CN110503082A - A kind of model training method and relevant apparatus based on deep learning - Google Patents
A kind of model training method and relevant apparatus based on deep learning Download PDFInfo
- Publication number
- CN110503082A CN110503082A CN201910818363.9A CN201910818363A CN110503082A CN 110503082 A CN110503082 A CN 110503082A CN 201910818363 A CN201910818363 A CN 201910818363A CN 110503082 A CN110503082 A CN 110503082A
- Authority
- CN
- China
- Prior art keywords
- training
- model
- eeg signal
- label
- electroencephalogram
- 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
- 238000012549 training Methods 0.000 title claims abstract description 217
- 238000000034 method Methods 0.000 title claims abstract description 65
- 238000013135 deep learning Methods 0.000 title claims abstract description 24
- 230000003340 mental effect Effects 0.000 claims abstract description 82
- 230000006870 function Effects 0.000 claims abstract description 66
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 23
- 238000001914 filtration Methods 0.000 claims description 25
- 230000009471 action Effects 0.000 claims description 24
- 230000008569 process Effects 0.000 claims description 22
- 238000005457 optimization Methods 0.000 claims description 21
- 238000012216 screening Methods 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 7
- 239000000284 extract Substances 0.000 claims description 6
- 238000001514 detection method Methods 0.000 claims description 5
- 230000002123 temporal effect Effects 0.000 claims description 5
- 238000001727 in vivo Methods 0.000 abstract description 5
- 230000000875 corresponding effect Effects 0.000 description 98
- 239000000523 sample Substances 0.000 description 75
- 210000004556 brain Anatomy 0.000 description 13
- 238000010586 diagram Methods 0.000 description 11
- 241001269238 Data Species 0.000 description 5
- 230000005611 electricity Effects 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 239000002699 waste material Substances 0.000 description 3
- 206010019468 Hemiplegia Diseases 0.000 description 2
- 230000008878 coupling Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 230000004424 eye movement Effects 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000003542 behavioural effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007635 classification algorithm Methods 0.000 description 1
- 230000001427 coherent effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 239000012497 inhomogeneous sample Substances 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 230000002085 persistent effect Effects 0.000 description 1
- 238000011897 real-time detection Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000005204 segregation Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
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
Landscapes
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The Mental imagery EEG signal that Different Individual acquires has been carried out information sharing by constructing disaggregated models based on multiple individuals by this application discloses a kind of model training method and relevant apparatus based on deep learning；And the accuracy rate of Mental imagery classification is further improved for the determination of model loss function by multiple sub- loss functions, reduce the difference of sample from multiple dimensions, and then reduces the problem that difference is big in a class as caused by local environment or the different state of mind in vivo；In addition individual difference alienation bring algorithm application limitation is also avoided, the relevant common information of Different Individual Mental imagery movement label is extracted, reduces the difference between individual, improve the accuracy of model training.
Description
Technical field
This application involves field of computer technology more particularly to a kind of model training methods and phase based on deep learning
Close device.
Background technique
Deep learning (Deep learning, DL) is a multi-field cross discipline, is related to probability theory, statistics, approaches
By multiple subjects such as, convextiry analysis, algorithm complexity theories.Specialize in the study row that the mankind were simulated or realized to computer how
To reorganize the existing structure of knowledge and being allowed to constantly improve the performance of itself, especially transporting to obtain new knowledge or skills
Dynamic imagination brain-computer interface (Motor imagery-Brain computer interface, MI-BCI) system, which has, widely answers
With prospect, the Rehabilitation training of the limbs such as hemiplegia inconvenience can be not only helped, control object, which is realized, takes care of oneself, and can also give pleasure to
Happy ordinary group, such as brain machine enhancing reality-virtualizing game etc..And the brain electricity based on Mental imagery (Motor imagery, MI)
Modulation recognition is the key link of MI-BCI, and decoded accuracy directly affects the performance of MI-BCI system.
General MI sorting algorithm is that a disaggregated model is individually trained for everyone, that is, extracts Imaginary Movement of doing exercises
The corresponding relationship of label and the EEG signals determined based on electroencephalogram is trained model, to carry out to the EEG signals of input
Classification obtains corresponding Imaginary Movement label of doing exercises.
But the above method needs successively individually to establish independent model, cause in the detection process of multiple individuals
Training process time-consuming is cumbersome, and due to the otherness of individual and the fluctuation of signal, causes classification performance not good enough, Jin Erying
Ring the accuracy of model training.
Summary of the invention
In view of this, the application first aspect provides a kind of model training method based on deep learning, mould can be applied to
In type training system or program process, specifically include: obtain multiple individuals electroencephalogram (Electroencephalograph,
EEG) and determine that corresponding EEG signal sample, the EEG signal sample are used to indicate corresponding Mental imagery movement label；
It determines the label information of multiple individuals and establishes the corresponding relationship with Mental imagery movement label, with true
Determine training dataset；
The multiple corresponding relationships concentrated based on the training data determine multiple sub- loss functions respectively, to be combined simultaneously
Determine that model loss function, multiple corresponding relationships that the training data is concentrated include the EEG signal sample of multiple individuals
It is dynamic with the corresponding relationship, the single individual EEG signal sample and the Mental imagery of Mental imagery movement label
Make the label information of label, the EEG signal sample and multiple individuals；
EEG decoded model is constructed according to the model loss function and is initialized；
Using the training dataset as input, by minimize the model loss function to electroencephalogram decoded model into
Row training, to obtain the disaggregated model of parameter optimization, the disaggregated model is used to export corresponding movement according to the electroencephalogram
The label information of Imaginary Movement label and corresponding individual.
Preferably, in some possible implementations of the application, it is described according to the training dataset to parameter optimization
The decoded model afterwards is trained, to obtain disaggregated model, comprising:
The training dataset is inputted at least one time convolutional layer, to obtain time training dataset, when described
Between training dataset be used to indicate the corresponding relationship that the training data is concentrated under different time；
The time training dataset is inputted at least one spatial convoluted layer, to obtain one-dimensional space training data
Collection, the one-dimensional space training dataset are used to indicate the corresponding relationship that the training data is concentrated under target action；
The time training dataset and the one-dimensional space training dataset are subjected to average pond, to obtain Chi Huate
Collection；
The feature in the pond feature set is extracted by least one full articulamentum, to obtain training characteristics collection；
The decoded model after parameter optimization is trained according to the training characteristics collection, to obtain disaggregated model.
Preferably, described to input the time training dataset at least in some possible implementations of the application
In one spatial convoluted layer, to obtain one-dimensional space training dataset, comprising:
Label, which is acted, according to the Mental imagery determines that multiple relevant spatial channels, the spatial channel refer to for filtering
Determine the corresponding signal of the target action；
The time training dataset is input in the multiple relevant spatial channel and is handled, it is one-dimensional to obtain
Space training dataset.
Preferably, described that EEG solution is constructed according to the model loss function in some possible implementations of the application
Code model is simultaneously initialized, comprising:
The training parameter is initialized according to preset algorithm；
The weight information for the corresponding relationship that the training data is concentrated is obtained, to determine according to the multiple sub- loss function
The model loss function；
The training parameter after initialization is optimized using gradient descent algorithm according to the model loss function.
Preferably, in some possible implementations of the application, the electroencephalogram EEG for obtaining multiple individuals and determination
Corresponding EEG signal sample, comprising:
Obtain the EEG of multiple individuals；
Channel corresponding with the EEG is selected according to preset rules, the EEG signal sample is obtained with screening, it is described pre-
If target action information setting of the rule based on EEG instruction.
Preferably, described corresponding with the EEG according to preset rules selection in some possible implementations of the application
Channel, with screening obtain the EEG signal sample, comprising:
The target action information is determined according to the EEG；
Select the corresponding frequency filtering of the target action information；
Corresponding channel is selected according to the frequency filtering, the EEG signal sample is obtained with screening.
Preferably, in some possible implementations of the application, after the screening obtains the EEG signal sample,
The method also includes:
Fluctuation information is determined according to the EEG signal sample；
If the fluctuation information meets preset condition, the EEG signal sample is passed through into exponentially weighted moving average meter
It calculates, to obtain the EEG signal sample after noise filtering, the preset condition is based on the fluctuation information and preset threshold
Size relation determines.
The application second aspect provides the device of another model training, comprising: acquiring unit, for obtaining multiple individuals
Electroencephalogram EEG and determine corresponding EEG signal sample, the EEG signal sample is used to indicate corresponding Mental imagery movement
Label；
Determination unit, for determining the label information of multiple individuals and establishing and Mental imagery movement label
Corresponding relationship, to determine training dataset；
The determination unit is also used to determine multiple sub- damages respectively based on multiple corresponding relationships that the training data is concentrated
Function is lost, to be combined and determine model loss function, multiple corresponding relationships that the training data is concentrated include multiple institutes
State the EEG signal sample of individual and the corresponding relationship of Mental imagery movement label, the single individual EEG signal sample
The label information of this and Mental imagery movement label, the EEG signal sample and multiple individuals；
Optimize unit, for constructing EEG decoded model according to the model loss function and being initialized；
Training unit is used for using the training dataset as input, by minimizing the model loss function to brain
Electrograph decoded model is trained, and to obtain the disaggregated model of parameter optimization, the disaggregated model is used for according to the electroencephalogram
Export the label information of corresponding Mental imagery movement label and corresponding individual.
Preferably, in some possible implementations of the application,
The training unit, specifically for inputting the training dataset at least one time convolutional layer, to obtain
Time training dataset, the time training dataset are used to indicate the corresponding pass that the training data under different time is concentrated
System；
The training unit, specifically for the time training dataset is inputted at least one spatial convoluted layer, with
One-dimensional space training dataset is obtained, the one-dimensional space training dataset is used to indicate the training dataset under target action
In corresponding relationship；
The training unit is specifically used for carrying out the time training dataset and the one-dimensional space training dataset
Average pond, to obtain pond feature set；
The training unit, specifically for extracting the feature in the pond feature set by least one full articulamentum,
To obtain training characteristics collection；
The training unit, specifically for being carried out according to the training characteristics collection to the decoded model after parameter optimization
Training, to obtain disaggregated model.
Preferably, in some possible implementations of the application, the training unit is specifically used for according to the movement
Imaginary Movement label determines that multiple relevant spatial channels, the spatial channel specify the target action corresponding for filtering
Signal；
The training unit, specifically for the time training dataset is input to the multiple relevant spatial channel
In handled, to obtain one-dimensional space training dataset.
Preferably, in some possible implementations of the application, the optimization unit is specifically used for according to preset algorithm
Initialize the training parameter；
The optimization unit, specifically for obtaining the weight information for the corresponding relationship that the training data is concentrated, with basis
The multiple sub- loss function determines the model loss function；
The optimization unit, specifically for according to the model loss function using gradient descent algorithm to initialization after
The training parameter optimizes.
Preferably, in some possible implementations of the application, the acquiring unit is specifically used for obtaining multiple individuals
EEG；
The acquiring unit, is specifically used for selecting channel corresponding with the EEG according to preset rules, obtains institute with screening
State EEG signal sample, the target action information setting that the preset rules are indicated based on the EEG.
Preferably, in some possible implementations of the application, the acquiring unit is specifically used for according to the EEG
Determine the target action information；
The acquiring unit is specifically used for selecting the corresponding frequency filtering of the target action information；
The acquiring unit, is specifically used for selecting corresponding channel according to the frequency filtering, obtains the EEG with screening
Sample of signal.
Preferably, in some possible implementations of the application, the acquiring unit is also used to be believed according to the EEG
Number sample determines fluctuation information；
The acquiring unit passes through the EEG signal sample if being also used to the fluctuation information meets preset condition
Exponentially weighted moving average calculates, and to obtain the EEG signal sample after noise filtering, the preset condition is based on the wave
The size relation of dynamic information and preset threshold determines.
The application third aspect provides a kind of electroencephalogram detection method based on brain-computer interface, specifically includes: being based on target
Object generates Mental imagery Training scene；Acquire multiple brains electricity that multiple individuals are generated according to the Mental imagery Training scene
Figure；The multiple electroencephalogram is pre-processed, and extracts multiple EEG signal samples in preset time period, the pre- place
The process of reason is determined based on the fluctuation information of the multiple electroencephalogram；The multiple EEG signal sample is inputted into classification mould
Type, to respectively obtain the label information and corresponding motion information label of the corresponding individual of the multiple EEG signal sample,
The disaggregated model is based on the described in any item model training methods based on deep learning of above-mentioned first aspect or first aspect
Gained.
The application fourth aspect provides a kind of computer equipment, comprising: memory, processor and bus system；It is described
Memory is for storing program code；The processor is used for according to the above-mentioned first aspect of instruction execution in said program code
Or the described in any item model training methods based on deep learning of first aspect.
The 5th aspect of the application provides a kind of computer readable storage medium, stores in the computer readable storage medium
There is instruction, when run on a computer, so that computer executes above-mentioned first aspect or first aspect is described in any item
Model training method based on deep learning.
As can be seen from the above technical solutions, the embodiment of the present application has the advantage that
It, can be well by the EEG signal progress information of different people acquisition by constructing disaggregated model based on multiple individuals
It is shared, solve the problems, such as first with model cannot using different personal datas and caused by data waste；And pass through multiple sub- damages
It loses function and the accuracy rate of Mental imagery classification is further improved for the determination of model loss function, reduce sample from multiple dimensions
This difference, and then reduce the problem that difference is big in a class as caused by local environment or the different state of mind in vivo；In addition
Individual difference alienation bring algorithm application limitation is also avoided, the relevant general character of Different Individual Mental imagery movement label is extracted
Information reduces the difference between individual, improves the accuracy of model training.
Detailed description of the invention
In order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
The embodiment of application for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is the network architecture diagram of model training systems operation；
Fig. 2 is a kind of model training process frame diagram；
Fig. 3 is a kind of flow chart of the model training method based on deep learning provided by the embodiments of the present application；
Fig. 4 is the flow chart of another model training method based on deep learning provided by the embodiments of the present application；
Fig. 5 is a kind of schematic diagram of interface display of model training provided by the embodiments of the present application；
Fig. 6 is the structural schematic diagram of model training apparatus provided by the embodiments of the present application；
Fig. 7 is the structural schematic diagram of another model training apparatus provided by the embodiments of the present application.
Specific embodiment
The embodiment of the present application provides a kind of model training method and relevant apparatus based on deep learning, can apply
It is corresponding especially by the electroencephalogram and determination that obtain multiple individuals in the training process of Mental imagery brain signal disaggregated model
EEG signal sample, the EEG signal sample are used to indicate corresponding Mental imagery movement label；Determine multiple individuals
Label information simultaneously establishes the corresponding relationship that label is acted with the Mental imagery, to determine training dataset；Based on the training
Multiple corresponding relationships in data set determine multiple sub- loss functions respectively, to be combined and determine model loss function, institute
The multiple corresponding relationships for stating training data concentration include the corresponding pass of the EEG signal sample with Mental imagery movement label
System, the label information of multiple individuals and the Mental imagery act between the corresponding relationship or multiple individuals of label
Corresponding relationship；EEG decoded model is constructed according to the model loss function and is initialized；According to the training dataset pair
The decoded model after parameter optimization is trained, to obtain disaggregated model.It, can be with based on multiple individual building disaggregated models
The EEG signal by different people acquisition carries out information sharing well, and solving cannot first be made with model using different personal datas
At data waste the problem of；And fortune is further improved for the determination of model loss function by multiple sub- loss functions
The accuracy rate of dynamic imagination classification reduces the difference of sample from multiple dimensions, and then reduction is in vivo due to local environment or difference
The big problem of difference in class caused by the state of mind；In addition individual difference alienation bring algorithm application limitation is also avoided, is extracted
The relevant common information of Different Individual Mental imagery movement label, reduces the difference between individual, improves model training
Accuracy.
The description and claims of this application and term " first ", " second ", " third ", " in above-mentioned attached drawing
The (if present)s such as four " are to be used to distinguish similar objects, without being used to describe a particular order or precedence order.It should manage
The data that solution uses in this way are interchangeable under appropriate circumstances, so that embodiments herein described herein for example can be to remove
Sequence other than those of illustrating or describe herein is implemented.In addition, term " includes " and " corresponding to " and their times
What is deformed, it is intended that cover it is non-exclusive include, for example, contain the process, method of a series of steps or units, system,
Product or equipment those of are not necessarily limited to be clearly listed step or unit, but may include be not clearly listed or for
The intrinsic other step or units of these process, methods, product or equipment.
It should be understood that model training method provided by the present application can be applied to the model of Mental imagery brain signal categorizing system
In training process, specifically, model training systems can be run in the network architecture as shown in Figure 1, as shown in Figure 1, being mould
The network architecture diagram of type training system operation, as figure shows, model training systems can obtain multiple individuals by brain-computer interface
Electroencephalogram information, and the data set obtained by Mental imagery database is associated as training set and electroencephalogram information,
And disaggregated model is trained, the disaggregated model that training obtains can be used for controlling external equipment, such as: electric wheelchair, machine
Device people, unmanned plane etc.；It can be used for carrying out behavioural analysis to the electroencephalogram information of limbs inconvenience person, it helped to restore movement energy
Power；It is understood that showing three individuals in Fig. 1, there can be more or fewer individuals to participate in actual scene
In sample collection procedure, depending on particular number is because of actual scene, herein without limitation；In addition, showing a picture number in Fig. 1
According to library, but in actual scene, there can also be the participation of multiple Mental imagery databases, especially in the interaction of more action datas
In scene, depending on specific Mental imagery quantity database is because of actual scene.
It is understood that above-mentioned model training systems can run on personal terminal, server can also be run on, also
It can be used as and run on third party device to provide the model training service of brain-computer interface, to obtain search report；Specific mould
Type training system can be to be run in above equipment in the form of a kind of program, can also be used as the Account Dept in above equipment
Part is run, and is also used as one kind of cloud service program and is not limited herein depending on specific operating mode is because of actual scene
It is fixed.
Deep learning is a multi-field cross discipline, is related to probability theory, statistics, Approximation Theory, convextiry analysis, algorithm complexity
The multiple subjects such as topology degree.Specialize in the learning behavior that the mankind were simulated or realized to computer how, with obtain new knowledge or
Technical ability reorganizes the existing structure of knowledge and is allowed to constantly improve the performance of itself, especially in Mental imagery brain machine interface system
It has a wide range of applications, can not only help the Rehabilitation training of the limbs such as hemiplegia inconvenience, control object is realized
It takes care of oneself, ordinary group, such as brain machine game etc. can also be entertained.And the eeg signal classification based on MI is the key that MI-BCI ring
Section, decoded accuracy directly affect the performance of MI-BCI system.
General MI sorting algorithm is that a disaggregated model is individually trained for everyone, that is, extracts Imaginary Movement of doing exercises
The corresponding relationship of label and the EEG signals determined based on electroencephalogram is trained model, to carry out to the EEG signals of input
Classification obtains corresponding Imaginary Movement label of doing exercises.
But the above method needs successively individually to establish independent model, cause in the detection process of multiple individuals
Training process time-consuming is cumbersome, and due to the otherness of individual and the fluctuation of signal, causes classification performance not good enough, Jin Erying
Ring the accuracy of model training.
To solve the above-mentioned problems, present applicant proposes a kind of model training method based on deep learning, this method is answered
For model training process frame shown in Fig. 2, as shown in Fig. 2, be a kind of model training process frame diagram, in figure to from
The EEG signal sample of multiple individuals carries out channel selecting, removes the signal unrelated with Mental imagery task, then carries out band logical filter
Wave processing, such as 8-32Hz, removal influence EEG signal sample bring by eye movement, ambient noise.Filtered signal is led to
It crosses exponentially weighted moving average operation and reduces the signal fluctuation as caused by noise, then input and carry out spy in convolutional neural networks model
Sign is extracted, and this feature contains Mental imagery correlated characteristic and individual special identity characteristic.Final mask classification provides defeated
The Mental imagery classification and the classification of signal institute source individual that the EEG signal sample entered is included.
It is understood that showing an EEG signal in figure, in actual scene, there can be more signal numbers
Or source, depending on particular number is because of actual scene, herein without limitation.
In conjunction with above-mentioned process frame, the model training method in the application based on deep learning will be introduced below,
Referring to Fig. 3, Fig. 3 is a kind of flow chart of the model training method based on deep learning provided by the embodiments of the present application, this Shen
Please embodiment at least include the following steps:
301, it obtains the electroencephalogram EEG of multiple individuals and determines corresponding EEG signal sample.
In the present embodiment, the EEG signal sample is used to indicate corresponding Mental imagery movement label, Mental imagery movement
Label is the data in Mental imagery database, such as: Mental imagery data set can be the hand exercise imagination in the present embodiment
Data set is also possible to foot motion imagination data set, can also be the corresponding Mental imagery data set in other positions of human body；It can
With understanding, one in above-mentioned data set can be selected in practical applications, can also be selected more in above-mentioned data set
It is a, herein without limitation.
It is understood that multiple individuals can be multiple users to be detected, it is also possible to the brain telecommunications of multiple simulations
Number set, can also be being used in mixed way for multiple users and the EEG signals of multiple simulations.Wherein, the EEG of multiple individuals is corresponding
EEG signal sample can be the brain electricity curve in certain time period, be also possible to real-time brain electricity curve and collect, specific data
Depending on the form of acquisition answers actual scene, herein without limitation.
In a kind of possible scene, since acquisition EEG is mainly for speculating Mental imagery label, other limbs at this time
Movement may have an impact EEG, such as: eye movement, lip moves or extraneous environmental noise, can be filtered at this time to EEG
Processing；Specifically, obtaining the EEG of multiple individuals first；Then channel corresponding with the EEG is selected according to preset rules, with
Screening obtains the EEG signal sample, wherein the target action information setting that preset rules are indicated based on the EEG, such as:
EEG instruction is movement imagination information, then the corresponding shape information of the general movement imagination information of selection；Further, for wave
The screening of shape information can be realized by the setting of frequency filtering, specifically, determining that the target action is believed according to the EEG
Breath；Then the corresponding frequency filtering of the target action information is selected；Then corresponding channel is selected according to the frequency filtering,
The EEG signal sample is obtained with screening, such as the filtering channel of setting 8-32Hz handles electroencephalogram waveform.
Optionally, certain deviation may be generated for the process of above-mentioned filtering, due to environmental factor further to subtract
Certain range can be arranged, specifically, true according to EEG signal sample first in few fluctuation generated by ambient noise for fluctuation
Determine fluctuation information；If the fluctuation information meets preset condition, the EEG signal sample is passed through into exponentially weighted moving average
It calculates, to obtain the EEG signal sample after noise filtering, wherein state preset condition and be based on the fluctuation information and preset
The size relation of threshold value determines, such as the fluctuation range of fluctuation information instruction is greater than 5% to meet preset condition.
302, it determines the label information of multiple individuals and establishes the corresponding relationship with Mental imagery movement label,
To determine training dataset；
In the present embodiment, by the data of all multiple individuals on the basis of original Mental imagery tag along sort, increase more
The label information of individual.At this point, being effectively utilized the training data of Different Individual, the size of training set is increased.
In addition, be to conduct the information between multiple individuals mutually, can extract the EEG signals of Different Individual with
The label information of individual and the motion information label unrelated with individual, wherein motion information label obtains more general for training
The Mental imagery classifier unrelated with individual difference, individual label information can be used for identifying the identity label of Different Individual.
303, the multiple corresponding relationships concentrated based on the training data determine multiple sub- loss functions respectively, to carry out group
Merge and determines model loss function.
In the present embodiment, multiple corresponding relationships that the training data is concentrated include the EEG signal sample of multiple individuals
Corresponding relationship, the single individual EEG signal sample and the Mental imagery of this and Mental imagery movement label
Act the corresponding relationship of label, the corresponding relationship of the EEG signal sample and the label information of multiple individuals.Wherein, respectively
A corresponding relationship has corresponded to a sub- loss function, in a kind of possible scene, to reduce the sample difference in each corresponding relationship
Different, the corresponding relationship of EEG signal sample and Mental imagery movement label for multiple individuals can use normalizing
Change figure penalties function (softmax loss)；It is corresponding with the label information of multiple individuals for EEG signal sample to close
System can use softmax loss；And label is acted for the single individual EEG signal sample and the Mental imagery
The relationship of corresponding relationship, i.e., the inhomogeneity EEG signal sample of same individual can be using comparison loss function (contrastive
loss)。
Specifically, it is contemplated that a internal uniformity signal difference, the contrastive loss that the present embodiment proposes is as reduction
The loss L of uniformity signal differencecont, to increase the robustness of algorithm.L is lost with another identification simultaneouslysubAssist director
Be engaged in Mental imagery classification Lmi.Identification loss can help to improve the accuracy rate of Mental imagery classification task, can also be rationally
The training data using different people.
In conclusion the model loss of the technical program is broadly divided into three parts: first is that, for judging Mental imagery classification
The softmax loss based on Mental imagery movement label, i.e. the corresponding of EEG signal sample and Mental imagery movement label close
System；Second is that input the individual softmax loss based on individual marking information belonging to EEG signal for judging, i.e., individual
The corresponding relationship of label information and Mental imagery movement label；Third is that for increasing an internal inter-class separability
Contrastive loss, i.e., the correspondence of the similar EEG signal sample motor message label between inhomogeneity of single individual
Relationship.The loss function of whole network model are as follows:
L=wmiLmi+wsubLsub+wcontLcont
Wherein LmiWith LsubCalculation formula be
Wherein, C is sample to number, siIt is i-th of value of output vector s, indicates that corresponding with vector s sample belongs to the
The probability of i classification；tiIndicate i-th of value in the corresponding relationship of sample.
LcontIt is the loss function for constraining the characteristic distance of sample pair, by minimizing this loss function, can makes same
Feature distribution between the sample pair of class is close as far as possible；The characteristic distance of inhomogeneous sample pair is more than certain boundary, according to
Above-mentioned rule, mathematic(al) representation are as follows:
Wherein, y is the whether matched label of two samples, and y=1 represents that two samples are similar or matching, y=0 then generation
Table mismatches；D is the European characteristic distance between two samples, and N is total sample to number, and margin is the boundary of sample.Measurement
Loss function between class at a distance from feature in class can equally be generalized for unknown losses, such as can use different characteristic distances
Measured (such as COS distance)；Different metric functions, such as triple loss function (triplet loss) etc., this
Outside, wmi、wsubAnd wcontFor the weight for respectively corresponding to sub- loss function, can be configured by related personnel.
304, EEG decoded model is constructed according to the model loss function and is initialized；
It, can be first according at the beginning of preset algorithm for the initialization procedure of the training parameter of decoded model in the present embodiment
The beginningization training parameter, the preset algorithm include Ze Weier initialization (xavier)；Then the training dataset is obtained
In corresponding relationship weight information, i.e. wmi、wsubAnd wcont, to determine the model according to the multiple sub- loss function
Loss function；Finally the training parameter after initialization is carried out using gradient descent algorithm according to the model loss function
Optimization.Such as: Mental imagery can be set and classify, the loss weight of identification and contrastive loss are respectively as follows:
1,1,0.2, and it is 1 that margin, which is arranged, in contrastive loss.
305, using the training dataset as input, mould is decoded to electroencephalogram by minimizing the model loss function
Type is trained, to obtain the disaggregated model of parameter optimization.
In the present embodiment, the disaggregated model is used to export corresponding Mental imagery movement label or right according to the EEG
The label information for the individual answered, it can obtain the label information of motion information label and individual.
Specifically, the training for decoded model may include steps of:
Step1: the training dataset is inputted at least one time convolutional layer, to obtain time training dataset,
The time training dataset is used to indicate the corresponding relationship that the training data is concentrated under different time.
Step2: the time training dataset is inputted at least one spatial convoluted layer, to obtain one-dimensional space training
Data set, the one-dimensional space training dataset are used to indicate the corresponding relationship that the training data is concentrated under target action.
Optionally, label being acted according to the Mental imagery and determining multiple relevant spatial channels, the space is logical
The corresponding signal of the target action is specified for filtering in road；Then the time training dataset is input to the multiple phase
It is handled in the spatial channel of pass, to obtain one-dimensional space training dataset.
Step3: the time training dataset and the one-dimensional space training dataset are subjected to average pond, to obtain
Pond feature set.
Step4: the feature in the pond feature set is extracted by least one full articulamentum, to obtain training characteristics
Collection.
Step5: the decoded model after parameter optimization is trained according to the training characteristics collection, to be classified
Model.
In conjunction with above-described embodiment it is found that can well be adopted different people by constructing disaggregated model based on multiple individuals
The EEG signal of collection carries out information sharing, solve first with model cannot using different personal datas and caused by asking of wasting of data
Topic；And the accurate of Mental imagery classification is further improved for the determination of model loss function by multiple sub- loss functions
Rate reduces the difference of sample from multiple dimensions, and then reduces a class as caused by local environment or the different state of mind in vivo
The big problem of interior difference；In addition individual difference alienation bring algorithm application limitation is also avoided, Different Individual movement is extracted and thinks
As the relevant common information of movement label, reduces the difference between individual, improve the accuracy of model training.
The training step for disaggregated model is indicated in conjunction with above-described embodiment, in the following, to specific under the training step
Training method is illustrated, as shown in figure 4, Fig. 4 is another model instruction based on deep learning provided by the embodiments of the present application
Practice the flow chart of method, the embodiment of the present application includes at least following procedure:
In the present embodiment, since the signal acquisition process of EEG can be real-time perfoming, there is regular hour characteristic,
And information collection can be carried out by different channels based on different Mental imagery movement labels.Therefore believed according to the EEG of acquisition
Number time and space characteristic, devise time convolutional layer (conv temporal) for the time with for spatial channel
Spatial convoluted layer (conv spatial), specific model parameter is referred to content shown in table 1.
For being directed to the time convolutional layer of time, time-based sample size, such as the sample frequency of signal are as set
For 1000Hz, input of the 1st second data of EEG signal as network model, number of sampling points 1000 are intercepted；In addition, right
20 channels relevant to Mental imagery, including FCz, FC1, FC2, FC3, FC4, FC5, FC6 have been selected in spatial convoluted layer choosing,
Cz, C1, C2, C3, C4, C5, C6, CP1, CP2, CP3, CP4, CP5, CP6, i.e., the signal of input network is having a size of 20x 1000；
Wherein the step-length (stride) of first two layers of time convolutional layer and spatial convoluted layer is all 1.
Network parameter table involved in 1 training process of table
| Model level title | Export specification | Time and space network parameter |
| Time convolutional layer | 20x 976 | 1x 25,30,stride 1 |
| Spatial convoluted layer | 1x 976 | 20x 1,30,stride 1 |
| Average pond layer | 1x 61 | 1x 75,stride 15 |
| Extensive layer | 1830 | - |
| Full articulamentum | 64 | 64 |
| The full articulamentum of Mental imagery | 32 | 32 |
| The full articulamentum of individual | 32 | 32 |
| Output 1 | 2 | 2 |
| Output 2 | 2 | 2 |
After relevant parameter is set, convolution operation is carried out to signal in time dimension and Spatial Dimension respectively, EEG is believed in simulation
Number time filtering and spatial filtering operation.20 channels of input are filtered simultaneously in the convolution of Spatial Dimension, are exported
One-dimensional space feature.Then time training dataset and one-dimensional space training dataset are subjected to average pond, to obtain pond
Then feature set extracts training characteristics by full articulamentum, export movement letter finally by Mental imagery branch, multitask branch
Label is ceased, individual tag information is provided by individual segregation branch.
It is understood that a time convolutional layer and a spatial convoluted layer are shown in Fig. 4, in actual scene,
There can be the participation of multiple time convolutional layers or spatial convoluted layer；It, can also be sufficient enough in data for full articulamentum number
In the case where increase full articulamentum number to expand the classification capacity of model, depending on specific quantity answers actual scene, herein not
It limits.
In conjunction with above-described embodiment it is found that by extracting more people MI features and individual special features using deep learning；Pass through
The Classification Loss and comparison loss function for constraining MI feature, are further reduced the difference of internal similar EEG signal, improve fortune
Dynamic imagination sorting algorithm accuracy and robustness；By the classification to individual special features, the identification based on EEG is realized.
To sum up, the method for model training provided by the present application is a kind of fortune using multi-task learning based on deep learning
Dynamic imagination eeg signal classification algorithm, had both broken the screen of each independent training pattern of individual need as caused by individual difference
Barrier, while realizing the function of identification.
By the method for above-mentioned model training, which can be used in following scene, that is, be based on brain-computer interface
Electroencephalogram detection process, be primarily based on target object generate Mental imagery Training scene；Then multiple individuals are acquired according to institute
State multiple electroencephalograms of Mental imagery Training scene generation；And the multiple electroencephalogram is pre-processed, and when extracting default
Between multiple EEG signal samples in section, the pretreated process determined based on the fluctuation information of the multiple electroencephalogram；
The multiple EEG signal sample is finally inputted into disaggregated model provided by the above embodiment, to respectively obtain the multiple brain
The label information and corresponding motion information label of the corresponding individual of electrical picture signal sample.
In a kind of possible display mode, display mode as described in Figure 5 can be used, Fig. 5 is the embodiment of the present application
A kind of schematic diagram of interface display of the model training provided.The interface may include the EEG real-time detection of multiple individuals, and show
Corresponding acquisition channel inquires button by clicking, and can show the corresponding movement label of EEG signal, and shows current
The case where acquisition state, can targetedly analyze each individual according to the movement label.
It is understood that the coherent element in step corresponding in above-mentioned Fig. 3 and Fig. 4 embodiment may be displayed on this
In interface, depending on particular content answers actual scene, herein without limitation.
For the above scheme of better implementation the embodiment of the present application, correlation for implementing the above scheme is also provided below
Device.Referring to Fig. 6, Fig. 6 is the structural schematic diagram of model training apparatus provided by the embodiments of the present application, model training apparatus
600 include:
Acquiring unit 601, it is described for obtaining the electroencephalogram EEG of multiple individuals and determining corresponding EEG signal sample
EEG signal sample is used to indicate corresponding Mental imagery movement label；
Determination unit 602 is marked for determining the label information of multiple individuals and establishing with Mental imagery movement
The corresponding relationship of label, to determine training dataset；
The determination unit 602 is also used to determine respectively based on multiple corresponding relationships that the training data is concentrated multiple
Sub- loss function, to be combined and determine that model loss function, multiple corresponding relationships that the training data is concentrated include more
The corresponding relationship of the EEG signal sample of a individual and Mental imagery movement label, the single individual EEG believe
The label letter of the corresponding relationship of number sample and Mental imagery movement label, the EEG signal sample and multiple individuals
The corresponding relationship of breath；
Optimize unit 603, for constructing EEG decoded model according to the model loss function and being initialized；
Training unit 604, for passing through the minimum model loss function pair using the training dataset as input
Electroencephalogram decoded model is trained, and to obtain the disaggregated model of parameter optimization, the disaggregated model is used for according to the brain electricity
Figure exports the label information of corresponding Mental imagery movement label and corresponding individual.
Preferably, in some possible implementations of the application,
The training unit 604, specifically for the training dataset is inputted at least one time convolutional layer, with
To timing training set, the temporal aspect collection is used to indicate the temporal information of the training dataset；
The training unit 604, specifically for the timing training set is inputted at least one spatial convoluted layer, with
To one-dimensional space feature set, the one-dimensional space feature set is used to indicate the corresponding time and space letter of the timing training dataset
Breath；
The training unit 604, it is extensive to obtain specifically for the one-dimensional space feature set is carried out average pond
Pond feature set；
The training unit 604, specifically for extracting the extensive pond feature set by least one full articulamentum
In feature, to obtain training characteristics collection；
The training unit 604, specifically for according to the training characteristics collection to the decoded model after parameter optimization
It is trained, to obtain disaggregated model.
Preferably, in some possible implementations of the application, the training unit 604 is specifically used for according to
Mental imagery movement label determines that multiple relevant spatial channels, the spatial channel specify the target action pair for filtering
The signal answered；
The training unit 604, specifically for the time training dataset is input to the multiple relevant space
It is handled in channel, to obtain one-dimensional space training dataset.
Preferably, in some possible implementations of the application, the optimization unit 603 is specifically used for according to default
Training parameter described in algorithm initialization, the preset algorithm include xavier；
The optimization unit 603, specifically for obtaining the weight information for the corresponding relationship that the training data is concentrated, with root
The model loss function is determined according to the multiple sub- loss function；
The optimization unit 603 is specifically used for according to the model loss function using gradient descent algorithm to initialization
The training parameter afterwards optimizes.
Preferably, in some possible implementations of the application, the acquiring unit is specifically used for obtaining multiple individuals
EEG；
The acquiring unit, is specifically used for selecting channel corresponding with the EEG according to preset rules, obtains institute with screening
State EEG signal sample, the target action information setting that the preset rules are indicated based on the EEG.
Preferably, in some possible implementations of the application, the acquiring unit 601 is specifically used for according to
EEG determines the target action information；
The acquiring unit 601 is specifically used for selecting the corresponding frequency filtering of the target action information；
The acquiring unit 601 is specifically used for selecting corresponding channel according to the frequency filtering, is obtained with screening described
EEG signal sample.
Preferably, in some possible implementations of the application, the acquiring unit 601 is also used to according to the EEG
Sample of signal determines fluctuation information；
The acquiring unit 601, if being also used to the fluctuation information meets preset condition, by the EEG signal sample
It is calculated by exponentially weighted moving average, to obtain the EEG signal sample after noise filtering, the preset condition is based on institute
The size relation for stating fluctuation information and preset threshold determines.
The present embodiment is by constructing disaggregated model based on multiple individuals, the EEG signal that can well acquire different people
Carry out information sharing, solve the problems, such as first with model cannot using different personal datas and caused by data waste；And pass through
Multiple sub- loss functions further improve the accuracy rate of Mental imagery classification for the determination of model loss function, from multiple dimensions
Degree reduces the difference of sample, and then reduces in a class as caused by local environment or the different state of mind in vivo that difference is big to ask
Topic；In addition individual difference alienation bring algorithm application limitation is also avoided, Different Individual Mental imagery movement label phase is extracted
The common information of pass reduces the difference between individual, improves the accuracy of model training.
The embodiment of the present application also provides a kind of model training apparatus, referring to Fig. 7, Fig. 7 is that the embodiment of the present application provides
Another model training apparatus structural schematic diagram, which can generate ratio because configuration or performance are different
Biggish difference may include one or more central processing units (central processing units, CPU) 722
(for example, one or more processors) and memory 732, one or more storage application programs 742 or data 744
Storage medium 730 (such as one or more mass memory units).Wherein, memory 732 and storage medium 730 can be with
It is of short duration storage or persistent storage.The program for being stored in storage medium 730 may include that (diagram does not have one or more modules
Mark), each module may include to the series of instructions operation in model training apparatus.Further, central processing unit
722 can be set to communicate with storage medium 730, and a series of fingers in storage medium 730 are executed on model training apparatus 700
Enable operation.
Model training apparatus 700 can also include one or more power supplys 726, one or more wired or nothings
Wired network interface 750, one or more input/output interfaces 758, and/or, one or more operating systems 741,
Such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM etc..
The step as performed by model training apparatus can be based on the model training apparatus shown in Fig. 7 in above-described embodiment
Structure.
A kind of computer readable storage medium is also provided in the embodiment of the present application, is stored in the computer readable storage medium
There is model training instruction, when run on a computer, so that computer is executed as earlier figures 2 to embodiment illustrated in fig. 5 is retouched
Step performed by model training apparatus in the method stated.
A kind of computer program product including model training instruction is also provided in the embodiment of the present application, when it is in computer
When upper operation, so that computer is executed as performed by the model training apparatus into the method described in embodiment illustrated in fig. 5 of earlier figures 2
The step of.
The embodiment of the present application also provides a kind of model training systems, the model training systems may include Fig. 6 and be retouched
State model training apparatus described in the model training apparatus or Fig. 7 in embodiment.
It is apparent to those skilled in the art that for convenience and simplicity of description, the system of foregoing description,
The specific work process of device and unit, can refer to corresponding processes in the foregoing method embodiment, and details are not described herein.
In several embodiments provided herein, it should be understood that disclosed system, device and method can be with
It realizes by another way.For example, the apparatus embodiments described above are merely exemplary, for example, the unit
It divides, only a kind of logical function partition, there may be another division manner in actual implementation, such as multiple units or components
It can be combined or can be integrated into another system, or some features can be ignored or not executed.Another point, it is shown or
The mutual coupling, direct-coupling or communication connection discussed can be through some interfaces, the indirect coupling of device or unit
It closes or communicates to connect, can be electrical property, mechanical or other forms.
The unit as illustrated by the separation member may or may not be physically separated, aobvious as unit
The component shown may or may not be physical unit, it can and it is in one place, or may be distributed over multiple
In network unit.It can select some or all of unit therein according to the actual needs to realize the mesh of this embodiment scheme
's.
It, can also be in addition, each functional unit in each embodiment of the application can integrate in one processing unit
It is that each unit physically exists alone, can also be integrated in one unit with two or more units.Above-mentioned integrated list
Member both can take the form of hardware realization, can also realize in the form of software functional units.
If the integrated unit is realized in the form of SFU software functional unit and sells or use as independent product
When, it can store in a computer readable storage medium.Based on this understanding, the technical solution of the application is substantially
The all or part of the part that contributes to existing technology or the technical solution can be in the form of software products in other words
It embodies, which is stored in a storage medium, including some instructions are used so that a computer
Equipment (can be personal computer, model training apparatus or the network equipment etc.) executes side described in each embodiment of the application
The all or part of the steps of method.And storage medium above-mentioned includes: USB flash disk, mobile hard disk, read-only memory (read-only
Memory, ROM), random access memory (random access memory, RAM), magnetic or disk etc. is various to deposit
Store up the medium of program code.
The above, above embodiments are only to illustrate the technical solution of the application, rather than its limitations；Although referring to before
Embodiment is stated the application is described in detail, those skilled in the art should understand that: it still can be to preceding
Technical solution documented by each embodiment is stated to modify or equivalent replacement of some of the technical features；And these
It modifies or replaces, the spirit and scope of each embodiment technical solution of the application that it does not separate the essence of the corresponding technical solution.
Claims (10)
1. a kind of model training method based on deep learning characterized by comprising
It obtains the electroencephalogram of multiple individuals and determines that corresponding EEG signal sample, the EEG signal sample are used to indicate
Corresponding Mental imagery acts label；
It determines the label information of multiple individuals and establishes the corresponding relationship with Mental imagery movement label, to determine instruction
Practice data set；
The multiple corresponding relationships concentrated based on the training data determine multiple sub- loss functions respectively, to be combined and determine
Model loss function, multiple corresponding relationships that the training data is concentrated include multiple individuals EEG signal sample and
Corresponding relationship, the single individual EEG signal sample and the Mental imagery of the Mental imagery movement label move
Make the corresponding relationship of label, the corresponding relationship of the EEG signal sample and the label information of multiple individuals；
Electroencephalogram decoded model is constructed according to the model loss function and is initialized；
Using the training dataset as input, electroencephalogram decoded model is instructed by minimizing the model loss function
Practice, to obtain the disaggregated model of parameter optimization, the disaggregated model is used to export corresponding Mental imagery according to the electroencephalogram
Act the label information of label and corresponding individual.
2. the method according to claim 1, wherein it is described according to the training dataset to parameter optimization after
The decoded model is trained, to obtain disaggregated model, comprising:
The training dataset is inputted at least one time convolutional layer, to obtain timing training set, the temporal aspect collection
For indicating the temporal information of the training dataset；
The timing training set is inputted at least one spatial convoluted layer, to obtain one-dimensional space feature set, the one-dimensional sky
Between feature set be used to indicate the corresponding time and space information of the timing training dataset；
The one-dimensional space feature set is subjected to average pond, to obtain extensive pond feature set；
The feature in the extensive pond feature set is extracted by least one full articulamentum, to obtain training characteristics collection；
The decoded model after parameter optimization is trained according to the training characteristics collection, to obtain disaggregated model.
3. according to the method described in claim 2, it is characterized in that, described input at least one for the time training dataset
In spatial convoluted layer, to obtain one-dimensional space training dataset, comprising:
Label is acted according to the Mental imagery and determines multiple relevant spatial channels, and the spatial channel is for filtering specified institute
State the corresponding signal of target action；
The time training dataset is input in the multiple relevant spatial channel and is handled, to obtain the one-dimensional space
Training dataset.
4. the method according to claim 1, wherein described construct electroencephalogram solution according to the model loss function
Code model is simultaneously initialized, comprising:
The training parameter is initialized according to preset algorithm；
The weight information for the corresponding relationship that the training data is concentrated is obtained, according to the multiple sub- loss function determination
Model loss function；
The training parameter after initialization is optimized using gradient descent algorithm according to the model loss function.
5. method according to claim 1-4, which is characterized in that the electroencephalogram for obtaining multiple individuals is simultaneously true
Fixed corresponding EEG signal sample, comprising:
Obtain the electroencephalogram of multiple individuals；
Channel corresponding with the electroencephalogram is selected according to preset rules, the EEG signal sample is obtained with screening, it is described
The target action information setting that preset rules are indicated based on the electroencephalogram.
6. according to the method described in claim 5, it is characterized in that, described corresponding with the electroencephalogram according to preset rules selection
Channel, with screening obtain the EEG signal sample, comprising:
The target action information is determined according to the electroencephalogram；
The label information of the individual is determined according to the electroencephalogram；
Select the corresponding frequency filtering of the target action information；
Corresponding channel is selected according to the frequency filtering, the EEG signal sample is obtained with screening.
7. according to the method described in claim 5, it is characterized in that, it is described screening obtain the EEG signal sample after,
The method also includes:
Fluctuation information is determined according to the EEG signal sample；
If the fluctuation information meets preset condition, the EEG signal sample is passed through into exponentially weighted moving average meter
It calculates, to obtain the EEG signal sample after noise filtering, the preset condition is based on the fluctuation information and default threshold
The size relation of value determines.
8. a kind of model training apparatus based on deep learning characterized by comprising
Acquiring unit, for obtaining the electroencephalogram of multiple individuals and determining corresponding EEG signal sample, the electroencephalogram letter
Number sample is used to indicate corresponding Mental imagery movement label；
Determination unit, for determining the label information of multiple individuals and establishing corresponding with Mental imagery movement label
Relationship, to determine training dataset；
The determination unit is also used to determine multiple sub- loss letters respectively based on multiple corresponding relationships that the training data is concentrated
Number, to be combined and determine that model loss function, multiple corresponding relationships that the training data is concentrated include multiple described
The corresponding relationship of the EEG signal sample of body and Mental imagery movement label, the single individual EEG signal
The label information of sample and Mental imagery movement label, the EEG signal sample and multiple individuals；
Optimize unit, for constructing electroencephalogram decoded model according to the model loss function and being initialized；
Training unit is used for using the training dataset as input, by minimizing the model loss function to electroencephalogram
Decoded model is trained, and to obtain the disaggregated model of parameter optimization, the disaggregated model is used to be exported according to the electroencephalogram
The label information of corresponding Mental imagery movement label and corresponding individual.
9. a kind of electroencephalogram detection method based on brain-computer interface characterized by comprising
Mental imagery Training scene is generated based on target object；
Acquire multiple electroencephalograms that multiple individuals are generated according to the Mental imagery Training scene；
The multiple electroencephalogram is pre-processed, and extracts multiple EEG signal samples in preset time period, it is described pre-
The process of processing is determined based on the fluctuation information of the multiple electroencephalogram；
The multiple EEG signal sample is inputted into disaggregated model, it is corresponding to respectively obtain the multiple EEG signal sample
Individual label information and corresponding motion information label, the disaggregated model be based on claim 1 to 7 it is described in any item
Obtained by model training method based on deep learning.
10. a kind of computer readable storage medium, it is stored with instruction in the computer readable storage medium, when it is in computer
When upper operation, so that computer executes the described in any item model training sides based on deep learning of the claims 1 to 7
Method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910818363.9A CN110503082B (en) | 2019-08-30 | 2019-08-30 | Model training method based on deep learning and related device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910818363.9A CN110503082B (en) | 2019-08-30 | 2019-08-30 | Model training method based on deep learning and related device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110503082A true CN110503082A (en) | 2019-11-26 |
| CN110503082B CN110503082B (en) | 2024-03-12 |
Family
ID=68590904
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910818363.9A Active CN110503082B (en) | 2019-08-30 | 2019-08-30 | Model training method based on deep learning and related device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110503082B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111428580A (en) * | 2020-03-04 | 2020-07-17 | 威海北洋电气集团股份有限公司 | Individual signal identification algorithm and system based on deep learning |
| CN111544855A (en) * | 2020-04-30 | 2020-08-18 | 天津大学 | Pure idea control intelligent rehabilitation method based on distillation learning and deep learning and application |
| CN111728609A (en) * | 2020-08-26 | 2020-10-02 | 腾讯科技（深圳）有限公司 | Electroencephalogram signal classification method, classification model training method, device and medium |
| CN111787351A (en) * | 2020-07-01 | 2020-10-16 | 百度在线网络技术（北京）有限公司 | Information query method, device, equipment and computer storage medium |
| CN112926553A (en) * | 2021-04-25 | 2021-06-08 | 北京芯盾时代科技有限公司 | Training method and device for motion detection network |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20050177058A1 (en) * | 2004-02-11 | 2005-08-11 | Nina Sobell | System and method for analyzing the brain wave patterns of one or more persons for determining similarities in response to a common set of stimuli, making artistic expressions and diagnosis |
| KR101345216B1 (en) * | 2012-07-18 | 2013-12-26 | 포항공과대학교 산학협력단 | Mehtod and apparatus for analyzing multi-subject electroencephalograph |
| US20170286809A1 (en) * | 2016-04-04 | 2017-10-05 | International Business Machines Corporation | Visual object recognition |
| US20180218256A1 (en) * | 2017-02-02 | 2018-08-02 | Qualcomm Incorporated | Deep convolution neural network behavior generator |
| US20180247107A1 (en) * | 2015-09-30 | 2018-08-30 | Siemens Healthcare Gmbh | Method and system for classification of endoscopic images using deep decision networks |
| CN108776788A (en) * | 2018-06-05 | 2018-11-09 | 电子科技大学 | A kind of recognition methods based on brain wave |
| CN109034365A (en) * | 2018-07-06 | 2018-12-18 | 电子科技大学 | The training method and device of deep learning model |
| CN109492666A (en) * | 2018-09-30 | 2019-03-19 | 北京百卓网络技术有限公司 | Image recognition model training method, device and storage medium |
| CN109645993A (en) * | 2018-11-13 | 2019-04-19 | 天津大学 | A kind of methods of actively studying of the raising across individual brain-computer interface recognition performance |
| CN109784211A (en) * | 2018-12-26 | 2019-05-21 | 西安交通大学 | A kind of Mental imagery Method of EEG signals classification based on deep learning |
| CN110163246A (en) * | 2019-04-08 | 2019-08-23 | 杭州电子科技大学 | The unsupervised depth estimation method of monocular light field image based on convolutional neural networks |
-
2019
- 2019-08-30 CN CN201910818363.9A patent/CN110503082B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20050177058A1 (en) * | 2004-02-11 | 2005-08-11 | Nina Sobell | System and method for analyzing the brain wave patterns of one or more persons for determining similarities in response to a common set of stimuli, making artistic expressions and diagnosis |
| KR101345216B1 (en) * | 2012-07-18 | 2013-12-26 | 포항공과대학교 산학협력단 | Mehtod and apparatus for analyzing multi-subject electroencephalograph |
| US20180247107A1 (en) * | 2015-09-30 | 2018-08-30 | Siemens Healthcare Gmbh | Method and system for classification of endoscopic images using deep decision networks |
| US20170286809A1 (en) * | 2016-04-04 | 2017-10-05 | International Business Machines Corporation | Visual object recognition |
| US20180218256A1 (en) * | 2017-02-02 | 2018-08-02 | Qualcomm Incorporated | Deep convolution neural network behavior generator |
| CN108776788A (en) * | 2018-06-05 | 2018-11-09 | 电子科技大学 | A kind of recognition methods based on brain wave |
| CN109034365A (en) * | 2018-07-06 | 2018-12-18 | 电子科技大学 | The training method and device of deep learning model |
| CN109492666A (en) * | 2018-09-30 | 2019-03-19 | 北京百卓网络技术有限公司 | Image recognition model training method, device and storage medium |
| CN109645993A (en) * | 2018-11-13 | 2019-04-19 | 天津大学 | A kind of methods of actively studying of the raising across individual brain-computer interface recognition performance |
| CN109784211A (en) * | 2018-12-26 | 2019-05-21 | 西安交通大学 | A kind of Mental imagery Method of EEG signals classification based on deep learning |
| CN110163246A (en) * | 2019-04-08 | 2019-08-23 | 杭州电子科技大学 | The unsupervised depth estimation method of monocular light field image based on convolutional neural networks |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111428580A (en) * | 2020-03-04 | 2020-07-17 | 威海北洋电气集团股份有限公司 | Individual signal identification algorithm and system based on deep learning |
| CN111544855A (en) * | 2020-04-30 | 2020-08-18 | 天津大学 | Pure idea control intelligent rehabilitation method based on distillation learning and deep learning and application |
| CN111544855B (en) * | 2020-04-30 | 2021-08-31 | 天津大学 | Pure idea control intelligent rehabilitation method based on distillation learning and deep learning and application |
| CN111787351A (en) * | 2020-07-01 | 2020-10-16 | 百度在线网络技术（北京）有限公司 | Information query method, device, equipment and computer storage medium |
| CN111787351B (en) * | 2020-07-01 | 2022-09-06 | 百度在线网络技术（北京）有限公司 | Information query method, device, equipment and computer storage medium |
| CN111728609A (en) * | 2020-08-26 | 2020-10-02 | 腾讯科技（深圳）有限公司 | Electroencephalogram signal classification method, classification model training method, device and medium |
| EP4101371A4 (en) * | 2020-08-26 | 2023-08-02 | Tencent Technology (Shenzhen) Company Limited | Electroencephalogram signal classifying method and apparatus, electroencephalogram signal classifying model training method and apparatus, and medium |
| CN112926553A (en) * | 2021-04-25 | 2021-06-08 | 北京芯盾时代科技有限公司 | Training method and device for motion detection network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110503082B (en) | 2024-03-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110503082A (en) | A kind of model training method and relevant apparatus based on deep learning | |
| CN113693613B (en) | Electroencephalogram signal classification method, electroencephalogram signal classification device, computer equipment and storage medium | |
| CN109165692B (en) | User character prediction device and method based on weak supervised learning | |
| CN110353675A (en) | The EEG signals emotion identification method and device generated based on picture | |
| Arora et al. | AutoFER: PCA and PSO based automatic facial emotion recognition | |
| CN110532996A (en) | The method of visual classification, the method for information processing and server | |
| CN112396106B (en) | Content recognition method, content recognition model training method, and storage medium | |
| CN110070029A (en) | A kind of gait recognition method and device | |
| CN108319928A (en) | A kind of deep learning model and application based on Multi-objective PSO optimization | |
| CN106909938A (en) | Viewing angle independence Activity recognition method based on deep learning network | |
| CN113221663A (en) | Real-time sign language intelligent identification method, device and system | |
| CN110135497A (en) | Method, the method and device of Facial action unit intensity estimation of model training | |
| CN110458235A (en) | Movement posture similarity comparison method in a kind of video | |
| CN110048978A (en) | A kind of signal modulate method | |
| CN110135244A (en) | It is a kind of based on brain-machine cooperative intelligent expression recognition method | |
| CN112233102A (en) | Method, device and equipment for identifying noise in image sample set and storage medium | |
| Dar et al. | Efficient-SwishNet based system for facial emotion recognition | |
| CN111753683A (en) | Human body posture identification method based on multi-expert convolutional neural network | |
| Wu et al. | An efficient binary convolutional neural network with numerous skip connections for fog computing | |
| CN114708637A (en) | Face action unit detection method based on meta-learning | |
| DeOliveira et al. | Har-ctgan: a mobile sensor data generation tool for human activity recognition | |
| CN109472307A (en) | A kind of method and apparatus of training image disaggregated model | |
| Fachruddin et al. | Network and layer experiment using convolutional neural network for content based image retrieval work | |
| Liu et al. | Design of sports competition assistant evaluation system based on big data and action recognition algorithm | |
| Li | Facial expression recognition by DenseNet-121 |
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 |