CN104459668A - Radar target recognition method based on deep learning network - Google Patents

Abstract

The invention discloses a radar target recognition method based on a deep learning network. The method mainly solves the problems that in the prior art, when a radar high-resolution range image is recognized, the recognition rate is low and robustness is poor. According to the technical scheme, firstly, an obtained original radar high-resolution range image is divided into a training set and a test set; secondly, preprocessing is performed on the training set; thirdly, a cost function of a first-layer network is built, gradient is calculated through a backward conduction algorithm, and optimal parameters are solved through a gradient descend method; fourthly, output of an upper layer serves as input of a lower layer, and optimal parameters of all the layers are obtained through training layer by layer; fifthly, output of the last layer serves as input of a linear support vector machine, and an optimal parameter of the linear support vector machine is obtained through training; sixthly, a test set sample is recognized by means of the optimal parameter of each layer and the optimal parameter of the linear support vector machine. The method improves the recognition rate of the radar high-resolution range image, and meanwhile stability of target recognition is enhanced on the situation that the training sample is incomplete.

Description

Based on the radar target identification method of degree of deep learning network

Technical field

The invention belongs to Radar Technology field, relating to radar utilizes High Range Resolution HRRP extract clarification of objective and identify, particularly utilize multilayer self-correcting scrambler SCAE to unmarked HRRP data compression coding, can be used for all kinds of radar to the identification of target HRRP data and the dimension reducing HRRP data.

Background technology

Along with aircraft, the diversified development of warship and these weapons of war of tank equipment, radar target recognition institute facing challenges is also more and more severeer.The conflict of localised region several times since the nineties in 20th century shows, people propose the requirement of stronger robotization and Geng Gao discrimination to radar target recognition.Therefore the development of radar target recognition must constantly bring forth new ideas to adapt to the challenge that new technology and new environment bring radar target recognition.

High Range Resolution HRRP includes the abundant structural information of target, such as target size, scattering point structure etc.Therefore 20 end of the centurys, some scholars propose the technology utilizing HRRP to complete radar target automatically to identify, see [S.P.Jacobs.Automatic target recognition using high-resolution radar range profiles.PhD dissertation, Washington Univ., St.Louis, MO, 1999]. after entering 21 century, this technology causes the extensive concern of people in field of radar.

Most importantly in radar target recognition extract clarification of objective exactly, it is that very important basis has been established in follow-up classification and identification.Many documents prove some features extracted from HRRP, and such as FFT amplitude characteristic and various higher-order spectrum feature can complete classification below and identification mission more effectively.But these methods have a common characteristic: the feature that original HRRP data can not be extracted automatically, and spend a large amount of labours in the process of feature extraction.This just have impact on target recognition speed and accuracy rate.These features are all based on shallow-layer linear structure in addition, are sometimes difficult to find the nonlinear relationship between data.

After multilayer neural network occurs, there has been proposed degree of depth study and non-supervisory feature learning thought, its structure and thought have imitated the brain visual cortex of the mankind, are considered to the leap that area of pattern recognition is huge.But this method is complicated due to its training, greatly limit its development.2006, Hinton teaches in the one section of article delivered on Science and proposes utilization " contrast difference " successively initialized training method, open the second wave of degree of depth study, see [G.Hinton and R.salakhutdinov.Reducing the dimensionality of data with neural networks.Science, vol.313 (5786), pp.504-507,2006]. some deep layer network models afterwards, such as dark belief network DBNs and multilayer denoising scrambler SDAE, is widely used.

The height of radar target recognition rate depends primarily on the quality to raw data feature extraction, and existing recognition methods is all shallow structure, from the statistical property of data, proposes various feature extracting method.These methods are inherently seen, are all artificial encoding to data, have the data dependency of height, do not have universality.Can be wasted in feature extraction by the plenty of time, causing can not real-time processing data simultaneously.These method robustnesss are poor in addition, and due to lack of training samples under the complex environment in for example battlefield, the discrimination of these methods can reduce greatly.

Summary of the invention

The object of the invention is to the deficiency for above-mentioned prior art, a kind of radar target identification method based on degree of deep learning network is proposed, greatly to improve the accuracy of subsequent classification identification, and when number of samples is less, increase the discrimination robustness of model.

Technical scheme of the present invention is achieved in that

The degree of depth network structure proposed based on optical imagery can the feature of self study data, but radar Range Profile HRRP data are different from optical imagery, easily be subject to time shift, the impact of angle on target and echo amplitude susceptibility, can not, directly as the input of existing deep layer network, need to carry out pre-service to original radar range profile data.From radar signal angle, the radar range profile data be more or less the same to observation angle are averaged and can reduce angle on target susceptibility, improve signal to noise ratio (S/N ratio) simultaneously and reduce the impact of exceptional value.Mean distance picture is averaged all radar range profile data within the scope of certain angle and obtains, and therefore can describe this section of radar range profile data from average.

The present invention, on the basis of original depth confidence degree of depth network DBNs structure, adds mean distance picture as correction term, sets up original radar range profile data and the unified degree of depth network structure of average Range Profile.New construction can make the single observation original radar range profile data mean distance picture corresponding with it reach optimum matching, thus improves recognition accuracy, and when number of samples is less, increases the discrimination robustness of model.

Performing step of the present invention is as follows:

(1) original Radar High Range Resolution S={s is obtained ₁, s ₂..., s _n..., s _n, wherein s _n=[s _n1, s _n2..., s _ni..., s _nD] ^trepresent the n-th Range Profile in S, [] ^tthe transposition of representing matrix, s _nirepresent the n-th Range Profile all scattering point echo vector field homoemorphism in i-th range unit, n=1,2 ..., N, N represent all Range Profile numbers of acquisition, i=1,2 ..., D, D represent range unit number;

(2) utilize center of gravity method of aliging that the center of the N number of Range Profile in original Radar High Range Resolution S is moved to its center of gravity place, obtain the data X={x after moving ₁, x ₂..., x _n..., x _n, wherein x _nrepresent s _nthe data obtained after movement;

(3) according to the difference of observation angle, the data X after movement is divided into P part, i.e. X={H ₁, H ₂..., H _j..., H _p, j=1,2 ..., P, P determine have T sample, i.e. H in every portion according to experimental data _j={ x _{j, 1}, x _{j, 2}..., x _j,t..., x _j,T, wherein x _j,trepresent t sample in jth part; Ask H _jin the mean distance of all samples as m _j, obtain mean distance image set and close M={m ₁, m ₂..., m _j..., m _p;

(4) using the input of the data X after movement and average Range Profile set M as ground floor network, the cost function of definition ground floor network is: L ¹(X, M; Θ ¹), wherein Θ ¹={ W ¹, b ¹, c ¹represent all parameters of ground floor, W ¹represent ground floor network weight, b ¹represent ground floor output offset amount, c ¹represent ground floor Input Offset Value;

(5) backward conduction algorithm is utilized to calculate the cost function L of ground floor network ¹(X, M; Θ ¹) middle Θ ¹the gradient of parameter, and utilize gradient descent method constantly to minimize the cost function L of ground floor network ¹(X, M; Θ ¹), obtain ground floor Θ ¹the optimal estimation of parameter and using the data X after movement and average Range Profile set M as the input of ground floor, use optimal estimation parameter as the parameter of ground floor network, calculate the output sample data G of ground floor network and export mean distance image set conjunction R;

(6) for number of plies k>=2, using the input data Y of the output sample data G of kth-1 layer network as kth layer network, the output mean distance image set of kth-1 layer network is closed the input mean distance image set conjunction V of R as kth layer network, provide the cost function L of kth layer network ^k(Y, V; Θ ^k), the method according to step (5) obtains kth layer parameter Θ ^k={ W ^k, b ^k, c ^koptimal estimation

(7) repeat step (6) successively to train, until last one deck, i.e. U layer, obtain the optimal estimation parameter of every one deck k=1,2 ..., U, and set the output sample data of last one deck to be the coding of degree of depth network to original Radar High Range Resolution S as F, F;

(8) original Radar High Range Resolution S={s is supposed ₁, s ₂..., s _n..., s _nin N number of sample from a different Q target, define the label vector L=[l of original Radar High Range Resolution S ₁, l ₂..., l _n..., l _n] ^t, wherein l _n=1,2 ..., Q, represents the n-th sample s _ntarget class number;

(9) the parameter Φ={ W of linear SVM is established _svm, b _svm, using the input of the coding F of original Radar High Range Resolution S and the tag set L of original Radar High Range Resolution as linear SVM, and the parameter Φ of training linear support vector machine, obtain the optimal estimation Φ of linear SVM parameter Φ _opt, wherein W _svmrepresent the weights of linear SVM, b _svmrepresent the biased of linear SVM;

(10) measuring distance decent e=[e is established ₁, e ₂..., e _i..., e _d] ^t, using the input of decent for a measuring distance e as U layer depth network, wherein e _irepresent all scattering point echo vector field homoemorphism in measuring distance decent e i-th range unit, the layer of the U altogether optimal estimation parameter that utilization obtains above k=1,2 ..., U, calculates the output of U layer network

$e_{\mathrm{out}}^U=\frac{1}{1+E^{-[W_{\mathrm{opt}}^U{e}_{\mathrm{out}}^{U-1}+b_{\mathrm{opt}}^U]}},$

Wherein, the output of U-1 layer, the best initial weights of U layer network, be the optimum output offset of U layer network, E is Euler's constant, and value is E=2.7183;

(11) by the output of U layer network as the input of linear SVM, utilize the optimal estimation parameter Φ of linear vector machine _optthe classification information l of measuring distance decent e is calculated with one-to-many policing algorithm _e, obtain target recognition result.

The present invention compared with prior art has the following advantages:

1. present invention improves over the existing method utilizing shallow-layer network structure to extract original Radar High Range Resolution feature, introduce degree of depth network architecture, make the original Radar High Range Resolution feature extracted by degree of depth network architecture have structured message more, improve the discrimination of original Radar High Range Resolution.

2. the present invention considers that the method for existing extraction original Radar High Range Resolution feature does not consider the relation between the mean distance that single original radar range profile is corresponding with it, when building network model, add mean distance picture as correction term, make new construction that the single original radar range profile mean distance picture corresponding with it can be made to reach optimum matching, thus when number of samples is less, increase the discrimination robustness of model.

3. the present invention utilizes degree of depth network architecture to extract Radar High Range Resolution feature, for large batch of Radar High Range Resolution data, can feature in automatic learning data, and improve operation efficiency.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is the wherein one deck structural drawing in the degree of depth network model that proposes of the present invention;

Fig. 3 is the one-piece construction figure of the degree of depth network model that the present invention proposes;

Fig. 4 is the radar actual measurement scene graph obtaining original Radar High Range Resolution;

Fig. 5 is under in the end one deck exports dimension different situations, differently to the discrimination comparison diagram of Radar High Range Resolution;

Fig. 6 exports to each layer of network the result figure carrying out dimensionality reduction by principal component analysis (PCA);

Fig. 7 is the result figure be reconstructed original Radar High Range Resolution sample by degree of depth network structure of the present invention.

Embodiment

The present invention utilizes mean distance picture to set up degree of depth network model as correction term, and using the output of last one deck as the coding to original Radar High Range Resolution, for follow-up Classification and Identification.

With reference to Fig. 1, specific implementation step of the present invention is as follows:

Step 1, carries out pre-service to original Radar High Range Resolution.

1.1) original Radar High Range Resolution S={s is obtained ₁, s ₂..., s _n..., s _n, wherein s _n=[s _n1, s _n2..., s _ni..., s _nD] ^trepresent the n-th Range Profile in S, [] ^tthe transposition of representing matrix, s _nirepresent the n-th Range Profile all scattering point echo vector field homoemorphism in i-th range unit, n=1,2 ..., N, N represent all Range Profile numbers of acquisition, i=1,2 ..., D, D represent range unit number;

1.2) the n-th sample s in original Radar High Range Resolution S is calculated _ncenter of gravity W _n:

$W_n=\frac{\underset{i=1}{\overset{D}{\mathrm{\Σ}}}i{\left|s_{\mathrm{ni}}\right|}^2}{\underset{i=1}{\overset{D}{\mathrm{\Σ}}}{\left|s_{\mathrm{ni}}\right|}^2},i=\mathrm{1,2},...,D;$

1.3) by the n-th sample s in original Radar High Range Resolution S _ncentral region to its center of gravity W _n, obtain the data X={x after moving ₁, x ₂..., x _n..., x _nin n-th move rear sample x _n=[x _n1, x _n2..., x _ni..., x _nD] the value of i-th range unit:

$x_{\mathrm{ni}}=\mathrm{IFFT}\left\{\mathrm{FFT}\right\{s_{\mathrm{ni}}\left\}{e}^{-j\left\{\mathrm{\φ}\right[W_n]-\mathrm{\φ}[C_n\left]\right\}a}\right\},$

Wherein, FFT is Fourier transform, and IFFT is inverse Fourier transform, the n-th sample s in original Radar High Range Resolution S _ncenter, φ [W _n] be s _nthe phase place corresponding to center of gravity, φ [C _n] be s _nthe phase place corresponding to center, a is center C _nplace unit and center of gravity W _ndistance between cells.

Step 2, is divided into P part by the data X after movement, calculates every increment mean distance picture originally.

2.1) the original Radar High Range Resolution S corresponding to data X after mobile is that radar is 0 ° ~ 360 ° in view angle respectively and acquires continuously, according to the rule of every 10 degree a group, the data X after movement is divided into P part, i.e. X={H ₁, H ₂..., H _j..., H _p, j=1,2 ..., P, P=36, have T sample, i.e. H in every portion _j={ x _{j, 1}, x _{j, 2}..., x _j,t..., x _j,T, wherein x _j,trepresent t sample in jth part;

2.2) jth part Range Profile H is supposed _j={ x _{j, 1}, x _{j, 2}..., x _j,t..., x _j,Tin the prior probability of T sample be equal, define H _ja middle T sample and its mean distance are as m _jbetween Euclidean distance:

$d_j^2=\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\left|\right|x_{j,t}-m_j\left|\right|}_2^2,$

Wherein, symbol || || ₂represent 2 norms asking matrix;

2.3) least-squares algorithm is utilized to minimize Euclidean distance obtain jth part Range Profile H _jmean distance picture:

$m_j=\frac{1}{T}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{x}_{j,t}.$

Step 3, with reference to Fig. 2, the cost function L of definition ground floor network ¹(X, M; Θ ¹), and utilize gradient descent method to minimize the cost function L of ground floor network ¹(X, M; Θ ¹), obtain ground floor Θ ¹the optimal estimation of parameter

3.1) using the input of the data X after movement and average Range Profile set M as ground floor network, the cost function L of definition ground floor network ¹(X, M; Θ ¹) be:

$L^1(X,M;{\mathrm{\Θ}}^1)=\frac{1}{T\·P}\underset{j=1}{\overset{P}{\mathrm{\Σ}}}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{(m_j-{\hat{x}}_{j,t})}^2{\mathrm{\Σ}}_j^{-1}(m_j-{\hat{x}}_{j,t})+\mathrm{\αKL}\left(\widehat{\mathrm{\ρ}}\right|\left|\mathrm{\ρ}\right)+\frac{\mathrm{\β}}{2}{\left|\right|W^1\left|\right|}_2^2,$

Wherein, equation right side Section 1 is jth part Range Profile H _jin t sample x _j.treconstruct data with H _jin the mean distance of all samples as m _jbetween mahalanobis distance, it features reconstruct data with mean distance as m _jbetween similarity degree, ${\hat{x}}_{j,t}=f(W^{1^T}\·f(W^1{x}_{j.t}+b^1)+c^1),f\left(x\right)=\frac{1}{1+E^{-x}},$ E is Euler's constant, and value is E=2.7183, jth part Range Profile H _jcovariance matrix, W ¹the weights of ground floor network, b ¹the output offset of ground floor network, c ¹the input being ground floor network is biased.

Equation right side Section 2 $\mathrm{KL}\left(\widehat{\mathrm{\ρ}}\right|\left|\mathrm{\ρ}\right)=\underset{j=1}{\overset{\left|{\widehat{\mathrm{\ρ}}}_j\right|}{\mathrm{\Σ}}}\mathrm{\ρ}\log \frac{\mathrm{\ρ}}{{\widehat{\mathrm{\ρ}}}_j}+(1-\mathrm{\ρ})\log \frac{1-\mathrm{\ρ}}{1-{\widehat{\mathrm{\ρ}}}_j}$ Represent vector and the KL distance between ρ, it features the openness of output layer coding, wherein it is vector i-th, ρ is the sparse value preset, and α is the sparse penalty factor weight preset;

Equation right side Section 3 is a regularization term, is also weight attenuation term, its objective is the amplitude reducing weight, prevents Expired Drugs, wherein || and W ¹|| ₂ground floor network weight W ¹2 norms, β is the regularization weight preset;

3.2) random to ground floor network weight W ¹, ground floor network output offset b ¹and the biased c of ground floor network input ¹a Gradient Descent initial value, is respectively and

3.3) backward conduction algorithm is utilized to calculate ground floor network weight W ¹gradient ground floor network output offset b ¹gradient and the biased c of ground floor network input ¹gradient

$\▿W^1=\frac{1}{T\·P}\underset{j=1}{\overset{P}{\mathrm{\Σ}}}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}[x_{j,t}{\·\left({\mathrm{\δ}}_{j,t}^2\right)}^T+{\mathrm{\δ}}_{j,t}^3\·{\left(h_{j,t}\right)}^T]+\mathrm{\β}{W}_{\mathrm{ori}}^1,$

$\▿b^1=\frac{1}{T\·P}\underset{j=1}{\overset{P}{\mathrm{\Σ}}}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{f}^{\′}\left(z_2\right)\⊗(W_{\mathrm{ori}}^{1T}\·{\mathrm{\δ}}^3+\mathrm{\α}\·t),$

$\▿c^1=\frac{1}{T\·P}\underset{j=1}{\overset{P}{\mathrm{\Σ}}}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{\mathrm{\Σ}}_j^{-1}(f\left(z_3\right)-m_j)\⊗f\left(z_3\right),$

Wherein, $z_3=W_{\mathrm{ori}}^{1T}{h}_{j,t}+c_{\mathrm{ori}}^1,h_{j,t}=f\left(z_2\right),z_2=W_{\mathrm{ori}}^1{x}_{j,t}+b_{\mathrm{ori}}^1,{\mathrm{\δ}}^2=f^{\′}\left(z_2\right)\⊗(W_{\mathrm{ori}}^{1T}\·{\mathrm{\δ}}^3+\mathrm{\α}\·t),$ vector t=[t ₁, t ₂..., t _i..., t _d] ^ti-th element be element multiplication on representing matrix correspondence position;

3.4) gradient descent algorithm undated parameter is utilized and initial value:

$W_{\mathrm{new}}^1:=W_{\mathrm{ori}}^1-{\mathrm{\α}}_W\▿W^1,$

$b_{\mathrm{new}}^1:=b_{\mathrm{ori}}^1-{\mathrm{\α}}_b\▿b^1,$

$c_{\mathrm{new}}^1:=c_{\mathrm{ori}}^1-{\mathrm{\α}}_c\▿c^1,$

Wherein, represent the weights after the renewal of ground floor network, represent the output offset after the renewal of ground floor network, represent that the input after the renewal of ground floor network is biased; α _wrepresent the Gradient Descent speed of ground floor network weight, α _brepresent the Gradient Descent speed of ground floor network output offset, α _crepresent the Gradient Descent speed that the input of ground floor network is biased;

3.5) judge whether to terminate gradient descent procedures: if then terminate gradient descent procedures, perform step 4, wherein δ _w, δ _band δ _cthree the parameter W preset ¹, b ¹and c ¹iteration upgrade outage threshold, its value is δ _w=δ _b=δ _c=10 ^-3; If weights after then ground floor network being upgraded output offset after ground floor network upgrades and the input after the renewal of ground floor network is biased as the initial value of next gradient descent procedures, even return step (3.3).

Step 4, calculates the output sample data G of ground floor network and exports mean distance image set conjunction R.

4.1) the output sample data G={g of ground floor network is established _1,1, g _1,2..., g _j,t..., g _p,T, wherein g _j,trepresent jth part t sample x _j,toutput; If export mean distance image set to close R={r ₁, r ₂..., r _j..., r _p, wherein r _jrepresent jth part Range Profile H _joutput mean distance picture, j=1,2 ..., P, t=1,2 ..., T;

4.2) jth part t sample x is calculated _j,toutput g _j,tand jth part Range Profile H _joutput mean distance as r _j:

$g_{j,t}=f(W_{\mathrm{opt}}^1{x}_{j,t}+b_{\mathrm{opt}}^1),$

$r_j=f(W_{\mathrm{opt}}^1{m}_j+b_{\mathrm{opt}}^1).$

Step 5, computational grid kth layer parameter Θ ^k={ W ^k, b ^k, c ^koptimal estimation

5.1) the cost function L of kth layer network is defined ^k(Y, V; Θ ^k):

With reference to Fig. 3, the input Y of kth layer network is made to equal the output G of kth-1 layer network, k>=2, i.e. Y=G={g _1,1, g _1,2..., g _j,t..., g _p,T, wherein g _j,trepresent jth part t sample x _j,toutput; The input mean distance of kth layer network is made to equal the output mean distance of kth-1 layer network as R as V, i.e. V=R={r ₁, r ₂..., r _j..., r _p, wherein r _jrepresent jth part Range Profile H _joutput mean distance picture; Set up the cost function L of kth layer network ^k(Y, V; Θ ^k) be:

$L^k(Y,V;{\mathrm{\Θ}}^k)=\frac{1}{T\·P}\underset{j=1}{\overset{P}{\mathrm{\Σ}}}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{(r_j-{\hat{g}}_{j,t})}^T(r_j-{\hat{g}}_{j,t})+\mathrm{\αKL}\left(\widehat{\mathrm{\ρ}}\right|\left|\mathrm{\ρ}\right)+\frac{\mathrm{\β}}{2}{\left|\right|W^k\left|\right|}_2^2,$

Wherein, ${\hat{g}}_{j,t}=f(W^{k^T}\·f(W^k{g}_{j.t}+b^k)+c^k),\mathrm{KL}\left(\widehat{\mathrm{\ρ}}\right|\left|\mathrm{\ρ}\right)=\underset{i=1}{\overset{D}{\mathrm{\Σ}}}\mathrm{\ρ}\log \frac{\mathrm{\ρ}}{{\widehat{\mathrm{\ρ}}}_i}+(1-\mathrm{\ρ})\log \frac{1-\mathrm{\ρ}}{1-{\widehat{\mathrm{\ρ}}}_i},$ it is vector i-th, ρ is the sparse value preset, and α is the sparse penalty factor weight preset, W ^kthe weights of kth layer network, b ^kthe output offset of kth layer network, c ^kthe input being kth layer network is biased, || W ^k|| ₂kth layer network weights W ^k2 norms, β is the regularization weight preset;

5.2) random to kth layer network weights W ^k, kth layer network output offset b ^kand the biased c of kth layer network input ^ka Gradient Descent initial value, is respectively and

5.3) backward conduction algorithm is utilized to calculate kth layer network weights W ^kgradient kth layer network output offset b ^kgradient and the biased c of kth layer network input ^kgradient

$\▿W^k=\frac{1}{T\·P}\underset{j=1}{\overset{P}{\mathrm{\Σ}}}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}[g_{j,t}{\·\left({\mathrm{\δ}}_{j,t}^2\right)}^T+{\mathrm{\δ}}_{j,t}^3\·{\left(h_{j,t}\right)}^T]+\mathrm{\β}{W}_{\mathrm{ori}}^k,$

$\▿b^k=\frac{1}{T\·P}\underset{j=1}{\overset{P}{\mathrm{\Σ}}}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}{f}^{\′}\left(z_2\right)\⊗(W_{\mathrm{ori}}^{\mathrm{kT}}\·{\mathrm{\δ}}^3+\mathrm{\α}\·t),$

$\▿c^k=\frac{1}{T\·P}\underset{j=1}{\overset{P}{\mathrm{\Σ}}}\underset{t=1}{\overset{T}{\mathrm{\Σ}}}(f\left(z_3\right)-r_j)\⊗f\left(z_3\right),$

Wherein, $z_3=W_{\mathrm{ori}}^{\mathrm{kT}}{h}_{j,t}+c_{\mathrm{ori}}^k,h_{j,t}=f\left(z_2\right),z_2=W_{\mathrm{ori}}^k{g}_{j,t}+b_{\mathrm{ori}}^k,{\mathrm{\δ}}^2=f^{\′}\left(z_2\right)\⊗(W_{\mathrm{ori}}^{\mathrm{kT}}\·{\mathrm{\δ}}^3+\mathrm{\α}\·t),$ ${\mathrm{\δ}}^3=(f\left(z_3\right)-r_j)\⊗f^{\′}\left(z_3\right),$ Vector t=[t ₁, t ₂..., t _i..., t _d] ^ti-th element be $t_i=-\frac{\mathrm{\ρ}}{{\widehat{\mathrm{\ρ}}}_i}+\frac{1-\mathrm{\ρ}}{1-{\widehat{\mathrm{\ρ}}}_i},$ element multiplication on representing matrix correspondence position;

5.4) gradient descent algorithm undated parameter is utilized and initial value:

$W_{\mathrm{new}}^k:=W_{\mathrm{ori}}^k-{\mathrm{\α}}_W\▿W^k,$

$b_{\mathrm{new}}^k:=b_{\mathrm{ori}}^k-{\mathrm{\α}}_b\▿b^k,$

$c_{\mathrm{new}}^k:=c_{\mathrm{ori}}^k-{\mathrm{\α}}_c\▿c^k,$

Wherein, represent the weights after the renewal of kth layer network, represent the output offset after the renewal of kth layer network, represent that the input after the renewal of kth layer network is biased; represent the Gradient Descent speed of kth layer network weights, represent the Gradient Descent speed of kth layer network output offset, represent the Gradient Descent speed that the input of kth layer network is biased;

5.5) judge whether to terminate gradient descent procedures: if ${\left|\right|\&dtri;W^k\left|\right|}_2<{\mathrm{\&delta;}}_W,{\left|\right|\&dtri;b^k\left|\right|}_2<{\mathrm{\&delta;}}_b,{\left|\right|\&dtri;c^k\left|\right|}_2<{\mathrm{\&delta;}}_c,$ Then terminate gradient descent procedures, perform step 6, wherein δ _w, δ _band δ _cbe three the parameter W preset respectively ^k, b ^kand c ^kiteration upgrade outage threshold, its value is δ _w=δ _b=δ _c=10 ^-3; If weights after then kth layer network being upgraded output offset after kth layer network upgrades and the input after the renewal of kth layer network is biased as the initial value of next gradient descent procedures, even $b_{\mathrm{ori}}^k=b_{\mathrm{new}}^k,c_{\mathrm{ori}}^k=c_{\mathrm{new}}^k,$ Return step (5.3).

Step 6, the output sample data of the last one deck of computational grid are F.

6.1) repeat step 5 successively to train, until last one deck, namely k=U, U represent the label of last one deck, obtain the optimal estimation parameter of every one deck k=1,2 ..., U;

6.2) the output sample data of the last one deck of computational grid are that F, F are the coding of degree of depth network to original Radar High Range Resolution S:

$F=\frac{1}{1+E^{-[W_{\mathrm{opt}}^U{q}_{\mathrm{in}}^{U-1}+b_{\mathrm{opt}}^U]}},$

Wherein, the input of U layer, the best initial weights of U layer network, be the optimum output offset of U layer network, E is Euler's constant, and value is E=2.7183.

Step 7, utilizes the coding F of original Radar High Range Resolution S, training linear support vector machine.

7.1) radar is while the original Radar High Range Resolution S of acquisition, can obtain the label vector L=[l of original Radar High Range Resolution S ₁, l ₂..., l _n..., l _n] ^t, wherein l _n=1,2 ..., Q, represents the n-th sample s _ntarget class number, Q expressive notation vector L in other total number of target class;

7.2) the parameter Φ={ W of linear SVM is established _svm, b _svm, wherein W _svmrepresent the weights of linear SVM, b _svmrepresent the biased of linear SVM, using the input of the coding F of original Radar High Range Resolution S and the tag set L of original Radar High Range Resolution as linear SVM;

Because linear SVM is the classifier technique that a kind of theory is perfect with application, so there is the linear SVM function kit write based on Matlab platform in a large number, because LibSVM kit is that solving speed is the fastest, the linear SVM function kit that classification results is best, so the present invention uses the function TrainSVM (F in LibSVM function kit, L) the parameter Φ of training linear support vector machine, obtains the optimal estimation Φ of linear SVM parameter Φ _opt.

Step 8, calculates the classification information l of measuring distance decent e _e.

8.1) measuring distance decent e=[e is established ₁, e ₂..., e _i..., e _d] ^t, wherein e _iall scattering point echo vector field homoemorphism in i-th range unit of expression measuring distance decent e, using the input of decent for a measuring distance e as U layer depth network, the U layer network optimal estimation parameter that utilization obtains above k=1,2 ..., U, calculates the output of U layer network

$e_{\mathrm{out}}^U=\frac{1}{1+E^{-[W_{\mathrm{opt}}^U{e}_{\mathrm{out}}^{U-1}+b_{\mathrm{opt}}^U]}},$

Wherein, the output of U-1 layer, the best initial weights of U layer network, be the optimum output offset of U layer network, E is Euler's constant, and value is E=2.7183;

8.2) by the output of U layer network as the input of linear SVM, utilize the optimal estimation parameter Φ of linear vector machine _opt, calculate the classification information of measuring distance decent e:

Owing to calculating the function of classification information in linear SVM function library LibSVM be write according to one-to-many algorithm, so solving speed is very fast, and radar needs real-time processing data fast, so the function write according to one-to-many policing algorithm in select linear support vector machine function library LibSVM of the present invention calculate the classification information of measuring distance decent e $l_e=\mathrm{TestSVM}(e_{\mathrm{out}}^U,{\mathrm{\&Phi;}}_{\mathrm{opt}}),$ Thus obtain target recognition result.

Effect of the present invention is further illustrated by the following emulation of the measured data to three class aircrafts:

1. obtain the radar measured data of original high-resolution Range Profile S, actual measurement scene is with reference to Fig. 4, wherein Fig. 4 a is the actual measurement scene graph of Yark-42 aircraft, Fig. 4 b is the actual measurement scene graph of Cessna Citation S/ II aircraft, Fig. 4 c is the actual measurement scene graph of An-26 aircraft, concrete aircraft and radar parameter as shown in the table:

When emulation, the original high-resolution Range Profile S of acquisition is divided into two classes: Te is gathered in training set Tr and test, follows two principles below during classification: 1) train the position angle of High Range Resolution in set Tr will cover all tests and gather the position angle of High Range Resolution in Te; 2) for same type of airplane, in test set Te, the position angle of High Range Resolution is different from the position angle that High Range Resolution in Tr is gathered in training.

2. emulate content:

Emulation 1, with reference to Fig. 4,2 ~ 5 sections that select Yark-42 aircraft, 6 ~ 7 sections of Cessna Citation S/ II aircraft and 5 ~ 6 sections of the An-26 aircraft original Radar High Range Resolution obtained are as training sample Tr, and remaining is as test sample book Te.Use whole training sample, utilize stack correction own coding device SCAE of the present invention and existing linear discriminant analysis LDA, k to walk odd value analysis K-SVD, principal component analysis (PCA) PCA, degree of depth confidence network DBNs and stacking denoising own coding device SDAE, Classification and Identification emulation is carried out to original Radar High Range Resolution.Simulation result is as shown in table 1 below.

Table 1 uses whole training sample, and each method is to the discrimination of original Radar High Range Resolution

The discriminatory analysis of table 1 neutral line LDA, k walk odd value analysis K-SVD, principal component analysis (PCA) PCA is single layer structure, and stack correction own coding device SCAE of the present invention, degree of depth confidence network DBNs and stacking denoising autoencoder network SDAE are sandwich constructions.Take three-decker in this emulation, if the dimension that ground floor exports is 1500, the dimension that the second layer exports is 500, and the dimension that third layer exports is 50.

As can be seen from the table: the first, compared with single-layer model, the model with sandwich construction has higher discrimination to original Radar High Range Resolution; The second, in all methods, the stack correction own coding device SCAE that the present invention proposes has the highest discrimination, reaches 92.03%.

Emulation 2, change the dimension of the coding F of original radar range profile S, utilize degree of depth network SCAE of the present invention and existing linear discriminant analysis LDA, k to walk odd value analysis K-SVD, principal component analysis (PCA) PCA, degree of depth confidence network DBNs and stacking denoising autoencoder network SDAE, Classification and Identification emulation is carried out to original Radar High Range Resolution.Simulation result as shown in Figure 5.

As can be seen from Figure 5, the dimension of coding F reduces, and except the inventive method, methodical discrimination declines all to some extent, and especially when the dimension of the F that encodes is less than 30, this decline is particularly meditated obviously.And the discrimination of stack correction own coding device SCAE of the present invention reduces along with the dimension reduction of coding F hardly, this illustrates that the present invention has stronger coding dimension robustness.

Emulation 3, with reference to Fig. 4, the 2nd section that selects Yark-42 aircraft, the 6th section of Cessna Citation S/ II aircraft and 5 sections of the An-26 aircraft original Radar High Range Resolution obtained are as training sample Tr, and remaining is as test sample book Te.Use part training sample, utilize stack correction own coding device SCAE of the present invention and existing linear discriminant analysis LDA, k to walk odd value analysis K-SVD, principal component analysis (PCA) PCA, degree of depth confidence network DBNs and stacking denoising autoencoder network SDAE, Classification and Identification emulation is carried out to original Radar High Range Resolution.Simulation result is as shown in table 2:

Table 2 uses part training sample, and each method is to the discrimination of original Radar High Range Resolution

The discriminatory analysis of table 2 neutral line LDA, k walk odd value analysis K-SVD, principal component analysis (PCA) PCA is single layer structure, and stack correction own coding device SCAE of the present invention, degree of depth confidence network DBNs and stacking denoising autoencoder network SDAE are sandwich constructions.Take three-decker in this emulation, if the dimension that ground floor exports is 1300, the dimension that the second layer exports is 200, and the dimension that third layer exports is 50.

As can be seen from Table 2: when training sample number deficiency, stack correction own coding device SCAE of the present invention still has the highest discrimination, reaches 85.64%.This illustrates that stack correction own coding device SCAE of the present invention has number of samples robustness.

Emulation 4, by principal component analysis (PCA), dimensionality reduction is carried out to the output of stack correction own coding device SCAE of the present invention last one deck and original Radar High Range Resolution, obtain the two-dimensional visualization figure of the output of SCAE last one deck and original Radar High Range Resolution, simulation result as shown in Figure 6.Wherein Fig. 6 a represents the two-dimensional visualization figure of complete training sample Tr in original Radar High Range Resolution, Fig. 6 b represents the two-dimensional visualization figure of test sample book Te in original Radar High Range Resolution, Fig. 6 c represents that in original Radar High Range Resolution, complete training sample Tr encodes through stack correction own coding device SCAE, the two-dimensional visualization figure that last one deck exports, Fig. 6 d represents that in original Radar High Range Resolution, test sample book Te encodes through stack correction own coding device SCAE, the two-dimensional visualization figure that last one deck exports.

As can be seen from Figure 6, in the two-dimensional visualization figure of original Radar High Range Resolution, the data of all categories mix, and cannot separate; In the two-dimensional visualization figure that the last one deck of SCAE exports, the data of all categories have had visible separatrix substantially.This illustrates that stack correction own coding device SCAE of the present invention has Encoding preferably to original Radar High Range Resolution.

Emulation 5, utilize the coding F of stack correction own coding device SCAE of the present invention to original Radar High Range Resolution S to be reconstructed emulation, simulation result as shown in Figure 7.Wherein Fig. 7 a is the original Radar High Range Resolution restructuring graph of Yark-42 aircraft in training sample Tr, Fig. 7 c is the restructuring graph of the original Radar High Range Resolution of Cessna Citation S/ II aircraft in training sample Tr, Fig. 7 e is the original Radar High Range Resolution restructuring graph of An-26 aircraft in training sample Tr, Fig. 7 b is the original Radar High Range Resolution restructuring graph of Yark-42 aircraft in test sample book Te, Fig. 7 d is the original Radar High Range Resolution restructuring graph of Cessna Citation S/ II aircraft in test sample book Te, Fig. 7 e is the original Radar High Range Resolution restructuring graph of An-26 aircraft in test sample book Te.

As can be seen from Figure 7, the coding F of stack correction own coding device SCAE of the present invention to original Radar High Range Resolution S has stronger re-configurability, and this has Encoding preferably from a side illustration stack correction of the present invention own coding device SCAE to original Radar High Range Resolution.

Claims (7)

1., based on a radar target identification method for degree of deep learning network, comprise the steps:

(1) original Radar High Range Resolution S={s is obtained ₁, s ₂..., s _n..., s _n, wherein s _n=[s _n1, s _n2..., s _ni..., s _nD] ^trepresent the n-th Range Profile in S, [] ^tthe transposition of representing matrix, s _nirepresent the n-th Range Profile all scattering point echo vector field homoemorphism in i-th range unit, n=1,2 ..., N, N represent all Range Profile numbers of acquisition, i=1,2 ..., D, D represent range unit number;

(2) utilize center of gravity method of aliging that the center of the N number of Range Profile in original Radar High Range Resolution S is moved to its center of gravity place, obtain the data X={x after moving ₁, x ₂..., x _n..., x _n, wherein x _nrepresent s _nthe data obtained after movement;

(3) according to the difference of observation angle, the data X after movement is divided into P part, i.e. X={H ₁, H ₂..., H _j..., H _p, j=1,2 ..., P, P determine have T sample, i.e. H in every portion according to experimental data _j={ x _{j, 1}, x _{j, 2}..., x _j,t..., x _j,T, wherein x _j,trepresent t sample in jth part; Ask H _jin the mean distance of all samples as m _j, obtain mean distance image set and close M={m ₁, m ₂..., m _j..., m _p;

(4) using the input of the data X after movement and average Range Profile set M as ground floor network, the cost function of definition ground floor network is: L ¹(X, M; Θ ¹), wherein Θ ¹={ W ¹, b ¹, c ¹represent all parameters of ground floor, W ¹represent ground floor network weight, b ¹represent ground floor output offset amount, c ¹represent ground floor Input Offset Value;

(5) backward conduction algorithm is utilized to calculate the cost function L of ground floor network ¹(X, M; Θ ¹) middle Θ ¹the gradient of parameter, and utilize gradient descent method constantly to minimize the cost function L of ground floor network ¹(X, M; Θ ¹), obtain ground floor Θ ¹the optimal estimation of parameter and using the data X after movement and average Range Profile set M as the input of ground floor, use optimal estimation parameter as the parameter of ground floor network, calculate the output sample data G of ground floor network and export mean distance image set conjunction R;

(6) for number of plies k>=2, using the input data Y of the output sample data G of kth-1 layer network as kth layer network, the output mean distance image set of kth-1 layer network is closed the input mean distance image set conjunction V of R as kth layer network, provide the cost function L of kth layer network ^k(Y, V; Θ ^k), the method according to step (5) obtains kth layer parameter Θ ^k={ W ^k, b ^k, c ^koptimal estimation

(7) repeat step (6) successively to train, until last one deck, i.e. U layer, obtain the optimal estimation parameter of every one deck k=1,2 ..., U, and set the output sample data of last one deck to be the coding of degree of depth network to original Radar High Range Resolution S as F, F;

(8) original Radar High Range Resolution S={s is supposed ₁, s ₂..., s _n..., s _nin N number of sample from a different Q target, define the label vector L=[l of original Radar High Range Resolution S ₁, l ₂..., l _n..., l _n] ^t, wherein l _n=1,2 ..., Q, represents the n-th sample s _ntarget class number;

(9) the parameter Φ={ W of linear SVM is established _svm, b _svm, using the input of the coding F of original Radar High Range Resolution S and the tag set L of original Radar High Range Resolution as linear SVM, and the parameter Φ of training linear support vector machine, obtain the optimal estimation Φ of linear SVM parameter Φ _opt, wherein W _svmrepresent the weights of linear SVM, b _svmrepresent the biased of linear SVM;

(10) measuring distance decent e=[e is established ₁, e ₂..., e _i..., e _d] ^t, using the input of decent for a measuring distance e as U layer depth network, wherein e _irepresent all scattering point echo vector field homoemorphism in measuring distance decent e i-th range unit, the layer of the U altogether optimal estimation parameter that utilization obtains above k=1,2 ..., U, calculates the output of U layer network

$e_{\mathrm{out}}^U=\frac{1}{1+E^{-[W_{\mathrm{opt}}^U{e}_{\mathrm{out}}^{U-1}+b_{\mathrm{opt}}^U]}},$

Wherein, the output of U-1 layer, the best initial weights of U layer network, be the optimum output offset of U layer network, E is Euler's constant, and value is E=2.7183;

(11) by the output of U layer network as the input of linear SVM, utilize the optimal estimation parameter Φ of linear vector machine _optthe classification information l of measuring distance decent e is calculated with one-to-many policing algorithm _e, obtain target recognition result.

2. according to claim 1 based on the radar target identification method of degree of deep learning network, the center of the N number of Range Profile in original Radar High Range Resolution S is moved to its center of gravity place by the center of gravity method of aliging that utilizes wherein described in step (2), obtains the data X={x after moving ₁, x ₂..., x _n..., x _n, n=1,2 ..., N, carries out as follows:

(2.1) s is established _n=[s _n1, s _n2..., s _ni..., s _nD] ^tthe n-th sample, wherein s in original Radar High Range Resolution S _nirepresent the n-th Range Profile s _nall scattering point echo vector field homoemorphism in i-th range unit;

(2.2) the n-th sample s in original Radar High Range Resolution S is calculated _ncenter of gravity W _n:

$W_n=\frac{\underset{i=1}{\overset{D}{\mathrm{\&Sigma;}}}i{\left|s_{\mathrm{ni}}\right|}^2}{\underset{i=1}{\overset{D}{\mathrm{\&Sigma;}}}{\left|s_{\mathrm{ni}}\right|}^2},i=\mathrm{1,2},...,D;$

(2.3) by the n-th sample s in original Radar High Range Resolution S _ncentral region to its center of gravity W _n, obtain the data X={x after moving ₁, x ₂..., x _n..., x _nin n-th move rear sample x _n=[x _n1, x _n2..., x _ni..., x _nD] value of i-th range unit:

$x_{\mathrm{ni}}=\mathrm{IFFT}\left\{\mathrm{FFT}\right\{s_{\mathrm{ni}}\left\}{e}^{-j\left\{\mathrm{\&phi;}\right[W_n]-\mathrm{\&phi;}[C_n\left]\right\}a}\right\},$

Wherein, FFT is Fourier transform, and IFFT is inverse Fourier transform, the n-th sample s in original Radar High Range Resolution S _ncenter, φ [W _n] be s _nthe phase place corresponding to center of gravity, φ [C _n] be s _nthe phase place corresponding to center, a is center C _nplace unit and center of gravity W _ndistance between cells.

3., according to claim 1 based on the radar target identification method of degree of deep learning network, wherein described in step (3), ask H _jin the mean distance of all samples as m _j, carry out as follows:

(3.1) jth part Range Profile H is supposed _j={ x _{j, 1}, x _{j, 2}..., x _j,t..., x _j,Tin the prior probability of T sample be equal, define H _ja middle T sample and its mean distance are as m _jbetween Euclidean distance:

$d_j^2=\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\left|\right|x_{j,t}-m_j\left|\right|}_2^2,$

Wherein, symbol || || ₂represent 2 norms asking matrix.

(3.2) least-squares algorithm is utilized to minimize Euclidean distance obtain jth part Range Profile H _jmean distance picture:

$m_j=\frac{1}{T}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{x}_{j,t}.$

4. the radar target identification method based on degree of deep learning network according to claim 1, wherein the cost function L of step (4) definition ground floor network ¹(X, M; Θ ¹) be expressed as:

$L^1(X,M;{\mathrm{\&Theta;}}^1)=\frac{1}{T\&CenterDot;P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{(m_j-{\hat{x}}_{j,t})}^T{\mathrm{\&Sigma;}}_j^{-1}(m_j-{\hat{x}}_{j,t})+\mathrm{\&alpha;KL}\left(\widehat{\mathrm{\&rho;}}\right|\left|\mathrm{\&rho;}\right)+\frac{\mathrm{\&beta;}}{2}{\left|\right|W^1\left|\right|}_2^2$

Wherein, in the Section 1 of the equation right side ${\hat{x}}_{j,t}=f(W^{1^T}\&CenterDot;f(W^1{x}_{j,t}+b^1)+c^1)$ To jth part t sample x _j.treconstruct data, $f\left(x\right)=\frac{1}{1+E^{-x}},$ E is Euler's constant, ${\mathrm{\&Sigma;}}_j=\frac{1}{N}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}(x_{j,t}-m_j){(x_{j,t}-m_j)}^T$ Jth part Range Profile H _jcovariance matrix; In Section 2 $\mathrm{KL}\left(\widehat{\mathrm{\&rho;}}\right|\left|\mathrm{\&rho;}\right)=\underset{i=1}{\overset{D}{\mathrm{\&Sigma;}}}\mathrm{\&rho;}\log \frac{\mathrm{\&rho;}}{{\widehat{\mathrm{\&rho;}}}_i}+(1-\mathrm{\&rho;})\log \frac{1-\mathrm{\&rho;}}{1-{\widehat{\mathrm{\&rho;}}}_i},$ Wherein $\widehat{\mathrm{\&rho;}}=\frac{1}{T\&CenterDot;P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}f(W^1{x}_{j,t}+b^1),$ it is vector i-th, ρ is the sparse value preset, and α is the sparse penalty factor weight preset; In Section 3 || W ¹|| ₂ground floor network weight W ¹2 norms, β is the regularization weight preset.

5. the radar target identification method based on degree of deep learning network according to claim 1, the backward conduction algorithm that utilizes wherein described in step (5) calculates the cost function L of ground floor network ¹(X, M; Θ ¹) middle Θ ¹the gradient of parameter, and utilize gradient descent method constantly to minimize the cost function L of ground floor network ¹(X, M; Θ ¹), carry out as follows:

(5.1) random to ground floor network weight W ¹, ground floor network output offset b ¹and the biased c of ground floor network input ¹a Gradient Descent initial value, is respectively and

(5.2) backward conduction algorithm is utilized to calculate ground floor network weight W ¹gradient ground floor network output offset b ¹gradient and the biased c of ground floor network input ¹gradient

${\&dtri;W}^1=\frac{1}{T\&CenterDot;P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}[x_{j,t}\&CenterDot;{\left({\mathrm{\&delta;}}_{j,t}^2\right)}^T+{\mathrm{\&delta;}}_{j,t}^3\&CenterDot;{\left(h_{j,t}\right)}^T]+\mathrm{\&beta;}{W}_{\mathrm{ori}}^1,$

$\&dtri;b^1=\frac{1}{T\&CenterDot;P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{f}^{\&prime;}\left(z_2\right)\&CircleTimes;(W_{\mathrm{ori}}^{1T}\&CenterDot;{\mathrm{\&delta;}}^3+\mathrm{\&alpha;}\&CenterDot;t),$

$\&dtri;c^1=\frac{1}{T\&CenterDot;P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{\mathrm{\&Sigma;}}_j^{-1}(f\left(z_3\right)-m_j)\&CircleTimes;f\left(z_3\right),$

Wherein, $z_3=W_{\mathrm{ori}}^{1T}{h}_{j,t}+c_{\mathrm{ori}}^1,$ h _j，t＝f(z ₂)， $z_2=W_{\mathrm{ori}}^1{x}_{j,t}+b_{\mathrm{ori}}^1,$ ${\mathrm{\&delta;}}^2=f^{\&prime;}\left(z_2\right)\&CircleTimes;(W_{\mathrm{ori}}^{1T}\&CenterDot;{\mathrm{\&delta;}}^3+\mathrm{\&alpha;}\&CenterDot;t),{\mathrm{\&delta;}}^3={\mathrm{\&Sigma;}}_j^{-1}(f\left(z_3\right)-m_j)\&CircleTimes;f^{\&prime;}\left(z_3\right),$ Vector t=[t ₁, t ₂..., t _i..., t _d] ^ti-th element be element multiplication on representing matrix correspondence position;

(5.3) gradient descent algorithm is utilized to upgrade following parameter:

$W_{\mathrm{new}}^1:=W_{\mathrm{ori}}^1-{\mathrm{\&alpha;}}_W\&dtri;W^1,$

$b_{\mathrm{new}}^1:=b_{\mathrm{ori}}^1-{\mathrm{\&alpha;}}_b\&dtri;b^1,$

$c_{\mathrm{new}}^1:=c_{\mathrm{ori}}^1-{\mathrm{\&alpha;}}_c\&dtri;c^1,$

Wherein, represent the weights after the renewal of ground floor network, represent the output offset after the renewal of ground floor network, represent that the input after the renewal of ground floor network is biased; α _wrepresent the Gradient Descent speed of ground floor network weight, α _brepresent the Gradient Descent speed of ground floor network output offset, α _crepresent the Gradient Descent speed that the input of ground floor network is biased;

(5.4) by above-mentioned parameter and as the initial parameter value of next Gradient Descent, even $W_{\mathrm{ori}}^1=W_{\mathrm{new}}^1,b_{\mathrm{ori}}^1=b_{\mathrm{new}}^1,c_{\mathrm{ori}}^1=c_{\mathrm{new}}^1,$ And the gradient of three parameters is calculated according to step (5.2) with

(5.5) step (5.3) and (5.4) are constantly repeated, until and wherein δ _w, δ _band δ _cbe that three parameter iterations preset upgrade the threshold value stopped, generally value is δ _w=δ _b=δ _c=10 ^-3.

6. the radar target identification method based on degree of deep learning network according to claim 1, output sample data G and the output mean distance image set of the calculating ground floor network wherein described in step (5) close R, carry out as follows:

(5a) the output sample data G={g of ground floor network is established _1,1, g _1,2..., g _j,t..., g _p,T, wherein g _j,trepresent jth part t sample x _j,toutput; If export mean distance image set to close R={r ₁, r ₂..., r _j..., r _p, wherein r _jrepresent jth part Range Profile H _joutput mean distance picture, j=1,2 ..., P, t=1,2 ..., T;

(5b) jth part t sample x is calculated _j,toutput g _j,tand jth part Range Profile H _joutput mean distance as r _j:

$g_{j,t}=f(W_{\mathrm{opt}}^1{x}_{j,t}+b_{\mathrm{opt}}^1),$

$r_j=f(W_{\mathrm{opt}}^1{m}_j+b_{\mathrm{opt}}^1).$

7. the radar target identification method based on degree of deep learning network according to claim 1, the cost function L setting up kth layer network wherein described in step (6) ^k(Y, V; Θ ^k), k>=2, carry out as follows:

(6.1) the input Y of kth layer network is made to equal the output G of kth-1 layer network, i.e. Y=G={g _1,1, g _1,2..., g _j,t..., g _p,T, wherein g _j,trepresent jth part t sample x _j,toutput; The input mean distance of kth layer network is made to equal the output mean distance of kth-1 layer network as R as V, i.e. V=R={r ₁, r ₂..., r _j..., r _p, wherein r _jrepresent jth part Range Profile H _joutput mean distance picture;

(6.2) the cost function L of kth layer network is set up ^k(Y, V; Θ ^k):

$L^k(Y,V;{\mathrm{\&Theta;}}^k)=\frac{1}{T\&CenterDot;P}\underset{j=1}{\overset{P}{\mathrm{\&Sigma;}}}\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}{(r_j-{\hat{g}}_{j,t})}^T(r_j-{\hat{g}}_{j,t})+\mathrm{\&alpha;KL}\left(\widehat{\mathrm{\&rho;}}\right|\left|\mathrm{\&rho;}\right)+\frac{\mathrm{\&beta;}}{2}{\left|\right|W^k\left|\right|}_2^2$

Wherein, ${\hat{g}}_{j,t}=f(W^{k^T}\&CenterDot;f(W^k{g}_{j,t}+b^k)+c^k),\mathrm{KL}\left(\widehat{\mathrm{\&rho;}}\right|\left|\mathrm{\&rho;}\right)=\underset{i=1}{\overset{D}{\mathrm{\&Sigma;}}}\mathrm{\&rho;}\log \frac{\mathrm{\&rho;}}{{\widehat{\mathrm{\&rho;}}}_i}+(1-\mathrm{\&rho;})\log \frac{1-\mathrm{\&rho;}}{1-{\widehat{\mathrm{\&rho;}}}_i},$ it is vector i-th, ρ is the sparse value preset, and α is the sparse penalty factor weight preset, W ^kthe weights of kth layer network, b ^kthe output offset of kth layer network, c ^kthe input being kth layer network is biased.