CN106447032B - The method for quick predicting of cerebral neuron action potential sequence - Google Patents
The method for quick predicting of cerebral neuron action potential sequence Download PDFInfo
- Publication number
- CN106447032B CN106447032B CN201610749057.0A CN201610749057A CN106447032B CN 106447032 B CN106447032 B CN 106447032B CN 201610749057 A CN201610749057 A CN 201610749057A CN 106447032 B CN106447032 B CN 106447032B
- Authority
- CN
- China
- Prior art keywords
- spike
- network
- neuron
- prediction
- module
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Theoretical Computer Science (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Neurology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The method for quick predicting of cerebral neuron action potential sequence is related to brain simulation field, in particular for the accurately and quickly calculating of neuron action potential sequence in extensive big brain modeling.By introducing two modules, one is used to classify, another is for predicting.Then a large amount of sample data is acquired to train two modules, and final realize only relies on a few film potential value for being spaced apart from each other biggish non-spike, can quickly and correctly predict the corresponding feature of spike.Its problem slow the technical problem to be solved is that neuron action potential calculating speed during extensive brain modeling, method in the present invention can greatly improve the calculating speed of action potential, quite high precision is maintained, the simulation of extensive brain neuron network is highly suitable for.
Description
Technical field
The present invention relates to brain simulation fields, in particular for neuron action potential sequence in extensive big brain modeling
Computational problem.
Background technique
Brain is formed by counting the neuron counted with hundreds of millions, these neurons constitute the substantially single of Neural information processing
Member, the discharge activities that neuron is shown just represent the neuron to the coding mode of environmental stimuli.How accurate simulation
These discharge activities are the key that carry out one step of brain emulation.In addition, the number of neuron is very large in brain, and
Functional group can be formed by complicated Synaptic junction between neuron, constitute the organic component of brain.In brain
In simulation, in order to the brain area of approaching to reality, it usually needs guarantee that neuron number has certain scale in network, number
Expand the demand in terms of just bringing simulation velocity, it is big to meet that also just there is an urgent need to a kind of not only quick but also accurate calculation methods
The demand of brain modeling.
Summary of the invention
The technical problem to be solved by the present invention is to speed is slow in terms of simulating extensive neuroid and accuracy is low
Under problem.
In order to solve this problem, the technical solution adopted by the present invention is divided into following three aspects:
A) sampling is carried out to neuron spike sequence and characteristic index is extracted, form spike sample database;
B) according to three continuous sampled points (continuous voltage value V of three of 1ms, 2ms and 3ms before spike peak1、
V2And V3), the peak, minimum and width of neuron spike are predicted using prediction network module;
C) any given three continuous sampled voltage point (Vt1、Vt2And Vt3, it is divided into 1ms), first pass through sorter network mould
Block judges whether these three points can be used to predict the feature of next spike.If the output result of sorter network reach or
More than preset threshold value, means that available predictions network is further to specific feature and predicted;If sorter network
Output result be less than threshold value, then with the step-length of 0.05ms to Vt1、Vt2And Vt3It is updated, then does classification judgement again.
According to the first aspect of the invention, with the following method:
Firstly, utilizing the experiment provided in online disclosed neuron model of ion channel library and the pertinent literature delivered
Data collect neuron model of ion channel, and the discharge mode that these models are shown, which can be covered in nervous system, to be seen
80% or more the electric discharge phenomena observed.
Secondly, being directed to each model, uses step-length for the fourth order Runge-Kutta numerical algorithm of 0.01ms, obtain different width
It spends under stimulus of direct current, the voltage value in film potential sequence, and extracts the peak (V of single spike in these sequencesmax), most
Low value (Vmin) and the width (difference of peak time corresponding with minimum, it may be assumed that Dt=tmin-tmax) feature.Meanwhile recording VmaxBefore
Voltage value V when 1ms, 2ms and 3ms1、V2And V3.Multiple groups are collected by changing electric current stimulation amplitude for each model
[Vmax,Vmin,Dt,V1,V2,V3] data.Then the corresponding spike sample database A of single model is constructed, for subsequent to pre- survey grid
Network module is trained and tests, and trained and amount of test data ratio is 7:3.
Finally, the fourth order Runge-Kutta algorithm for being 0.01ms also with step-length, obtains the V of previous spikeminLater
To the V of next spikemaxBefore in this section neuron voltage value, and interval is therefrom chosen with the sliding window of 0.01ms
For three successive value VV of 1ms1、VV2And VV3, constitute database B, training and test data as sorter network module.Its
In, mark VmaxThe VV of preceding 1ms, 2ms, 3ms1、VV2、VV3(with V in sample database A1、V2、V3Selection it is consistent) be positive sample, table
Show that they can be used to predict the feature of spike, other VV1,VV2,VV3For negative sample, indicate that they can not be used to predict
The feature of spike.Trained and amount of test data ratio is 7:3.
According to the second aspect of the invention, with the following method:
Be trained first: the characteristics of according to sample database A has been collected, constructing corresponding prediction network module, here I
Given a forecast using 3 layers of artificial neural network, here number of plies value 3 be not be fixed, other neural networks for being greater than 3 layers
Be also be applicable in, but the number of plies excessively will cause network training time it is elongated.In addition, not being here one for the module of prediction yet
If fixed artificial neural network, other prediction techniques are equally applicable, such as: method (linear regression) based on regression analysis and
Method (decision tree etc.) based on tree.
In the prediction module that the present invention is constructed, input layer has 3 units, respectively corresponds V1,V2And V3;Middle layer is (hidden
Containing layer) there are 10 units, for extracting the characteristic information in input layer, 10 unit numbers and on-fixed here, between 5-15
Unit number is all that feasible, too small unit number will lead to the feature that can not effectively extract in preceding layer, and excessive unit number
Influence to feature extraction effect is little;Output layer has 3 units, and output result needs and true value Vmax,VminAnd Dt
It compares, therefore whole process belongs to the study for having supervision.Then with backpropagation (backpropagation) algorithm to net
Connection weight in network between different units is updated, until network training is completed (to set the number of iterations as 1000, instruct at this time
It is very small by (< 10 to practice error-5), it is more than that the number of iterations determines that training is completed).
Followed by test: for trained prediction network module, being input to network using the data in test set
In, examine the difference in the output and test set of network between true value (to be described with mean square error, mean squared
Error, MSE), to analyze the test effect of network.
According to the third aspect of the present invention, with the following method:
It is the training of sorter network module first: the characteristics of according to sample database B has been collected, constructs corresponding classification mould
Block, we still use 3 layers of artificial neural network to classify to do here.Certainly, other neural networks for being greater than 3 layers are also applicable
, only the number of plies excessively will cause network training time it is elongated.Similarly, the module here for classification is also not necessarily wanted
It is artificial neural network, other classification methods are equally applicable, such as: method (Bayesian network) based on bayesian theory,
Method (binary tree etc.) based on tree and the method based on support vector machines.
In sorter network module: input layer has 3 units, respectively corresponds VV1,VV2And VV3;Middle layer has 10 units,
For extracting the characteristic information in input layer, similarly, 10 unit numbers and on-fixed here, the unit number between 5-15 is all
Feasible, too small unit number will lead to the feature that can not effectively extract in preceding layer, and excessive unit number is to feature extraction
The influence of effect is little;Output layer has 2 units, for judging three point (VV under the input1,VV2And VV3) whether
It is sufficient to the judgement corresponding three characteristic value (V of spike belowmax,VminAnd Dt).2 results of network output are 1 He respectively
0, wherein 1 indicates to be enough to judge, 0 indicates not judging.When the value of output be greater than threshold value 0.5 when, be labeled as 1, otherwise label for
0.Herein, threshold value 0.5 is derived from 0 and 1 average value, shows that exporting the probability of result appearance 0 and 1 is 50%.Other thresholds
Value, such as: 0.4 or 0.6 is theoretically and feasible, low when only the accuracy of final classification network can be than 0.5.
Followed by the test of sorter network module: for trained sorter network module, using in test set
Data are input in network, examine in the output and test set of network between true value difference (with the absolute value of the difference of the two come
Description), to analyze the test effect of network.
Finally, sorter network module and prediction network module are combined: by the step-length of 0.05ms, obtaining in a certain stimulation
The voltage value of the lower non-part spike of neuron, and three successive value V of 1ms are divided between extractingt1、Vt2、Vt3.First use categorization module
Judged, if output result is 1, by Vt1、Vt2、Vt3It is input in prediction module, to three spies of neuron spike
Value indicative (Vmax,VminAnd Dt) predicted;If exporting result is 0, by Vt1、Vt2、Vt3It is carried out with the step-length of 0.05ms forward
It updates, then re-starts classification judgement.Finally, the information of the part spike can be with non-in neuron action potential sequence
Value (the V of the part spiket1、Vt2、Vt3) accurately predict.Due to Vt1、Vt2、Vt3Value be at the step-length 0.05ms
It arrives, therefore entirely the solution of action potential sequence and calculating ratio are nearly 5 times fast with quadravalence Long Gekuta method at 0.01ms,
Very high accuracy is still maintained simultaneously.
The beneficial effects of the invention are as follows substantially increasing in extensive brain neuron network model, neuron action potential
Quick calculating and accurate obtain.
Detailed description of the invention
Fig. 1 is a kind of Simplification on Forecasting Method block diagram of quick cerebral neuron action potential sequence
Fig. 2 is sample collection schematic diagram in neuron action potential sequence
Fig. 3 is the framework that categorization module network and prediction module network are classified and predicted to neuron spike feature
Figure
Several neuron models and corresponding discharge waveform Fig. 4 used when being trained acquisition and test data
Fig. 5 is the result schematic diagram that spike feature is solved and predicted using 0.05ms step-length
Specific embodiment
Technical solution of the present invention is described in detail below in conjunction with the drawings and specific embodiments in the present invention.
V in the present inventionmax,VminWith Dt respectively indicate single spike peak, minimum and width (peak with most
Low value corresponds to the difference of time, it may be assumed that Dt=tmin-tmax), V1,V2And V3Three continuous voltage values of 1ms are divided between expression, equally
Ground, VV1、VV2、VV3And Vt1、Vt2、Vt3Also three continuous voltage values of 1ms are divided between respectively indicating.
Fig. 1 is a kind of Simplification on Forecasting Method block diagram of quick cerebral neuron action potential sequence, wherein 10 expressions pair
The action potential sequence of given neuron carries out sample collection;20 indicate the sample signal based on acquisition, build prediction module net
Network, and be trained and test;30 indicate the sample signal based on acquisition, build categorization module network, and be trained and survey
Examination, then enters data into prediction module network according to classification results, is predicted.
Fig. 2 is sample collection schematic diagram in neuron action potential sequence, chooses interval according to the step-length sliding of 0.01ms
For continuous three voltage value V of 1ms1,V2And V3, and the corresponding three characteristic value V of single spikemax,VminAnd Dt.Here,
0.01ms step-length can ensure that obtain sufficiently accurate data.
Fig. 3 is the architecture diagram that categorization module network and prediction module network predict neuron spike feature.
Firstly, the characteristics of according to sample data has been collected, the prediction module network of 3 layers of building, wherein input layer has 3 lists
Position, respectively corresponds V1、V2And V3;Middle layer (hidden layer) has 10 units, for extracting the characteristic information in input layer;Output
Layer has 3 units, output result and true value Vmax、Vmin(study for having supervision) is compared with Dt, then uses backpropagation
Algorithm is updated the connection weight between unit in network, until network training completes (iteration 1000 times or more).
Secondly, being input in network for trained prediction module network using the data in test set, examine
Difference (being portrayed with mean square error MSE) in the output of network and test set between true value, to analyze the test effect of network.
Again, the module network for classification of same 3 layers of building, wherein input layer has 3 units, respectively corresponds VV1、
VV2And VV3;Middle layer has 10 units, for extracting the characteristic information in input layer;Output layer has 2 units, for judging
Three point (VV under the input1、VV2And VV3) whether it is sufficient to the judgement corresponding three characteristic value (V of spike belowmax,
VminAnd Dt).2 results of output are 1 and 0 respectively, wherein 1 indicates to be enough to judge, 0 indicates not judging.When the value of output is big
When threshold value 0.5, it is labeled as 1, otherwise label is.And then, using back-propagation algorithm to the connection between unit in network
Weight is updated, until network training completes (iteration 1000 times or more).
Then, it for trained categorization module network, is input in network, is examined using the data in test set
Difference (being portrayed with the absolute value of the difference of the two) in the output of network and test set between true value, to analyze the test of network
Effect.
Finally, integrating categorization module network and prediction module network.For some model, pass through the time step of 0.05ms
It is long, the voltage value of the non-part spike of neuron under a certain stimulation is obtained, and three successive value V of 1ms are divided between obtainingt1、
Vt2And Vt3.It is judged by sorter network first, if output result is greater than or equal to threshold value, just by Vt1、Vt2With
Vt3It is introduced into prediction network, to three characteristic value (V of neuron spikemax、VminAnd Dt) predicted.If output knot
Fruit is less than threshold value, then by Vt1、Vt2、Vt3It is updated with the step-length of 0.05ms forward, then re-starts classification judgement.In this way,
The information of the part spike can use the value (V of the non-part spike in neuron action potential sequencet1、Vt2And Vt3) come accurately
Ground predicts, to accelerate the solution and calculating process of entire action potential sequence.
Fig. 4 is several neuron models and corresponding discharge waveform used when collecting trained and test data.
In nervous system, there are many kinds of the types of neuron, and the discharge waveform showed is also varied.
In order to preferably examine our algorithm, we have collected 9 kinds of model of ion channel, under the stimulation of certain amplitude, these moulds
Type can show 10 kinds of typical discharge waveforms, these waveforms cover the electric discharge behavior of 80% or more neuron in brain.
Fig. 5 is the result schematic diagram that spike feature is predicted using 0.05ms step-length.
In actual operation, the voltage sample data (V before spike is obtained using the iteration step length of 0.05mst1、Vt2
And Vt3).Judge that can it be used to predict the feature of spike by being input to sorter network module, if output result is 1,
It is then input to prediction network module and does specific prediction;If exporting result is 0, next group of electricity is done forward with 0.05ms step-length
Pressure-like sheet (Vt1、Vt2And Vt3) collection, judged using sorter network.
Curve is the film potential curve being calculated under 0.01ms step-length with 4 rank Long Gekuta algorithms in Fig. 5, be used as pair
According to.Star mark is the truthful data point being calculated under 0.01ms step-length with 4 rank Long Gekuta algorithms, is successively marked
It is the V of 3ms, 2ms and 1ms before spike peak value3,V2And V1And the peak value and minimum of spike.Triangle mark is net
The predicted value of network.It is very close to can be seen that predicted value and true value from the data in figure.This indicates that, although take
Step-length (0.05ms) is 5 times of typical 0.01ms, but model still can calculate to a nicety out several characteristic values of spike very much.
Due to using relatively large step-length, 5 times of promotion has also just been obtained in calculating speed.
Claims (3)
1. the method for quick predicting of cerebral neuron action potential sequence, feature comprise the steps of:
(1) collection of neuron model of ion channel and action potential data sample
Firstly, collecting neuron models and complete ion channel data, the model of collection is wanted to reflect neural in nervous system
The electric discharge behavior of 80% or more member;
Secondly, being directed to each model, uses step-length for the fourth order Runge-Kutta numerical algorithm of 0.01ms, it is straight to obtain different amplitudes
Under stream stimulation, the voltage value in film potential sequence, and extract the peak V of single spike in these sequencesmax, minimum
VminAnd width;Width is the difference of peak time corresponding with minimum, it may be assumed that Dt=tmin-tmax;Meanwhile recording VmaxPreceding 1ms,
Voltage value V when 2ms and 3ms1、V2And V3;Multiple groups [V is collected by changing electric current stimulation amplitude for each modelmax,
Vmin,Dt,V1,V2,V3] data;Then the corresponding spike sample database A of single model is constructed, for subsequent to prediction network module
It is trained and tests;
Finally, the fourth order Runge-Kutta algorithm for being 0.01ms also with step-length, obtains the V of previous spikeminIt arrives down later
The V of one spikemaxBefore in this section neuron voltage value, and therefrom with the sliding window of 0.01ms choose between be divided into
Three successive value VV of 1ms1、VV2And VV3, constitute database B, training and test data as sorter network module;Wherein,
Mark VmaxThe VV of preceding 1ms, 2ms, 3ms1、VV2、VV3For positive sample, indicate that they can be used to predict the feature of spike, it is other
VV1、VV2、VV3For negative sample, indicate that they can not be used to predict the feature of spike;
(2) training and test of prediction module
The characteristics of according to the sample data collected, network module of the building for prediction;Then, back-propagation algorithm pair is utilized
Connection weight in network between neuron is updated, until network training is completed;Finally, with test data set to prediction mould
The performance of block is tested;
(3) training and test of categorization module
The characteristics of according to the sample data collected, network module of the building for classification;Then, it is calculated also with backpropagation
Method is updated the connection weight between neuron in network, until network training is completed;Finally, with test data set to point
The performance of generic module is tested;
(4) according in (3) sorter network module as a result, utilize in (2) trained prediction network module to the several of spike
A characteristic feature is predicted;It is specific as follows: by the step-length of 0.05ms, to obtain the non-portion spike of neuron under a certain stimulation
The voltage value divided, and three successive value V of 1ms are divided between extractingt1、Vt2、Vt3;First judged with categorization module, if output
It as a result is 1, then by Vt1、Vt2、Vt3It is input in prediction module, to three characteristic value (V of neuron spikemax、VminAnd Dt)
It is predicted;If exporting result is 0, by Vt1、Vt2、Vt3It is updated with the step-length of 0.05ms forward, is then re-started
Classification judgement;
2 results of sorter network output are 1 and 0 respectively, wherein 1 expression input data can be used for predicting the feature of spike,
0 expression input data may not be usable for the feature of prediction spike;When the value of output is greater than threshold value, it is labeled as 1, otherwise is marked
It is 0;Threshold value is set as a numerical value in [0.4,0.6].
2. the method for quick predicting of cerebral neuron action potential sequence according to claim 1, it is characterised in that: it uses
The prediction technique employed in the network module of prediction include the method for artificial neural network, the method based on regression analysis or
Method based on tree.
3. the method for quick predicting of cerebral neuron action potential sequence according to claim 1, it is characterised in that: it uses
The classification method employed in the network module of classification includes the method for artificial neural network, the side based on bayesian theory
Method, the method based on tree or the method based on support vector machines.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610749057.0A CN106447032B (en) | 2016-09-09 | 2016-09-09 | The method for quick predicting of cerebral neuron action potential sequence |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610749057.0A CN106447032B (en) | 2016-09-09 | 2016-09-09 | The method for quick predicting of cerebral neuron action potential sequence |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106447032A CN106447032A (en) | 2017-02-22 |
CN106447032B true CN106447032B (en) | 2018-12-25 |
Family
ID=58181523
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610749057.0A Active CN106447032B (en) | 2016-09-09 | 2016-09-09 | The method for quick predicting of cerebral neuron action potential sequence |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106447032B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109325596B (en) * | 2018-09-17 | 2024-03-26 | 中国传媒大学 | Nerve synapse plasticity calculation method based on calcium concentration |
CN108960217A (en) * | 2018-09-25 | 2018-12-07 | 北京工业大学 | The hardware structure of action potential classification in a kind of brain-computer interface |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP4975287B2 (en) * | 2005-08-08 | 2012-07-11 | パナソニック株式会社 | Prediction device |
CN105283757A (en) * | 2013-03-15 | 2016-01-27 | 豪夫迈·罗氏有限公司 | Methods of failsafing electrochemical measurements of an analyte as well as devices, apparatuses and systems incorporating the same |
CN105930899A (en) * | 2011-08-16 | 2016-09-07 | 高通股份有限公司 | Method and apparatus for neural temporal coding, learning and recognition |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10531806B2 (en) * | 2013-12-17 | 2020-01-14 | University Of Florida Research Foundation, Inc. | Brain state advisory system using calibrated metrics and optimal time-series decomposition |
-
2016
- 2016-09-09 CN CN201610749057.0A patent/CN106447032B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP4975287B2 (en) * | 2005-08-08 | 2012-07-11 | パナソニック株式会社 | Prediction device |
CN105930899A (en) * | 2011-08-16 | 2016-09-07 | 高通股份有限公司 | Method and apparatus for neural temporal coding, learning and recognition |
CN105283757A (en) * | 2013-03-15 | 2016-01-27 | 豪夫迈·罗氏有限公司 | Methods of failsafing electrochemical measurements of an analyte as well as devices, apparatuses and systems incorporating the same |
Also Published As
Publication number | Publication date |
---|---|
CN106447032A (en) | 2017-02-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108062561A (en) | A kind of short time data stream Forecasting Methodology based on long memory network model in short-term | |
Qu et al. | Series arc fault detection of indoor power distribution system based on LVQ-NN and PSO-SVM | |
CN109933881A (en) | A kind of Fault Diagnosis of Power Electronic Circuits method based on optimization deepness belief network | |
CN106779151B (en) | A kind of line of high-speed railway wind speed multi-point multi-layer coupling prediction method | |
CN110674752A (en) | Hidden Markov model-based tool wear state identification and prediction method | |
CN107274011A (en) | The equipment state recognition methods of comprehensive Markov model and probability net | |
CN104155574A (en) | Power distribution network fault classification method based on adaptive neuro-fuzzy inference system | |
CN111881971A (en) | Power transmission line fault type identification method based on deep learning LSTM model | |
CN103840988A (en) | Network traffic measurement method based on RBF neural network | |
CN109100627A (en) | A kind of power equipment partial discharges fault diagnostic method based on end-to-end mode | |
CN110263979A (en) | Method and device based on intensified learning model prediction sample label | |
CN106203531A (en) | A kind of transmission line fault sorting technique based on the sparse autocoder of convolution | |
CN110070102B (en) | Method for establishing sequence-to-sequence model for identifying power quality disturbance type | |
CN112396234A (en) | User side load probability prediction method based on time domain convolutional neural network | |
CN113901977A (en) | Deep learning-based power consumer electricity stealing identification method and system | |
CN110455512B (en) | Rotary mechanical multi-integration fault diagnosis method based on depth self-encoder DAE | |
CN106447032B (en) | The method for quick predicting of cerebral neuron action potential sequence | |
CN115673596B (en) | Welding abnormity real-time diagnosis method based on Actor-Critic reinforcement learning model | |
Lin et al. | Short-term metro passenger flow prediction based on random forest and LSTM | |
ChikkaKrishna et al. | Short-term traffic prediction using fb-prophet and neural-prophet | |
CN115859837A (en) | Fan blade dynamic impact detection method and system based on digital twin modeling | |
CN115965135A (en) | New energy prediction error modeling method and system based on naive Bayes classification | |
CN115712871A (en) | Power electronic system fault diagnosis method combining resampling and integrated learning | |
CN113554716A (en) | Knowledge distillation-based tile color difference detection method and device | |
CN110098944A (en) | A method of protocol data flow is predicted based on FP-Growth and RNN |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |