CN108983187A - Online radar target identification method based on EWC - Google Patents
Online radar target identification method based on EWC Download PDFInfo
- Publication number
- CN108983187A CN108983187A CN201810757440.XA CN201810757440A CN108983187A CN 108983187 A CN108983187 A CN 108983187A CN 201810757440 A CN201810757440 A CN 201810757440A CN 108983187 A CN108983187 A CN 108983187A
- Authority
- CN
- China
- Prior art keywords
- high range
- range resolution
- batch
- pth
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computer Networks & Wireless Communication (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a kind of online radar target identification method based on EWC, belongs to Radar Technology field, main thought are as follows: determines the original Radar High Range Resolution training data S of pth batchpTarget category lpWith the original Radar High Range Resolution test data T of pth batchpTarget category Tlp;P=1,2 ..., P, P > 1;Convolutional neural networks model is established, trained convolutional neural networks are obtained;Then S is obtainedpThe Fisher information matrix of middle m data;Determine the 1st crowd of original Radar High Range Resolution test data T1Target category Tl1To P crowdes of original Radar High Range Resolution test data TPTarget category TlPAnd the prediction target category l ' of the 1st batch of original Radar High Range Resolution test data1To the prediction target category l ' of P batches of original Radar High Range Resolution test datasP;And then it obtains the 1st classification and identifies correct target to theA classification identifies correct target,By what is obtained at this timeA classification identifies that correct target is a kind of online radar target recognition result based on EWC.
Description
Technical field
The invention belongs to Radar Technology field, in particular to a kind of online radar target identification method based on EWC, i.e. base
Consolidate the online radar target identification method of (Elastic Weight Consolidation, EWC) in elastic weight, is suitable for
On-line study radar target recognition task.
Background technique
With the development of modern war advanced technology, the demand of Technology of Radar Target Identification is further strong;Radar high-resolution
Range Profile (High Resolution Range Profile, HRRP) is the target scattering idea obtained with wideband-radar signal
The amplitude wave-shape for the vector sum that echo projects on radar ray, High Range Resolution HRRP sample reflects to be regarded in certain radar
When angle, the Radar Cross Section of scatterer (such as head, wing, tail rudder, air inlet, engine) in target
(RCS) along the distribution situation of radar line of sight (RLOS), the relative geometrical relation of scattering point is embodied;Therefore, HRRP sample includes
Target structural information abundant, such as target size, scattering point structure etc. are very valuable to target recognition and classification.
Object detection method based on deep learning develops very rapidly in recent years, and convolutional neural networks are as deep learning
One of, become the research hotspot of current speech analysis and field of image recognition, the weight of convolutional neural networks shares network
Structure is allowed to be more closely similar to biological neural network, reduces the complexity of network model, reduces the quantity of weight, and advantage exists
What is showed when the input of network is multidimensional image becomes apparent, and image is allow to avoid tradition directly as the input of network
Complicated feature extraction and data reconstruction processes in recognizer;Convolutional network is one to identify two-dimensional shapes and special designing
A multilayer perceptron, this network structure have height constant translation, scaling, inclination or the deformation of his total form
Property.
Currently, needing online to identify the data constantly obtained due to the particularity of radar;With the increasing of data
Add, many algorithms can generate forgetting to previous data characteristics in the training identification process to new data, cause to previous number
According to recognition capability decline rapidly.
Summary of the invention
In view of the deficiency of the prior art, it is an object of the invention to propose a kind of online radar based on EWC
Target identification method, this kind is based on the online radar target identification method of EWC using High Range Resolution HRRP to the spy of target
Sign is extracted and is identified, consolidates (Elastic Weight Consolidation, EWC) in particular with elastic weight to prevent
The only forgetting in current data training process to previous data characteristics, so that radar is gone back while identifying to current data training
It can guarantee the recognition capability to previous data, and retain the feature of previous data.
Technical thought of the invention: the data after carrying out Short Time Fourier Transform by HRRP data set, training are end-to-end
Convolutional neural networks model, and EWC is added in the training of every batch data and improves network model to the memory capability of data characteristics,
The recognition capability to previous data characteristics is kept while guarantee to current data recognition capability.
To reach above-mentioned technical purpose, the present invention is realised by adopting the following technical scheme.
A kind of online radar target identification method based on EWC, comprising the following steps:
Step 1, the original Radar High Range Resolution training data S of pth batch is determinedpWith pth batch original radar high-resolution away from
From as test data Tp, and determine the original Radar High Range Resolution training data S of pth batchpTarget category lpWith pth batch
Original Radar High Range Resolution test data TpTarget category Tlp;P=1,2 ..., P, P > 1;
Step 2, convolutional neural networks model is established, and according to the original Radar High Range Resolution training data S of pth batchp,
Obtain trained convolutional neural networks;
Step 3, according to trained convolutional neural networks, the original Radar High Range Resolution training data of pth batch is obtained
SpThe Fisher information matrix of middle m data;m≥1;
Step 4, according to the original Radar High Range Resolution training data S of pth batchpThe Fisher information square of middle m data
Battle array, determines the updated convolutional neural networks model M of pth '+1 batch datap'+1;P'=1,2,3 ..., P-1, p'+1=2,
The initial value that the initial value of 3 ..., P, p' are 1, p'+1 is 2;
Step 5, it enables the value of p' add 1, repeats step 4, until p'=P-1, p'+1=P, and then obtain the update of P batch data
Convolutional neural networks model M afterwardsP, the value of p' is then initialized as 1;
Step 6, the 1st crowd of original Radar High Range Resolution test data T is determined1Target category Tl1It is original to P batches
Radar High Range Resolution test data TPTarget category TlP, and according to the updated convolutional neural networks mould of P batch data
Type MP, obtain the prediction target category l of the 1st batch of original Radar High Range Resolution test data1To P batches of original radar high scores
Distinguish the prediction target category l ' of Range Profile test dataP;
Step 7, if leWith TleEqual, e=1,2 ..., P, then explanation has identified e batches of original radar high-resolution distances
It as the target in training data, and is denoted as the e' classification and identifies correct target, the initial value of e' is 1, and the value of e' is enabled to add 1;
If leWith TleIt is unequal, then illustrate the target category identification mistake of e batches of original Radar High Range Resolution test datas, gives up
The secondary result;
It enables e take 1 to P respectively, and then obtains the 1st classification and identify correct target to theA classification identifies correct target,By what is obtained at this timeA classification identifies that correct target is a kind of online radar target recognition result based on EWC.
The invention has the following advantages over the prior art:
First, the present invention solves the disadvantage that traditional neural network is unable to timing sequence process multiple tasks, and proposing practical has
The method of effect makes the importance that ensure that previous tasks during timing training pattern, so that learning the same of new task
When maintain Memorability and recognition capability to previous task.
Second, the present invention extracts Radar High Range Resolution feature using depth network architecture, for large batch of
For Radar High Range Resolution data, can feature in automatic learning data, the high dimensional feature of especially data known
Not, operation efficiency is improved.
Detailed description of the invention
Invention is further described in detail with reference to the accompanying drawings and detailed description.
Fig. 1 is a kind of online radar target identification method implementation flow chart based on EWC of the invention;
Fig. 2 a is the actual measurement scene figure of Yark-42 aircraft;
Fig. 2 b is the actual measurement scene figure of II aircraft of Cessna Citation S/;
Fig. 2 c is the actual measurement scene figure of An-26 aircraft;
Fig. 3 is performance change curve graph of the present invention to three classes aircraft task A identification;
Fig. 4 is performance change curve graph of the present invention to three classes aircraft task B identification.
Specific embodiment
It referring to Fig.1, is a kind of online radar target identification method implementation flow chart based on EWC of the invention;Wherein institute
State the online radar target identification method based on EWC, comprising the following steps:
Step 1, training sample and test sample are obtained, to data initialization.
Determine high resolution radar, the high resolution radar receives target echo data in its detection range, then from described
N number of data are randomly selected in target echo data as the original Radar High Range Resolution training data S of pth batchp, in the mesh
It is removed in mark echo data and randomly selects N' data again outside N number of data of extraction, as the original radar high-resolution distance of pth batch
As test data Tp, p=1,2 ..., P, P indicates to obtain original Radar High Range Resolution training data and original radar high score
Distinguish total lot number of Range Profile test data.
The original Radar High Range Resolution training data S of (1a) pth batchp={ s1,s2,…,sn,…,sN, wherein snIt indicates
The original Radar High Range Resolution training data S of pth batchpIn n-th of Range Profile, sn=[sn1,sn2,…,sni,…,snD]T,
[·]TThe transposition of representing matrix, sniIndicate the original Radar High Range Resolution training data S of pth batchpIn n-th of Range Profile exist
Value in i-th of distance unit, n=1,2 ..., N, N indicate the original Radar High Range Resolution training data S of pth batchpIncluding
Range Profile total number, i.e. pth crowd original Radar High Range Resolution training data SpIncluding training sample total number, i=
1,2 ..., D, D indicate the original Radar High Range Resolution training data S of pth batchpIn each High Range Resolution distance for including
Unit total number (i.e. single sample vector dimension).
(1b) calculates the original Radar High Range Resolution training data S of pth batchpIn n-th of High Range Resolution snCenter of gravity
Wn:
Pth is criticized original Radar High Range Resolution training data S by (1c)pIn n-th of High Range Resolution snCenter remove
Move to its center of gravity Wn, and value x of n-th of High Range Resolution at i-th of distance unit after movement is calculatedni, expression
Formula are as follows:
Wherein, FFT indicates Fast Fourier Transform (FFT), and IFFT indicates inverse fast fourier transform, sniIndicate the original thunder of pth batch
Up to High Range Resolution training data SpIn value of n-th of High Range Resolution in i-th of distance unit, CnIndicate that pth batch is original
Radar High Range Resolution training data SpIn n-th of High Range Resolution snCenter,φ[Wn] indicate pth batch original
Beginning Radar High Range Resolution training data SpIn n-th of High Range Resolution snCenter of gravity WnCorresponding phase, φ [Cn] table
Show the original Radar High Range Resolution training data S of pth batchpIn n-th of High Range Resolution snCenter CnCorresponding phase
Position, a indicate the original Radar High Range Resolution training data S of pth batchpIn n-th of High Range Resolution snCenter CnPlace away from
From unit and the original Radar High Range Resolution training data S of pth batchpIn n-th of High Range Resolution snCenter of gravity WnPlace away from
From the distance between unit, e indicates that exponential function, j indicate imaginary unit.
(1d) enables i take 1 to D, repeats (1c), so respectively obtain it is mobile after n-th of High Range Resolution at the 1st
Value x at distance unitn1Value x of n-th of High Range Resolution at the D distance unit after to movementnD, after being denoted as movement
N-th of High Range Resolution xn, xn=[xn1,xn2,…,xni,…,xnD], the value of i is then initialized as 1.
(1e) enables n take 1 to N respectively, repeats (1c) and (1d), so respectively obtain it is mobile after the 1st high-resolution away from
From as x1N-th High Range Resolution x after to movementN, it is denoted as the original Radar High Range Resolution training data of pth batch after movement
Xp, Xp={ x1,x2,…,xn,…,xN, xn=[xn1,xn2,…,xni,…,xnD]。
With 1,2 ..., D is as abscissa, with x1,x2,…,xn,…,xNAs ordinate, by the original thunder of pth after movement batch
Up to High Range Resolution training data XpIt is depicted as 2 d plane picture, p-th of sample echo waveform figure is denoted as, according to p-th of sample
This echo waveform figure is to the original Radar High Range Resolution training data S of pth batchpTarget category is added, the original thunder of pth batch is denoted as
Up to High Range Resolution training data SpTarget category lp。
The original Radar High Range Resolution test data T of (1f) pth batchp, Tp={ t1,t2,…,tn′,…,tN′, wherein
tn′Indicate the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a Range Profile, tn=[tn′1,tn′2,…,
tn′i′,…,tn′D′]T, []TThe transposition of representing matrix, sn′i′Indicate the original Radar High Range Resolution test data T of pth batchp
In value of the n-th ' a Range Profile in the i-th ' a distance unit, n '=1,2 ..., N ', N ' expression pth batch original radar high-resolution away from
From as test data TpIncluding Range Profile total number, i.e. pth crowd original Radar High Range Resolution test data TpIncluding instruction
Practice sample total number, i '=1,2 ..., the original Radar High Range Resolution test data T of D ', D ' expression pth batchpIn it is each high
The distance unit total number (i.e. single sample vector dimension) that resolution distance picture includes.
(1g) calculates the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution tn′Weight
Heart Wn′:
Pth is criticized original Radar High Range Resolution test data T by (1h)pIn the n-th ' a High Range Resolution tn′Center
It moves to its center of gravity Wn′, and value x of the n-th ' a High Range Resolution at the i-th ' a distance unit after movement is calculatedn′i′',
Its expression formula are as follows:
Wherein, FFT indicates Fast Fourier Transform (FFT), and IFFT indicates inverse fast fourier transform, tn′i′Indicate that pth batch is original
Radar High Range Resolution test data TpIn value of the n-th ' a High Range Resolution in the i-th ' a distance unit, Cn′Indicate pth
Criticize original Radar High Range Resolution test data TpIn the n-th ' a High Range Resolution tn′Center,φ[Wn′] table
Show the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution tn′Center of gravity Wn′Corresponding phase
Position, φ [Cn′] indicate the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution tn′Center
Cn′Corresponding phase, a indicate the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution tn′
Center Cn′Place distance unit and the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution
tn′Center of gravity Wn′The distance between place distance unit, e indicate that exponential function, j indicate imaginary unit.
(1i) enables i ' take 1 to D ', repeats (1h), so respectively obtain it is mobile after the n-th ' a High Range Resolution the
Value x at 1 distance unitn′1' to value x of the n-th ' a High Range Resolution at a distance unit of D ' after mobilen′D′', note
For after movement the n-th ' a High Range Resolution xn′', xn'=[xn′1′,xn′2′,…,xn′i′′,…,xn′D′'], then by the value of i '
It is initialized as 1.
(1j) enables n ' take 1 to N ' respectively, repeats (1h) and (1i), and then respectively obtains the 1st high-resolution after movement
Range Profile x1' to a High Range Resolution x of N ' after mobileN′', the original Radar High Range Resolution of pth batch is surveyed after being denoted as movement
Try data Tp', Tp'={ x1′,x2′,…,xn′′,…,xN′', xn'=[xn′1′,xn′2′,…,xn′i′′,…,xn′D′′]。
With 1,2 ..., D is as abscissa, with xn′1′,xn′2′,…,xn′i′′,…,xn′D′' it is used as ordinate, after movement
The original Radar High Range Resolution test data T of pth batchp' it is depicted as two-dimensional surface curve graph, it is bent to be denoted as p-th of sample echo
Line chart, according to p-th of sample echo curve figure to the original Radar High Range Resolution test data T of pth batchpAdd target class
Not, it is denoted as the original Radar High Range Resolution test data T of pth batchpTarget category Tlp。
Step 2, convolutional neural networks model is established.
The convolutional neural networks model is made of three-layer coil lamination and two layers of full articulamentum, and construction step is as follows:
(2a) constructs first layer convolutional layer: the first layer convolutional layer be used for batch original radar high-resolution of pth after movement away from
From as training data XpOne-dimensional convolution is carried out, includes C in first layer convolutional layer1A convolution kernel, and by the C of first layer1A convolution kernel
It is denoted asFor criticizing original Radar High Range Resolution training data X with pth after movementpCarry out convolution;Size setting
For M1×1×C1, wherein M1Indicate the size of each convolution kernel window in first layer convolutional layer, 1≤M1≤D。
The convolution step-length that first layer convolutional layer is arranged is L1, 1≤L1≤ D-1 is usually arranged to reduce down-sampling process
L1=2;The original Radar High Range Resolution training data X of pth batch after will be mobilepWith the C in first layer convolutional layer1A convolution kernel
Convolution is carried out respectively, obtains first layer convolutional layer C1It is after a convolution as a result, and being denoted as the C of first layer convolutional layer1A characteristic patternCalculation formula is as follows:
Wherein,Indicate the C of first layer convolutional layer1A characteristic pattern, XpAfter indicating mobile the original radar high-resolution of pth batch away from
From as training data,Indicate the C in first layer convolutional layer1A convolution kernel,Indicate complete 1 biasing of first layer convolutional layer, * table
Show convolution operation, f () indicates activation primitive, f (z1)=max (0, z1),Maximum is sought in max () expression
Value Operations.
(2b) constructs second layer convolutional layer: second layer convolutional layer includes C2A convolution kernel, and by the C of second layer convolutional layer2It is a
Convolution kernel is defined asFor the C with first layer convolutional layer1A characteristic patternCarry out convolution, the C of second layer convolutional layer2A volume
Product coreIt is dimensioned to M2×C1×C2, wherein M2For the size of convolution kernel window each in second layer convolutional layer,The convolution step-length that second layer convolutional layer is arranged is L2,L is set in the present embodiment2=2.
By the C of first layer convolutional layer1A characteristic patternWith the C of second layer convolutional layer2A convolution kernelConvolution is carried out respectively,
Obtain second layer convolutional layer C2It is after a convolution as a result, and being denoted as the C of second layer convolutional layer2A characteristic patternIts calculation formula
It is as follows:
Wherein,Indicate complete 1 biasing of second layer convolutional layer, * indicates convolution operation, and f () indicates activation primitive, f (z2)
=max (0, z2),
(2c) constructs third layer convolutional layer: the third layer convolutional layer is used for the C to second layer convolutional layer2A characteristic patternInto
Row convolution, the convolution kernel for defining the third layer convolutional layer areThe convolution kernel of third layer convolutional layer includes C3A convolution kernel, and the
The convolution kernel of three-layer coil laminationBe dimensioned to M3×C2×C3, wherein M3Indicate each convolution kernel in third layer convolutional layer
The size of window,The convolution step-length that third layer convolutional layer is arranged is L3,It is set in the present embodiment
Set L3=2.
By the C of second layer convolutional layer2A characteristic patternWith the convolution kernel of third layer convolutional layerConvolution is carried out respectively, is obtained
To third layer convolutional layer C3It is after a convolution as a result, and being denoted as the C of third layer convolutional layer3A characteristic patternIts calculation formula is such as
Under:
Wherein,Indicate complete 1 biasing of third layer convolutional layer, * indicates convolution operation, and f () indicates activation primitive, f (z3)
=max (0, z3),
(2d) constructs the 4th layer of full articulamentum: first by the C of third layer convolutional layer3A characteristic patternIt elongates respectively and is transformed to grow
Degree isColumn vector, and then obtain elongating transformed C3A column vector, each column vector includeA neuron from
And it obtains elongating transformedA neuron;4th layer of full articulamentum is provided with h neuron, for that will elongate transformation
C afterwards3The weight matrix of a column vector and the 4th layer of full articulamentumWith complete 1 biasing of the 4th layer of full articulamentumIt carries out non-
Linear process transformation, the data result after obtaining the 4th layer of full articulamentum nonlinear transformationIts calculation expression are as follows:
Wherein,Indicate transformed by elongatingH neuron phase of a neuron and the 4th layer of full articulamentum
The weight matrix of connection,Indicating complete 1 biasing of the 4th layer of full articulamentum, representing matrix is multiplied, and f () indicates activation primitive,
f(z4)=max (0, z4),
(2e) constructs the full articulamentum of layer 5: the full articulamentum of layer 5 is provided with a neuron of h ', for complete by the 4th layer
Data result after 4th layer of full articulamentum nonlinear transformation of articulamentum outputWith the weight square of the full articulamentum of the layer 5
Battle arrayWith complete 1 biasing of the full articulamentum of layer 5Carry out linear transformation, the number after obtaining the full articulamentum linear transformation of layer 5
According to result
Wherein, the data result after the full articulamentum linear transformation of the layer 5Its calculation expression are as follows:
Wherein, W5It indicates by a neuron phase of h ' of the h neuron and the full articulamentum of layer 5 of the 4th layer of full articulamentum
The h connected and composed × h ' dimension matrix,Indicate complete 1 biasing of the full articulamentum of layer 5.
Data result after obtaining the full articulamentum linear transformation of layer 5Afterwards, illustrate that convolutional neural networks building is completed,
It is denoted as trained convolutional neural networks.
Step 3, the Fisher information matrix of EWC is calculated.
Data result after the full articulamentum linear transformation of layer 5In randomly select m data, calculate separately extraction
The parameter that is connected to all convolutional layers and entirely of m dataSingle order local derviation
Number, and its square is calculated to each first-order partial derivative result and is summed, the single order derived function quadratic sum of m data is obtained, each
The single order derived function quadratic sum of data is all the Fisher information matrix of corresponding data, and then obtains the original radar high score of pth batch
Distinguish Range Profile training data SpThe Fisher information matrix of middle m data, the wherein original Radar High Range Resolution training of pth batch
Data SpIn j-th of data Fisher information matrix be Fpj, formula is as follows:
Wherein,Data result after indicating the full articulamentum linear transformation of layer 5In j-th of data, j=
1 ..., m, FpjIndicate the original Radar High Range Resolution training data S of pth batchpIn j-th of data Fisher information matrix,
Convolutional neural networks model is updated for next batch data.
Step 4, the original Radar High Range Resolution training data S in pth '+1 batch is obtainedp+1, the original radar in pth '+1 batch it is high
Resolution distance is as training data Sp'+1Target category lp'+1;P'=1,2,3 ..., P-1, p'+1=2,3 ..., P's, p' is initial
The initial value that value is 1, p'+1 is 2.
Then the original Radar High Range Resolution training data S in pth '+1 batch is calculatedp'+1EWC loss function LOSSp'+1
Are as follows:
Wherein, λ is weight coefficient, and value is usually (0,1);Qp'+1For the original Radar High Range Resolution instruction in pth '+1 batch
Practice data Sp'+1Parameter variation value, calculation formula are as follows:
The original Radar High Range Resolution training data S in pth '+1 batch is utilized by Back Propagation Algorithmp'+1EWC loss
Function LOSSp'+1Training is updated to trained convolutional neural networks, obtains the updated convolution mind of pth '+1 batch data
Through network model Mp'+1。
Step 5, it enables the value of p' add 1, repeats step 4, until p'=P-1, p'+1=P, updated until obtaining P batch data
Convolutional neural networks model M afterwardsP, the value of p' is then initialized as 1.
Step 6, for the original Radar High Range Resolution test data T of pth after movement batchp' and the original radar height of pth batch
Resolution distance is as test data TpTarget category Tlp, enable the value of p take 1 to P respectively, and then the 1st batch of original thunder after being moved
Up to High Range Resolution test data T1' P criticizes original Radar High Range Resolution test data T to after movingp' and the 1st
Criticize original Radar High Range Resolution test data T1Target category Tl1Number is tested to P batches of original Radar High Range Resolutions
According to TPTarget category TlP。
And the 1st crowd of original Radar High Range Resolution test data T after moving1' high to P batches of original radars after movement
Resolution distance is inputted as test data into trained convolutional neural networks, utilizes the updated convolutional Neural of P batch data
Network model MP, respectively correspond to obtain the prediction target category l of the 1st batch of original Radar High Range Resolution test data1' to P
Criticize the prediction target category l ' of original Radar High Range Resolution test dataP。
Step 7, if le' and TleEqual, e=1,2 ..., P illustrate e batches of original Radar High Range Resolution test datas
Target category identification it is correct, that is, think to have identified the target in e batches of original Radar High Range Resolution training datas, note
Correct target is identified for the e' classification, and the initial value of e' is 1, and the value of e' is enabled to add 1;If le' and TleIt is unequal, then illustrate
The target category of e batches of original Radar High Range Resolution test datas identifies mistake, gives up the secondary result.
It enables e take 1 to P respectively, and then obtains the 1st classification and identify correct target to theA classification identifies correct target,By what is obtained at this timeA classification identifies that correct target is a kind of online radar target recognition result based on EWC.
Effect of the invention is further illustrated by emulating below to the measured data of three classes aircraft:
The radar measured data of original Radar High Range Resolution is obtained, actual measurement scene is referring to Fig. 2 a, Fig. 2 b and Fig. 2 c institute
Show, wherein Fig. 2 a is the actual measurement scene figure of Yark-42 aircraft, and Fig. 2 b is the actual measurement scene of II aircraft of Cessna Citation S/
Figure, Fig. 2 c is the actual measurement scene figure of An-26 aircraft.
When emulation, the original high-resolution Range Profile that will acquire is divided into two classes: training set Tr and test set Te,
Wherein training set Tr includes TrAAnd TrB。TrAFor the training data of task A, TrBFor the training data of task B;Also, training
Set TrAAnd TrBThe respectively different azimuth information data of High Range Resolution.
The identification situation of observation mission A and task B, as a result as shown in Figure 3 and Figure 4, Fig. 3 are that the present invention appoints three classes aircraft
The performance change curve graph of business A identification, Fig. 4 are performance change curve graph of the present invention to three classes aircraft task B identification.
From experimental result as can be seen that model can not only have good knowledge to current new task when new task arrives
Other ability, as shown in figure 4, and still can have good recognition capability to previous tasks, such as shown in Figure 3.
In conclusion emulation experiment demonstrates correctness of the invention, validity and reliability.
Obviously, various changes and modifications can be made to the invention without departing from essence of the invention by those skilled in the art
Mind and range;In this way, if these modifications and changes of the present invention belongs to the range of the claims in the present invention and its equivalent technologies
Within, then the present invention is also intended to include these modifications and variations.
Claims (7)
1. a kind of online radar target identification method based on EWC, which comprises the following steps:
Step 1, the original Radar High Range Resolution training data S of pth batch is determinedpIt is surveyed with the original Radar High Range Resolution of pth batch
Try data Tp, and determine the original Radar High Range Resolution training data S of pth batchpTarget category lpWith the original thunder of pth batch
Up to High Range Resolution test data TpTarget category Tlp;P=1,2 ..., P, P > 1;
Step 2, convolutional neural networks model is established, and according to the original Radar High Range Resolution training data S of pth batchp, obtain
Trained convolutional neural networks;
Step 3, according to trained convolutional neural networks, the original Radar High Range Resolution training data S of pth batch is obtainedpMiddle m
The Fisher information matrix of a data;m≥1;
Step 4, according to the original Radar High Range Resolution training data S of pth batchpThe Fisher information matrix of middle m data, really
Determine the updated convolutional neural networks model M of pth '+1 batch datap'+1;P'=1,2,3 ..., P-1, p'+1=2,3 ..., P, p'
Initial value be 1, p'+1 initial value be 2;
Step 5, it enables the value of p' add 1, repeats step 4, until p'=P-1, p'+1=P, and then it is updated to obtain P batch data
Convolutional neural networks model MP, the value of p' is then initialized as 1;
Step 6, the 1st crowd of original Radar High Range Resolution test data T is determined1Target category Tl1To P batches of original radars
High Range Resolution test data TPTarget category TlP, and according to the updated convolutional neural networks model M of P batch dataP,
Obtain the prediction target category l ' of the 1st batch of original Radar High Range Resolution test data1To P batch original radar high-resolution away from
From the prediction target category l ' as test dataP;
Step 7, if l 'eWith TleEqual, e=1,2 ..., P, then explanation has identified e batches of original Radar High Range Resolution instructions
Practice the target in data, and be denoted as the e' classification and identify correct target, the initial value of e' is 1, and the value of e' is enabled to add 1;If l 'e
With TleIt is unequal, then illustrate the target category identification mistake of e batches of original Radar High Range Resolution test datas, gives up this
Secondary result;
It enables e take 1 to P respectively, and then obtains the 1st classification and identify correct target to theA classification identifies correct target,By what is obtained at this timeA classification identifies that correct target is a kind of online radar target recognition result based on EWC.
2. a kind of online radar target identification method based on EWC as described in claim 1, which is characterized in that in step 1
In, the original Radar High Range Resolution training data S of the pth batchpWith the original Radar High Range Resolution test data of pth batch
Tp, determination process are as follows:
Determine that high resolution radar, the high resolution radar receive target echo data in its detection range, then from the target
N number of data are randomly selected in echo data as the original Radar High Range Resolution training data S of pth batchp, returned in the target
N number of data that wave number removes extraction in randomly select N' data outside again, survey as the original Radar High Range Resolution of pth batch
Try data Tp, p=1,2 ..., P, P indicate to obtain original Radar High Range Resolution training data and original radar high-resolution away from
From total lot number as test data;
The original Radar High Range Resolution training data S of the determining pth batchpTarget category lp, determination process are as follows:
The original Radar High Range Resolution training data S of (1a) pth batchp={ s1,s2,…,sn,…,sN, wherein snIndicate pth
Criticize original Radar High Range Resolution training data SpIn n-th of Range Profile, sn=[sn1,sn2,…,sni,…,snD]T, []T
The transposition of representing matrix, sniIndicate the original Radar High Range Resolution training data S of pth batchpIn n-th of Range Profile at i-th
Value in distance unit, n=1,2 ..., N, N indicate the original Radar High Range Resolution training data S of pth batchpIncluding distance
As total number, the i.e. original Radar High Range Resolution training data S of pth batchpIncluding training sample total number, i=1,2 ...,
D, D indicate the original Radar High Range Resolution training data S of pth batchpIn each High Range Resolution distance unit for including it is total
Number;
(1b) calculates the original Radar High Range Resolution training data S of pth batchpIn n-th of High Range Resolution snCenter of gravity Wn:
Pth is criticized original Radar High Range Resolution training data S by (1c)pIn n-th of High Range Resolution snCentral region extremely
Its center of gravity Wn, and value x of n-th of High Range Resolution at i-th of distance unit after movement is calculatedni, expression formula are as follows:
Wherein, FFT indicates Fast Fourier Transform (FFT), and IFFT indicates inverse fast fourier transform, sniIndicate that the original radar of pth batch is high
Resolution distance is as training data SpIn value of n-th of High Range Resolution in i-th of distance unit, CnIndicate the original radar of pth batch
High Range Resolution training data SpIn n-th of High Range Resolution snCenter,φ[Wn] indicate the original thunder of pth batch
Up to High Range Resolution training data SpIn n-th of High Range Resolution snCenter of gravity WnCorresponding phase, φ [Cn] indicate the
P crowdes of original Radar High Range Resolution training data SpIn n-th of High Range Resolution snCenter CnCorresponding phase, a table
Show the original Radar High Range Resolution training data S of pth batchpIn n-th of High Range Resolution snCenter CnPlace distance unit
With the original Radar High Range Resolution training data S of pth batchpIn n-th of High Range Resolution snCenter of gravity WnPlace distance unit
The distance between, e indicates that exponential function, j indicate imaginary unit;
(1d) enables i take 1 to D, repeats (1c), so respectively obtain it is mobile after n-th of High Range Resolution in the 1st distance
Value x at unitn1Value x of n-th of High Range Resolution at the D distance unit after to movementnD, it is denoted as after movement n-th
High Range Resolution xn, xn=[xn1,xn2,…,xni,…,xnD], the value of i is then initialized as 1;
(1e) enables n take 1 to N respectively, repeats (1c) and (1d), and then respectively obtains the 1st High Range Resolution x after movement1
N-th High Range Resolution x after to movementN, it is denoted as the original Radar High Range Resolution training data X of pth batch after movementp, Xp=
{x1,x2,…,xn,…,xN, xn=[xn1,xn2,…,xni,…,xnD];
With 1,2 ..., D is as abscissa, with x1,x2,…,xn,…,xNIt is as ordinate, the original radar of pth after movement batch is high
Resolution distance is as training data XpIt is depicted as 2 d plane picture, is denoted as p-th of sample echo waveform figure, is returned according to p-th of sample
Wave waveform diagram is to the original Radar High Range Resolution training data S of pth batchpTarget category is added, it is high to be denoted as the original radar of pth batch
Resolution distance is as training data SpTarget category lp。
3. a kind of online radar target identification method based on EWC as described in claim 1, which is characterized in that in step 1
In, the original Radar High Range Resolution test data T of the pth batchpTarget category Tlp, determination process are as follows:
The original Radar High Range Resolution test data T of (1f) pth batchp, Tp={ t1,t2,…,tn′,…,tN′, wherein tn′It indicates
The original Radar High Range Resolution test data T of pth batchpIn the n-th ' a Range Profile, tn=[tn′1,tn′2,…,tn′i′,…,
tn′D′]T, []TThe transposition of representing matrix, sn′i′Indicate the original Radar High Range Resolution test data T of pth batchpIn it is the n-th ' a
Value of the Range Profile in the i-th ' a distance unit, n '=1,2 ..., the original Radar High Range Resolution test of N ', N ' expression pth batch
Data TpIncluding Range Profile total number, i.e. pth crowd original Radar High Range Resolution test data TpIncluding training sample it is total
Number, i '=1,2 ..., the original Radar High Range Resolution test data T of D ', D ' expression pth batchpIn each high-resolution distance
As the distance unit total number for including;
(1g) calculates the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution tn′Center of gravity
Wn′:
Pth is criticized original Radar High Range Resolution test data T by (1h)pIn the n-th ' a High Range Resolution tn′Central region
To its center of gravity Wn′, and value x of the n-th ' a High Range Resolution at the i-th ' a distance unit after movement is calculatedn′i′', table
Up to formula are as follows:
Wherein, FFT indicates Fast Fourier Transform (FFT), and IFFT indicates inverse fast fourier transform, tn′i′Indicate the original radar of pth batch
High Range Resolution test data TpIn value of the n-th ' a High Range Resolution in the i-th ' a distance unit, Cn′Indicate pth batch original
Beginning Radar High Range Resolution test data TpIn the n-th ' a High Range Resolution tn′Center,φ[Wn′] indicate the
P crowdes of original Radar High Range Resolution test data TpIn the n-th ' a High Range Resolution tn′Center of gravity Wn′Corresponding phase,
φ[Cn′] indicate the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution tn′Center Cn′
Corresponding phase, a indicate the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution tn′'s
Center Cn′Place distance unit and the original Radar High Range Resolution test data T of pth batchpIn the n-th ' a High Range Resolution tn′
Center of gravity Wn′The distance between place distance unit, e indicate that exponential function, j indicate imaginary unit;
(1i) enables i ' take 1 to D ', repeats (1h), so respectively obtain it is mobile after the n-th ' a High Range Resolution at the 1st
Value x at distance unitn′1' to value x of the n-th ' a High Range Resolution at a distance unit of D ' after mobilen′D′', it is denoted as
The n-th ' a High Range Resolution x after movementn′', xn'=[xn′1′,xn′2′,…,xn′i′′,…,xn′D′'], it then will be at the beginning of the value of i '
Beginning turns to 1;
(1j) enables n ' take 1 to N ' respectively, repeats (1h) and (1i), and then respectively obtains the 1st high-resolution distance after movement
As x1' to a High Range Resolution x of N ' after mobileN′', it is denoted as the original Radar High Range Resolution test number of pth batch after movement
According to Tp', Tp'={ x1′,x2′,…,xn′′,…,xN′', xn'=[xn′1′,xn′2′,…,xn′i′′,…,xn′D′′];
With 1,2 ..., D is as abscissa, with xn′1′,xn′2′,…,xn′i′′,…,xn′D′' it is used as ordinate, by pth after movement batch
Original Radar High Range Resolution test data Tp' it is depicted as two-dimensional surface curve graph, it is denoted as p-th of sample echo curve figure,
According to p-th of sample echo curve figure to the original Radar High Range Resolution test data T of pth batchpTarget category is added, is denoted as
The original Radar High Range Resolution test data T of pth batchpTarget category Tlp。
4. a kind of online radar target identification method based on EWC as claimed in claim 2, which is characterized in that in step 2
In, the trained convolutional neural networks are obtained after being trained to the convolutional neural networks model of foundation as a result, institute
The convolutional neural networks model for stating foundation includes three-layer coil lamination and two layers of full articulamentum, training process are as follows:
The convolution step-length that first layer convolutional layer is arranged in (2a) is L1, 1≤L1≤ D-1 includes C in first layer convolutional layer1A convolution
Core, and by the C of first layer1A convolution kernel is denoted as Be dimensioned to M1×1×C1, wherein M1Indicate first layer convolution
The size of each convolution kernel window, 1≤M in layer1≤D;
The original Radar High Range Resolution training data X of pth batch after will be mobilepWith the C in first layer convolutional layer1A convolution kernel point
Not carry out convolution, obtain first layer convolutional layer C1It is after a convolution as a result, and being denoted as the C of first layer convolutional layer1A characteristic pattern
Calculation formula is as follows:
Wherein,Indicate the C of first layer convolutional layer1A characteristic pattern,Indicate complete 1 biasing of first layer convolutional layer, * indicates convolution
Operation, f () indicate activation primitive, f (z1)=max (0, z1),Maximum value behaviour is sought in max () expression
Make;
(2b) second layer convolutional layer includes C2A convolution kernel, and by the C of second layer convolutional layer2A convolution kernel is defined asThe second layer
The C of convolutional layer2A convolution kernelIt is dimensioned to M2×C1×C2, wherein M2For convolution kernel window each in second layer convolutional layer
Size,The convolution step-length that second layer convolutional layer is arranged is L2,
By the C of first layer convolutional layer1A characteristic patternWith the C of second layer convolutional layer2A convolution kernelConvolution is carried out respectively, is obtained
To second layer convolutional layer C2It is after a convolution as a result, and being denoted as the C of second layer convolutional layer2A characteristic patternIts calculation formula is such as
Under:
Wherein,Indicate complete 1 biasing of second layer convolutional layer, f (z2)=max (0, z2),
The convolution kernel that (2c) defines the third layer convolutional layer isThe convolution kernel of third layer convolutional layer includes C3A convolution kernel, and
The convolution kernel of third layer convolutional layerBe dimensioned to M3×C2×C3, wherein M3Indicate each convolution in third layer convolutional layer
The size of core window,The convolution step-length that third layer convolutional layer is arranged is L3,
By the C of second layer convolutional layer2A characteristic patternWith the convolution kernel of third layer convolutional layerConvolution is carried out respectively, obtains
Three-layer coil lamination C3It is after a convolution as a result, and being denoted as the C of third layer convolutional layer3A characteristic patternIts calculation formula is as follows:
Wherein,Indicate complete 1 biasing of third layer convolutional layer, f (z3)=max (0, z3),
(2d) is by the C of third layer convolutional layer3A characteristic patternIt elongates respectively and is transformed to length and isColumn vector, and then
To the transformed C of elongation3A column vector, each column vector includeA neuron is transformed to obtain elongatingIt is a
Neuron;
4th layer of full articulamentum is provided with h neuron, for that will elongate transformed C3A column vector and the 4th layer of full articulamentum
Weight matrixWith complete 1 biasing of the 4th layer of full articulamentumNonlinear Processing transformation is carried out, the 4th layer of full articulamentum is obtained
Data result after nonlinear transformationIts calculation expression are as follows:
Wherein,Indicate transformed by elongatingA neuron is connected with h neuron of the 4th layer of full articulamentum
Weight matrix,Indicate complete 1 biasing of the 4th layer of full articulamentum, representing matrix is multiplied, and f () indicates activation primitive, f (z4)
=max (0, z4),
The full articulamentum of (2e) layer 5 is provided with a neuron of h ', for connecting the 4th layer of the 4th layer of full articulamentum output entirely
Data result after layer nonlinear transformationWith the weight matrix of the full articulamentum of the layer 5With the full articulamentum of layer 5
Complete 1 biasingCarry out linear transformation, the data result after obtaining the full articulamentum linear transformation of layer 5Its calculation expression
Are as follows:
Wherein, W5Indicate the structure that is connected by h neuron of the 4th layer of full articulamentum with a neuron of h ' of the full articulamentum of layer 5
At h × h ' dimension matrix,Indicate complete 1 biasing of the full articulamentum of layer 5;
Data result after obtaining the full articulamentum linear transformation of layer 5Afterwards, illustrate that convolutional neural networks building is completed, be denoted as
Trained convolutional neural networks.
5. a kind of online radar target identification method based on EWC as claimed in claim 4, which is characterized in that in step 3
In, the original Radar High Range Resolution training data S of the pth batchpThe Fisher information matrix of middle m data, process are as follows:
Data result after the full articulamentum linear transformation of layer 5In randomly select m data, calculate separately the m of extraction
The parameter that a data are connected to all convolutional layers and entirelyFirst-order partial derivative, and
Its square is calculated to each first-order partial derivative result and is summed, the single order derived function quadratic sum of m data is obtained, each data
Single order derived function quadratic sum is all the Fisher information matrix of corresponding data, and then obtains the original radar high-resolution distance of pth batch
As training data SpThe Fisher information matrix of middle m data, wherein pth criticizes original Radar High Range Resolution training data Sp
In j-th of data Fisher information matrix be Fpj, formula is as follows:
Wherein,Data result after indicating the full articulamentum linear transformation of layer 5In j-th of data, j=1 ..., m,
FpjIndicate the original Radar High Range Resolution training data S of pth batchpIn j-th of data Fisher information matrix.
6. a kind of online radar target identification method based on EWC as claimed in claim 5, which is characterized in that in step 4
In, the updated convolutional neural networks model M of pth '+1 batch datap'+1, determination process are as follows:
The original Radar High Range Resolution training data S in pth '+1 batch is obtained firstp+1, the original radar high-resolution in pth '+1 batch away from
From as training data Sp'+1Target category lp'+1;P'=1,2,3 ..., P-1, p'+1=2, the initial value of 3 ..., P, p' are 1,
The initial value of p'+1 is 2;
Then the original Radar High Range Resolution training data S in pth '+1 batch is calculatedp'+1EWC loss function LOSSp'+1Are as follows:
Wherein, λ is weight coefficient, and value is usually (0,1);Qp'+1For the original Radar High Range Resolution training number in pth '+1 batch
According to Sp'+1Parameter variation value, calculation formula are as follows:
The original Radar High Range Resolution training data S in pth '+1 batch is utilized eventually by Back Propagation Algorithmp'+1EWC loss
Function LOSSp'+1Training is updated to trained convolutional neural networks, obtains the updated convolution mind of pth '+1 batch data
Through network model Mp'+1。
7. a kind of online radar target identification method based on EWC as described in claim 3 or 6, which is characterized in that step 6
Determination process are as follows:
For the original Radar High Range Resolution test data T of pth after movement batchp' and the original Radar High Range Resolution of pth batch
Test data TpTarget category Tlp, enable the value of p take 1 to P respectively, so after being moved the 1st batch of original radar high-resolution away from
From as test data T '1P crowdes of original Radar High Range Resolution test data T after to movementp' and the 1st batch of original radar
High Range Resolution test data T1Target category Tl1To P crowdes of original Radar High Range Resolution test data TPTarget
Classification TlP;
And the 1st crowd of original Radar High Range Resolution test data T ' after moving1P batches of original radar high-resolution after to movement
Range Profile test data is inputted into trained convolutional neural networks, utilizes the updated convolutional neural networks of P batch data
Model MP, respectively correspond to obtain the prediction target category l ' of the 1st batch of original Radar High Range Resolution test data1To P batches of originals
The prediction target category l ' of beginning Radar High Range Resolution test dataP。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810757440.XA CN108983187B (en) | 2018-07-11 | 2018-07-11 | Online radar target identification method based on EWC |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810757440.XA CN108983187B (en) | 2018-07-11 | 2018-07-11 | Online radar target identification method based on EWC |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108983187A true CN108983187A (en) | 2018-12-11 |
CN108983187B CN108983187B (en) | 2022-07-15 |
Family
ID=64536851
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810757440.XA Active CN108983187B (en) | 2018-07-11 | 2018-07-11 | Online radar target identification method based on EWC |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108983187B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113126052A (en) * | 2021-03-08 | 2021-07-16 | 西安电子科技大学 | High-resolution range profile target identification online library building method based on stage-by-stage segmentation training |
CN113171102A (en) * | 2021-04-08 | 2021-07-27 | 南京信息工程大学 | ECG data classification method based on continuous deep learning |
CN114246563A (en) * | 2021-12-17 | 2022-03-29 | 重庆大学 | Intelligent heart and lung function monitoring equipment based on millimeter wave radar |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080169939A1 (en) * | 2007-01-11 | 2008-07-17 | Dickens Charles E | Early warning control system for vehicular crossing safety |
CN104459668A (en) * | 2014-12-03 | 2015-03-25 | 西安电子科技大学 | Radar target recognition method based on deep learning network |
US20160358024A1 (en) * | 2015-06-03 | 2016-12-08 | Hyperverge Inc. | Systems and methods for image processing |
CN107563411A (en) * | 2017-08-07 | 2018-01-09 | 西安电子科技大学 | Online SAR target detection method based on deep learning |
CN107728142A (en) * | 2017-09-18 | 2018-02-23 | 西安电子科技大学 | Radar High Range Resolution target identification method based on two-dimensional convolution network |
CN107784320A (en) * | 2017-09-27 | 2018-03-09 | 电子科技大学 | Radar range profile's target identification method based on convolution SVMs |
-
2018
- 2018-07-11 CN CN201810757440.XA patent/CN108983187B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20080169939A1 (en) * | 2007-01-11 | 2008-07-17 | Dickens Charles E | Early warning control system for vehicular crossing safety |
CN104459668A (en) * | 2014-12-03 | 2015-03-25 | 西安电子科技大学 | Radar target recognition method based on deep learning network |
US20160358024A1 (en) * | 2015-06-03 | 2016-12-08 | Hyperverge Inc. | Systems and methods for image processing |
CN107563411A (en) * | 2017-08-07 | 2018-01-09 | 西安电子科技大学 | Online SAR target detection method based on deep learning |
CN107728142A (en) * | 2017-09-18 | 2018-02-23 | 西安电子科技大学 | Radar High Range Resolution target identification method based on two-dimensional convolution network |
CN107784320A (en) * | 2017-09-27 | 2018-03-09 | 电子科技大学 | Radar range profile's target identification method based on convolution SVMs |
Non-Patent Citations (5)
Title |
---|
KIRKPATRICK J等: "Overcoming catastrophic forgetting in neural networks", 《PROCEEDINGS OF THE NATIONAL ACADEMY OF SCIENCES》 * |
QIAN SONG: "Zero-Shot Learning of SAR Target Feature Space With Deep Generative Neural Networks", 《IEEE GEOSCIENCE AND REMOTE SENSING LETTERS》 * |
徐彬: "基于注意循环神经网络模型的雷达高分辨率距离像目标识别", 《电子与信息学报》 * |
曾贤灏等: "基于Fisher的线性判别回归分类算法", 《安阳工学院学报》 * |
王婷婷: "面向连续状态的神经网络强化学习研究", 《中国优秀博硕士学位论文全文数据库(硕士)》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113126052A (en) * | 2021-03-08 | 2021-07-16 | 西安电子科技大学 | High-resolution range profile target identification online library building method based on stage-by-stage segmentation training |
CN113171102A (en) * | 2021-04-08 | 2021-07-27 | 南京信息工程大学 | ECG data classification method based on continuous deep learning |
CN113171102B (en) * | 2021-04-08 | 2022-09-02 | 南京信息工程大学 | ECG data classification method based on continuous deep learning |
CN114246563A (en) * | 2021-12-17 | 2022-03-29 | 重庆大学 | Intelligent heart and lung function monitoring equipment based on millimeter wave radar |
CN114246563B (en) * | 2021-12-17 | 2023-11-17 | 重庆大学 | Heart and lung function intelligent monitoring equipment based on millimeter wave radar |
Also Published As
Publication number | Publication date |
---|---|
CN108983187B (en) | 2022-07-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108229404B (en) | Radar echo signal target identification method based on deep learning | |
CN109086700B (en) | Radar one-dimensional range profile target identification method based on deep convolutional neural network | |
CN110334741B (en) | Radar one-dimensional range profile identification method based on cyclic neural network | |
CN107728142B (en) | Radar high-resolution range profile target identification method based on two-dimensional convolutional network | |
CN106355151B (en) | A kind of three-dimensional S AR images steganalysis method based on depth confidence network | |
CN107563411B (en) | Online SAR target detection method based on deep learning | |
CN108872988A (en) | A kind of inverse synthetic aperture radar imaging method based on convolutional neural networks | |
CN108983187A (en) | Online radar target identification method based on EWC | |
CN108960330A (en) | Remote sensing images semanteme generation method based on fast area convolutional neural networks | |
CN104732243A (en) | SAR target identification method based on CNN | |
CN108008385A (en) | Interference environment ISAR high-resolution imaging methods based on management loading | |
CN104459668A (en) | Radar target recognition method based on deep learning network | |
CN109683161A (en) | A method of the inverse synthetic aperture radar imaging based on depth ADMM network | |
CN112052762A (en) | Small sample ISAR image target identification method based on Gaussian prototype | |
CN112965062B (en) | Radar range profile target recognition method based on LSTM-DAM network | |
CN108957418A (en) | A kind of radar target identification method based on Recognition with Recurrent Neural Network model | |
CN109507666A (en) | The sparse frequency band imaging method of ISAR based on off-network variation bayesian algorithm | |
CN109343060A (en) | ISAR imaging method and system based on deep learning time frequency analysis | |
CN108646247A (en) | Inverse synthetic aperture radar imaging method based on Gamma process linear regression | |
CN106680776A (en) | Low side-lobe wave form design method insensitive to Doppler information | |
CN115061126A (en) | Radar cluster target behavior identification method based on multi-dimensional parameter neural network | |
CN110082738A (en) | Radar target identification method based on Gaussian Mixture and tensor Recognition with Recurrent Neural Network | |
He et al. | Automatic recognition of ISAR images based on deep learning | |
CN108919228A (en) | one-dimensional radar data processing method and system | |
CN108387880A (en) | Detection method of small target under a kind of sea clutter background based on multiple dimensioned oriented Lyapunov indexes |
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 |