CN107607942A - Based on the large scale electromagnetic scattering of deep learning model and the Forecasting Methodology of back scattering - Google Patents
Based on the large scale electromagnetic scattering of deep learning model and the Forecasting Methodology of back scattering Download PDFInfo
- Publication number
- CN107607942A CN107607942A CN201710767562.2A CN201710767562A CN107607942A CN 107607942 A CN107607942 A CN 107607942A CN 201710767562 A CN201710767562 A CN 201710767562A CN 107607942 A CN107607942 A CN 107607942A
- Authority
- CN
- China
- Prior art keywords
- mrow
- target area
- neural networks
- convolutional neural
- msup
- 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
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000013136 deep learning model Methods 0.000 title claims abstract description 21
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 81
- 238000012549 training Methods 0.000 claims abstract description 53
- 230000007246 mechanism Effects 0.000 claims abstract description 23
- 238000012360 testing method Methods 0.000 claims description 28
- 238000002310 reflectometry Methods 0.000 claims description 27
- 238000003384 imaging method Methods 0.000 claims description 23
- 230000006870 function Effects 0.000 claims description 17
- 230000005540 biological transmission Effects 0.000 claims description 13
- 238000003874 inverse correlation nuclear magnetic resonance spectroscopy Methods 0.000 claims description 8
- 230000007935 neutral effect Effects 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 abstract description 6
- 230000005672 electromagnetic field Effects 0.000 abstract 1
- 238000013528 artificial neural network Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 4
- 238000005259 measurement Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 238000004088 simulation Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000009434 installation Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007123 defense Effects 0.000 description 1
- 238000012938 design process Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 239000004744 fabric Substances 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 108090000623 proteins and genes Proteins 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 230000008685 targeting Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a kind of based on the large scale electromagnetic scattering of deep learning model and the Forecasting Methodology of back scattering.The present invention proposes two kinds of convolutional neural networks based on physical mechanism respectively for electromagnetic scattering and Inverse Problems in Electromagnetics;Existing convolutional neural networks framework is generalized to complex field from real number field, and assigns its corresponding physical meaning;The Forecasting Methodology of scattering and the back scattering of the present invention is applied to conventional various radar systems;The convolutional neural networks of the present invention are applied to training algorithms all at present;Network is convolution operation, thus has extremely strong applicability, suitable for any large scale electromagnetic field scape.It is proposed by the invention to solve electromagnetic scattering using based on the convolutional neural networks of physical mechanism and having that computational efficiency is high during inverse Problem, the characteristics such as time delay is low, to realize that the real-time solution of large scale electromagnetic scattering and inverse Problem is laid a good foundation.
Description
Technical field
The present invention relates to electromagnetic scattering and back scattering technology, and in particular to a kind of large scale electricity based on deep learning model
Magnetic scattering and the Forecasting Methodology of back scattering.
Background technology
With the rapid development of economic society, large scale electromagnetic scattering and back scattering have become remote sensing, electromagnetism stealth and
Key, the Basic Problems of the various fields such as medical treatment.
Over-the-horizon targeting capability is considered as the key factor that can a country win modernized war.Due to current
The radar that army equips has the characteristics that operating distance is remote, detecting precision is high, all weather operations more, thus how to design high motor-driven
Property, high stealth aircraft becomes the important topic of national defense industry research.In high speed stealthy aircraft design process, such as
An important factor for what quick obtaining target scattering data turns into flow of research.In the traditional meter of academia and industrial quarters generally use
Calculate electromagnetic method (numerical method such as such as Finite-Difference Time-Domain Method, moment method) or using based on the business simulation for calculating electromagnetic method
Software obtains target scattering data.Although the advantages that these methods have technology maturation, and precision is high, when target area is larger
When, then traditional calculations electromagnetic method needs to consume a large amount of computing resources, the shortcomings of calculating overlong time, has seriously tied down aircraft
The progress of design.For improve computational efficiency, researcher it is also proposed various numerical approximation methods (such as Born approximation method,
High-frequency approximation method etc.), these methods more ignore between target internal and target electromagnetic coupled effect, cause computational accuracy compared with
It is low.
On the other hand, Inverse Problems in Electromagnetics (such as radar imagery) has nowadays been widely used in geographical science, medical science
And other various military and civilian scenes.Among the past few decades, Inverse Problems in Electromagnetics is also every country research work
Person's priority fields of study.Because Inverse Problems in Electromagnetics has uncomfortable fixed, the extremely challenging feature such as non-linear and pathosis,
Although numerous algorithms have been proposed (such as in scientist:Back-projection algorithm, Born's iterative method, distorted born iterative method method, gene are calculated
Method etc.), but still can not effectively meet the needs of engineer applied.When calculating scene increase, conventional method can meet with meter
Calculation amount is big, the problem of can not meeting real-time application requirement.
In recent years, deep learning has become the important research direction of artificial intelligence field, is widely used in pattern knowledge
Not, classify, the field such as detection, and achieve unprecedented success.However, deep learning network is applied to real number field more at present, and
Electromagnetic scattering problems and Inverse Problems in Electromagnetics belong to complex field problem, and deep learning framework does not have accurate thing at present
Meaning support is managed, it is difficult to directly apply to electromagnetic scattering and Inverse Problems in Electromagnetics to cause existing deep learning framework.
How under the conditions of existing hardware system, propose that precision is high, calculate fast solution electromagnetic scattering and inverse Problem
Algorithm have become those skilled in the art be badly in need of solve extremely challenging key technology difficulty.
The content of the invention
In order to solve above-mentioned key technology difficulty, the present invention proposes the deep learning model based on physical mechanism respectively
The prediction of the Forecasting Methodology of large scale electromagnetic scattering and the large scale THE INVERSE ELECTROMAGNETIC SCATTERING of deep learning model based on physical mechanism
Method;The present invention directly applies to the big chi of prediction by training small yardstick electromagnetic scattering and back scattering data, and then by the network
Spend electromagnetic scattering and inverse Problem.
It is an object of the present invention to propose a kind of prediction side of the large scale electromagnetic scattering based on deep learning model
Method.
The Forecasting Methodology of the large scale electromagnetic scattering based on deep learning model of the present invention, comprises the following steps:
1) training sample is obtained:
Electromagnetism radar system includes T emitter, M receiver, emitter successively to target area transmission signal, and by
Whole receivers receive the scattered field of target area, and contrast function is the reflectivity of target area and corresponding target area
Scattered field forms one group of sample, and multigroup sample data is two parts by random division:Wherein Part I sample be used to train
Convolutional neural networks are as training sample;Part II sample is used in testing for the generalization ability of convolutional neural networks as test
Sample;
2) convolutional neural networks are built:
Convolutional neural networks are built based on physical mechanism in series;
3) convolutional neural networks based on physical mechanism are trained:
A) output using the scattered field of the target area of the training sample obtained in step 1) as convolutional neural networks, will
Input of the reflectivity of the target area of training sample as convolutional neural networks, training convolutional neural networks;
B) output using the scattered field of the target area of the test sample obtained in step 1) as convolutional neural networks, will
Input of the reflectivity of the target area of test sample as convolutional neural networks, convolutional neural networks are examined, if error exists
In critical field, then convolutional neural networks training is completed, into step 4), the scope if error is above standard, and return to step
A) re -training network, until error, in critical field, training network terminates, into step 4);
4) the scattering prediction of large scale electromagnetism scene:
Using the reflectivity of the target area of large scale electromagnetism scene as the input of convolutional neural networks in step 3), then roll up
The output of product neutral net is the scattered field of the target area of the corresponding large scale electromagnetism scene of prediction.
Wherein, in step 1), electromagnetism radar system includes T emitter, and M receiver, emitter is successively to target
Field emission signal, and by the scattered field of whole receivers reception target area;When t-th of emitter transmission signal, m-th
The scattered field for the target area that receiver receivesFor:
And
Wherein, T and M respectively >=2 natural number, t=1,2 ..., T;M=1,2 ..., M;DinvRepresent target area;r
With r ' ∈ DinvSite and source point, r are represented respectivelytFor the position where t-th of emitter, rmFor the position where m-th of receiver
Put;E(t)(r ') represents the resultant field of target area during t-th of emitter transmission signal;Contrast function (the i.e. reflection of target area
Rate) beWherein k and k0=ω/c represents the wave number in target and free space respectively, and ω represents center frequency
Rate, c represent the light velocity;Two-dimentional Green's function is represented,Represent the first kind zeroth order Chinese
Gram function, i represent imaginary unit.
In step 2), a convolutional neural networks of the N layers based on physical mechanism, wherein n-th layer network output table are built
It is shown as:
E(n)=Ein+A(n)χ(E(n-1)) (3)
Wherein, EinRepresent corresponding in-field during generation training sample;E(n-1)Represent that the (n-1)th layer network exports, and be the
The input of n-layer network, n are natural number and 1≤n≤N-1;A(n)Represent the convolution kernel of n-th layer network;χ represents the mesh of training sample
Mark the reflectivity in region;E(n)Represent the output of n-th layer network;The input of initial network is E(0)=Ein;N-th layer network exports table
It is shown as:
E(N)=A(N)χ(E(N-1)) (4)
Wherein, A(N)=G (rm,rt) represent n-th layer network convolution kernel, G (rm,rt) represent rtLocate emitter to rmPlace connects
The Green's function of receipts machine.
In step 3), critical field≤0.02.
It is another object of the present invention to provide a kind of the pre- of large scale THE INVERSE ELECTROMAGNETIC SCATTERING based on deep learning model
Survey method.
The Forecasting Methodology of the large scale THE INVERSE ELECTROMAGNETIC SCATTERING based on deep learning model of the present invention, comprises the following steps:
1) training sample is obtained:
Electromagnetism radar system includes T emitter, M receiver, emitter successively to target area transmission signal, and by
Whole receivers receive the scattered field of target area, and contrast function is the reflectivity of target area and corresponding target area
Scattered field forms one group of sample, and multigroup sample data is two parts by random division:Wherein Part I sample be used to train
Convolutional neural networks are as training sample;Part II sample is used in testing for the generalization ability of convolutional neural networks as test
Sample;
2) convolutional neural networks are built:
Convolutional neural networks are built based on physical mechanism in series;
3) convolutional neural networks based on physical mechanism are trained:
A) scattered field of the target area of the training sample obtained in step 1) is subjected to rear orientation projection's imaging, rear orientation projection
Input of the imaging results as convolutional neural networks, using the reflectivity of the target area of training sample as convolutional neural networks
Output, training convolutional neural networks;
B) output using the reflectivity of the target area of the test sample obtained in step 1) as convolutional neural networks, will
Input of the rear orientation projection's imaging results of the target area of test sample as convolutional neural networks, convolutional neural networks are examined,
If error, in critical field, convolutional neural networks training is completed, into step 4), the scope if error is above standard,
Then return to step a) re -trainings network, until error, in critical field, training network terminates, into step 4);
4) the back scattering prediction of large scale electromagnetism scene:
Using rear orientation projection's imaging results of large scale electromagnetism scene as the input of convolutional neural networks in step 3), then roll up
The output of product neutral net is the imaging results of the target area of the corresponding large scale electromagnetism scene of prediction.
Wherein, in step 2), N layer series connection convolutional neural networks of the structure one based on physical mechanism, wherein single laminate roll
Product neutral net is divided into three sublayers:First sublayer is input layer, characterizes observational variable, such as rear orientation projection's imaging results;The
Two sublayers are characterized layer, characterize the induced-current of reconstruct;Third layer is output layer, characterizes the reflectivity of the target area of reconstruct.
The observational variable of input layer with K convolution kernel convolution, obtains K eigenmatrix of the second sublayer, that is, characterized respectively
Induced-current under different visual angles.K eigenmatrix connect entirely or multilayer convolution by way of obtain the target area of output layer
The reflectivity in domain.
Advantages of the present invention:
The present invention proposes two kinds of convolution god based on physical mechanism respectively for electromagnetic scattering and Inverse Problems in Electromagnetics
Through network;Existing convolutional neural networks framework is generalized to complex field from real number field, and assigns its corresponding physical meaning;This hair
Bright scattering and the Forecasting Methodology of back scattering are applied to conventional various radar systems;The convolutional neural networks of the present invention are applied to
Training algorithms all at present;Network is convolution operation, thus has extremely strong applicability, suitable for any large scale electromagnetism
Scene.It is proposed by the invention to have using based on the convolutional neural networks of physical mechanism when solving electromagnetic scattering and inverse Problem
Have that computational efficiency is high, the characteristics such as time delay is low, to realize that base has been established in the real-time solution of large scale electromagnetic scattering and inverse Problem
Plinth.
Brief description of the drawings
Fig. 1 is that the Forecasting Methodology of large scale electromagnetic scattering based on deep learning model and the back scattering of the present invention is applicable
Exemplary two dimensional electromagnetic scattering and the schematic diagram of back scattering system;
Fig. 2 is the large scale electromagnetic scattering prediction method based on deep learning model of the present invention based on physical mechanism
The structured flowchart of convolutional neural networks;
Fig. 3 be the present invention based on the large scale electromagnetic scattering prediction method of deep learning model under small yardstick scene
The comparison figure of prediction effect;
Fig. 4 is the Forecasting Methodology of the small yardstick electromagnetic scattering based on deep learning model of the present invention under large scale scene
Prediction effect comparison figure;
Fig. 5 is the Forecasting Methodology of the large scale THE INVERSE ELECTROMAGNETIC SCATTERING based on deep learning model of the present invention based on physical machine
Make the structured flowchart of the convolutional neural networks of structure;
Fig. 6 is the Forecasting Methodology of the small yardstick THE INVERSE ELECTROMAGNETIC SCATTERING based on deep learning model of the present invention in small yardstick scene
Under convolutional neural networks prediction effect figure;
Fig. 7 is the Forecasting Methodology of the large scale THE INVERSE ELECTROMAGNETIC SCATTERING based on deep learning model of the present invention in large scale scene
Under convolutional neural networks prediction effect figure.
Embodiment
Below in conjunction with the accompanying drawings, by specific embodiment, the present invention is expanded on further.
Embodiment one
In the present embodiment, the structure of two-dimensional simulation system is as shown in figure 1, radar system uses bistatic.
The Forecasting Methodology of the large scale electromagnetic scattering based on deep learning model of the present embodiment, comprises the following steps:
1) training sample is obtained:
Electromagnetism radar system includes T=1 emitter, M=19 receiver, is R=10 λ away from origin radius, uniformly divides
Cloth is on the semicircle between -90 ° to 90 °.The incident plane wave signal of target area is 4GHz simple signals, target area (L=
1.4 λ) uniformly it is split into the λ of 196 0.1 λ × 0.1 square net.Emitter successively to target area transmission signal, and
The scattered field of target area is received by whole receivers, contrast function is the reflectivity of target area and corresponding target area
Scattered field formed one group of sample.10000 groups of sample datas are two parts by random division:Wherein 5000 groups of samples of Part I
It is used for training network as training sample;5000 groups of samples of Part II are used in testing for the generalization ability of network as test
Sample.
2) convolutional neural networks are built:
Convolutional neural networks are built based on physical mechanism in series as shown in Fig. 2 being set to 5 layer networks.
Preceding four layers of convolution kernel Initialize installation is the random matrix of the Gaussian Profile of 14 × 14 zero-mean.
3) convolutional neural networks based on physical mechanism are trained:
Using the scattered field of the target area of the training sample obtained in step 1) as the output of convolutional neural networks, will instruct
Practice input of the reflectivity of the target area of sample as convolutional neural networks, training convolutional neural networks.Using back-propagating
Algorithm updates network parameter.During back-propagating, cost function is expressed as:
Wherein EtrueRepresent measurement scattered field.
The gradient of each layer network parameter can be expressed as:
The error between the cost function quantitative analysis measurement scattered field and prediction scattered field in formula (8) is utilized simultaneously:
WhereinWithThe measurement scattered field and prediction scattered field of s group test samples are represented respectively, and M represents to receive
The number of machine, the sum of S table test samples.
To illustrate the validity of the convolutional neural networks based on physical mechanism, the present embodiment will use traditional Born approximation side
Prediction scattering field data caused by method does further comparison with neural network forecast scattering field data.Traditional Born approximation method is predicted
Scattering field data can be obtained by formula (9):
As shown in figure 3, the measurement scattered field in test sample and the scattering using neural network forecast of the invention in step 1)
Field error is minimum, and error E RR is only 0.0016;And the scattered field obtained by Born approximation is then with testing scattered field difference very
Greatly, only part trend is identical.
4) the scattering prediction of large scale electromagnetism scene:
Imaging region is increased into the square area that the length of side is L=2.8 λ, then imaging region is divided into 784 0.1 λ
× 0.1 λ square net.The target area of large scale increases four times greater compared to the target area of step 1) Small and Medium Sized.It is other
Parameter constant, produce 5000 groups of test samples.Rolled up using the reflectivity of the target area of large scale electromagnetism scene as in step 3)
The input of product neutral net, then the output of convolutional neural networks is the target area of the corresponding large scale electromagnetism scene of prediction
Scattered field.As shown in figure 4, network measure scattered field and prediction scattered field are basically identical, error E RR is smaller, and ERR is only 0.017.
Embodiment two
In the present embodiment, the structure of two-dimensional simulation system is as shown in figure 1, radar system uses bistatic.
The Forecasting Methodology of the large scale THE INVERSE ELECTROMAGNETIC SCATTERING based on deep learning model of the present embodiment, comprises the following steps:
1) training sample is obtained:
System architecture such as embodiment one, centered on origin, transmitter and receiver is rotated to 45 ° of acquisitions, one group of number simultaneously
According to the total that rotates a circle obtains eight groups of data as a sample data.Caused 10000 groups of sample stochastic averaginas are divided into two
Part, wherein 5000 groups of samples of Part I are used for training network as training sample, and 5000 groups of samples of Part II are used to survey
Network efficiency is tried as test sample.
2) convolutional neural networks are built:
The convolutional neural networks built in series based on physical mechanism are as shown in figure 5, be individual layer by network settings
Network, and be eight convolution kernels.Convolution kernel Initialize installation is the random matrix of the Gaussian Profile of 14 × 14 zero-mean.
3) convolutional neural networks based on physical mechanism are trained:
The scattered field of the target area of the training sample obtained in step 1) is subjected to rear orientation projection's imaging, rear orientation projection into
As input of the result as convolutional neural networks, using the reflectivity of the target area of training sample as the defeated of convolutional neural networks
Go out, training convolutional neural networks.
And weigh neural network forecast error E RR with the cost function in formula (10):
WhereinWithThe reflectivity and prediction target area in the real goal region of s group test samples are represented respectively
The reflectivity in domain.P represent imaging region grid number, S be test sample sum, s=1 ..., S.
As shown in fig. 6, rear orientation projection's imaging results noise is larger, objective fuzzy, can not be imaged by human eye Direct Recognition
Target, mean error are up to ERR=0.023.And passing through training network using rear orientation projection's result as input, its training network is defeated
The prediction imaging results gone out are then clear in structure, and noise is smaller, and mean error is only ERR=0.009.
4) the back scattering prediction of large scale electromagnetism scene:
Imaging region is increased into the square area that the length of side is L=2.8 λ, then imaging region is divided into 784 0.1 λ
× 0.1 λ square net.The target area of large scale increases four times greater compared to the target area of step 1) Small and Medium Sized.It is other
Parameter constant, produce 5000 groups of test samples.Using the target area rear orientation projection imaging results of large scale test sample as step
It is rapid 3) in convolutional neural networks input, then network output be prediction large scale target area imaging results.Can by Fig. 7
Know, under large scale scene, neural network forecast imaging results are much better than rear orientation projection's imaging results.Rear orientation projection and neural network forecast knot
The mean error of fruit is respectively 0.044 and 0.020.
It is finally noted that the purpose for publicizing and implementing example is that help further understands the present invention, but this area
Technical staff be appreciated that:Without departing from the spirit and scope of the invention and the appended claims, it is various to replace and repair
It is all possible for changing.Therefore, the present invention should not be limited to embodiment disclosure of that, and the scope of protection of present invention is to weigh
The scope that sharp claim defines is defined.
Claims (6)
- A kind of 1. Forecasting Methodology of the large scale electromagnetic scattering based on deep learning model, it is characterised in that the Forecasting Methodology Comprise the following steps:1) training sample is obtained:Electromagnetism radar system includes T emitter, and M receiver, emitter is successively to target area transmission signal, and by whole Receiver receives the scattered field of target area, and contrast function is reflectivity and the scattering of corresponding target area of target area Field forms one group of sample, and multigroup sample data is two parts by random division:Wherein Part I sample is used for training convolutional Neutral net is as training sample;Part II sample is used in testing for the generalization ability of convolutional neural networks as test specimens This;2) convolutional neural networks are built:Convolutional neural networks are built based on physical mechanism in series;3) convolutional neural networks based on physical mechanism are trained:A) using the scattered field of the target area of the training sample obtained in step 1) as the output of convolutional neural networks, will train Input of the reflectivity of the target area of sample as convolutional neural networks, training convolutional neural networks;B) using the scattered field of the target area of the test sample obtained in step 1) as the output of convolutional neural networks, will test Input of the reflectivity of the target area of sample as convolutional neural networks, convolutional neural networks are examined, if error is in standard In the range of, then convolutional neural networks training is completed, into step 4), the scope if error is above standard, and return to step a) weights New training network, until error, in critical field, training network terminates, into step 4);4) the scattering prediction of large scale electromagnetism scene:Input using the reflectivity of the target area of large scale electromagnetism scene as convolutional neural networks in step 3), then convolution is refreshing Output through network is the scattered field of the target area of the corresponding large scale electromagnetism scene of prediction.
- 2. Forecasting Methodology as claimed in claim 1, it is characterised in that in step 1), electromagnetism radar system includes T transmitting Machine, M receiver, emitter successively to target area transmission signal, and by whole receivers receive target area scattered field; When t-th of emitter transmission signal, the scattered field for the target area that m-th of receiver receivesFor:<mrow> <msubsup> <mi>E</mi> <mrow> <mi>s</mi> <mi>c</mi> <mi>a</mi> </mrow> <mrow> <mo>(</mo> <mi>t</mi> <mi>m</mi> <mo>)</mo> </mrow> </msubsup> <mrow> <mo>(</mo> <msub> <mi>r</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <msubsup> <mi>k</mi> <mn>0</mn> <mn>2</mn> </msubsup> <munder> <mo>&Integral;</mo> <msub> <mi>D</mi> <mrow> <mi>i</mi> <mi>n</mi> <mi>v</mi> </mrow> </msub> </munder> <mi>G</mi> <mrow> <mo>(</mo> <msub> <mi>r</mi> <mi>m</mi> </msub> <mo>,</mo> <msub> <mi>r</mi> <mi>t</mi> </msub> <mo>)</mo> </mrow> <mi>&chi;</mi> <mrow> <mo>(</mo> <mi>r</mi> <mo>)</mo> </mrow> <msup> <mi>E</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msup> <mrow> <mo>(</mo> <mi>r</mi> <mo>)</mo> </mrow> <mi>d</mi> <mi>r</mi> </mrow>And<mrow> <msup> <mi>E</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msup> <mrow> <mo>(</mo> <mi>r</mi> <mo>)</mo> </mrow> <mo>=</mo> <msubsup> <mi>E</mi> <mrow> <mi>i</mi> <mi>n</mi> </mrow> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msubsup> <mrow> <mo>(</mo> <mi>r</mi> <mo>)</mo> </mrow> <mo>+</mo> <msubsup> <mi>k</mi> <mn>0</mn> <mn>2</mn> </msubsup> <munder> <mo>&Integral;</mo> <msub> <mi>D</mi> <mrow> <mi>i</mi> <mi>n</mi> <mi>v</mi> </mrow> </msub> </munder> <mi>G</mi> <mrow> <mo>(</mo> <mi>r</mi> <mo>,</mo> <msup> <mi>r</mi> <mo>&prime;</mo> </msup> <mo>)</mo> </mrow> <mi>&chi;</mi> <mrow> <mo>(</mo> <msup> <mi>r</mi> <mo>&prime;</mo> </msup> <mo>)</mo> </mrow> <msup> <mi>E</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msup> <mrow> <mo>(</mo> <msup> <mi>r</mi> <mo>&prime;</mo> </msup> <mo>)</mo> </mrow> <msup> <mi>dr</mi> <mo>&prime;</mo> </msup> </mrow>Wherein, T and M respectively >=2 natural number, t=1,2 ..., T;M=1,2 ..., M;DinvRepresent target area;R and r ' ∈DinvSite and source point, r are represented respectivelytFor the position where t-th of emitter, rmFor the position where m-th of receiver;E(t)(r ') represents the resultant field of target area during t-th of emitter transmission signal;The reflectivity of target area isWherein k and k0=ω/ c represents the wave number in target and free space respectively, and ω represents centre frequency, c tables Show the light velocity;Two-dimentional Green's function is represented,First kind zeroth order Hunk function is represented, I represents imaginary unit.
- 3. Forecasting Methodology as claimed in claim 1, it is characterised in that in step 2), one N layer of structure is based on physical mechanism Convolutional neural networks, wherein n-th layer network output be expressed as:E(n)=Ein+A(n)χ(E(n-1))Wherein, EinRepresent corresponding in-field during generation training sample;E(n-1)Represent that the (n-1)th layer network exports, and be n-th layer net The input of network, n are natural number and 1≤n≤N-1;A(n)Represent the convolution kernel of n-th layer network;χ represents the target area of training sample The reflectivity in domain;E(n)Represent the output of n-th layer network;The input of initial network is E(0)=Ein;The output of n-th layer network represents For:E(N)=A(N)χ(E(N-1))Wherein, A(N)=G (rm,rt) represent n-th layer network convolution kernel, G (rm,rt) represent rtLocate emitter to rmLocate receiver Green's function.
- A kind of 4. Forecasting Methodology of the large scale THE INVERSE ELECTROMAGNETIC SCATTERING based on deep learning model, it is characterised in that the prediction side Method comprises the following steps:1) training sample is obtained:Electromagnetism radar system includes T emitter, and M receiver, emitter is successively to target area transmission signal, and by whole Receiver receives the scattered field of target area, and contrast function is reflectivity and the scattering of corresponding target area of target area Field forms one group of sample, and multigroup sample data is two parts by random division:Wherein Part I sample is used for training convolutional Neutral net is as training sample;Part II sample is used in testing for the generalization ability of convolutional neural networks as test specimens This;2) convolutional neural networks are built:Convolutional neural networks are built based on physical mechanism in series;3) convolutional neural networks based on physical mechanism are trained:A) scattered field of the target area of the training sample obtained in step 1) is subjected to rear orientation projection's imaging, rear orientation projection's imaging As a result the input as convolutional neural networks, using the reflectivity of the target area of training sample as the defeated of convolutional neural networks Go out, training convolutional neural networks;B) using the reflectivity of the target area of the test sample obtained in step 1) as the output of convolutional neural networks, will test Input of the rear orientation projection's imaging results of the target area of sample as convolutional neural networks, convolutional neural networks are examined, if Error in critical field, then complete, and into step 4), the scope if error is above standard, returns by convolutional neural networks training Step a) re -training networks are returned, until error, in critical field, training network terminates, into step 4);4) the back scattering prediction of large scale electromagnetism scene:Input using rear orientation projection's imaging results of large scale electromagnetism scene as convolutional neural networks in step 3), then convolution is refreshing Output through network is the imaging results of the target area of the corresponding large scale electromagnetism scene of prediction.
- 5. a kind of Forecasting Methodology as claimed in claim 4, it is characterised in that in step 1), electromagnetism radar system includes T Emitter, M receiver, emitter receive dissipating for target area successively to target area transmission signal by whole receivers Penetrate field;When t-th of emitter transmission signal, the scattered field for the target area that m-th of receiver receivesFor:<mrow> <msubsup> <mi>E</mi> <mrow> <mi>s</mi> <mi>c</mi> <mi>a</mi> </mrow> <mrow> <mo>(</mo> <mi>t</mi> <mo>,</mo> <mi>m</mi> <mo>)</mo> </mrow> </msubsup> <mrow> <mo>(</mo> <msub> <mi>r</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <msubsup> <mi>k</mi> <mn>0</mn> <mn>2</mn> </msubsup> <munder> <mo>&Integral;</mo> <msub> <mi>D</mi> <mrow> <mi>i</mi> <mi>n</mi> <mi>v</mi> </mrow> </msub> </munder> <mi>G</mi> <mrow> <mo>(</mo> <msub> <mi>r</mi> <mi>m</mi> </msub> <mo>,</mo> <msub> <mi>r</mi> <mi>t</mi> </msub> <mo>)</mo> </mrow> <mi>&chi;</mi> <mrow> <mo>(</mo> <mi>r</mi> <mo>)</mo> </mrow> <msup> <mi>E</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msup> <mrow> <mo>(</mo> <mi>r</mi> <mo>)</mo> </mrow> <mi>d</mi> <mi>r</mi> </mrow>And<mrow> <msup> <mi>E</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msup> <mrow> <mo>(</mo> <mi>r</mi> <mo>)</mo> </mrow> <mo>=</mo> <msubsup> <mi>E</mi> <mrow> <mi>i</mi> <mi>n</mi> </mrow> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msubsup> <mrow> <mo>(</mo> <mi>r</mi> <mo>)</mo> </mrow> <mo>+</mo> <msubsup> <mi>k</mi> <mn>0</mn> <mn>2</mn> </msubsup> <munder> <mo>&Integral;</mo> <msub> <mi>D</mi> <mrow> <mi>i</mi> <mi>n</mi> <mi>v</mi> </mrow> </msub> </munder> <mi>G</mi> <mrow> <mo>(</mo> <mi>r</mi> <mo>,</mo> <msup> <mi>r</mi> <mo>&prime;</mo> </msup> <mo>)</mo> </mrow> <mi>&chi;</mi> <mrow> <mo>(</mo> <msup> <mi>r</mi> <mo>&prime;</mo> </msup> <mo>)</mo> </mrow> <msup> <mi>E</mi> <mrow> <mo>(</mo> <mi>t</mi> <mo>)</mo> </mrow> </msup> <mrow> <mo>(</mo> <msup> <mi>r</mi> <mo>&prime;</mo> </msup> <mo>)</mo> </mrow> <msup> <mi>dr</mi> <mo>&prime;</mo> </msup> </mrow>Wherein, T and M respectively >=2 natural number, t=1,2 ..., T;M=1,2 ..., M;DinvRepresent target area;R and r ' ∈DinvSite and source point, r are represented respectivelytFor the position where t-th of emitter, rmFor the position where m-th of receiver;E(t)(r ') represents the resultant field of target area during t-th of emitter transmission signal;The reflectivity of target area isWherein k and k0=ω/c represents the wave number in target and free space respectively, and ω represents centre frequency, c tables Show the light velocity;Two-dimentional Green's function is represented,First kind zeroth order Hunk function is represented, I represents imaginary unit.
- 6. a kind of Forecasting Methodology as claimed in claim 4, it is characterised in that in step 2), structure one is based on physical machine The N layers series connection convolutional neural networks of system, wherein individual layer convolutional neural networks are divided into three sublayers:First sublayer is input layer, table Levy observational variable;Second sublayer is characterized layer, characterizes the induced-current of reconstruct;Third layer is output layer, characterizes the target of reconstruct The reflectivity in region.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710767562.2A CN107607942B (en) | 2017-08-31 | 2017-08-31 | Based on the large scale electromagnetic scattering of deep learning model and the prediction technique of back scattering |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710767562.2A CN107607942B (en) | 2017-08-31 | 2017-08-31 | Based on the large scale electromagnetic scattering of deep learning model and the prediction technique of back scattering |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107607942A true CN107607942A (en) | 2018-01-19 |
CN107607942B CN107607942B (en) | 2019-09-13 |
Family
ID=61056780
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710767562.2A Expired - Fee Related CN107607942B (en) | 2017-08-31 | 2017-08-31 | Based on the large scale electromagnetic scattering of deep learning model and the prediction technique of back scattering |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107607942B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111428407A (en) * | 2020-03-23 | 2020-07-17 | 杭州电子科技大学 | Electromagnetic scattering calculation method based on deep learning |
CN111444601A (en) * | 2020-03-23 | 2020-07-24 | 杭州电子科技大学 | AI learning type electromagnetic scattering calculation method suitable for any incident field |
CN111609787A (en) * | 2020-05-28 | 2020-09-01 | 杭州电子科技大学 | Two-step phase-free imaging method for solving electromagnetic backscattering problem based on neural network |
CN111610374A (en) * | 2020-05-28 | 2020-09-01 | 杭州电子科技大学 | Scattered field phase recovery method based on convolutional neural network |
CN112528869A (en) * | 2020-12-14 | 2021-03-19 | 北京航空航天大学杭州创新研究院 | Phase-free data imaging method based on complex neural network |
CN113177356A (en) * | 2021-04-28 | 2021-07-27 | 北京航空航天大学 | Target electromagnetic scattering characteristic rapid prediction method based on deep learning |
CN113705031A (en) * | 2021-06-15 | 2021-11-26 | 西安电子科技大学 | Nano antenna array electromagnetic performance prediction method based on deep learning |
CN114741951A (en) * | 2022-03-11 | 2022-07-12 | 上海师范大学 | Medium target electromagnetic detection method based on convolutional neural network |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103712955A (en) * | 2014-01-02 | 2014-04-09 | 李云梅 | Class-II water atmospheric correction method based on neural network quadratic optimization |
CN106124449A (en) * | 2016-06-07 | 2016-11-16 | 中国科学院合肥物质科学研究院 | A kind of soil near-infrared spectrum analysis Forecasting Methodology based on degree of depth learning art |
WO2017004626A1 (en) * | 2015-07-01 | 2017-01-05 | The Board Of Trustees Of The Leland Stanford Junior University | Systems and methods for providing reinforcement learning in a deep learning system |
US20170116520A1 (en) * | 2015-10-23 | 2017-04-27 | Nec Laboratories America, Inc. | Memory Efficient Scalable Deep Learning with Model Parallelization |
KR20170096282A (en) * | 2016-02-15 | 2017-08-24 | 한국과학기술원 | Deep learning type classification method with feature-based weighting |
-
2017
- 2017-08-31 CN CN201710767562.2A patent/CN107607942B/en not_active Expired - Fee Related
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103712955A (en) * | 2014-01-02 | 2014-04-09 | 李云梅 | Class-II water atmospheric correction method based on neural network quadratic optimization |
WO2017004626A1 (en) * | 2015-07-01 | 2017-01-05 | The Board Of Trustees Of The Leland Stanford Junior University | Systems and methods for providing reinforcement learning in a deep learning system |
US20170116520A1 (en) * | 2015-10-23 | 2017-04-27 | Nec Laboratories America, Inc. | Memory Efficient Scalable Deep Learning with Model Parallelization |
KR20170096282A (en) * | 2016-02-15 | 2017-08-24 | 한국과학기술원 | Deep learning type classification method with feature-based weighting |
CN106124449A (en) * | 2016-06-07 | 2016-11-16 | 中国科学院合肥物质科学研究院 | A kind of soil near-infrared spectrum analysis Forecasting Methodology based on degree of depth learning art |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111428407A (en) * | 2020-03-23 | 2020-07-17 | 杭州电子科技大学 | Electromagnetic scattering calculation method based on deep learning |
CN111444601A (en) * | 2020-03-23 | 2020-07-24 | 杭州电子科技大学 | AI learning type electromagnetic scattering calculation method suitable for any incident field |
CN111444601B (en) * | 2020-03-23 | 2023-06-20 | 杭州电子科技大学 | AI learning type electromagnetic scattering calculation method suitable for arbitrary incident field |
CN111609787B (en) * | 2020-05-28 | 2021-10-01 | 杭州电子科技大学 | Two-step phase-free imaging method for solving electromagnetic backscattering problem based on neural network |
CN111610374A (en) * | 2020-05-28 | 2020-09-01 | 杭州电子科技大学 | Scattered field phase recovery method based on convolutional neural network |
CN111610374B (en) * | 2020-05-28 | 2022-08-05 | 杭州电子科技大学 | Scattered field phase recovery method based on convolutional neural network |
CN111609787A (en) * | 2020-05-28 | 2020-09-01 | 杭州电子科技大学 | Two-step phase-free imaging method for solving electromagnetic backscattering problem based on neural network |
CN112528869A (en) * | 2020-12-14 | 2021-03-19 | 北京航空航天大学杭州创新研究院 | Phase-free data imaging method based on complex neural network |
CN112528869B (en) * | 2020-12-14 | 2023-04-25 | 北京航空航天大学杭州创新研究院 | Phase-free data imaging method based on complex neural network |
CN113177356A (en) * | 2021-04-28 | 2021-07-27 | 北京航空航天大学 | Target electromagnetic scattering characteristic rapid prediction method based on deep learning |
CN113177356B (en) * | 2021-04-28 | 2021-10-15 | 北京航空航天大学 | Target electromagnetic scattering characteristic rapid prediction method based on deep learning |
CN113705031A (en) * | 2021-06-15 | 2021-11-26 | 西安电子科技大学 | Nano antenna array electromagnetic performance prediction method based on deep learning |
CN114741951A (en) * | 2022-03-11 | 2022-07-12 | 上海师范大学 | Medium target electromagnetic detection method based on convolutional neural network |
Also Published As
Publication number | Publication date |
---|---|
CN107607942B (en) | 2019-09-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107607942B (en) | Based on the large scale electromagnetic scattering of deep learning model and the prediction technique of back scattering | |
Chen et al. | False-alarm-controllable radar detection for marine target based on multi features fusion via CNNs | |
CN106355151B (en) | A kind of three-dimensional S AR images steganalysis method based on depth confidence network | |
CN106468770B (en) | Nearly optimal radar target detection method under K Distribution Clutter plus noise | |
CN108182450A (en) | A kind of airborne Ground Penetrating Radar target identification method based on depth convolutional network | |
CN104020451B (en) | Outer transmitter-based radar target track processing method based on clustering | |
CN108008385A (en) | Interference environment ISAR high-resolution imaging methods based on management loading | |
CN108828547A (en) | The high method of the low Elevation of metre wave radar based on deep neural network | |
CN104318593B (en) | Simulation method and system of radar sea clusters | |
Li et al. | IncepTCN: A new deep temporal convolutional network combined with dictionary learning for strong cultural noise elimination of controlled-source electromagnetic data | |
CN109934101A (en) | Radar clutter recognition method based on convolutional neural networks | |
CN105243280A (en) | Time domain physical optics algorithm based on CPU (Central Processing Unit) and GPU (Graphics Processing Unit) hybrid asynchronous parallel way | |
CN110232342A (en) | Sea situation level determination method and device based on convolutional neural networks | |
CN107392863A (en) | SAR image change detection based on affine matrix fusion Spectral Clustering | |
Wei et al. | Intra-pulse modulation radar signal recognition based on Squeeze-and-Excitation networks | |
CN114117912A (en) | Sea clutter modeling and inhibiting method under data model dual drive | |
CN107607945A (en) | A kind of scanning radar forword-looking imaging method based on spatial embedding mapping | |
CN108387880A (en) | Detection method of small target under a kind of sea clutter background based on multiple dimensioned oriented Lyapunov indexes | |
CN107942326B (en) | Two-dimensional active millimeter wave imaging method with high universality | |
CN106353743B (en) | It is matched with the nearly optimal radar target detection method of equivalent shapes parameter | |
CN107833180A (en) | A kind of method using complex field neutral net rapid solving nonlinear electromagnetic inverse Problem | |
CN105954739A (en) | Knowledge-aided nonparametric constant false alarm detection method | |
CN113567975B (en) | Human body rapid security inspection method based on vortex electromagnetic wave mode scanning | |
CN106842201A (en) | A kind of Ship Target ISAR chiasmal image method of discrimination based on sequence image | |
Li et al. | Generalized Hough transform and ANN for subsurface cylindrical object location and parameters inversion from GPR data |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190913 |
|
CF01 | Termination of patent right due to non-payment of annual fee |