CN108319928A - A kind of deep learning model and application based on Multi-objective PSO optimization - Google Patents
A kind of deep learning model and application based on Multi-objective PSO optimization Download PDFInfo
- Publication number
- CN108319928A CN108319928A CN201810169310.4A CN201810169310A CN108319928A CN 108319928 A CN108319928 A CN 108319928A CN 201810169310 A CN201810169310 A CN 201810169310A CN 108319928 A CN108319928 A CN 108319928A
- Authority
- CN
- China
- Prior art keywords
- particle
- target
- collection
- deep learning
- parameter set
- 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
-
- 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
- G06F2218/16—Classification; Matching by matching signal segments
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Signal Processing (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
A kind of deep learning model based on Multi-objective PSO optimization：EEG signals are acquired, the EEG signals of acquisition are pre-processed, obtain the p-channel brain teledata sample set that signal is p dimensions；A data as deep learning model, that is, convolutional neural networks after p dimension brain teledata samples normalizations are inputted, the classes of instructions of the imagination will be corresponded to as the output of convolutional neural networks last layer；Establish preliminary convolutional neural networks；Preliminary convolutional neural networks are optimized and revised using multi-objective particle swarm optimization algorithm, obtain deep learning model；Multiple target movement auxiliary is realized using multi-objective particle swarm optimization algorithm.The present invention using multi-objective particle swarm optimization algorithm solve manually to deep learning model be adjusted the local optimum being likely to occur, under efficiency, need the problems such as priori.Its output of the deep learning model of structure can be as the signal of the plurality of devices such as control machinery arm or ectoskeleton.
Description
Technical field
The present invention relates to a kind of deep learning models.More particularly to a kind of depth based on Multi-objective PSO optimization
Spend learning model and application.
Background technology
EEG signals be it is a kind of cerebral cortex can be people detected by neuron physiological activity reflection mode, according to
The concrete condition of EEG signals can extract a large amount of physiologic information from different regional neuronal current potentials.Different human brain works
Making accurate detection, identification, the classification of state can provide in brain control system etc. for certain special movement back-up needs
Theoretical foundation.Therefore the identification processing of EEG signals has very important significance.
Recently as science and technology development, especially with brain-computer interface (brain-computer interface,
BCI) the development of technology so that the research of corresponding brain-computer interface technology becomes new research hotspot, including SSVEP,
The technologies such as Mental imagery.The wherein research of Mental imagery not only helps to improve the understanding to the related function of brain, moreover it is possible to
Contribute substantially to the research of brain control motion assistant system related field.The technology has have the subject of movement back-up needs
There is provided a kind of potential and outlook of more easily daily life means ideally enables to subject to pass through idea
Control external equipment carries out the degree of basic or even complicated life activity, is outwardly conveyed by consciousness or realizes them
Idea, improve subject spirit and material life experience.
In daily life, the movement of the four limbs of people supports most of life activity, whether the fortune of biped
The complex physiologic function of dynamic function or both hands, all has vital effect to the normal life of people.Whether life must
The material activity of palpus or the material base creation of cultural life, are required for corresponding limbs to carry out physiology support.However for
Some have special movement either may be since the movement or behavior of required realization have for the subject of activity need
Have an environmental restrictions, such as carry out the fault detect of strong corrosive environment, obviously cannot utilize in this case human body into
These movements of row and behavior.Therefore, the subject's carry out activity for there are movement back-up needs is helped to have using brain-computer interface technology
Very important meaning.
Deep learning as a kind of theoretical method emerging in recent years, it is unique and significantly to the feature of data
Practise, feed back and received the characteristics of realizing Supervised classification the extensive concern of people from all walks of life.Deep learning is as machine learning
One branch, algorithm belong to the further expansion of neural network derived from being simulated to human brain, it simulate the mechanism of human brain into
The multilayer of the explanation of row data and analytic learning, autonomous learning initial data indicates, deep layer nerve is constituted using many hidden layers
Network structure.Typical network structure has：Convolutional neural networks (CNN) generate confrontation network (GAN) and long memory network in short-term
(LSTM) etc..With the gradually development of deep learning method so that carried for various network overfitting problem solutions
Go out, also therefore can train and obtain the deep neural network with excellent results, compared to conventional feature extraction and classification
Algorithm, be obviously improved on accuracy performance, be accordingly used in the brain coron feature signal extraction to subject,
With remarkable result.
Particle Swarm Optimization (PSO) is to be equal to one kind that nineteen ninety-five develops by J.Kennedy and R.C.Eberhart to drill
Change computing technique, derives from a simplified social model, be one kind of swarm intelligence, be included into multiagent optimization system
In.PSO simulates flock of birds predation, and in residing background, the solution of each optimization problem is equal to a bird in space, we
Referred to as particle, all particles have the adaptive value and its speed of the decision of optimised function.The advantage of PSO algorithms is letter
List is easy to implement and without the adjusting of many parameters, has been widely used in function optimization, neural metwork training, mould at present
Paste system controls and the application field of other genetic algorithms, becomes an important branch of Natural computation.It is passed through using PSO algorithms
It crosses and the improved MOPSO algorithms (multi-objective particle swarm optimization algorithm) of specialization are excellent to the progress of deep learning network to be realized to multiple target
Change, can replace it is artificial continuously attempt to, independently carry out the network reference services of deep learning using computer, it is more efficient and
Effect is more preferable.
Meanwhile during carrying out brain control system motion auxiliary, the motion component of the eeg data of subject not office
It is limited to the target action for attempting to realize, and there is more complicated content of thought, simple detects target action EEG signals,
Ignore the action of other EEG signals, with each limbs organ in normal person's physiological activity work in coordination execution the case where not
Meet, the effect of brain control action auxiliary can be played, but perhaps advantage can not compared with manual operation machinery executes for effect
There is quantum jump.If carrying out optimizing auxiliary to its complicated EEG signals using MOPSO algorithms, simple a certain target is moved
It assists, is accurately changed to the auxiliary of a moving system entirety, the brain control action auxiliary of optimum efficiency may be implemented.
Invention content
The technical problem to be solved by the invention is to provide a kind of deep learnings based on Multi-objective PSO optimization
Model and application.
The technical solution adopted in the present invention is：A kind of deep learning model based on Multi-objective PSO optimization,
Include the following steps：
1) EEG signals are acquired, the EEG signals of acquisition are pre-processed, it is two-dimensional p-channel brain telecommunications to obtain signal
Set of data samples x (t)；
2) deep learning model, that is, convolutional Neural will be used as after p dimension brain teledata x (t) samples normalizations in step 1)
One data of network input, using the classes of instructions of the corresponding imagination as the output of convolutional neural networks last layer；
3) preliminary convolutional neural networks are established；
4) preliminary convolutional neural networks are optimized and revised using multi-objective particle swarm optimization algorithm, obtain deep learning mould
Type；
5) multi-objective particle swarm optimization algorithm is used to realize multiple target movement auxiliary.
Step 1) is to acquire EEG signals, the brain telecommunications using the 32 channel electrode caps that international 10-20 systems define
Number it is the EEG signals based on Mental imagery, acquired content is fixed limb motion Imaginary Movement.
The EEG signals to acquisition described in step 1) pre-process, including：Bandpass filtering, samples normalization and rejecting
There are the EEG signals of deformity, wherein the range of the bandpass filtering is according to the selection pair of the characteristic of Mental imagery inter-related task
The brain wave frequency range answered.
Step 3) includes：
(1) sample is chosen from eeg data sample set x (t) enter convolutional neural networks；
(2) it calculates the sample and enters the reality output that convolutional neural networks obtain, in this stage, information is from input layer
By transformation step by step, it is transmitted to output layer, this process is also the normal flow of network completion after training；
(3) difference of the ideal output of reality output and respective sample is calculated；
(4) weighed value adjusting is carried out according to the method for minimization error, obtains preliminary convolutional neural networks.
Step 4) includes：
(1) parameter group and the target component collection of preliminary convolutional neural networks, the preliminary convolutional Neural net are initialized
Network parameter group includes structure and corresponding each hyper parameter；
(2) it carries out Particle Swarm movement and obtains new target component collection；
(3) it is inputted the new target component collection obtained as the parameter of new convolutional neural networks, substitutes into eeg data
Sample set x (t) carries out desk checking, verifies the accuracy and efficiency of new convolutional neural networks, convolutional neural networks after verification,
As final convolutional neural networks, deep learning model is constituted.
(1) step includes assigning an initial value at random to parameter group, generates initial convolutional neural networks population of parameters
Best solution in initial convolutional neural networks parameter group P1 is stored in target component collection as initial optimum bit by body P1
Set archive collection A1.
(2) step includes that set the particle for making currently to be moved be j, carries out following process：
(2.1) dense information that target component concentrates particle is calculated, specifically target component collection space grid decile
At several regions, use the population that each region is included as the density information of particle；Include in region where particle
Population is more, and the area density values are bigger, otherwise smaller, is implemented as：
It calculates in parameter group iterations upper limit value t, the boundary in target component collection spaceMeter
Calculate the mould of gridWherein M refers to the region that grid is divided
Sum, round numbers, F1 tWithFor target function value；The particle concentrated in target component is traversed, the grain that target component is concentrated is calculated
The number of grid where subWherein Int is bracket function；It calculates
The density estimation value of gridding information and particle；
(2.2) particle P in setting parameter groupJ, tIt is to concentrate a best particle G in target componentJ, t, particle GJ, tMatter
Amount determines that the constringency performance and diversity of multi-objective particle swarm optimization algorithm, selection gist are that the particle that target component is concentrated is close
Spend information；Wherein j is the label of current kinetic particle, and t is the value of current iteration number；It is better than population of parameters with target component collection particle
The population of body carrys out the search potentiality of evaluation goal parameter set particle, more better than the target component collection particle of parameter group, mesh
Mark parameter set search potentiality are bigger, and algorithm is specific as follows：
It calculates target component and concentrates and be better than particle PJ, tParticle collection Aj, for integer k from 1 to AtThe population for including
In range, Aj=Aj+{AK, t|AK, tp PJ, t, AK, t∈At}；Then particle collection A is calculatedjThe particle collection G of middle density minimumj, Gj=
min{Density(Ak), k=1,2 ..., | Aj|, Ak∈Aj}；Wherein, AjIt is used to store target component collection AtIn be better than particle
PJ, tParticle collection, AjThe particle of middle density minimum is stored in particle collection GjIn；AK, tMiddle t refers to the t times iteration, same primary
It is A that iterations are omitted in iterationk；Density(Ak) it is to calculate particle AkDensity estimation value；
(2.3) in undated parameter group particle position and speed drawing in the global best particle G and best particle P of individual
Lower search optimal solution is led, is update target component collection；
(2.4) blocked operation for carrying out target component collection, it is more than defined amount to avoid population；
(2.5) particle information for exporting target component collection, becomes new target component collection.
A kind of application of the deep learning model based on Multi-objective PSO optimization, includes the following steps：
1) preliminary polynary limb action input parameter collection and target action parameter set, the preliminary polynary limbs are initialized
Action input parameter set is by obtained classification after the sample input deep learning model in eeg data sample set x (t)
Target action parameter set corresponding to information；
2) it carries out Particle Swarm movement and obtains new target action parameter set；
3) the new target action parameter set that will be obtained, compared with the target action parameter set that subject repeatedly acts
Compared with whether the method validation screened using threshold value, the new target action parameter set is repeatedly to carry out classification judgement before
The target action parameter set of the highest frequency gone out is then as final target action parameter set, to realize corresponding auxiliary
It acts, otherwise return to step 2).
Step 1) includes assigning an initial value at random to polynary limb action input parameter group, is generated initial polynary
Best solution in initial polynary limb action input parameter group P2 is stored in target by limb action input parameter group P2
Set of behavioural parameters achieves collection A2 as initial optimum position.
(2) step includes that set the particle for making currently to be moved be i, carries out following process：
(2.1) dense information for calculating particle in target action parameter set, specifically uses target action parameter set space
Grid is divided into several regions, uses the population that each region is included as the density information of particle；Region where particle
In include population it is more, the area density values are bigger, otherwise smaller, are implemented as：
It calculates in parameter group iterations upper limit value t, the boundary in target component collection spaceMeter
Calculate the mould of gridWherein M refers to the region that grid is divided
Sum, round numbers, F1 tWithFor target function value；The particle in target action parameter set is traversed, target action parameter is calculated
The number of grid where the particle of concentrationWherein Int is rounding letter
Number；Calculate the density estimation value of gridding information and particle；
(2.2) particle P in setting parameter groupJ, tIt is a best particle G in target action parameter setI, t, particle GI, t
Quality determine the constringency performance and diversity of multi-objective particle swarm optimization algorithm, selection gist is in target action parameter set
Particle density information；Wherein i is the label of current kinetic particle, and t is the value of current iteration number；With target action parameter set grain
Son carrys out the search potentiality of evaluation goal set of behavioural parameters particle better than the population of parameter group, dynamic better than the target of parameter group
It is more to make parameter set particle, target action parameter set search potentiality are bigger, and algorithm is specific as follows：
It calculates and is better than particle P in target action parameter setI, tParticle collection Ai, for integer k from 1 to AtThe grain for including
Within the scope of subnumber, Ai=Ai+{AK, t|AK, tp PI, t, AK, t∈At}；Then particle collection A is calculatediThe particle collection G of middle density minimumi,
Gi=min { Density (Ak), k=1,2 ..., | Ai|, Ak∈Ai}；Wherein, AiIt is used to store target action parameter set AtIn it is excellent
In particle PI, tParticle collection, AiThe particle of middle density minimum is stored in particle collection GiIn；AK, tMiddle t refers to the t times iteration,
It is A with iterations are omitted in an iterationk；Density(Ak) it is to calculate particle AkDensity estimation value；
(2.3) in undated parameter group particle position and speed drawing in the global best particle G and best particle P of individual
Lower search optimal solution is led, is update target action parameter set；
(2.4) blocked operation for carrying out target action parameter set, it is more than defined amount to avoid population；
(2.5) particle information for exporting target action parameter set, becomes new target action parameter set.
A kind of the deep learning model and application based on Multi-objective PSO optimization of the present invention, it is micro- using multiple target
Particle swarm optimization algorithm solve manually to deep learning model be adjusted the local optimum being likely to occur, under efficiency, needs
The problems such as priori.Thus its output of the deep learning model built can be used as control machinery arm or ectoskeleton etc. are a variety of to set
Standby signal, outstanding recognition accuracy and the high recognition efficiency that is brought by deep learning model are in contrast to traditional signal
Identification has clear advantage.Expansion identification is carried out by Multi-objective PSO simultaneously, it is more that single action, which is assisted in identifying,
Target action assists in identifying, and functionally has considerable degree of raising in action auxiliary, has broad application prospects.
Description of the drawings
Fig. 1 is a kind of flow chart of the deep learning model and application optimized based on Multi-objective PSO of the present invention.
Specific implementation mode
With reference to embodiment and attached drawing to a kind of deep learning based on Multi-objective PSO optimization of the present invention
Model and application are described in detail.
A kind of the deep learning model and application based on Multi-objective PSO optimization of the present invention, is to provide a kind of base
Deep learning network model is established in brain electricity EEG signal, PSO optimization algorithms is expanded and deep learning network is optimized and used
In the method for brain control equipment, carried out efficient deep using the advantage of MOPSO optimization algorithms to overcome the deficiencies in the prior art
Learning network structure is spent, while using MOPSO algorithms by single action auxiliary mode, optimal with major heading, auxiliary mark synchronizes
The complete action auxiliary system of form composition of optimization.
As shown in Figure 1, a kind of deep learning model based on Multi-objective PSO optimization of the present invention, including it is as follows
Step：
1) EEG signals are acquired, are the 32 channel electrode caps acquisition EEG signals for using " international 10-20 systems " to define, institute
The EEG signals stated are the EEG signals based on Mental imagery, and acquired content is fixed limb motion Imaginary Movement.It is described
Bandpass filtering range it is specific as follows：Mental imagery and the control signal of perception are the perception movement rhythm and pace of moving things.The perception movement rhythm and pace of moving things
Feature is mainly manifested in μ wave bands (frequency range 8-12Hz) and wave band (frequency range 18-26Hz).Therefore, bandpass filtering
Ranging from 1-30Hz.The EEG signals of acquisition are pre-processed, it is two-dimensional p-channel brain teledata sample set to obtain signal
X (t), the EEG signals to acquisition pre-process, and are to remove eye telecommunications manually using corresponding such as the methods of ICA
It is number equal to influence, and the frequency range progress bandpass filtering of corresponding task is determined by the perception rhythm and pace of moving things of Mental imagery, including：Band
Pass filter, samples normalization and rejecting have the EEG signals of deformity, wherein the range of the bandpass filtering is thought according to movement
As the characteristic of inter-related task chooses corresponding brain wave frequency range.
2) deep learning model, that is, convolutional Neural will be used as after p dimension brain teledata x (t) samples normalizations in step 1)
One data of network (CNN) input, using the classes of instructions of the corresponding imagination as the output of convolutional neural networks last layer；
3) preliminary convolutional neural networks are established；
Convolutional neural networks (Convolutional Neural Network, CNN) are the changes of multi-layer perception (MLP) (MLP)
Kind, common neocognitron includes two class neurons, that is, undertakes the sampling member of feature extraction and the convolution member of resistance, sampling member
Be related to two parameter receptive fields and threshold value, receptive field determine input connection number, threshold value is controlled for the anti-of feature subpattern
Answer degree.Perceptrons of the CNN substantially as a multilayer, it is unique in that it not only reduces the quantity of weights and makes
Network is easy to optimize, and also reduces the risk of over-fitting.The shared feature structure of its weights makes CNN networks compared to common
Perceptron is closer to the neural network of biology, reduces the complexity of network model.CNN is the nerve net of a multilayer
Network, every layer is made of multiple two dimensional surfaces, there is multiple independent neurons in each plane.Include simple member and complexity in network
Member, simple member form simple face, and simple face polymerization forms simple layer, are also identical definition in complicated member, complicated face and complicated layer.
The middle section of network is concatenated by simple layer and complexity, and input layer only only has one layer, it directly receives the two of input
Tie up EEG signals.In general, simple layer is feature extraction layer, and the input of neuron is mutual with one layer of local receptor field before
It is connected, realizes the feature for extracting the receptive field content, if this feature is successfully extracted, the position of this feature and other feature
The relationship of setting will be determined；Complicated layer is Feature Mapping layer, and each computation layer of network is made of multiple Feature Mappings, Mei Gete
It levies and is mapped as a plane, the weights of all neurons are identical in plane.Since the neuron weights on each mapping face are shared,
The free parameter number for reducing network reduces the complexity of network parameter selection.CNN networks are in addition to input layer and output layer
Except, also original two-dimentional EEG signals are directly inputted to input layer by intermediate convolutional layer, sample layer and full articulamentum,
The size of original eeg data determines that the size of input vector, neuron extract the local feature of eeg data, each nerve
Member is all connected with the local receptor field of preceding layer, by the sampling layer (simple layer) that is alternately present and convolutional layer (complicated layer) and most
Full articulamentum afterwards provides the output of network in output layer.Usually in CNN, what network training and right value update leaned on is reversed
Propagation algorithm (BP algorithm),
It is of the present invention to establish preliminary convolutional neural networks, including：
(1) sample is chosen from eeg data sample set x (t) enter convolutional neural networks；
(2) it calculates the sample and enters the reality output that convolutional neural networks obtain, in this stage, information is from input layer
By transformation step by step, it is transmitted to output layer, this process is also the normal flow of network completion after training；
(3) difference of the ideal output of reality output and respective sample is calculated；
(4) weighed value adjusting is carried out according to the method for minimization error, obtains preliminary convolutional neural networks.
4) preliminary convolutional neural networks are optimized and revised using multi-objective particle swarm optimization algorithm (MOPSO), obtains depth
Learning model；
Multi-objective particle swarm optimization algorithm (MOPSO) is that PSO algorithms are transported in the improvement of classical PSO optimization algorithms
For deep learning, it may appear that since parameter involved by deep learning is numerous, thus need the content optimized is equally many to ask
Topic, and the PSO optimization algorithms that this is single goal cann't be solved, and propose MOPSO algorithms thus.It is selected in single object optimization
When individual optimum position P, it is only necessary to which can be judged more preferably by being compared, but for deep learning network, it is all
For CNN, which the comparison of two particles can not immediately arrive at more preferably, if in the case that multiple targets are best simultaneously,
Certain particle has best effect, but in practice, different advantage places is usually had from each other.Likewise,
For group optimum position, for the PSO of single goal, there are one optimized individuals for meeting in population, but for more mesh
For mark optimization, optimal individual might have very much.MOPSO is for the solution of individual optimum position, cannot be tight
Lattice contrast which position it is better in the case of, randomly choose out a position best as history, and best for group
Position selects one according to the densely distributed degree of individual optimum position and is referred to as " leader " in the optimal collection of individual
Optimum position, this leader generally select a not intensive position particle and serve as.An adaptive mesh method is used herein
The update of group optimum position and individual optimum position is carried out, concrete thought is according to mesh generation, it is assumed that in each grid
Population n, b which grid represented, the selected probability of particle is in the gridThe more crowded selection of particle
Probability is lower.This is to ensure that unknown solution space region can be explored, and this point is outstanding in the optimization of CNN networks
Its is important.
It is of the present invention to optimize and revise preliminary convolutional Neural net using multi-objective particle swarm optimization algorithm (MOPSO)
Network obtains deep learning model, including：
(1) parameter group and the target component collection (the archive collection of optimum position) of preliminary convolutional neural networks, institute are initialized
The preliminary convolutional neural networks parameter group stated includes structure and corresponding each hyper parameter；
Including giving parameter group to assign an initial value at random, initial convolutional neural networks parameter group P1 is generated, is incited somebody to action
Best solution deposit target component collection in initial convolutional neural networks parameter group P1 is achieved as initial optimum position
Collect A1.It is used for judging herein in the standard that it is good and bad, accounts for first and judge that position is the parameter group hypencephalon electric data collecting point
Secondly the accuracy rate of class reaches requirement, structure with wide usage etc. for operation required time, parameter.
(2) it carries out Particle Swarm movement and obtains new target component collection；It is j including setting the particle for making currently to be moved, into
The following process of row：
(2.1) dense information that target component concentrates particle is calculated, specifically target component collection space grid decile
At several regions, use the population that each region is included as the density information of particle；Include in region where particle
Population is more, and the area density values are bigger, otherwise smaller, is implemented as：
It calculates in parameter group iterations upper limit value t, the boundary in target component collection spaceMeter
Calculate the mould of gridWherein M refers to the region that grid is divided
Sum, round numbers, F1 tWithFor target function value；The particle concentrated in target component is traversed, the grain that target component is concentrated is calculated
The number of grid where subWherein Int is bracket function；It calculates
The density estimation value of gridding information and particle；
(2.2) particle P in setting parameter groupJ, tIt is to concentrate a best particle G in target componentJ, t, particle GJ, tMatter
Amount determines that the constringency performance and diversity of multi-objective particle swarm optimization algorithm, selection gist are the particles that target component is concentrated
Density information；Wherein j is the label of current kinetic particle, and t is the value of current iteration number；It is better than parameter with target component collection particle
The population of group carrys out the search potentiality of evaluation goal parameter set particle, more better than the target component collection particle of parameter group,
Target component collection search potentiality are bigger, and algorithm is specific as follows：
It calculates target component and concentrates and be better than particle PJ, tParticle collection Aj, for integer k from 1 to AtThe population for including
In range, Aj=Aj+{AK, t|AK, tp PJ, t, AK, t∈At}；Then particle collection A is calculatedjThe particle collection G of middle density minimumj, Gj=
min{Density(Ak), k=1,2 ..., | Aj|, Ak∈Aj}；Wherein, AjIt is used to store target component collection AtIn be better than particle
PJ, tParticle collection, AjThe particle of middle density minimum is stored in particle collection GjIn；AK, tMiddle t refers to the t times iteration, same primary
It is A that iterations are omitted in iterationk；Density(Ak) it is to calculate particle AkDensity estimation value；
(2.3) in undated parameter group particle position and speed drawing in the global best particle G and best particle P of individual
Lower search optimal solution is led, is update target component collection；
(2.4) blocked operation for carrying out target component collection, it is more than defined amount to avoid population；
(2.5) particle information for exporting target component collection, becomes new target component collection.
(3) it is inputted the new target component collection obtained as the parameter of new convolutional neural networks, substitutes into eeg data
Sample set x (t) carries out desk checking, verifies the accuracy and efficiency of new convolutional neural networks, convolutional neural networks after verification,
As final convolutional neural networks, deep learning model is constituted.
5) multi-objective particle swarm optimization algorithm is used to realize multiple target movement auxiliary.
In the present invention, MOPSO algorithms, which also carry realization conversion single action auxiliary, becomes multiple target action auxiliary system
The effect of system.Its realize basic demand be：Ensure not influence and interfere the action miscellaneous function mainly run originally, and carries
The secondary action miscellaneous function realized for being as closely as possible to the normal human action of script, is assisted with safety action close to normal human action
Coordinate with brain signal.
The application of the deep learning model based on Multi-objective PSO optimization of the present invention, includes the following steps：
1) preliminary polynary limb action input parameter collection and target action parameter set, the preliminary polynary limbs are initialized
Action input parameter set is by obtained classification after the sample input deep learning model in eeg data sample set x (t)
Target action parameter set corresponding to information；Initialization group herein is the input parameter collection of another auxiliary movement device,
And the polynary limb action combined result that optimization aim is then actively formed as auxiliary mark by one.That is at this time we
It is the auxiliary device input value of the corresponding position for the action assisting device for being likely to require brain control system user as one
A input parameter collection is quickly found by MOPSO optimizations when carrying out action auxiliary and best suits current EEG signals trigger action
It exports it nearby or the input of relevant auxiliary device, formation one completely acts assistant system rather than some single pass
Section or muscle, using best combination of actions configuration as the output valve of parameter.Its criterion derives from repeated phase
The correlation for the action auxiliary system called in the same or similar movement imagination.It specifically includes：
An initial value is assigned at random to polynary limb action input parameter group, and it is defeated to generate initial polynary limb action
Enter parameter group P2, the best solution in initial polynary limb action input parameter group P2 is stored in target action parameter set
Collection A2 is achieved as initial optimum position.
2) it carries out Particle Swarm movement and obtains new target action parameter set；It is i including setting the particle for making currently to be moved,
Carry out following process：
(2.1) dense information for calculating particle in target action parameter set, specifically uses target action parameter set space
Grid is divided into several regions, uses the population that each region is included as the density information of particle；Region where particle
In include population it is more, the area density values are bigger, otherwise smaller, are implemented as：
It calculates in parameter group iterations upper limit value t, the boundary in target component collection spaceMeter
Calculate the mould of gridWherein M refers to the region that grid is divided
Sum, round numbers, F1 tWithFor target function value；The particle in target action parameter set is traversed, target action parameter is calculated
The number of grid where the particle of concentrationWherein Int is rounding letter
Number；Calculate the density estimation value of gridding information and particle；
(2.2) particle P in setting parameter groupJ, tIt is a best particle G in target action parameter setI, t, particle GI, t
Quality determine the constringency performance and diversity of multi-objective particle swarm optimization algorithm, selection gist is target action parameter set
In particle density information；Wherein i is the label of current kinetic particle, and t is the value of current iteration number；With target action parameter set
Particle carrys out the search potentiality of evaluation goal set of behavioural parameters particle better than the population of parameter group, is better than the target of parameter group
Set of behavioural parameters particle is more, and target action parameter set search potentiality are bigger, and algorithm is specific as follows：
It calculates and is better than particle P in target action parameter setI, tParticle collection Ai, for integer k from 1 to AtThe grain for including
Within the scope of subnumber, Ai=Ai+{AK, t|AK, tp PI, t, AK, t∈At}；Then particle collection A is calculatediThe particle collection G of middle density minimumi,
Gi=min { Density (Ak), k=1,2 ..., | Ai|, Ak∈Ai}；Wherein, AiIt is used to store target action parameter set AtIn it is excellent
In particle PI, tParticle collection, AiThe particle of middle density minimum is stored in particle collection GiIn；AK, tMiddle t refers to the t times iteration,
It is A with iterations are omitted in an iterationk；Density(Ak) it is to calculate particle AkDensity estimation value；
(2.3) in undated parameter group particle position and speed drawing in the global best particle G and best particle P of individual
Lower search optimal solution is led, is update target action parameter set；
(2.4) blocked operation for carrying out target action parameter set, it is more than defined amount to avoid population；
(2.5) particle information for exporting target action parameter set, becomes new target action parameter set.
3) the new target action parameter set that will be obtained, compared with the target action parameter set that subject repeatedly acts
Compared with whether the method validation screened using threshold value, the new target action parameter set is repeatedly to carry out classification judgement before
The target action parameter set of the highest frequency gone out is then as final target action parameter set, to realize corresponding auxiliary
It acts, otherwise return to step 2).
By for above-mentioned EEG signals carry out deep learning model foundation study processing, using MOPSO optimization algorithms into
The optimization design of row deep learning model, may finally be by above-mentioned approach application in the brain control equipment development of Mental imagery.
Its complete operational process is as follows：First, the EEG signals that target is acquired using electroencephalogramsignal signal collection equipment, with this
As input data set, a CNN is established, multiple-objection optimization is carried out to the network with MOPSO algorithms, it is accurate to make auxiliary with active
Rate is first object, and pair action accuracy rate, reaction efficiency etc. are the second target, and the classification for carrying out brain coron imaginary signals is distinguished
Know；Then, MOPSO auxiliary movement group's optimizer using the combination of priori cumulative information as criterion is used, it is excellent with this
Change device to carry out inputting optimization to the parameter of the systematicness auxiliary of Mental imagery action auxiliary；Finally, corresponding parameter is inputted excellent
Change result to be input in auxiliary movement device, so that it may accordingly to realize the correlation function of brain control equipment.
A kind of the deep learning model and application based on Multi-objective PSO optimization of the present invention, is first with brain
Electric data collecting equipment measures multi channel signals and passes through preconditioning technique, and eye is removed manually using such as the methods of ICA accordingly
The influences such as electric signal, and the frequency range of corresponding task is determined by the perception rhythm and pace of moving things of Mental imagery；By pretreated number
According to the training sample as deep learning network, CNN networks are inputted, will need to distinguish the brain instruction conduct recognized in actual task
The output classification of CNN is adjusted and more using MOPSO algorithms using the convolution sum pond operation training to numerous training samples
The parameter group of new network, optimizes the relevant parameter of network, and then a Mental imagery EEG signals can be obtained and distinguish
The deep learning model of knowledge, and single action auxiliary is extended using MOPSO algorithms, make complete action
Assistant system, in the exploitation of brain control equipment, deep learning network and follow-up system finally to be recognized the result for adjusting out
As the corresponding content of instruction classification of brain control equipment, the command adapted thereto that brain control equipment is carried out with this operates.
Above to the description of the present invention and embodiment, it is not limited to which this, the description in embodiment is only the reality of the present invention
One of mode is applied, it is without departing from the spirit of the invention, any not inventively to design and the technical solution
Similar structure or embodiment, belongs to protection scope of the present invention.
Claims (10)
1. a kind of deep learning model based on Multi-objective PSO optimization, which is characterized in that include the following steps：
1) EEG signals are acquired, the EEG signals of acquisition are pre-processed, it is two-dimensional p-channel brain teledata to obtain signal
Sample set x (t)；
2) deep learning model, that is, convolutional neural networks will be used as after p dimension brain teledata x (t) samples normalizations in step 1)
The input of a data, will the corresponding imagination classes of instructions as the output of convolutional neural networks last layer；
3) preliminary convolutional neural networks are established；
4) preliminary convolutional neural networks are optimized and revised using multi-objective particle swarm optimization algorithm, obtain deep learning model；
5) multi-objective particle swarm optimization algorithm is used to realize multiple target movement auxiliary.
2. a kind of deep learning model based on Multi-objective PSO optimization according to claim 1, feature exist
In step 1) is to acquire EEG signals using the 32 channel electrode caps that international 10-20 systems define, and the EEG signals are bases
In the EEG signals of Mental imagery, acquired content is fixed limb motion Imaginary Movement.
3. a kind of deep learning model based on Multi-objective PSO optimization according to claim 1, feature exist
In, the EEG signals to acquisition described in step 1) pre-process, including：Bandpass filtering, samples normalization and rejecting have abnormal
The EEG signals of type, wherein the range of the bandpass filtering is corresponding according to the selection of the characteristic of Mental imagery inter-related task
Brain wave frequency range.
4. a kind of deep learning model based on Multi-objective PSO optimization according to claim 1, feature exist
In step 3) includes：
(1) sample is chosen from eeg data sample set x (t) enter convolutional neural networks；
(2) it calculates the sample and enters the reality output that convolutional neural networks obtain, in this stage, information is passed through from input layer
Transformation step by step, is transmitted to output layer, this process is also the normal flow of network completion after training；
(3) difference of the ideal output of reality output and respective sample is calculated；
(4) weighed value adjusting is carried out according to the method for minimization error, obtains preliminary convolutional neural networks.
5. a kind of deep learning model based on Multi-objective PSO optimization according to claim 1, feature exist
In step 4) includes：
(1) parameter group and the target component collection of preliminary convolutional neural networks, preliminary convolutional neural networks ginseng are initialized
Number group includes structure and corresponding each hyper parameter；
(2) it carries out Particle Swarm movement and obtains new target component collection；
(3) it is inputted the new target component collection obtained as the parameter of new convolutional neural networks, substitutes into eeg data sample
Collect x (t) and carry out desk checking, verifies the accuracy and efficiency of new convolutional neural networks, convolutional neural networks after verification, as
Final convolutional neural networks constitute deep learning model.
6. a kind of deep learning model based on Multi-objective PSO optimization according to claim 5, feature exist
In, (1) step includes assigning an initial value at random to parameter group, generates initial convolutional neural networks parameter group P1,
Best solution deposit target component collection in initial convolutional neural networks parameter group P1 is deposited as initial optimum position
Shelves collection A1.
7. a kind of deep learning model based on Multi-objective PSO optimization according to claim 5, feature exist
Include that set the particle for making currently to be moved be j in, (2) step, carries out following process：
(2.1) dense information that target component concentrates particle is calculated, if specifically target component collection space is divided into grid
Dry region, uses the population that each region is included as the density information of particle；The particle for including in region where particle
Number is more, and the area density values are bigger, otherwise smaller, is implemented as：
It calculates in parameter group iterations upper limit value t, the boundary in target component collection spaceCalculate net
The mould of latticeWherein M refers to that the region that grid is divided is total
Number, round numbers, F1 tWithFor target function value；The particle concentrated in target component is traversed, the particle that target component is concentrated is calculated
The number of place gridWherein Int is bracket function；Calculate net
The density estimation value of lattice information and particle；
(2.2) particle P in setting parameter groupJ, tIt is to concentrate a best particle G in target componentJ, t, particle GJ, tQuality determine
The constringency performance and diversity of multi-objective particle swarm optimization algorithm are determined, selection gist is the particle density letter that target component is concentrated
Breath；Wherein j is the label of current kinetic particle, and t is the value of current iteration number；With target component collection particle better than parameter group
Population carrys out the search potentiality of evaluation goal parameter set particle, more better than the target component collection particle of parameter group, target ginseng
Manifold search potentiality are bigger, and algorithm is specific as follows：
It calculates target component and concentrates and be better than particle PJ, tParticle collection Aj, for integer k from 1 to AtThe population range for including
It is interior, Aj=Aj+{AK, t|AK, tp PJ, t, AK, t∈At}；Then particle collection A is calculatedjThe particle collection G of middle density minimumj, Gj=min
{Density(Ak), k=1,2 ..., | Aj|, Ak∈Aj}；Wherein, AjIt is used to store target component collection AtIn be better than particle PJ, t's
Particle collection, AjThe particle of middle density minimum is stored in particle collection GjIn；AK, tMiddle t refers to the t times iteration, in same an iteration
Omission iterations are Ak；Density(Ak) it is to calculate particle AkDensity estimation value；
(2.3) in undated parameter group the position and speed of particle under the guiding of global best particle G and the best particle P of individual
Optimal solution is searched, is update target component collection；
(2.4) blocked operation for carrying out target component collection, it is more than defined amount to avoid population；
(2.5) particle information for exporting target component collection, becomes new target component collection.
8. a kind of application of the deep learning model described in claim 1 based on Multi-objective PSO optimization, feature
It is, includes the following steps：
1) preliminary polynary limb action input parameter collection and target action parameter set, the preliminary polynary limb action are initialized
Input parameter collection is by obtained classification information after the sample input deep learning model in eeg data sample set x (t)
Corresponding target action parameter set；
2) it carries out Particle Swarm movement and obtains new target action parameter set；
3) the new target action parameter set that will be obtained, compared with the target action parameter set that subject repeatedly acts,
The method validation screened using threshold value, whether the new target action parameter set repeatedly carries out classification before is judged
The target action parameter set of highest frequency, be then as final target action parameter set, to realize corresponding auxiliary movement,
Otherwise return to step 2).
9. the application of the deep learning model according to claim 8 based on Multi-objective PSO optimization, feature
It is, step 1) includes assigning an initial value at random to polynary limb action input parameter group, generates initial polynary limb
Body action input parameter group P2 moves the best solution deposit target in initial polynary limb action input parameter group P2
Make parameter set and achieves collection A2 as initial optimum position.
10. the application of the deep learning model according to claim 8 based on Multi-objective PSO optimization, feature
It is, (2) step includes that set the particle for making currently to be moved be i, carries out following process：
(2.1) dense information for calculating particle in target action parameter set, specifically target action parameter set space grid
Several regions are divided into, use the population that each region is included as the density information of particle；Region Zhong Bao where particle
The population contained is more, and the area density values are bigger, otherwise smaller, is implemented as：
It calculates in parameter group iterations upper limit value t, the boundary in target component collection spaceCalculate net
The mould of latticeWherein M refers to that the region that grid is divided is total
Number, round numbers, F1 tWithFor target function value；The particle in target action parameter set is traversed, target action parameter set is calculated
In particle where grid numberWherein Int is rounding letter
Number；Calculate the density estimation value of gridding information and particle；
(2.2) particle P in setting parameter groupJ, tIt is a best particle G in target action parameter setI, t, particle GI, tMatter
Amount determines the constringency performance and diversity of multi-objective particle swarm optimization algorithm, and selection gist is the grain in target action parameter set
Sub- density information；Wherein i is the label of current kinetic particle, and t is the value of current iteration number；It is excellent with target action parameter set particle
Carry out the search potentiality of evaluation goal set of behavioural parameters particle in the population of parameter group, the target action better than parameter group is joined
Manifold particle is more, and target action parameter set search potentiality are bigger, and algorithm is specific as follows：
It calculates and is better than particle P in target action parameter setI, tParticle collection Ai, for integer k from 1 to AtThe population for including
In range, Ai=Ai+{AK, t|AK, tp PI, t, AK, t∈At}；Then particle collection A is calculatediThe particle collection G of middle density minimumi, Gi=
min{Density(Ak), k=1,2 ..., | Ai|, Ak∈Ai}；Wherein, AiIt is used to store target action parameter set AtIn be better than grain
Sub- PI, tParticle collection, AiThe particle of middle density minimum is stored in particle collection GiIn；AK, tMiddle t refers to the t times iteration, same
It is A that iterations are omitted in secondary iterationk；Density(Ak) it is to calculate particle AkDensity estimation value；
(2.3) in undated parameter group the position and speed of particle under the guiding of global best particle G and the best particle P of individual
Optimal solution is searched, is update target action parameter set；
(2.4) blocked operation for carrying out target action parameter set, it is more than defined amount to avoid population；
(2.5) particle information for exporting target action parameter set, becomes new target action parameter set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810169310.4A CN108319928B (en) | 2018-02-28 | 2018-02-28 | Deep learning method and system based on multi-target particle swarm optimization algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810169310.4A CN108319928B (en) | 2018-02-28 | 2018-02-28 | Deep learning method and system based on multi-target particle swarm optimization algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108319928A true CN108319928A (en) | 2018-07-24 |
| CN108319928B CN108319928B (en) | 2022-04-19 |
Family
ID=62900991
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810169310.4A Active CN108319928B (en) | 2018-02-28 | 2018-02-28 | Deep learning method and system based on multi-target particle swarm optimization algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108319928B (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109003680A (en) * | 2018-09-28 | 2018-12-14 | 四川大学 | Epilepsy data statistical approach and device |
| CN109063319A (en) * | 2018-07-27 | 2018-12-21 | 天津大学 | A kind of analogy method of bioecosystem neural network based |
| CN109596326A (en) * | 2018-11-30 | 2019-04-09 | 电子科技大学 | Rotary machinery fault diagnosis method based on optimization structure convolutional neural networks |
| CN110221684A (en) * | 2019-03-01 | 2019-09-10 | Oppo广东移动通信有限公司 | Apparatus control method, system, electronic device and computer readable storage medium |
| CN110414732A (en) * | 2019-07-23 | 2019-11-05 | 中国科学院地理科学与资源研究所 | A kind of trip Future Trajectory Prediction method, apparatus, storage medium and electronic equipment |
| CN110584597A (en) * | 2019-07-15 | 2019-12-20 | 天津大学 | Multi-channel electroencephalogram signal monitoring method based on time-space convolutional neural network and application |
| CN111436929A (en) * | 2019-01-17 | 2020-07-24 | 复旦大学 | Method for generating and identifying neurophysiological signals |
| WO2021007812A1 (en) * | 2019-07-17 | 2021-01-21 | 深圳大学 | Deep neural network hyperparameter optimization method, electronic device and storage medium |
| CN112605983A (en) * | 2020-12-01 | 2021-04-06 | 浙江工业大学 | Mechanical arm pushing and grabbing system suitable for intensive environment |
| CN112990224A (en) * | 2021-02-04 | 2021-06-18 | 郑州航空工业管理学院 | Particle swarm method for identifying brain function partition from fMRI data |
| CN116849942A (en) * | 2023-07-28 | 2023-10-10 | 中国医学科学院生物医学工程研究所 | Brain-control intelligent lifting and turning-over multifunctional medical care bed |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101711709A (en) * | 2009-12-07 | 2010-05-26 | 杭州电子科技大学 | Method for controlling electrically powered artificial hands by utilizing electro-coulogram and electroencephalogram information |
| CN104899436A (en) * | 2015-05-29 | 2015-09-09 | 北京航空航天大学 | Electroencephalogram signal time-frequency analysis method based on multi-scale radial basis function and improved particle swarm optimization algorithm |
| CN105447567A (en) * | 2015-11-06 | 2016-03-30 | 重庆科技学院 | BP neural network and MPSO algorithm-based aluminium electrolysis energy-saving and emission-reduction control method |
| CN105740887A (en) * | 2016-01-26 | 2016-07-06 | 杭州电子科技大学 | Electroencephalogram feature classification method based on PSO-SVM (Particle Swarm Optimization-Support Vector Machine) |
| WO2017014599A1 (en) * | 2015-07-22 | 2017-01-26 | 주식회사 씨코헬스케어 | Composition for stabilizing radiochemical purity of [18f]fluoro-dopa and method for preparing same |
| CN106821681A (en) * | 2017-02-27 | 2017-06-13 | 浙江工业大学 | A kind of upper limbs ectoskeleton control method and system based on Mental imagery |
-
2018
- 2018-02-28 CN CN201810169310.4A patent/CN108319928B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101711709A (en) * | 2009-12-07 | 2010-05-26 | 杭州电子科技大学 | Method for controlling electrically powered artificial hands by utilizing electro-coulogram and electroencephalogram information |
| CN104899436A (en) * | 2015-05-29 | 2015-09-09 | 北京航空航天大学 | Electroencephalogram signal time-frequency analysis method based on multi-scale radial basis function and improved particle swarm optimization algorithm |
| WO2017014599A1 (en) * | 2015-07-22 | 2017-01-26 | 주식회사 씨코헬스케어 | Composition for stabilizing radiochemical purity of [18f]fluoro-dopa and method for preparing same |
| CN105447567A (en) * | 2015-11-06 | 2016-03-30 | 重庆科技学院 | BP neural network and MPSO algorithm-based aluminium electrolysis energy-saving and emission-reduction control method |
| CN105740887A (en) * | 2016-01-26 | 2016-07-06 | 杭州电子科技大学 | Electroencephalogram feature classification method based on PSO-SVM (Particle Swarm Optimization-Support Vector Machine) |
| CN106821681A (en) * | 2017-02-27 | 2017-06-13 | 浙江工业大学 | A kind of upper limbs ectoskeleton control method and system based on Mental imagery |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109063319A (en) * | 2018-07-27 | 2018-12-21 | 天津大学 | A kind of analogy method of bioecosystem neural network based |
| CN109063319B (en) * | 2018-07-27 | 2023-04-07 | 天津大学 | Simulation method of biological ecosystem based on neural network |
| CN109003680A (en) * | 2018-09-28 | 2018-12-14 | 四川大学 | Epilepsy data statistical approach and device |
| CN109596326A (en) * | 2018-11-30 | 2019-04-09 | 电子科技大学 | Rotary machinery fault diagnosis method based on optimization structure convolutional neural networks |
| CN111436929B (en) * | 2019-01-17 | 2021-06-01 | 复旦大学 | Method for generating and identifying neurophysiological signals |
| CN111436929A (en) * | 2019-01-17 | 2020-07-24 | 复旦大学 | Method for generating and identifying neurophysiological signals |
| CN110221684A (en) * | 2019-03-01 | 2019-09-10 | Oppo广东移动通信有限公司 | Apparatus control method, system, electronic device and computer readable storage medium |
| CN110584597A (en) * | 2019-07-15 | 2019-12-20 | 天津大学 | Multi-channel electroencephalogram signal monitoring method based on time-space convolutional neural network and application |
| WO2021007812A1 (en) * | 2019-07-17 | 2021-01-21 | 深圳大学 | Deep neural network hyperparameter optimization method, electronic device and storage medium |
| CN110414732A (en) * | 2019-07-23 | 2019-11-05 | 中国科学院地理科学与资源研究所 | A kind of trip Future Trajectory Prediction method, apparatus, storage medium and electronic equipment |
| CN110414732B (en) * | 2019-07-23 | 2020-09-18 | 中国科学院地理科学与资源研究所 | Travel future trajectory prediction method and device, storage medium and electronic equipment |
| CN112605983A (en) * | 2020-12-01 | 2021-04-06 | 浙江工业大学 | Mechanical arm pushing and grabbing system suitable for intensive environment |
| CN112605983B (en) * | 2020-12-01 | 2022-04-19 | 浙江工业大学 | Mechanical arm pushing and grabbing system suitable for intensive environment |
| CN112990224A (en) * | 2021-02-04 | 2021-06-18 | 郑州航空工业管理学院 | Particle swarm method for identifying brain function partition from fMRI data |
| CN112990224B (en) * | 2021-02-04 | 2023-07-11 | 郑州航空工业管理学院 | Particle swarm method for identifying brain function division from fMRI data |
| CN116849942A (en) * | 2023-07-28 | 2023-10-10 | 中国医学科学院生物医学工程研究所 | Brain-control intelligent lifting and turning-over multifunctional medical care bed |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108319928B (en) | 2022-04-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108319928A (en) | A kind of deep learning model and application based on Multi-objective PSO optimization | |
| Jaafra et al. | Reinforcement learning for neural architecture search: A review | |
| Martín et al. | Evodeep: a new evolutionary approach for automatic deep neural networks parametrisation | |
| Karlik et al. | A fuzzy clustering neural network architecture for multifunction upper-limb prosthesis | |
| Kasabov | Evolving connectionist systems: Methods and applications in bioinformatics, brain study and intelligent machines | |
| CN110110707A (en) | Artificial intelligence CNN, LSTM neural network dynamic identifying system | |
| CN108345846A (en) | A kind of Human bodys' response method and identifying system based on convolutional neural networks | |
| US20200334451A1 (en) | Motion behavior pattern classification method, system and device | |
| Mici et al. | A self-organizing neural network architecture for learning human-object interactions | |
| CN109919245A (en) | Deep learning model training method and device, training equipment and storage medium | |
| Gallagher | Multi-layer perceptron error surfaces: visualization, structure and modelling | |
| Baysal et al. | Multi-objective symbiotic organism search algorithm for optimal feature selection in brain computer interfaces | |
| CN109977394A (en) | Text model training method, text analyzing method, apparatus, equipment and medium | |
| CN110503082A (en) | A kind of model training method and relevant apparatus based on deep learning | |
| CN114330541A (en) | Road traffic accident risk prediction deep learning algorithm | |
| Moore et al. | Validation of a convolutional neural network model for spike transformation using a generalized linear model | |
| Wei et al. | Binary multi-objective particle swarm optimization for channel selection in motor imagery based brain-computer interfaces | |
| Nanni et al. | Building ensemble of deep networks: convolutional networks and transformers | |
| CN114781441A (en) | EEG motor imagery classification method and multi-space convolution neural network model | |
| Dinesh et al. | Reliable evaluation of neural network for multiclass classification of real-world data | |
| Zhu et al. | Support vector machine optimized using the improved fish swarm optimization algorithm and its application to face recognition | |
| Chandrasekaran et al. | Topology constraint free fuzzy gated neural networks for pattern recognition | |
| Liu | Deep spiking neural networks | |
| Lakshmi | A review on deep learning algorithms in healthcare | |
| Liu et al. | Multiple objects tracking based on snake model and selective attention mechanism |
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 |