CN1261193A - Training method of basic artificial nerve network and device for automatical training of artificial nerve network - Google Patents
Training method of basic artificial nerve network and device for automatical training of artificial nerve network Download PDFInfo
- Publication number
- CN1261193A CN1261193A CN 99124209 CN99124209A CN1261193A CN 1261193 A CN1261193 A CN 1261193A CN 99124209 CN99124209 CN 99124209 CN 99124209 A CN99124209 A CN 99124209A CN 1261193 A CN1261193 A CN 1261193A
- Authority
- CN
- China
- Prior art keywords
- network
- node
- output
- neural network
- artificial neural
- 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
Images
Abstract
A training method of basic artificial nerve network and a device for automatically training basic artificial nerve network are disclosed for full-automatically quantatative detection of epileptic discharge in electroencephalogram. The data acquisition device and computer are connected in bidirectional mode. The input latch area and parameter latch area of artificial nerve nertwork BP are created in the memory of computer. It particularly includes basic nerve network, patient's ED mode selection program module and special training program module.
Description
The present invention relates to a kind of method and apparatus that detects the epilepsy state of an illness, specifically, is special-purpose Artificial Neural Network model of a kind of full-automatic training patient and device.
As everyone knows, epilepsy is a kind of common disease, and morbidity rate accounts for 0.4% of population, and patient's morbidity is lost consciousness often and fallen down to the ground, and signs such as whole body tic are the important diseases of harm people ' s health.Whether at first to correctly diagnose epilepsy, making it has the disease of similar symptom to make a distinction with other, take correct treatment measure, method commonly used at present is the electroencephalogram of record patient, analyze electroencephalogram by the doctor again, see whether epileptiform discharge (Epileptiform Discharges is wherein arranged, be called for short ED), main objective basis as the epilepsy diagnosis, about this respect, the applicant submits an application for a patent for invention (application number CN 98122894.1) that is called " automatically judging the device that has or not epileptiform discharge in the electroencephalogram " on Dec 30th, 1998 to Patent Office of the People's Republic of China, mainly solves the qualitative judgement of " having " and " not having " ED in this part application.This is crucial for taking correct treatment route.For the patient that ED is arranged, be judged as epileptic patient as comprehensive other illnesss, then the doctor just needs to understand patient's psychomotor concrete condition, the frequent degree of showing effect how, the quantity of ED and when, outbreak or the like in what situations, with decision how much etc. with what medicine, dosage, after one stage of treatment, the effect that the doctor need know treatment how, effective, invalid, effective bad or the like, consider whether continue to carry on the treatment with similar medicine, whether dose will increase and decrease, and still uses another kind of medicine instead.The state of an illness of epileptics can mainly be seen on the quantity of ED distributed with its time.Unless patient's outbreak is very frequent, then in hospital, also can see ED in the conventional EEG(electroencephalography), generally need make dynamic eeg recording, be exactly patient's brain of 24 hours of continuous recording, it is very time-consuming, effort coming the manual analysis Active electroencephalogram (EEG) by the doctor, because the data volume that 24 hour records obtain is very huge,, come interpretation by the doctor again if show page by page, each part Active electroencephalogram (EEG) just need cost doctor several hrs the time ask, but also can not get quantitative results.It is a qualitative impression, if do quantitative measurment, the doctor will do a lot of measurements and statistical work more so, also to plot chart to the result, such time perhaps the doctor can only handle a patient's dynamic eeg recording, the brain electricity analytical doctor who has wide experience is also few in the big city at present, the nature small and medium-sized cities then still less, even do not have, and vast rural area, frontier area, do not have this respect professional especially, so epileptic patient can not obtain effectively, treatment targetedly.
Press for the robotization of computing machine for the interpretation of dynamic brain electricity, this work international academic community has passed through the effort in more than 20 year, effect is obviously improved, especially in recent years use " artificial neural network " to detect ED, especially this work is pushed to a new height, the discrimination of ED obviously improves, false positive rate descends greatly, up to the present wait at " Chinese biological engineering in medicine journal " 1998 with Zhang Tong, 17 (1): the technical scheme level of " the automatic detection and the classification of electroencephalogram epilepsy wave " that 1-11 delivers is the highest, they adopt by different level, multi-method integration approach is with automatic adaptive prediction, wavelet transformation, artificial neural network, fuzzy recognition system, signal processing technologies such as expert system combine, detect epilepsy wave and obtained good result, can also carry out somatotype epilepsy wave.But this method is quite complicated, needs with high speed and high capacity computing machine, and cost is higher, also is unfavorable for quick identification.This paper apart from practical, particularly go to apply to vast small city, border areas very big distance still arranged, most importantly, apply as if the concrete grammar of this article gained, parameter, network connection weight etc. are fixed up.It is the effect that can not obtain.Its reason is because the zone that ED takes place in the epileptics human brain is different, has plenty of diffuse type, has plenty of focal type, what have occurs in this position, what have then occurs in another position, the waveform of ED is also different, spike (ripple is wide about 70ms) is arranged, sharp wave is arranged, and (ripple is wide 100～200ms), and spike and slow wave complex or the like is arranged.If all patients are detected with a fixing artificial neural network, then the result of Jian Ceing naturally can be not fine to all patients, for the effect that will obtain, answer " concrete condition concrete analysis ", at each patient, the ED ripple of selecting this patient comes the specialized training network as the ED pattern, obtains the neural network that its parameters is particularly suitable for this patient, measure this patient again, effect will be improved naturally greatly like this.But who selects this patient's ED ripple as the ED pattern? if " automatically ", this task should be born by computing machine so.Can train the device of the special-purpose artificial neural network of patient not appear in the newspapers so far automatically, trace it to its cause be because:
One, the special-purpose artificial neural network that will can train this patient is at first wanted and can be picked out the ED ripple in this patient's electroencephalogram, and this task will be finished automatically by computing machine.
Two, want to pick out varied representational ED, train the network that comes out to varied this patient's ED ripple good identification to be arranged like this.Even same patient, the ED of he (she) also has significant change, this is because of the variation with patient condition, the road number that the waveform of ED on the multichannel brain electric amplifier own occurs etc. may change to some extent, because spontaneous brain electricity is superimposed upon on the ED, the ED waveform is changed especially, therefore require to pick out representational multiple ED pattern, this representative being difficult to is grasped, and generally discusses solution by experienced brain electricity doctor and veteran engineers.
Three, allow computing machine select the ED pattern, be difficult to avoid finding fault with people.A non-ED selected becomes the ED pattern, is used for the training of human artificial neural networks, and what kind of result can this cause? if bad result should manage to alleviate or avoid.
In sum, because above-mentioned all difficulties, this class " does not have the doctor to participate in, trains the device of the special-purpose artificial neural network of patient fully by computing machine automatically " also not see up to now report.
The device that the purpose of this invention is to provide the special-purpose artificial neural network of full-automatic training patient of a kind of basic artificial neural network training method and release doctor participation.
The present invention is achieved in that basic artificial neural network training method of the present invention, and its step comprises:
(1) lid of setting up in a computing machine is earlier as shown in Figure 3 won artificial neural network topological structure and rich non-ED algorithm and the BP algorithm of resident lid, it is characterized in that:
(2) resident in this computing machine
1. determine the algorithm of peak width by the average peak width among one group of ED of a patient;
2. by the ED pattern depart from M times of standard deviation calculate non-ED pattern algorithm and
3. depart from and adopt the algorithm that absolute value is diminished;
(3) be that the patient's of diffuse type spike and slow wave complex ED trains this neural network with arbitrary ED, generate each parameter value of its connection weight and amount of bias, form a basic artificial network., more particularly, above-mentioned said M=10, and said one group of ED is meant 15 ED fragments choosing from 16 road EEG(electroencephalography).
The special-purpose artificial neuron device of the full-automatic training patient that said method according to the present invention is made, it comprises into data acquisition unit and computing machine that two-way circuit connects, the input end of this data acquisition unit is connected with a multichannel eeg amplifier output terminal, this computing machine then is connected with printer, in the internal storage of this computing machine, set up the input working area of artificial neural network, the parameter working storage, and resident BP (feedforward contrary propagate learning algorithm) algorithm routine arranged, each road EEG signals has corresponding input node in this neural network, this artificial neural network is made up of two three layers of BP sub-networks, be called left sub-network and right sub-network, the output of left side sub-network and right sub-network is after weighted mean, form network output O, a left side, right sub-network respectively has three layers: input layer, hidden layer and output layer, input layer data-storing district is set respectively in the internal memory of this computing machine, hidden layer data-storing district and output layer data storage area are for a BP sub-network, W
JiI node of expression input layer is to the connection weight of j node of hidden layer, W
LjJ node of expression hidden layer is to the connection weight of l node of output layer, and the node of hidden layer and output layer contains bias; θ
jThe bias of j node of expression hidden layer, θ
lThe bias of l node of expression output layer; The input and output excitation function of hidden layer node and output layer node adopts " S " type function:
f(X)＝1/(1+exp(-X))；
An input layer has ni node, and hidden layer has nj node, and output layer has the sub-network of nl node, and Pi represents the input value of i node of sub-network, H
INjThe input value of j node of expression hidden layer, H
OUTjThe activation value of j node of expression hidden layer, O
INlThe input value of l node of expression output layer, O
OUTlThe output valve of l node of expression output layer, its feedforward is calculated by following feedforward formula;
H
OUTj＝1/(1+exp(-H
INj))
O
OUTl＝1/(1+exp(-O
INl))
Output layer has only a node, and the output of left and right two sub-networks is expressed as O respectively
lAnd O
r, the weighting coefficient of left sub-network is φ, then the weighting coefficient of right sub-network is 1-φ, the output of network:
O=φ O
l+ (1-φ) O
r, be characterized in that the patient ED pattern that resident above-mentioned basic artificial neural network and back will illustrate in this calculator memory storage chooses program module and the special-purpose neural network procedure module of training.Wherein patient ED pattern is chosen program module as shown in Figure 4.Train special-purpose neural network procedure module as shown in Figure 5.After EEG signals input apparatus of the present invention, choose the fragment of in patient's brain electricity, finding out satisfactory some after the routine processes, artificial neural network is trained as the ED pattern through the ED pattern.Training obtains being fit to the special-purpose artificial neural network that this patient ED detects after finishing, and this device can download to every connection weight of special-purpose artificial neural network, the value of laying particular stress on, weighting coefficient etc. in the Intelligent Dynamic eeg recording instrument, and its limit record, limit are discerned.
The above-mentioned electrode that is connected with the multichannel eeg amplifier, it settles the international 10-20 of employing system of system, scalp one pole 16 passage interrecord structures; And the contrary error function prevalue E=0.01 that propagates the learning algorithm network that should feedover, study factor η=0.01; Factor of momentum α=0.05, hidden layer node number H=16, left sub-network weighting coefficient φ=0.5, the standard deviation multiple M=10 of ED and non-ED pattern.
The patient that this device is handled is the patient who analyzed and be judged to be ED with " automatically judge and have or not the epileptiform discharge device in the electroencephalogram ".
Advantage of the present invention:
1, realize the special-purpose artificial neural network of the full-automatic training patient who does not need the doctor to participate in, its result is near the detection level that the experience doctor is arranged.
2, save doctor's loaded down with trivial details, arduous labor, saved time, laborsaving, therefore be specially adapted to vast samll cities and towns, rural area and backcountry, because the there does not just have experienced brain electricity doctor at all.After patient gets epileptics, must to big and medium-sized cities could treat, expensive time-consuming again.
3, being suitable for Intelligent Dynamic eeg recording instrument is used, after special-purpose artificial neural network trains, can download in the Intelligent Dynamic eeg recording instrument, make it limit record, limit identification ED, record is finished also identification and is finished, can save later recognition time, and can produce alerting signal etc., the injury that helps patient to avoid showing effect and cause according to recognition result.
4, because this device, trains the dedicated network that is fit to this patient for each patient, be used further to later on this patient is detected, therefore very strong adaptive faculty is arranged, can in the zone of different regions, different nationalities, use.
Accompanying drawing of the present invention is simply described as follows:
Fig. 1 uses the present invention to constitute the block schematic diagram of the system of ED discharge in the full-automatic detection by quantitative electroencephalogram with multichannel eeg amplifier and Intelligent Dynamic eeg recording instrument etc.
Fig. 2 is the data acquisition unit circuit theory diagrams among the present invention.
Fig. 3 is BP topology of networks figure of the present invention.
Fig. 4 is the process flow diagram that patient ED pattern is chosen program among the present invention.
Fig. 5 is the process flow diagram of the special-purpose neural network procedure of training among the present invention.
Provide better embodiment of the present invention according to Fig. 1～Fig. 5 below, and described in detail, enabling that architectural feature of the present invention, function are described better, rather than be used for limiting claim protection domain of the present invention.
See also Fig. 1 and Fig. 2, as shown in the figure, apparatus of the present invention comprise into data acquisition unit 2 and the computing machine 3 that two-way circuit connects, and be connected with a multichannel eeg amplifier 1 by data acquisition unit 2, the output of this computing machine 3 can be recorded on the floppy disk, inserts other system and uses or download in the Intelligent Dynamic eeg recording instrument 4.In the present embodiment, multichannel eeg amplifier 1 is 16 road eeg amplifiers; Adopt Japanese photoelectricity company (NIHON KOHDEN) 4217 type electroence phalographs record electroencephalogram during experiment, international 10-20 system scalp one pole 16 trace records are adopted in the arrangement of recording electrode, after high pass 0.3Hz and low pass 60Hz filtering, trace on paper, the analog-to-digital figure place that EEG signals after amplifying is sent into data acquisition unit 2, this data acquisition unit 2 is 10bit simultaneously, frequency acquisition is 200 times/second, and computing machine 3 adopts 486 microcomputers.
See also Fig. 3, it shows the topological structure of artificial neural network of the present invention, it is the method for rich to cover (Gabor) etc., BP network after network parameter is optimized, it is made up of two three layers of BP sub-networks, is called left sub-network and right sub-network, respectively has three layers: input layer, hidden layer and output layer, for a BP sub-network, W
JiI node of expression input layer is to the connection weight of j node of hidden layer, W
LjJ node of expression hidden layer is to the connection weight of l node of output layer, and the node of hidden layer and output layer contains bias.θ
jThe bias of j node of expression hidden layer, θ
lThe bias of l node of expression output layer.The input and output excitation function of hidden layer node and output layer node adopts " S " type function:
f(X)＝1/(1+exp(-X))
The error function prevalue E=0.01 of this network; Study factor η=0.01; Factor of momentum α=0.05; Input layer and hidden layer node number are 16, and output layer node number is 1.Left side sub-network weighting coefficient φ=0.5; Standard deviation multiple M=10 between ED and the non-ED pattern.
If n is arranged for an input layer
iIndividual node, hidden layer has n
jIndividual node, output layer has n
lThe sub-network of individual node, Pi represents the input value of i node of sub-network, H
INjInput value for j node of hidden layer.H
OUTjThe activation value of j node of expression hidden layer.O
INlThe input value of l node of expression output layer.O
OUTlThe output valve of l node of expression output layer, feedforward of the present invention is calculated by following formula;
H
OUTj＝1/(1+exp(-H
INj))，
O
OUTl＝1/(1+exp(-O
INl))。
Output layer has only a node, about the output of two sub-networks be expressed as O respectively
lAnd O
rThe weighting coefficient of left side sub-network is φ, and the weighting coefficient of right sub-network is 1-φ, and then network is output as:
O＝φO
l+(1-φ)O
r
We have set up the algorithm of above-mentioned BP network in the memory field of computing machine 3,16 tunnel EEG signals are sent into corresponding 16 input node Pl of BP network
1, Pl
2Pl
16Pr
1, Pr
2Pr
16The network output valve is a certain determined value between 0～1.
About training to basic neural network
At first to from this patient's electroencephalogram, pick out the ED ripple, as the ED pattern artificial neural network is trained, we utilize the method for intersection identification, promptly choose the ED pattern from a patient's brain electricity, network is trained, and the network that trains is discerned different patients' ED.
We have selected three representative patients to experimentize, and they are that ED is the patient A of diffuse type spike and slow wave complex, and ED is that the patient B and the ED of diffuse type sharp wave is the patient C of focal type one spike.With one of them patient's ED training of human artificial neural networks, again all the other two people's brain electricity is discerned then, these three patients, the waveform of ED has the light wave of spike, and the space distribution of ED has diffuse type and focal type, therefore has representative preferably.In order to be fit to intersect the requirement of discerning, we improve covering rich method.One, peak width is to determine according to one group of ED pattern in the feature extraction, makes peak width more representative; Two, for any ED pattern of choosing, the rich method of press cover is calculated and is produced a non-ED pattern, and non-ED pattern departs from M times of standard deviation of this ED pattern, and the M value that we select is 10, and is more much bigger than the M value that lid is rich; Three, when calculating non-ED pattern, we adopt the method that disappears mutually, promptly can be regarded as non-ED pattern with the absolute value of the network input value M times of standard deviation that diminish.Even above-mentioned three improvement make brain wave patterns have certain degree to depart from the ED pattern, still can be by Network Recognition.
Experimental result shows: it is best 1, to go to discern other patient's effects again with ED (being the diffuse type spike and slow wave complex) the training of human artificial neural networks of patient A; 2, for the brain electricity that does not have artifact to disturb, when threshold value was 0.86, the false positive rate of intersection identification was near zero; 3, self identification is better than intersecting identification, and this point has proved that the present invention trains special-purpose artificial neural network with this patient's ED, and the ED effect that is used for discerning this patient again is best.
According to above-mentioned experimental result, the artificial neural network that selected patient's by a diffuse type spike and slow wave complex ED trains out is an elemental network, removes to select each patient's ED with elemental network.
About choosing to ED pattern in patient's brain electricity
Choosing of ED pattern is with elemental network this patient's brain electricity to be identified as the basis among the present invention, and when threshold value was decided to be 0.86, the ED that identifies was genuine, and false positive rate is close to 0; When discrimination improved when threshold value descends, false positive rate also increased gradually.
Our experiment shows, threshold value is decided to be at 0.86 o'clock, the ED that identifies carries out training effect and bad as the ED pattern to network, trace it to its cause is that the ED pattern of at this moment selecting all is very typical ED, and among the actual ED, because be subjected to the factors such as influence of spontaneous brain electricity, waveform becomes and has not been true to type, and this class ED quantity is a lot.By the network that typical ED trains out these ED are just discerned not come out, so discrimination is obviously descended.
Our experiment shows, should select diversified representational ED to make the ED pattern, training network, and discrimination will be high like this.
Choosing the ED pattern among the present invention is to import basic artificial neural network with this ripple, the peak value PV of its output valve
iFor foundation.
Our experiment shows PV
i＜0.82 without exception be not chosen for is because PV
lBe lower than 0.82, false positive is quite high, just is easy to non-ED is elected to be the ED pattern, and the Network Recognition rate that training is come out descends.
Our experiment shows PV
i＞0.92 ED seldom, the patient who has even do not have.Therefore choose PV
iIn following scope: 0.82≤PV
i≤ 0.92.
After having compared five different distributions patterns, our experiment shows that two little patterns broad in the middle are best, i.e. PV
i0.89～0.87 (centre) wants multiselect, PV
iWhen 0.92 changes, choose number and reduce, work as PV
IWhen 0.82 changes, choose number and also reduce.
Our experiment shows, the effect of choosing 15 patterns altogether is best, and it has taken into account discrimination than higher and short two aspects of training time.Xuan Ding distribution pattern is PV at last
iValue is 0.92 1,0.90 2, and (0.89～0.87) 8,0.86 2,0.84 1,0.82 1.
In choosing ED mould mode like this, the possibility that non-ED is falsely dropped to the ED pattern is little, and our experiment shows, and is in 15 ED patterns choosing, if sneaked into one or two non-ED, little to the discrimination influence of the artificial neural network of training gained.
According to above-mentioned result of study, the process flow diagram of the program of choosing of ED pattern as shown in Figure 4.As seen from the figure, execution in step 50 behind the program start, and patient's EEG signals is imported basic artificial neural network → step 51, differentiate network output and reach peak value? if do not reach peak value, return execution in step 50; If reach peak value, then → step 52, differentiate peak value PV
i〉=0.82? if, PV
i＜0.82, then execution in step 531, if PV
i〉=0.82, just → step 53, differentiate PV
iWhether less than 0.92, if less than 0.92, execution in step 531; If be not less than 0.92, then → step 14, deposit this section eeg data in data working area → step 55 as candidate ED pattern, differentiate in the signal wave that is elected to be the ED pattern this PV
iIs value chosen number and is reached setting? if reached the regulation number, just execution in step 531, if do not reach the regulation number, then → and step 56, this section ripple to be chosen and the ED pattern, its eeg data dumps to ED pattern working area, → step 57 is differentiated the PV that regulation is chosen
iDistribute and whether all be full? if be not full, just execution in step 531, if be full, then → and step 59, expression ED pattern is chosen work and is finished, and enters the step 60 in the program block 6.
Above-mentioned steps 531, arrive 160 minutes access time differentiating regulation? if reached 160 minutes, → step 533, as less than 160 minutes, then → step 532, does differentiating the EEG signals input finish? do not finish as input, return execution in step 50, repeat abovementioned steps, as finishing, then execution in step 533, establish PV
i=0.82, → step 534 is differentiated this PV
iIs value chosen number and whether has been reached setting? if reached setting, just execution in step 537; If do not reach setting, then → step 535, differentiate this PV
iIs there there candidate ED pattern near the value? if there is not candidate ED pattern, just execution in step 537, if there is candidate ED pattern, then → and step 536, this candidate is elected to be the ED pattern, and returns execution in step 534, repeat above-mentioned steps; If when step 535, there is not candidate ED pattern, then execution in step 537, make PV
i=PV
I+ 0.01, → step 58 is differentiated PV
iValue＞0.92? if be not more than 0.92, just return execution in step 534; If greater than 0.92, then → step 59, finish the ED pattern and choose work, and enter
Also to point out, use when of the present invention, Tathagata a ripple, its PV
i=0.90, at this moment if PV
i2 of the less thaies that=0.90 ripple is chosen, then this ripple is selected as the ED pattern, if PV
i=0.90 ripple has been chosen 2, and this ripple just wouldn't be chosen.If ripple required in the above-mentioned distribution is all selected, then the ED pattern is chosen end-of-job.If patient's EEG signals has been imported the access time that finishes or stipulate and arrived, the distribution of regulation is not full yet, then as far as possible with PV
iClose (PV
i± 0.02) ripple replaces, as PV
i=0.86 also lack 1, PV is arranged among the candidate
i=0.92,0.88,0.85 each several, then choose PV
iOne of=0.85 as the ED pattern, if still can not reach the requirement of specified distribution after substituting, what is then selected still finish with regard to what, lectotype.
Artificial neural network about training patient special use
The method of training is to form through optimization according to the rich method of lid.The feature of ED is the synchronous sharp wave of multiple tracks, the left side of this ripple claims left side branch half-peak breadth from the lowest point to the summit, the right side is right branch half-peak breadth from the crest of peak to the lowest point, because the superposition of spontaneous brain electricity etc. makes slightly change before and after the summit position, we calculate the position of peak valley with method such as average, obtain the right side, left side branch half-peak breadth, calculate the average left and right branch half-peak breadth W of each ED pattern again
lAnd W
rThe variable quantity of i road signal is through being normalized into the input quantity P for left and right network at interval
LipAnd P
RipBetween each ED pattern of i road EEG signals standard deviation S is arranged
LiAnd S
Ri, the rich method of press cover can be calculated a non-ED pattern by an ED pattern, and they have identical W
lAnd W
r, and their left and right sub-network input quantity Q
LipAnd Q
RipCalculate and get by following formula.
Q
lip＝P
lip-MS
li Q
rip＝P
rip+MS
li
Through our optimization, the M value is 10, and the former gets and subtracts in the above-mentioned formula, and the latter gets and adds, and these are rich different with lid.
The training of network is carried out against broadcasting learning algorithm according to known feedforward, and left and right sub-network respectively connects weight and is initialized as random number in-0.3 to+0.3 scope.Then with the network input quantity P of each pattern
Lip, P
RipOr Q
Lip, l
RipFan-in network, the network output of calculating each pattern by the feedforward formula of network.
The hope output valve of ED pattern is 1, and the hope output valve of non-ED pattern is 0, poor between the hope output valve of each pattern and the real output value, after square again addition promptly get " error ".
Whether " error " gives the value of putting E less than error function.If less than E, then training finishes.If more than or equal to E, then press the contrary propagation algorithm of known feedforward by the difference of hope output and actual output, revise connection weight and bias.The amount that wherein has study factor η and factor of momentum α control to revise, through we experimental study E=0.01, η=0.01, α=0.05 makes training speed fast, the discrimination height.
The weighting coefficient φ of left side sub-network our experiments show that φ=0.5 effect is better.
The subroutine flow chart of the artificial neural network of training patient special use as shown in Figure 5.
See also Fig. 5, it illustrates the structural drawing of the special-purpose neural network subroutine module of training, and it comprises step 60, extract the characteristic quantity → step 61 of ED pattern, with M=10, go out a non-ED pattern → step 62, the connection weight W of randomization decider artificial neural networks from each ED mode computation
Ji, W
Lj, bias θ
j, θ
lDeng, → step 63 is calculated the network input quantity Pl and the Pr → step 64 of each ED pattern and Ge Fei ED pattern, with the network input quantity input artificial neural network of each pattern, calculate the network output of each pattern by the feedforward formula of network, → step 65 is calculated " error " → step 66, do you differentiate error≤E? if not, return step 64 after the → step 67, if then → execution in step 68 training finish the W of gained
Ji, W
Lj, θ
jAnd θ
lAnd φ constitutes this patient's special-purpose artificial neural network.Above-mentioned step 67 is by the contrary learning algorithm of propagating of known feedforward, revises connection weight W with η=0.01 and α=0.05
Ji, W
LjWith bias θ
jAnd θ
l
Also be noted that:
1, step 70～73 are with reference to the method for the rich Gabor of lid, and wherein M is that non-ED mode standard difference multiple is got M=10 after we optimize.
2, the ideal network output valve of ED pattern is 1.0, and the ideal network output valve of non-ED pattern is 0, the difference of each pattern ideal network output valve and real network output valve again square, addition is " error " again.
3, E is the error function prevalue, gets E=0.01 through optimization, makes training speed fast, the discrimination height.
4, η is the study factor, and α is a factor of momentum, gets η=0.01 through optimization, and α=0.05 makes training speed fast, the discrimination height.
5, φ is left sub-network weighting derivative, optimize φ=0.5, the discrimination height.
Claims (6)
1, a kind of basic artificial neural network training method, its step comprises:
(1) lid of setting up in a computing machine is earlier as shown in Figure 3 won artificial neural network topological structure and rich non-ED algorithm and the BP algorithm of resident lid, it is characterized in that:
(2) resident in this computing machine
1. determine the algorithm of peak width by the average peak width among one group of ED of a patient;
2. by the ED pattern depart from M times of standard deviation calculate non-ED pattern algorithm and
3. depart from and adopt the algorithm that absolute value is diminished;
(3) be that the patient's of diffuse type spike and slow wave complex ED trains this neural network with arbitrary ED, generate each parameter value of its connection weight and amount of bias, form a basic artificial network.
2, artificial neural network training method according to claim 1 is characterized in that M=10.
3, basic artificial neural network training method according to claim 1 is characterized in that said one group of ED is meant 15 ED fragments choosing from 16 road EEG(electroencephalography).
4, a kind of device of the automatic training of human artificial neural networks of making according to claim 1 or 2 or 3 described basic artificial neural network coaching methods, comprise into data acquisition unit and computing machine that two-way circuit connects, the input end of this data acquisition unit is connected with a multichannel eeg amplifier output terminal, this computing machine then is connected with printer, in the internal storage of this computing machine, set up the input working area of artificial neural network, the parameter working storage, and resident the contrary study algorithm routine of propagating of feedforward arranged, each road EEG signals has corresponding input node in this neural network, this artificial neural network is made up of two three layers of BP sub-networks, be called left sub-network and right sub-network, the output of left side sub-network and right sub-network is after weighted mean, form network output O, a left side, right sub-network respectively has three layers: input layer, hidden layer and output layer, input layer data-storing district is set respectively in the internal memory of this computing machine, hidden layer data-storing district and output layer data storage area are for a BP sub-network, W
JiI node of expression input layer is to the connection weight of j node of hidden layer, W
LjJ node of expression hidden layer is to the connection weight of l node of output layer, and the node of hidden layer and output layer contains bias; θ
jThe bias of j node of expression hidden layer, θ
lThe bias of l node of expression output layer; The input and output excitation function of hidden layer node and output layer node adopts " S " type function:
f(X)＝1/(1+exp(-X))；
An input layer has ni node, and hidden layer has nj node, and output layer has the sub-network of nl node, and Pi represents the input value of i node of sub-network, H
INjThe input value of j node of expression hidden layer, H
OUTjThe activation value of j node of expression hidden layer, O
INlThe input value of l node of expression output layer, O
OUTlThe output valve of l node of expression output layer, its feedforward is calculated by following feedforward formula;
H
OUTj＝1/(1+exp(-H
INj))
O
OUTl＝1/(1+exp(-O
INl))
Output layer has only a node, and the output of left and right two sub-networks is expressed as O respectively
lAnd O
r, the weighting coefficient of left sub-network is φ, then the weighting coefficient of right sub-network is 1-φ, the output of network:
O=φ O
l+ (1-φ) O
r, it is characterized in that: also in this calculator memory reservoir, set up basic artificial neural network and patient ED pattern and choose program module and the special-purpose neural network module of training.
5, the device of automatic training of human artificial neural networks according to claim 4 is characterized in that said patient ED pattern chooses program module, and its flowage structure as shown in Figure 4.
6, the device of automatic training of human artificial neural networks according to claim 4 is characterized in that the special-purpose neural network procedure module of said training, and its flowage structure as shown in Figure 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB991242092A CN1148700C (en) | 1999-12-03 | 1999-12-03 | Training method of basic artificial nerve network and device for automatical training of artificial nerve network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CNB991242092A CN1148700C (en) | 1999-12-03 | 1999-12-03 | Training method of basic artificial nerve network and device for automatical training of artificial nerve network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN1261193A true CN1261193A (en) | 2000-07-26 |
| CN1148700C CN1148700C (en) | 2004-05-05 |
Family
ID=5283295
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CNB991242092A Expired - Fee Related CN1148700C (en) | 1999-12-03 | 1999-12-03 | Training method of basic artificial nerve network and device for automatical training of artificial nerve network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN1148700C (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN100580654C (en) * | 2003-08-07 | 2010-01-13 | 索尼株式会社 | Information processing apparatus and method |
| CN107301454A (en) * | 2016-04-15 | 2017-10-27 | 北京中科寒武纪科技有限公司 | The artificial neural network reverse train apparatus and method for supporting discrete data to represent |
| CN110751260A (en) * | 2018-07-24 | 2020-02-04 | 北京三星通信技术研究有限公司 | Electronic device, task processing method and neural network training method |
| CN114947885A (en) * | 2022-07-05 | 2022-08-30 | 华南师范大学 | Radio frequency thermosetting lesion postoperative effect assessment method |
-
1999
- 1999-12-03 CN CNB991242092A patent/CN1148700C/en not_active Expired - Fee Related
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN100580654C (en) * | 2003-08-07 | 2010-01-13 | 索尼株式会社 | Information processing apparatus and method |
| CN107301454A (en) * | 2016-04-15 | 2017-10-27 | 北京中科寒武纪科技有限公司 | The artificial neural network reverse train apparatus and method for supporting discrete data to represent |
| CN107301454B (en) * | 2016-04-15 | 2021-01-22 | 中科寒武纪科技股份有限公司 | Artificial neural network reverse training device and method supporting discrete data representation |
| CN110751260A (en) * | 2018-07-24 | 2020-02-04 | 北京三星通信技术研究有限公司 | Electronic device, task processing method and neural network training method |
| CN114947885A (en) * | 2022-07-05 | 2022-08-30 | 华南师范大学 | Radio frequency thermosetting lesion postoperative effect assessment method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN1148700C (en) | 2004-05-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108776788A (en) | A kind of recognition methods based on brain wave | |
| Molnar et al. | Interpretation of spontaneous spike discharge patterns of neurons in the cochlear nucleus | |
| CN104548347A (en) | Pure idea nerve muscle electrical stimulation control and nerve function evaluation system | |
| CN105249964B (en) | Appraisal procedure is rebuild based on magneticencephalogram and the multi-modal brain function of diffusion tensor | |
| CN109480833A (en) | The pretreatment and recognition methods of epileptic's EEG signals based on artificial intelligence | |
| CN102499676A (en) | Effective time sequence and electrode recombination based electroencephalograph signal categorizing system and method | |
| CN111976733B (en) | Method and system for continuously predicting steering intention of driver | |
| Scott et al. | Neural correlates of auditory short-term memory in rostral superior temporal cortex | |
| CN112488002B (en) | Emotion recognition method and system based on N170 | |
| CN109871831A (en) | A kind of emotion identification method and system | |
| CN111563581B (en) | Brain muscle function network construction method based on wavelet coherence | |
| CN1148700C (en) | Training method of basic artificial nerve network and device for automatical training of artificial nerve network | |
| CN113558639A (en) | Motor intention brain muscle network analysis method based on glange causal relation and graph theory | |
| CN112957056A (en) | Method and system for extracting muscle fatigue grade features by utilizing cooperative network | |
| CN116340824A (en) | Electromyographic signal action recognition method based on convolutional neural network | |
| Downer et al. | Hierarchical differences in population coding within auditory cortex | |
| CN114081505A (en) | Electroencephalogram signal identification method based on Pearson correlation coefficient and convolutional neural network | |
| CN113378737A (en) | Implanted brain-computer interface neuron spike potential classification method | |
| CN110738093B (en) | Classification method based on improved small world echo state network electromyography | |
| CN112244868A (en) | Epilepsia electroencephalogram signal classification method based on ANFIS | |
| Martens et al. | Separating burst from background spikes in multichannel neuronal recordings using return map analysis | |
| Weeks et al. | The degree of potentiation is associated with synaptic number during the maintenance of long-term potentiation in the rat dentate gyrus | |
| CN113558637B (en) | Music perception brain network construction method based on phase transfer entropy | |
| CN1127940C (en) | Method and equipment for automatically determining epileptic discharge threshould recognized by artificial nerve network | |
| CN101057793A (en) | Realtime control method for artificial hand |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C06 | Publication | ||
| PB01 | Publication | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| C56 | Change in the name or address of the patentee |
Owner name: SHANGHAI INST. OF LIFE SCIENCE, CAS Free format text: FORMER NAME OR ADDRESS: SHANGHAI PHYSIOLOGY INST., CHINESE ACADEMY OF SCIENCES |
|
| CP03 | Change of name, title or address |
Address after: 200031 No. 320, Yueyang Road, Shanghai, Xuhui District Patentee after: Shanghai Institute of life Sciences, Chinese Academy of Sciences Address before: 200031 No. 320, Yueyang Road, Shanghai, Xuhui District Patentee before: Shanghai Physiology Inst., of the Chinese Academy of Sciences |
|
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: GR Ref document number: 1065270 Country of ref document: HK |
|
| C19 | Lapse of patent right due to non-payment of the annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |