;
Step 04: determining the quantizating index of the quality evaluation of single test data;
Step 05: the comprehensive score Cs of training sample quality is calculated according to the quantizating index of the step 4;
Step 06: according to the numerical value of the comprehensive score Cs of the step 5, the quality of training set being assessed;
Step 07: being commented according to gained Acc={ Acc (k), k=1 ..., I }, quantizating index { Med, Max, Min } and synthesis
Divide Cs, the single test sample in training set is in optimized selection.
The step of optimum choice includes:
Step 07-1: the single test sample that Acc (k) is significantly lower than intermediate value Med is rejected from training set D;
Step 07-2: the single test sample architecture for selecting from training set D Acc (k) value to be apparently higher than intermediate value Med is new
Training set;
Step 07-3: if the comprehensive score C of training set DSValue is obvious relatively low, then it is integrally invalid can to mark this training set.
The factor for influencing the MI-EEG quality of data includes following 2 kinds.
1, noise artifacts.
The mu/beta rhythm and pace of moving things ingredient that Mental imagery induces usually is submerged in spontaneous EEG.In addition, non-neururgic
Electro physiology artefact (eye electricity, myoelectricity and electrocardio etc.), environment electromagnetics interference and connecting fault between electrode and scalp etc., can
Biggish negative effect is generated to the MI-EEG quality of data.
2, invalid single MI-EEG training sample.
The collection process typical time of MI-EEG is longer, therefore asking easily occurs mental fatigue or energy and do not collect medium in subject
Topic, this will lead to subject when carrying out single test, cannot be accurately finished the Mental imagery task of specified type, such feelings
MI-EEG data are collected under condition belongs to invalid training sample.
In order to improve the signal-to-noise ratio of multichannel EEG, airspace filter technology is played in the realization of MI-BCI system and is extremely closed
The effect of key.Currently used airspace filter technology have ICA and common space mode (common spatial pattern,
CSP).ICA and CSP is both needed to be designed using training sample, and the quality of performance and training sample is closely related.ICA is not
The unsupervised design method of MI type label is relied on, theoretically any EEG data section is used equally for setting for ICA spatial filter
Meter, therefore the artifacts in data are the main reason for influencing ICA performance of filter.CSP is then a kind of dependence MI label letter
Breath has supervision airspace filter design method, therefore the accuracy pair of the artifacts and label information in MI-EEG data
There is influence in the performance of CSP filter.
Since ICA and CSP airspace filter technology have different demands to training sample quality, while in view of ICA/CSP is empty
Domain filtering technique itself has the ability of certain anti-artifacts, and therefore, the Time-domain Statistics for only relying only on MI-EEG signal are special
Sign is unable to judge accurately the quality of data, although for example, eye movement artefact amplitude is very big, and in the numerous appearance of MI-EEG signal intermediate frequency,
After conventional time domain and airspace filter pretreatment, its negative effect to BCI system performance can be effectively relieved, therefore not
Can simply by the MI-EEG data segment containing eye movement artifacts it is qualitative be low quality sample, and rejected.On the contrary, because by
Examination person's fatigue or energy invalid single test data, time domain waveform and feature and Non Apparent Abnormality caused by not concentrating, but such as
This kind of single training sample is used for the design of CSP filter by fruit, then may result in CSP airspace filter performance it is obvious under
Drop.
Based on above-mentioned analysis, analysis of time-domain characteristic method is relied solely on, it is difficult to all-sidedly and accurately assess MI-EEG data matter
Amount.For this purpose, the invention proposes a kind of new MI-EEG Data Quality Assessment Methodologies: utilizing training set D={ [xi,yi], i=1,
L, I } in single MI-EEG data x design ICA spatial filter, then test its applied to training set D classifying quality (know
Not rate).Closely related in view of ICA performance of filter and the quality of its design sample x, therefore, gained discrimination can be used for list
The quality evaluation of secondary sample x.The specific implementation process of method is as shown in Figure 2.
Table 2 is every quantizating index proposed by the present invention.According to gained Acc={ Acc (k), k=1 ..., I } and its corresponding
Quantizating index { Med, Max, Min } and comprehensive score Cs, the single test sample in training set can be in optimized selection.
Selection strategy has following several:
1) the single test sample that Acc (k) is significantly lower than intermediate value Med is rejected from training set D.
2) Acc (k) value is selected to be apparently higher than the new training set of single test sample architecture of intermediate value Med from training set D.
If 3) the comprehensive score C of training set DSValue is obvious relatively low, then it is integrally invalid can to mark this training set.
If the design of MI-BCI system uses CSP filter, need further to judge having for selected single test sample
Effect property, that is, judge single test data xiInitial form label value yiWith valuationIt is whether consistent.For this purpose, the present invention passes through ratio
Compared with label matrixEvery a lineWith original tag [y1,L,yk] otherness, construct recognition matrix YD。
YD={ dk,i, k=1, L, I;I=1, L, I } (6)
In formula (6):
The evaluation process of the quality of data according to fig. 2, in order to detect the degree that single test data are affected by artifacts, instruction
Practice each single test data concentrated all to be respectively used to design different ICA filters, and is applied in training set whole I
The type of sports of single test identifies.Each list after completing step and walking 1 whole test process to step 3, in training set
Secondary test is checked out I times, and gained type identification result is placed on YDEach column in.YDA line be then a certain ICA filtering
Device is applied to the classification results of entire training set.According to the pass of the factors such as aforementioned artifacts and the state of mind and the quality of data
System, YDMatrix column vector can be used for analyzing and judging the reliability of MI label.Row vector then reflects single test data by puppet
The degree of mark interference.Therefore, it is based on recognition matrix YD, can also be by Y other than it can obtain several quantization quality indexD's
Visualization processing, intuitively shows the quality information of each single test data, provides reference for the reasonable selection of training sample.
For MI-EEG Data Quality Assessment Methodology shown in Fig. 2, Fig. 3 a, Fig. 3 b, Fig. 3 c, Fig. 3 d are given to four MI-
EEG training set D1, D2, D3, D4Quality test results: Acc={ Acc (k), k=1 ..., I } and each quantizating index.
Training set D1And D2Contain I=75 three classes MI single test (right-hand man and foot Mental imagery), every class single test
Sample size is 25.Training set D3And D4Contain I=150 three classes MI single test (right-hand man and foot Mental imagery), every class list
Secondary test sample quantity is 50.
Result can see according to Fig.3, training set D1Overall quality index CsValue only 65.8, and contain in Acc
There is the discrimination significantly lower than intermediate value, shows that corresponding single MI-EEG data receive more serious interference.In addition, corresponding to
The maximum discrimination score of single test sample also only has 78.7%.Therefore, if by D1For spatial filter design and classifier
Training, then MI-BCI system performance is difficult to ensure.In comparison, D3And D4Overall quality index more than 90, highest identification
Rate reaches 99%, shows that the quality of this two group data set is very high.After rejecting a small amount of lower single test sample of discrimination,
D3, D4It can be applied to the MI-BCI system design based on ICA or CSP.D2Indices it is moderate, the corresponding knowledge of each single test
Rate gap is not little, and more outstanding is that extremely low discrimination is less in Acc, shows that each single test quality of data is relatively uniform.
Similar D2The training set of performance indicator is suitble to the design of ICA airspace filter.
In discrimination result shown in Fig. 3, the corresponding single MI-EEG data of anomalous identification rate usually contain serious
Artifacts, for example, Acc (43)=34.7% and Acc (46)=33.3% are significantly lower than discrimination intermediate value in Fig. 3 a
70.7%.Their corresponding two single MI-EEG (x43And x46) waveform is as shown in figures 4 a and 4b, it can be seen that x43Whole
There is bursty interference in lead, and x46C3Then there is electrode connecting fault in lead.Therefore, using x43And x46Design
The performance of two groups of ICA spatial filters receives serious influence, and then leads to the appearance of anomaly classification result.
Fig. 5 is four training set (D1, D2, D3, D4) difference matrix { YD(j), visual presentation (two j=1 ..., 4 }
It is worth image), gray scale " 0 "/" 1 " (black/white) respectively indicates original tag and identification label consistency/inconsistency.According to front to knowledge
Other matrix YDDefinition and its characteristic introduction, can the quality to four groups of training sets more fully analyzed.On the whole,
Training set D1Quality it is worst because containing most " 1 " in its recognition matrix image, wherein 43 and 46 liang of rows are first containing " 1 "
It is plain most, show the 43rd and 46 single test data (x43And x46) receive more serious interference.In addition Y is observedD(1) image
Column, it has also been discovered that, the initial form label and estimated result of more single test are inconsistent, with similar characteristics training
Collection generally can not be directly used in the design of ICA/CSP spatial filter.In comparison, training set D2Quality be better than D1.Pass through
Observe the row and column of image, it is possible to find D2In without by the single test data of severe jamming, but have several (10 or so) it is former
Beginning label and the possible inconsistent single test of estimation label.It is similarly not difficult to obtain, D3And D4The quality of data to be substantially better than D1
And D2。
In MI-BCI system realization, quality and quantity to training sample have higher requirements, but since MI is instructed
The acquisition experiment for practicing sample is limited to factors (for example, test environment, subject's experience and the state of mind etc.), therefore training
The quality of sample is often difficult to ensure, it is also relatively fewer to test the sample size that can be acquired every time.In the MI-BCI to have registered
It in research, generallys use manual type and checks MI-EEG data waveform, to reject the low quality data for containing obvious artifacts
Section, but not only low efficiency, error are big for manual type, but also are difficult to detect by because the factors such as subject's state of mind and experience are led
The engineering noise training sample of cause.
Solution proposed by the invention can realize the quality evaluation to single test sample each in training set automatically, and
Give the every quantizating index and TOP SCORES of reflection training set quality.Operand involved in mentioned method is very low, therefore
With good practicability.Characteristic of the invention is converted to " the direct evaluation problem of training sample quality " to " based on single
The Performance Evaluation problem of the ICA filter of secondary experimental design ", this thinking can comprehensively consider influence the quality of data it is different because
Element, gained quality assessment result have preferable accuracy and comprehensive.
It is obvious to a person skilled in the art that invention is not limited to the details of the above exemplary embodiments, Er Qie
In the case where without departing substantially from spirit or essential attributes of the invention, the present invention can be realized in other specific forms.Therefore, no matter
From the point of view of which point, the present embodiments are to be considered as illustrative and not restrictive, and the scope of the present invention is by appended power
Benefit requires rather than above description limits, it is intended that all by what is fallen within the meaning and scope of the equivalent elements of the claims
Variation is included within the present invention.Any reference signs in the claims should not be construed as limiting the involved claims.
In addition, it should be understood that although this specification is described in terms of embodiments, but not each embodiment is only wrapped
Containing an independent technical solution, this description of the specification is merely for the sake of clarity, and those skilled in the art should
It considers the specification as a whole, the technical solutions in the various embodiments may also be suitably combined, forms those skilled in the art
The other embodiments being understood that.
Claims (8)
- The training set method for evaluating quality of 1.MI-BCI system, characterized in that include the following steps:Step 1: on-test, the single test data acquisition of Mental imagery EEG obtain the MI-EEG data of single test;Step 2: using the MI-EEG data of the single test, designing ICA spatial filter;Step 3: single discrimination Acc (k) and overall discrimination Acc are calculated according to ICA spatial filter;Step 4: determining the quantizating index of the quality evaluation of single test data;Step 5: the comprehensive score Cs of training sample quality is calculated according to the quantizating index of the step 4;Step 6: according to the numerical value of the comprehensive score Cs of the step 5, the quality of training set being assessed.
- 2. the training set method for evaluating quality of MI-BCI system according to claim 1, characterized in that in the step 1, Single test data acquisition are as follows: computer issues prompt tone in each on-test, and prompting subject's single test is It will start, computer display will appear the type prompts symbol of arrow shaped Mental imagery after 1 second kind, and left/right direction arrow mentions Show that subject carries out left/right hand Mental imagery, down arrow then prompts double-legged Mental imagery;The type prompts MI arrow is in computer screen Time T is shown on curtainMIt is 5-6 seconds, then computer display blank screen 3-4 seconds, indicates that single test terminates;Complete single examination Testing time T is 10 seconds, the time interval of adjacent single test 2-3 seconds.
- 3. the training set method for evaluating quality of MI-BCI system according to claim 1, characterized in that in the step 2, The design process of the ICA spatial filter are as follows: selection single test data carry out independent component analysis, obtain and constitute scalp The source signal ingredient and ICA mixed model A and separation matrix W of EEG;The isolated component airspace for being included using ICA mixed model A Feature determines from N number of output isolated component and moves relevant isolated component ingredient;Selection and the isolated component ingredient pair It answers the row vector of separation matrix W as MI spatial filter, is denoted as respectively: wl, wr, wf, and save.
- 4. the training set method for evaluating quality of MI-BCI system according to claim 3, characterized in that in the step 3, Using MI spatial filter to training set D={ [xi,yi], i=1, L, I } in whole I single test data carry out airspace filter Then wave carries out feature extraction and classification;And combine the Mental imagery type and the corresponding MI type of single test of single test Label calculates the discrimination Acc (k) of single test, and then obtains overall discrimination Acc.
- 5. the training set method for evaluating quality of MI-BCI system according to claim 4, characterized in that in the step 4, The quantizating index includes Med, Max and Min;Wherein, Med=median (Acc), Max=Max (Acc), Min=Min (Acc)。
- 6. the training set method for evaluating quality of MI-BCI system according to claim 5, characterized in that in the step 5, The calculation formula of comprehensive score Cs are as follows:Wherein w1, w2, w3For weight coefficient.
- 7. a kind of single training sample optimization method of -6 appraisal procedure according to claim 1, characterized in that including walking as follows It is rapid:Step 01: on-test, the single test data acquisition of Mental imagery EEG obtain single MI-EEG data;Step 02: utilizing the single MI-EEG data, design ICA spatial filter;Step 03: overall discrimination Acc is calculated according to ICA spatial filter;Step 04: determining the quantizating index of the quality evaluation of single test data;Step 05: the comprehensive score Cs of training sample quality is calculated according to the quantizating index of the step 4;Step 06: according to the numerical value of the comprehensive score Cs of the step 5, the quality of training set being assessed;Step 07: according to gained Acc={ Acc (k), k=1 ..., I }, quantizating index { Med, Max, Min } and comprehensive score Cs, Single test sample in training set is in optimized selection.
- 8. the optimization method of single training sample according to claim 7, characterized in that the step of optimum choice wraps It includes:Step 07-1: the single test sample that Acc (k) is significantly lower than intermediate value Med is rejected from training set D;Step 07-2: Acc (k) value is selected to be apparently higher than the new training of single test sample architecture of intermediate value Med from training set D Collection;Step 07-3: if the comprehensive score C of training set DSValue is obvious relatively low, then it is integrally invalid can to mark this training set.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910677300.6A CN110516711B (en) | 2019-07-25 | 2019-07-25 | Training set quality evaluation method of MI-BCI system and optimization method of single training sample |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910677300.6A CN110516711B (en) | 2019-07-25 | 2019-07-25 | Training set quality evaluation method of MI-BCI system and optimization method of single training sample |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110516711A true CN110516711A (en) | 2019-11-29 |
CN110516711B CN110516711B (en) | 2022-10-21 |
Family
ID=68623576
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910677300.6A Active CN110516711B (en) | 2019-07-25 | 2019-07-25 | Training set quality evaluation method of MI-BCI system and optimization method of single training sample |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110516711B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116226481A (en) * | 2022-12-30 | 2023-06-06 | 北京视友科技有限责任公司 | Electroencephalogram-based experimental data screening method, system and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2012153965A2 (en) * | 2011-05-09 | 2012-11-15 | 광주과학기술원 | Brain-computer interface device and classification method therefor |
CN108710895A (en) * | 2018-04-23 | 2018-10-26 | 安徽大学 | Mental imagery Method of EEG signals classification based on independent component analysis |
CN109144266A (en) * | 2018-08-29 | 2019-01-04 | 安徽大学 | A kind of brain-computer interface lead optimization method based on independent component analysis |
-
2019
- 2019-07-25 CN CN201910677300.6A patent/CN110516711B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2012153965A2 (en) * | 2011-05-09 | 2012-11-15 | 광주과학기술원 | Brain-computer interface device and classification method therefor |
CN108710895A (en) * | 2018-04-23 | 2018-10-26 | 安徽大学 | Mental imagery Method of EEG signals classification based on independent component analysis |
CN109144266A (en) * | 2018-08-29 | 2019-01-04 | 安徽大学 | A kind of brain-computer interface lead optimization method based on independent component analysis |
Non-Patent Citations (1)
Title |
---|
刘锦等: "单次样本对的CSP滤波器设计及其在脑电训练样本优化中的应用", 《信号处理》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116226481A (en) * | 2022-12-30 | 2023-06-06 | 北京视友科技有限责任公司 | Electroencephalogram-based experimental data screening method, system and storage medium |
CN116226481B (en) * | 2022-12-30 | 2023-11-21 | 北京视友科技有限责任公司 | Electroencephalogram-based experimental data screening method, system and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN110516711B (en) | 2022-10-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110765920B (en) | Motor imagery classification method based on convolutional neural network | |
CN106886792B (en) | Electroencephalogram emotion recognition method for constructing multi-classifier fusion model based on layering mechanism | |
CN112137627B (en) | Intelligent human factor evaluation and training method and system | |
CN110309797A (en) | Merge the Mental imagery recognition methods and system of CNN-BiLSTM model and probability cooperation | |
CN109770924A (en) | A kind of tired classification method based on Hadamard product building brain function network and Method Using Relevance Vector Machine | |
KR101247748B1 (en) | Apparatus for analysing focus and nonfocus states and method therof | |
CN102200833B (en) | Speller brain-computer interface (SCI) system and control method thereof | |
CN110584660B (en) | Electrode selection method based on brain source imaging and correlation analysis | |
Craston et al. | The attentional blink reveals serial working memory encoding: Evidence from virtual and human event-related potentials | |
CN108960182A (en) | A kind of P300 event related potential classifying identification method based on deep learning | |
JPS63226340A (en) | Method and apparatus for displaying timewise relation between position and internal area of brain nerve activity | |
CN109255309B (en) | Electroencephalogram and eye movement fusion method and device for remote sensing image target detection | |
CN110192874B (en) | Lie detection method based on multi-lead electroencephalogram signal margin factors | |
CN108959895A (en) | A kind of EEG signals EEG personal identification method based on convolutional neural networks | |
CN106166065A (en) | A kind of wearable electrocardio health interacting platform based on social networks and its implementation | |
CN106175757A (en) | Behaviour decision making prognoses system based on brain wave | |
CN101382837B (en) | Computer mouse control device of compound motion mode | |
CN113349780A (en) | Method for evaluating influence of emotional design on online learning cognitive load | |
CN105700687A (en) | Single-trial electroencephalogram P300 component detection method based on folding HDCA algorithm | |
CN106073767B (en) | Phase synchronization measurement, coupling feature extraction and the signal recognition method of EEG signal | |
CN113397482B (en) | Human behavior analysis method and system | |
CN111714089B (en) | Drug addiction evaluation system based on multi-stimulus short video event related potential | |
CN110516711A (en) | The training set method for evaluating quality of MI-BCI system and the optimization method of single training sample | |
CN108491792A (en) | Office scene human-computer interaction Activity recognition method based on electro-ocular signal | |
CN107169434A (en) | One kind possesses the electric personal identification method of exclusive brain |
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 |