CN109086658A - A kind of sensing data generation method and system based on generation confrontation network - Google Patents
A kind of sensing data generation method and system based on generation confrontation network Download PDFInfo
- Publication number
- CN109086658A CN109086658A CN201810589194.1A CN201810589194A CN109086658A CN 109086658 A CN109086658 A CN 109086658A CN 201810589194 A CN201810589194 A CN 201810589194A CN 109086658 A CN109086658 A CN 109086658A
- Authority
- CN
- China
- Prior art keywords
- data
- generator
- arbiter
- behavior
- truthful
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/20—Movements or behaviour, e.g. gesture recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/251—Fusion techniques of input or preprocessed data
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Psychiatry (AREA)
- Social Psychology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of based on the sensing data generation method for generating confrontation network, comprising: model construction step is constructed by neural network model with truthful data and generates confrontation network model, and it includes generator and arbiter which, which fights network model,；Model training step to fight game mechanism training generator and the arbiter, and is iterated, until the data obtained from the generator meet evaluation criterion；Data generation step generates generated data by the confrontation network model with the generator.
Description
Technical field
The present invention relates to general fit calculation, deep learning and machine learning fields, and in particular to one kind is based on generation confrontation net
The sensing data generation method and system of network.
Background technique
Confrontation network (generative adversarial networks, GANs) is generated in 2014 by Ian
GoodFellow is put forward for the first time, and this method frame is mainly by a generation network (generator) and a differentiation network (arbiter)
It constitutes, target is mainly that can accurately learn the generation of initial data distribution character come supplemental training one by differentiation network
Network.It generates network and differentiates continuous improvement self performance of the network in mutual game and confrontation, is i.e. generator is constantly evolved
To generate the generated data for being more nearly truthful data, arbiter also constantly evolves and promotes oneself to true, mantissa at the same time
According to discrimination capabilities.In the ideal situation, the two by fight game training method reach a kind of dynamic balance state and make from
The performance of body is optimal: generator by receive random noise (usually from be uniformly distributed or normal distribution) accurately learn
To truthful data distribution character and produce the generated data for being enough to mix the spurious with the genuine, and arbiter distinguished with 1/2 probability it is true
Real data and generated data.
Since self-generating confrontation network is suggested, correlative study proposes various types of generation confrontation moulds on this basis
The variant of type, be mainly used for image data generate field (such as: image generation, image repair, image conversion, image super-resolution,
Image goes to block), while also being tried to explore to be applied to the fields such as text generation, speech production, video estimation.But at present
The complete confrontation network architecture that generates is not applied to sensing data generation field also.Therefore, this invention takes the lead in generate
Confrontation network come be applied to sensing data generation related fields in, have it is comparable perspective, while also have it is important
Researching value and research significance.
With depth learning technology fast development and it is universal, various neural network models are able to study and use.Circulation
Neural network (recurrent neural network, RNN) is used exclusively for processing sequence data (x1,x2,…,xt)
Neural network model has all achieved many research achievements in research fields such as Language Modeling, machine translation, picture descriptions.So
And each hidden unit of traditional Recognition with Recurrent Neural Network only includes the shirtsleeve operations such as Tanh or ReLU, in practical applications
The defects of there are gradient disappearance/gradient explosions, so being difficult to handle long-term Dependence Problem, this results in it to lead in sensing data
The development and application in domain are very limited.
Develop on the basis of traditional Recognition with Recurrent Neural Network (RNN) and the shot and long term of evolution out remembers (long short-
Term memory, LSTM) network is a kind of special RNN model that long-term Dependence Problem can be effectively treated, it is particularly suitable for place
Reason and predicted time sequence data, have numerous applications in fields such as machine translation, image analysis, image recognitions.Shot and long term note
It is not both the unit for being added to and judging whether information is useful that it is maximum, which to recall network (LSTM) and Recognition with Recurrent Neural Network (RNN), each
Containing there are three doorway logics, respectively input gate, forgetting door and out gate in unit.So far, shot and long term memory network
(LSTM) one of most efficient method since solving the problems, such as long sequence time has been it.Therefore, the present invention is in processing long-term sequence
Sensing data when mainly use main functional modules of the shot and long term memory network (LSTM) as this method model.
Convolutional neural networks (convolutional neural networks, CNNs) are by partially connected and weight
It is the methods of shared come a possibility that efficiently reducing network parameter, and reduce over-fitting to a certain extent.In general, convolutional Neural
Network both can handle One-dimension Time Series, also can handle two-dimensional image data.It is most widely used in field of image recognition,
It is also field of image recognition one of method the most outstanding.Wherein, one-dimensional convolution is also referred to as convolution, is convolutional neural networks
(CNNs) one kind is mainly used for carrying out Neighborhood Filtering in one-dimensional input data.Present study is mainly by convolutional neural networks
(CNNs) it as a kind of auxiliary of shot and long term memory network (LSTM) and expands and constructs and generate confrontation network model.
To sum up, present study is mainly by shot and long term memory network (LSTM) and convolution neural network (CNNs) as master
Strategic point functional module can sufficiently learn the generation confrontation network model of sensing data distribution character to construct.It is fought generating
Under the confrontation game training mechanism of network, our main targets are to train to obtain one that original sensor data can be distributed
The generator that characteristic is sufficiently learnt, the generator can be continuously generated the generated data of specified type.
In order to verify the validity of method model proposed by the invention, it is raw for sensing data that we have proposed three kinds
At the evaluation criterion of quality: local data's assessment, global data assessment and memory independence data assessment.It is logical in order to verify simultaneously
The validity of this method model generated data generated is crossed, we utilize the sensing data and phase under true application scenarios
It closes recognizer and availability of data quality evaluation has been carried out to the generated data of three types.Shown by many experiments verifying
Proposed by the invention can effectively learn sensor number based on the sensing data generation method model for generating confrontation network
According to distribution characteristics, synthesized sensing data also can promote effectively based on involved in sensing data correlative study
The precision and effect of classification and identification algorithm.
Summary of the invention
In view of the above-mentioned problems, the present invention proposes a kind of sensing data generation method based on generation confrontation network, comprising:
Truthful data obtaining step acquires the data of human behavior under true application scenarios by sensor, as truthful data；Model
Training step is constructed, generator and arbiter are constructed by neural network model, confrontation game mechanism is passed through with the truthful data
The generator and the arbiter are trained, assesses when the arbiter from the data that the generator obtains and meets evaluation criterion
When, confrontation network model is made a living into the generator and arbiter combination；Generated data generation step fights net with the generation
Network model generates generated data；The generated data is mixed with the truthful data to obtain blended data, is led to by model verification step
It crosses machine learning classification recognition methods to identify the blended data, to verify the accuracy of generation confrontation network model.
Sensing data generation method of the present invention, in which:
The generator is
The arbiter is
Wherein, z is random noise data, and x is the truthful data, and G (z) indicates the generator by random noise data z
Be converted to the generated data；D (x) indicates true and false differentiation of the arbiter to truthful data x；D (G (z)) indicates the arbiter
True and false differentiation to the generated data G (z) generated by generator；pdata(x) data distribution of truthful data x is indicated；pz(z)
Indicate the data distribution of random noise data z.
Sensing data generation method of the present invention, wherein the evaluation criterion includes:
Local data's assessment, with the life during the dual training of the generator and the arbiter, in each trained batch
At the fitted trend between data and the initial data as evaluation criterion；
Global data assessment, the generation with the data variation trend and characteristic of the generated data, i.e., after training up
The Data Synthesis ability and synthetic effect of device are as evaluation criterion；
Remember the assessment of independence data, evades performance when facing impurity data with the generator, i.e., the generator avoids
Learn the data distribution characteristic of the impurity data as evaluation criterion；Wherein the impurity data be a certain classification data in be mingled with
The truthful data that closes on classification adjacent with a certain classification.
Sensing data generation method of the present invention, wherein the human behavior include static behavior, it is walking behavior, small
Running is, wherein constructing the generator and arbiter of the corresponding static behavior with one-dimensional convolutional neural networks；With two-way shot and long term
Memory network constructs the generator of the corresponding walking behavior, and the differentiation of the corresponding walking behavior is constructed with one-dimensional convolutional neural networks
Device；The generator of the corresponding behavior of trotting is constructed with shot and long term memory network, is somebody's turn to do so that the building of two-way shot and long term memory network is corresponding
It trots the arbiter of behavior.
System is generated based on the sensing data for generating confrontation network the invention further relates to a kind of, comprising:
Truthful data obtains module, for acquiring the data of human behavior under true application scenarios by sensor, as
Truthful data；
Model construction training module generates confrontation network model for constructing and training；Wherein pass through neural network model
Generator and arbiter are constructed, the generator and the arbiter are trained by fighting game mechanism with the truthful data,
When arbiter assessment meets evaluation criterion from the data that the generator obtains, made a living with the generator and arbiter combination
At confrontation network model；
Generated data generation module, for generating generated data with generation confrontation network model；
Model authentication module, for verifying the accuracy of generation confrontation network model；Wherein by the generated data and this
Truthful data is mixed to obtain blended data, is identified by machine learning classification recognition methods to the blended data, to test
Demonstrate,prove the accuracy of generation confrontation network model.
Sensing data of the present invention generates system, in which:
The generator is
The arbiter is
Z is random noise data, and x is the truthful data, and G (z) indicates that the generator is converted to random noise data z
The generated data；D (x) indicates true and false differentiation of the arbiter to truthful data x；D (G (z)) indicates the arbiter to conversion
The true and false differentiation of the generated data G (z)；pdata(x) data distribution of truthful data x is indicated；pz(z) random noise is indicated
The data distribution of data z.
Sensing data of the present invention generates system, wherein the evaluation criterion includes:
Local data's assessment, with the life during the dual training of the generator and the arbiter, in each trained batch
At the fitted trend between data and the initial data as evaluation criterion；
Global data assessment, the life with the data variation trend and characteristic of the generated data, i.e., after training up
The Data Synthesis ability and synthetic effect grown up to be a useful person are as evaluation criterion；
Remember the assessment of independence data, evades performance when facing impurity data with the generator, i.e., the generator avoids
Learn the data distribution characteristic of the impurity data as evaluation criterion；Wherein the impurity data be a certain classification data in be mingled with
The truthful data that closes on classification adjacent with a certain classification.
Sensing data of the present invention generates system, wherein the human behavior include static behavior, it is walking behavior, small
Running is, wherein constructing the generator and arbiter of the corresponding static behavior with one-dimensional convolutional neural networks；With two-way shot and long term
Memory network constructs the generator of the corresponding walking behavior, and the differentiation of the corresponding walking behavior is constructed with one-dimensional convolutional neural networks
Device；The generator of the corresponding behavior of trotting is constructed with shot and long term memory network, is somebody's turn to do so that the building of two-way shot and long term memory network is corresponding
It trots the arbiter of behavior.
The sensing data that the present invention is generated by this method model can effectively promote the phase based on sensing data
Close the precision and performance of the identification sorting algorithm of research field.In the various application fields based on sensing data, (such as behavior is known
Not, indoor positioning, health medical treatment etc.) original sensor data can be supplemented using this method and system, to effectively solve
It is certainly related based on the research of sensing data limited resource, environment and under the conditions of the small sample problem that is faced, it is final to promote
Further develop and apply into Related Research Domain.
Detailed description of the invention
Fig. 1 is the generator of the invention based on the sensing data generation method for generating confrontation network and arbiter confrontation
Schematic diagram is trained in game.
Fig. 2 is of the invention based on the sensing data synthetic method and system schematic that generate confrontation network.
Fig. 3 is the confrontation game training data table figure of arbiter and generator of the invention.
Fig. 4 is local data's assessment tables of data figure of three kinds of dissimilar sensor data of the invention.
Fig. 5 is the global data assessment tables of data figure of three kinds of dissimilar sensor data of the invention.
Fig. 6 is memory independence data assessment tables of data figure of the invention.
Fig. 7 is the classification and recognition tables of data figure after the generated data of addition different number of the invention.
Wherein, drawing illustration are as follows:
1: truthful data obtains module 2: model construction training module
3: generated data generation module 4: model authentication module
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with attached drawing, the present invention is mentioned
The sensing data synthetic method and system further description based on generation confrontation network out.It should be appreciated that this place
The specific implementation method of description is only used to explain the present invention, is not intended to limit the present invention.
Variability in view of the numerical value of sensing data at any time and to characteristics such as required precision height, so that adapting to
It is not directly applicable the generation of sensing data in traditional generation confrontation network method and model that image data generates, this
Also the corresponding difficulty for generating confrontation network method model of building is determined.Therefore, the present invention is directed to the characteristic of sensing data
Research effectively generates the forward-looking technology problem that confrontation network method model is a complexity and difficulty.Mainly remembered with shot and long term
Network (LSTM), convolution neural network (CNNs) and full Connection Neural Network (FCN), which are recalled, as basic function module carrys out structure
Build can the distribution characteristics to sensing data sufficiently learnt and can be generated the generation close to true sensing data
Fight network method model.
Fig. 1 is generator and arbiter confrontation game training based on the sensing data generation method for generating confrontation network
Schematic diagram.As shown in Figure 1, how the core based on the sensing data generation method model for generating confrontation network is by not
The neural network module of same type combines to construct the generator of adaptation sensing data feature with arbiter, while to guarantee to give birth to
It grows up to be a useful person and best performance state is finally attained by by confrontation game training with arbiter.Finally to guarantee trained generator
It fully grasps the distribution character of original sensor data and effective generated data can be generated as desired, thus for solution
Certainly an effective method and approach are provided based on the small sample problem that sensing data correlative study is faced.The present invention is directed to
The feature of sensing data is mainly using shot and long term memory network, convolution neural network and full Connection Neural Network come structure
It builds and effectively generates confrontation network model for dissimilar sensor data.
Fig. 2 is of the invention based on the sensing data synthesis system schematic diagram for generating confrontation network.As shown in Fig. 2, this
The sensing data based on generation confrontation network of invention generates system and includes:
Truthful data obtains module 1, the data of human behavior under true application scenarios is acquired by sensor, as true
Data；
Model construction training module 2 constructs generator and arbiter by neural network model, with truthful data by pair
Anti- game mechanism is trained generator and arbiter, when the data that arbiter assessment is obtained from generator meet evaluation criterion
When, confrontation network model is made a living into generator and arbiter combination；
Generated data generation module 3 generates generated data to generate confrontation network model；
Generated data is mixed with truthful data to obtain blended data, passes through machine learning classification by model authentication module 4
Recognition methods identifies blended data, to verify the accuracy for generating confrontation network model.
Although it should be appreciated that the sensing data generation side based on the generation confrontation network architecture proposed in the present invention
Method and system are based primarily upon shot and long term memory network (LSTM), convolution neural network (CNNs) and full Connection Neural Network
(FCNs), specific implementation method described herein is only used to explain the present invention, is not intended to limit the present invention.We can be with
Characteristic based on different sensing datas constructs diversified generation by different types of neural network module and fights network
Model, to further promote the generation quality of sensing data.
One, the specific embodiment party proposed by the invention based on the sensing data generation method for generating the confrontation network architecture
Formula and process are as follows:
1, the sensing data of human behavior under real scene is obtained as truthful data；
2, different neural network modules (such as shot and long term memory network, volume are selected according to the characteristic of different sensing datas
Product neural network, full Connection Neural Network) building generates confrontation network model framework, the wherein specific mesh of generator and sensor
Mark and mode are as follows:
(1) generator:
(a) training objective:
(b) be directed to different sensing datas (by it is static, walk, for three kinds of human behaviors of trotting) construct respectively
Different Maker models, such as: the generator of static behavior is mainly made of one-dimensional convolutional neural networks；The generation of walking behavior
Device is mainly made of two-way shot and long term memory network (Bi-direction LSTM)；Trot behavior generator mainly by length
Phase memory network (LSTM) is constituted.
(2) arbiter:
(a) training objective:
(b) be directed to different sensing datas (by it is static, walk, for three kinds of human behaviors of trotting) construct respectively
Different discrimination models, such as: the arbiter of static behavior is mainly made of one-dimensional convolutional neural networks；The arbiter of walking behavior
Mainly it is made of one-dimensional convolutional neural networks；Trot behavior arbiter mainly by two-way shot and long term memory network (Bi-
Direction LSTM) it constitutes；
Wherein, z is random noise；G (z) indicates that random noise data are converted to composite number by production network (generator)
According to；D (x) indicates true and false differentiation of the arbiter to truthful data x；D (G (z)) indicates arbiter to the true and false of generated data G (z)
Differentiate；pdata(x) data distribution of truthful data is indicated；pz(z) data distribution of random noise is indicated；
Being related to noise inputs in above process is mainly gaussian random noise, is also needed when constructing generator and arbiter
Want full Connection Neural Network layer；
By fighting game mechanism constantly repetitive exercise generator and arbiter.
3, generated data is generated by the generator for being able to train up；
4, generated data is mixed with truthful data to obtain blended data, by machine learning classification recognition methods to this
Blended data is identified, to verify the accuracy of generation confrontation network model
The present invention is not limited to the above methods, and the basic function component for constructing generator and arbiter may be full connection mind
Through network (Fully Connected Neural Networks), thresholding cycling element (Gated Recurrent Unit,
GRU) and the other kinds of neural network such as convolutional neural networks (Convolutional Neural Networks, CNNs) and
A combination thereof.
Two, total technical effect
Fig. 3 is the confrontation game training data table figure of arbiter and generator of the invention.As shown in figure 3, relative to making
With simple generation model come for the technology of the sensing data generated, the invention proposes one kind more sufficiently to learn
Original sensor data distribution character and the new method and approach that significantly more efficient sensing data can be generated.This method model
Stable physical training condition can be reached based on the distribution character of sensing data, can finally generator be allowed sufficiently to learn raw sensory
Device data, more stable curve are the loss of generator, fluctuate the slightly larger loss for arbiter.Confrontation proposed by the invention
Game training main purpose be one is obtained to be capable of the quasi- generation network for holding original sensor data distribution character, therefore
The number of training generator is 3 times of training arbiter number in practice, also results in the loss of arbiter training in certain journey
It is fluctuated on degree.But for entirety, generator and arbiter all tend towards stability.The method model proposed through the invention
Come generated data not instead of simply memory original sensor data or it is simple must be fitted original sensor data, be based on learning
More true, more multifarious sensing data is generated on the basis of habit initial data distribution characteristics, and then is promoted existing
The precision and performance of classification and identification algorithm in correlative study based on sensing data.
Currently, generating the generation that confrontation network method model is mainly used for image data, the standard of its performance quality is evaluated
Mainly vision turing test (Visual Turing Test), initial value (Inception Score) and improved initial
Score value (Modified Inception Score) etc..However, these can not be direct to image data effectively evaluating standard
Generation quality for evaluation sensor data.Therefore, distribution character and data characteristics of the present invention in research sensing data
On the basis of, propose three kinds it is novel to generated data effectively evaluating standard:
(1) local data assesses: it is mainly used for the performance that evaluation and test generates confrontation network generator in game training process,
Conjunction by observation generator to identical quantity is generated in the learning time section of the original sensor data of a trained batch
At the quality of data, the assessment of generator performance is marked in this, as during training, Fig. 4 is three kinds of different types of the invention
The local data of sensing data assesses tables of data figure, as shown in Figure 4.Generator can be preferable in confrontation game training process
Assurance original sensor data distribution characteristics, while avoiding excessive data acquisition system again
(2) global data is assessed: being mainly used for the performance that evaluation and test adequately generates network by confrontation game training, is root
The generated data of fixed number magnitude is generated according to actual needs, different types of sensing data characteristic can be anti-by the evaluating standard
Mirror difference.Fig. 5 is the global data assessment tables of data figure of three kinds of dissimilar sensor data of the invention.Fig. 5 reflection
It is the global data Evaluated effect of 3 kinds of different types of sensing datas, we can pass through the observation of evaluation criterion image
The quality of generated data.
(3) remember independence data assessment: being mainly used for evaluating and testing proposed by the invention based on the biography for generating confrontation network
Sensor data creation method model avoids the ability of over-fitting.In the Related Research Domain based on sensing data, often need
The time series datas such as sensor are split, but may mixed in certain type of sensing data in cutting procedure
The adjacent other kinds of sensing data in boundary.The adjacent sensing data in these boundaries affects the quality of data, if
Panel data model will mix these other kinds of only by simple study and fitting original sensor data
Sensing data learns together, so as to cause the second-rate sensing data of availability is generated.It would therefore be desirable to a kind of new
The evaluation criterion of type come evaluate and test method model proposed by the invention whether have it is this avoid study impurity data ability, also
It is memory independence data assessment standard.Fig. 6 is memory independence data assessment tables of data figure of the invention, as shown in fig. 6, on
The curve in face is represented containing the other kinds of sensing data (impurity data) that boundary is adjacent in certain sensing data, below
Curve represents the sensing data based on generation confrontation network method model synthesis proposed through the invention, can obviously see
The study to impurity data being effectively prevented from based on generation confrontation network method model is observed, that is, is avoided to original biography
The overfitting problem of sensor data.
Proposed by the invention generates the final mesh of the method model of sensing data based on the confrontation network architecture is generated
Mark is that generation can be with property amount height and close to the artificial synthesized sensing data of original sensor data, therefore needs one kind and comment
The standard of valence sensing data synthesis quality.The present invention passes through according to the characteristic of sensing data and in conjunction with practical application scene
The generated data of human behavior is generated to supplement original sensor data, and then row is verified based on traditional machine learning method
For the precision of identification, so that be verified the sensing data for generating confrontation network synthesis by Activity recognition classifying quality can
With property and validity.Fig. 7 is the classification and recognition tables of data figure after the generated data of addition different number of the invention, is such as schemed
Shown in 7, mark " 0 " and " all " in X-axis respectively represent entirely original sensor data and are entirely the sensor number of synthesis
According to intermediate digital representation respectively represents the Classification and Identification after the generated data for adding different number into original sensor data
Precision.
In order to be further verified generate confrontation the network architecture training generator synthesized by data validity and can
With property, (DecisionTree) when we include: support vector machines (SVM), decision using traditional machine learning method building,
K nearest neighbor (kNN), logistic regression (Logistic Regression), AdaBoost, random forest (Random Forest) etc. 6
Kind of classifier, for carrying out Classification and Identification to three kinds of common daily behaviors (static, walk, trot).Specific steps are as follows:
Firstly, verifying Classification and Identification by 6 kinds of common classifying identification methods on whole actual sensor data sets
Precision, such as following table the first row.
Then, every kind of behavior is directed to by the sufficient generator of training respectively and generates 200000 synthesis sensing datas,
Three behaviors amount to 600000 synthesis sensing datas, by these generated datas by 6 kinds of common classification and identification algorithms into
Row classification, accuracy of identification such as the 2nd row of following table.
Finally, this 6,000,000,000 data is all added in actual sensor data (194436 it is static, 204046
Item walking, 200691 trot) form a mixed data set, same method carries out Classification and Identification, and classification and recognition is such as
Shown in table 1
| Support vector machines | Decision tree | K nearest neighbor | Logistic regression | AdaBoost | Random forest | |
| Truthful data | 0.85714 | 0.79592 | 0.87714 | 0.87755 | 0.89796 | 0.87755 |
| Generated data | 1.00000 | 1.00000 | 1.00000 | 1.00000 | 1.00000 | 1.00000 |
| Blended data | 0.97938 | 0.97938 | 0.93814 | 0.95876 | 0.95876 | 0.97938 |
Table 1
The experimental results showed that, the generator synthesis generated by generating confrontation network mechanism training senses utensil above
There is ideal practical application effect, the sensing data of three kinds of synthesized common daily behaviors passes through classical machine
The recognition accuracy of Study strategies and methods is up to 100%.Synthesized data are added to truthful data simultaneously to concentrate, it also can be significant
Enhance classification and recognition, this is just proved from practical application, proposed by the invention based on the biography for generating confrontation network mechanism
Sensor data creation method can generate synthesis sensing data effective, can be high with property amount.
Meanwhile it being passed to further verify the synthesis that the method added into initial data and proposed through the invention generates
The effect of sensor data, respectively using trained generator generate 100000,300000,500000,700000,
900000,1100000, the sensing datas of 1300000 three kinds of common daily behaviors (static, walk, trot),
Then these data are respectively added to truthful data to concentrate, form respectively the hybrid sensor data for being mixed with different number
Collection, (DecisionTree), k nearest neighbor (k-NN), logistic regression when being separately input to support vector machines (SVM), decision
(Logistic Regression), AdaBoost, random forest (Random Forest) have waited the machine of the classics such as 6 kinds of classifiers
Device Study strategies and methods carry out Activity recognition and classification, accuracy of identification are as shown in Figure 6.
The experimental results showed that the increase of the quantity with the synthesis sensor being added in original sensor data, 6 kinds of warps
The behavior classification and recognition of the machine learning Activity recognition disaggregated model of allusion quotation is constantly promoted, finally close to ideal 100%
Activity recognition precision.
Description of the invention and application be it is illustrative, it is not intended to limit the scope of the present invention to the above embodiment.
The deformation and change of embodiments disclosed herein are possible, the embodiments for those skilled in the art
Replacement and equivalent various parts be well known.It will be apparent to those skilled in the art that do not depart from spirit of the invention or
In the case where substantive characteristics, the present invention can in other forms, structure, arrangement, ratio, and with other components, material and portion
Part is realized.In this way, if these modifications and changes of the present invention belongs to the range of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to include these modifications and variations.
Claims (10)
1. a kind of based on the sensing data generation method for generating confrontation network characterized by comprising
Truthful data obtaining step acquires the data of human behavior under true application scenarios by sensor, as truthful data；
Model construction training step constructs generator and arbiter by neural network model, passes through confrontation with the truthful data
Game mechanism is trained the generator and the arbiter, when the data satisfaction that arbiter assessment is obtained from the generator is commented
Price card is punctual, makes a living into confrontation network model with the generator and arbiter combination；
Generated data generation step generates generated data with generation confrontation network model.
2. sensing data generation method as described in claim 1, it is characterised in that:
The generator is
The arbiter is
Wherein, z is random noise data, and x is the truthful data, and G (z) indicates that the generator converts the random noise data z
For the generated data；D (x) indicates true and false differentiation of the arbiter to truthful data x；D (G (z)) indicate the arbiter to by
The true and false differentiation for the generated data G (z) that generator generates；pdata(x) data distribution of truthful data x is indicated；pz(z) it indicates
The data distribution of random noise data z.
3. sensing data generation method as described in claim 1, which is characterized in that further include:
The generated data is mixed with the truthful data to obtain blended data, passes through machine learning classification by model verification step
Recognition methods identifies the blended data, to verify the accuracy of generation confrontation network model.
4. sensing data generation method as described in claim 1, which is characterized in that the evaluation criterion includes:
Local data's assessment, with the generation number during the dual training of the generator and the arbiter, in each trained batch
According to the fitted trend between the initial data as evaluation criterion；
Global data assessment, with the data variation trend and characteristic of the generated data, i.e., generator after training up
Data Synthesis ability and synthetic effect are as evaluation criterion；
Remember the assessment of independence data, evades performance when facing impurity data with the generator, i.e., the generator avoids learning
The data distribution characteristic of the impurity data is as evaluation criterion；Wherein the impurity data be a certain classification data in be mingled with
The adjacent truthful data for closing on classification of a certain classification.
5. sensing data generation method as described in claim 1, which is characterized in that the human behavior include static behavior,
Walking behavior, behavior of trotting, wherein constructing the generator and arbiter of the corresponding static behavior with one-dimensional convolutional neural networks；With
Two-way shot and long term memory network constructs the generator of the corresponding walking behavior, constructs the corresponding walking with one-dimensional convolutional neural networks
The arbiter of behavior；The generator of the corresponding behavior of trotting is constructed, with shot and long term memory network with two-way shot and long term memory network
Construct the arbiter of the corresponding behavior of trotting.
6. a kind of generate system based on the sensing data for generating confrontation network characterized by comprising
Truthful data obtains module, for acquiring the data of human behavior under true application scenarios by sensor, as true
Data；
Model construction training module generates confrontation network model for constructing and training；Wherein constructed by neural network model
Generator and arbiter are trained the generator and the arbiter by fighting game mechanism with the truthful data, when this
When arbiter assessment meets evaluation criterion from the data that the generator obtains, made a living in pairs with the generator and arbiter combination
Anti- network model；
Generated data generation module, for generating generated data with generation confrontation network model.
7. sensing data as claimed in claim 6 generates system, it is characterised in that:
The generator is
The arbiter is
Wherein, z is random noise data, and x is the truthful data, and G (z) indicates that the generator converts the random noise data z
For the generated data；D (x) indicates true and false differentiation of the arbiter to truthful data x；D (G (z)) indicate the arbiter to by
The true and false differentiation for the generated data G (z) that generator generates；pdata(x) data distribution of truthful data x is indicated；pz(z) it indicates
The data distribution of random noise data z.
8. sensing data as claimed in claim 7 generates system, which is characterized in that further include:
Model authentication module, for verifying the accuracy of generation confrontation network model；It is wherein that the generated data is true with this
Data mixing identifies the blended data to obtain blended data, by machine learning classification recognition methods, is somebody's turn to do with verifying
Generate the accuracy of confrontation network model.
9. sensing data as claimed in claim 6 generates system, which is characterized in that the evaluation criterion includes:
Local data's assessment, with the generation number during the dual training of the generator and the arbiter, in each trained batch
According to the fitted trend between the initial data as evaluation criterion；
Global data assessment, the generator with the data variation trend and characteristic of the generated data, i.e., after training up
Data Synthesis ability and synthetic effect as evaluation criterion；
Remember the assessment of independence data, evades performance when facing impurity data with the generator, i.e., the generator avoids learning
The data distribution characteristic of the impurity data is as evaluation criterion；Wherein the impurity data be a certain classification data in be mingled with
The adjacent truthful data for closing on classification of a certain classification.
10. sensing data as described in claim 1 generates system, which is characterized in that the human behavior include static behavior,
Walking behavior, behavior of trotting, wherein constructing the generator and arbiter of the corresponding static behavior with one-dimensional convolutional neural networks；With
Two-way shot and long term memory network constructs the generator of the corresponding walking behavior, constructs the corresponding walking with one-dimensional convolutional neural networks
The arbiter of behavior；The generator of the corresponding behavior of trotting is constructed, with shot and long term memory network with two-way shot and long term memory network
Construct the arbiter of the corresponding behavior of trotting.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810589194.1A CN109086658B (en) | 2018-06-08 | 2018-06-08 | Sensor data generation method and system based on generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810589194.1A CN109086658B (en) | 2018-06-08 | 2018-06-08 | Sensor data generation method and system based on generation countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109086658A true CN109086658A (en) | 2018-12-25 |
| CN109086658B CN109086658B (en) | 2021-06-08 |
Family
ID=64839812
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810589194.1A Active CN109086658B (en) | 2018-06-08 | 2018-06-08 | Sensor data generation method and system based on generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109086658B (en) |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109813542A (en) * | 2019-03-15 | 2019-05-28 | 中国计量大学 | The method for diagnosing faults of air-treatment unit based on production confrontation network |
| CN109889452A (en) * | 2019-01-07 | 2019-06-14 | 中国科学院计算技术研究所 | Network context flow generation method and system based on condition production confrontation network |
| CN110046653A (en) * | 2019-03-22 | 2019-07-23 | 赣州好朋友科技有限公司 | A kind of white tungsten method for separating and system based on XRT ray |
| CN110133643A (en) * | 2019-05-22 | 2019-08-16 | 北京林业大学 | Root system of plant detection method and device |
| CN110533578A (en) * | 2019-06-05 | 2019-12-03 | 广东世纪晟科技有限公司 | Image translation method based on conditional countermeasure neural network |
| CN110769196A (en) * | 2019-10-17 | 2020-02-07 | 天津大学 | Video prediction method for discontinuous monitoring road section |
| CN110880020A (en) * | 2019-10-30 | 2020-03-13 | 西安交通大学 | Self-adaptive trans-regional base station energy consumption model migration and compensation method |
| CN111028941A (en) * | 2019-12-10 | 2020-04-17 | 南京医科大学 | TKA preoperative clinical staging intelligent evaluation method based on generation of antagonistic network |
| CN111383651A (en) * | 2018-12-29 | 2020-07-07 | Tcl集团股份有限公司 | Voice noise reduction method and device and terminal equipment |
| CN111597338A (en) * | 2020-05-19 | 2020-08-28 | 杭州英视信息科技有限公司 | Countermeasure support vector machine |
| CN111832229A (en) * | 2020-07-17 | 2020-10-27 | 中国人民解放军海军工程大学 | Vibration transmission system based on cycleGAN model and training method thereof |
| CN112462001A (en) * | 2020-11-17 | 2021-03-09 | 吉林大学 | Gas sensor array model calibration method for data amplification based on condition generation countermeasure network |
| WO2021046723A1 (en) * | 2019-09-10 | 2021-03-18 | 西门子股份公司 | Sensor model generation method and system, and sensor measurement method and system |
| CN112884076A (en) * | 2021-03-24 | 2021-06-01 | 中国科学院计算技术研究所 | Sensor data generation model and method for generating countermeasure network based on conditional expression |
| US20210216857A1 (en) * | 2018-09-17 | 2021-07-15 | Robert Bosch Gmbh | Device and method for training an augmented discriminator |
| US20210271591A1 (en) * | 2020-02-27 | 2021-09-02 | At&T Intellectual Property I, L.P. | Mock data generator using generative adversarial networks |
| CN113542221A (en) * | 2021-06-15 | 2021-10-22 | 四川英得赛克科技有限公司 | Method and system for judging tampering of sensor data of intelligent substation, electronic equipment and storage medium |
| CN113673680A (en) * | 2021-08-20 | 2021-11-19 | 上海大学 | Model verification method and system for automatically generating verification properties through countermeasure network |
| CN113768523A (en) * | 2021-11-11 | 2021-12-10 | 华南理工大学 | Method and system for prewarning stool based on countermeasure generation network |
| WO2023172460A1 (en) * | 2022-03-07 | 2023-09-14 | Applied Materials, Inc. | Synthetic time series data associated with processing equipment |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180101770A1 (en) * | 2016-10-12 | 2018-04-12 | Ricoh Company, Ltd. | Method and system of generative model learning, and program product |
| CN107909621A (en) * | 2017-11-16 | 2018-04-13 | 深圳市唯特视科技有限公司 | It is a kind of based on it is twin into confrontation network medical image synthetic method |
| CN107945118A (en) * | 2017-10-30 | 2018-04-20 | 南京邮电大学 | A kind of facial image restorative procedure based on production confrontation network |
-
2018
- 2018-06-08 CN CN201810589194.1A patent/CN109086658B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180101770A1 (en) * | 2016-10-12 | 2018-04-12 | Ricoh Company, Ltd. | Method and system of generative model learning, and program product |
| CN107945118A (en) * | 2017-10-30 | 2018-04-20 | 南京邮电大学 | A kind of facial image restorative procedure based on production confrontation network |
| CN107909621A (en) * | 2017-11-16 | 2018-04-13 | 深圳市唯特视科技有限公司 | It is a kind of based on it is twin into confrontation network medical image synthetic method |
Non-Patent Citations (3)
| Title |
|---|
| IAN J.GOODFELLOW等: ""Generative Adversarial Nets"", 《ARXIV:1406.2661V1》 * |
| 时光杂货店: ""生成对抗网络(GAN)简单梳理"", 《HTTP://BLOG.CSD N.NET/XG123321123/ARTICLE/DETAILS/78034859》 * |
| 王坤峰等: ""生成式对抗网络:从生成数据到创造智能"", 《自动化学报》 * |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US12050990B2 (en) * | 2018-09-17 | 2024-07-30 | Robert Bosch Gmbh | Device and method for training an augmented discriminator |
| US20210216857A1 (en) * | 2018-09-17 | 2021-07-15 | Robert Bosch Gmbh | Device and method for training an augmented discriminator |
| CN111383651A (en) * | 2018-12-29 | 2020-07-07 | Tcl集团股份有限公司 | Voice noise reduction method and device and terminal equipment |
| CN109889452A (en) * | 2019-01-07 | 2019-06-14 | 中国科学院计算技术研究所 | Network context flow generation method and system based on condition production confrontation network |
| CN109889452B (en) * | 2019-01-07 | 2021-06-11 | 中国科学院计算技术研究所 | Network background flow generation method and system based on condition generation type countermeasure network |
| CN109813542A (en) * | 2019-03-15 | 2019-05-28 | 中国计量大学 | The method for diagnosing faults of air-treatment unit based on production confrontation network |
| CN110046653A (en) * | 2019-03-22 | 2019-07-23 | 赣州好朋友科技有限公司 | A kind of white tungsten method for separating and system based on XRT ray |
| CN110046653B (en) * | 2019-03-22 | 2021-05-25 | 赣州好朋友科技有限公司 | White tungsten sorting method and system based on XRT rays |
| CN110133643A (en) * | 2019-05-22 | 2019-08-16 | 北京林业大学 | Root system of plant detection method and device |
| CN110533578A (en) * | 2019-06-05 | 2019-12-03 | 广东世纪晟科技有限公司 | Image translation method based on conditional countermeasure neural network |
| WO2021046723A1 (en) * | 2019-09-10 | 2021-03-18 | 西门子股份公司 | Sensor model generation method and system, and sensor measurement method and system |
| CN110769196A (en) * | 2019-10-17 | 2020-02-07 | 天津大学 | Video prediction method for discontinuous monitoring road section |
| CN110880020A (en) * | 2019-10-30 | 2020-03-13 | 西安交通大学 | Self-adaptive trans-regional base station energy consumption model migration and compensation method |
| CN110880020B (en) * | 2019-10-30 | 2022-10-25 | 西安交通大学 | Self-adaptive trans-regional base station energy consumption model migration and compensation method |
| CN111028941A (en) * | 2019-12-10 | 2020-04-17 | 南京医科大学 | TKA preoperative clinical staging intelligent evaluation method based on generation of antagonistic network |
| US20210271591A1 (en) * | 2020-02-27 | 2021-09-02 | At&T Intellectual Property I, L.P. | Mock data generator using generative adversarial networks |
| CN111597338A (en) * | 2020-05-19 | 2020-08-28 | 杭州英视信息科技有限公司 | Countermeasure support vector machine |
| CN111597338B (en) * | 2020-05-19 | 2023-04-18 | 杭州英视信息科技有限公司 | Countermeasure support vector machine |
| CN111832229A (en) * | 2020-07-17 | 2020-10-27 | 中国人民解放军海军工程大学 | Vibration transmission system based on cycleGAN model and training method thereof |
| CN112462001A (en) * | 2020-11-17 | 2021-03-09 | 吉林大学 | Gas sensor array model calibration method for data amplification based on condition generation countermeasure network |
| CN112462001B (en) * | 2020-11-17 | 2021-07-23 | 吉林大学 | Gas sensor array model calibration method for data amplification based on condition generation countermeasure network |
| CN112884076A (en) * | 2021-03-24 | 2021-06-01 | 中国科学院计算技术研究所 | Sensor data generation model and method for generating countermeasure network based on conditional expression |
| CN112884076B (en) * | 2021-03-24 | 2022-07-15 | 中国科学院计算技术研究所 | Sensor data generation model and method for generating confrontation network based on conditional expression |
| CN113542221B (en) * | 2021-06-15 | 2023-11-03 | 四川英得赛克科技有限公司 | Method and system for judging falsification of sensor data of intelligent substation, electronic equipment and storage medium |
| CN113542221A (en) * | 2021-06-15 | 2021-10-22 | 四川英得赛克科技有限公司 | Method and system for judging tampering of sensor data of intelligent substation, electronic equipment and storage medium |
| CN113673680A (en) * | 2021-08-20 | 2021-11-19 | 上海大学 | Model verification method and system for automatically generating verification properties through countermeasure network |
| CN113673680B (en) * | 2021-08-20 | 2023-09-15 | 上海大学 | Model verification method and system for automatically generating verification properties through an antagonism network |
| CN113768523B (en) * | 2021-11-11 | 2022-04-22 | 华南理工大学 | Method and system for prewarning stool based on countermeasure generation network |
| CN113768523A (en) * | 2021-11-11 | 2021-12-10 | 华南理工大学 | Method and system for prewarning stool based on countermeasure generation network |
| WO2023172460A1 (en) * | 2022-03-07 | 2023-09-14 | Applied Materials, Inc. | Synthetic time series data associated with processing equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109086658B (en) | 2021-06-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109086658A (en) | A kind of sensing data generation method and system based on generation confrontation network | |
| CN106682616B (en) | Method for recognizing neonatal pain expression based on two-channel feature deep learning | |
| Singh et al. | An Iris Recognition System Using CNN & VGG16 Technique | |
| CN109558896A (en) | Disease intelligent analysis method and system based on ultrasound omics and deep learning | |
| Prasetio et al. | The facial stress recognition based on multi-histogram features and convolutional neural network | |
| WO2019137538A1 (en) | Emotion representative image to derive health rating | |
| Cheng et al. | Classification of long sequential data using circular dilated convolutional neural networks | |
| Tsietso et al. | Multi-input deep learning approach for breast cancer screening using thermal infrared imaging and clinical data | |
| Salazar-Jurado et al. | Towards the generation of synthetic images of palm vein patterns: A review | |
| CN109598671A (en) | Image generating method, device, equipment and medium | |
| Battleday et al. | Modeling human categorization of natural images using deep feature representations | |
| Laves et al. | Quantifying the uncertainty of deep learning-based computer-aided diagnosis for patient safety | |
| Guest et al. | Levels of representation in a deep learning model of categorization | |
| Kumar et al. | Dorsal hand vein recognition using very deep learning | |
| Bandyopadhyay et al. | Alzheimer’s disease detection using ensemble learning and artificial neural networks | |
| Himel et al. | Vision intelligence for smart sheep farming: applying ensemble learning to detect sheep breeds | |
| Putri et al. | Enhancing Pneumonia Disease Classification using Genetic Algorithm-Tuned DCGANs and VGG-16 Integration | |
| De Waele et al. | Time-series-based feature selection and clustering for equine activity recognition using accelerometers | |
| CN110675312B (en) | Image data processing method, device, computer equipment and storage medium | |
| Avramova | Curriculum learning with deep convolutional neural networks | |
| Cui et al. | Research on broiler health status recognition method based on improved YOLOv5 | |
| Raja et al. | IRIS recognition system using neural network and genetic algorithm | |
| Jeyaraj et al. | A review of techniques for image classification to enhance online animal adoption speed | |
| CN114842972A (en) | Method, device, electronic equipment and medium for determining user state | |
| Lin et al. | DHI-GAN: improving dental-based human identification using generative adversarial networks |
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 |