CN114428234A - Radar high-resolution range profile noise reduction identification method based on GAN and self-attention - Google Patents
Radar high-resolution range profile noise reduction identification method based on GAN and self-attention Download PDFInfo
- Publication number
- CN114428234A CN114428234A CN202111592204.5A CN202111592204A CN114428234A CN 114428234 A CN114428234 A CN 114428234A CN 202111592204 A CN202111592204 A CN 202111592204A CN 114428234 A CN114428234 A CN 114428234A
- Authority
- CN
- China
- Prior art keywords
- layer
- noise
- range profile
- resolution range
- hrrp
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 41
- 230000009467 reduction Effects 0.000 title claims abstract description 41
- 238000012360 testing method Methods 0.000 claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 27
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 22
- 238000002592 echocardiography Methods 0.000 claims description 21
- 238000000605 extraction Methods 0.000 claims description 10
- 238000010606 normalization Methods 0.000 claims description 10
- 210000002569 neuron Anatomy 0.000 claims description 9
- 239000013598 vector Substances 0.000 claims description 6
- 230000005484 gravity Effects 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000009827 uniform distribution Methods 0.000 claims description 2
- 238000007781 pre-processing Methods 0.000 claims 1
- 238000013528 artificial neural network Methods 0.000 description 9
- 238000004088 simulation Methods 0.000 description 9
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000007246 mechanism Effects 0.000 description 6
- 230000002457 bidirectional effect Effects 0.000 description 4
- 230000000306 recurrent effect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000011946 reduction process Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 210000004556 brain Anatomy 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 241000370001 Hantavirus Liu Species 0.000 description 1
- 230000003042 antagnostic effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000003920 cognitive function Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Images
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
-
- 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
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/88—Radar or analogous systems specially adapted for specific applications
- G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Remote Sensing (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Radar, Positioning & Navigation (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Networks & Wireless Communication (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Electromagnetism (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a high-resolution range profile noise reduction identification method based on GAN and self-attention, which mainly solves the problems that the characteristics with strong separability are taken as noise to be eliminated and long-distance dependence information cannot be extracted in the existing radar high-resolution range profile noise reduction identification technology. The implementation scheme is as follows: 1) acquiring a target radar echo, and generating a training sample set and a test sample set; 2) constructing a radar high-resolution range profile noise reduction identification network which is composed of a generator, a discriminator and an identifier and is based on GAN and self attention; 3) training the noise reduction recognition network constructed in the step 2) by using a training sample set and adopting a back propagation algorithm; 4) and inputting the test sample set into a trained radar high-resolution range profile noise reduction and identification network to obtain a radar high-resolution range profile noise reduction and identification result. The method improves the target recognition rate in the low signal-to-noise ratio environment, and can be used for recognizing the airplane and satellite targets in the low signal-to-noise ratio environment.
Description
Technical Field
The invention belongs to the technical field of target identification, and further relates to an end-to-end radar high-resolution range profile HRRP noise reduction and target identification method which can be used for realizing feature extraction and identification of targets such as airplanes and satellites under the environment with low signal-to-noise ratio.
Background
The HRRP is the vector sum amplitude waveform of the projection of a target scattering point sub-echo in the radar sight direction acquired by a broadband radar signal, and contains important information of target size, structure and scattering point distribution, and the HRRP identification is an important research direction in the field of RATR (automatic target identification) of radar. In an actual application scenario, the HRRP of the target contains noise, and the presence of the noise interferes with the feature extraction process, thereby affecting the target identification result.
For a noisy HRRP, the existing target identification method generally adopts the following three schemes:
1) the method is characterized in that identification is directly carried out, the HRRP containing noise is not specially processed, and the HRRP is input into a target identification model established aiming at the noise-free HRRP to give an identification result, and the method has simple steps, but low adaptability and poor identification performance;
2) extracting robust features, namely extracting noise robust features of HRRP (high resolution ratio), such as bispectrum features of a plurality of radar signals, scattering coefficients and positions of main scattering centers of targets and the like according to physical characteristics of the HRRP, and further designing a classifier according to the features to identify the targets, but the method needs manual design to extract the features, and is limited in performance when the environment changes;
3) the method comprises the steps of firstly reducing noise and then identifying, respectively constructing a noise reduction model and an identification model, firstly using the noise reduction model to reduce noise of the noise-containing HRRP to generate the HRRP with high signal-to-noise ratio, and then using the identification model to extract and identify the characteristics of the HRRP with high signal-to-noise ratio.
Zhao, x.he, j.liang, t.wang, c.huang, in its published paper "Radar HRRP target registration view semi-collaborative multi-task deep network" (IEEE Access, 2019), aiming at the problem of HRRP target identification in low snr environment, a semi-supervised multitask identification framework is proposed, which includes two models of DUBDNet and RSRNet, respectively realizing noise reduction and identification of low snr HRRP, inputting noisy HRRP into dunet, obtaining noise-reduced HRRP, and then using RSRNet to complete target identification. However, the method trains the noise reduction model first and then trains the recognition model, the two processes are not influenced by each other, and the problem that the recognition rate is influenced by eliminating part of the characteristics with strong separability as noise in the noise reduction process may exist.
Nie, Y.Xiao, L.Huang, F.Lv, in the published paper "Time-Frequency Analysis and Target registration of HRRP" (2021, compatibility), a Recognition method based on CN-LSGAN, STFT and CNN is proposed, CN-LSGAN is used for reducing noise of HRRP with low signal-to-noise ratio, STFT is used for transforming the HRRP after noise reduction to obtain a two-dimensional Time-Frequency diagram capable of representing information of Time domain and Frequency domain of a signal at the same Time, and CNN is used for carrying out feature extraction on the two-dimensional Time-Frequency diagram to further obtain a Target Recognition result. However, the method also processes the noise reduction and the identification separately, is not an end-to-end model, and still has the problem that part of the feature with strong separability is taken as noise elimination.
In recent years, deep learning is widely applied to automatic target recognition ATR based on HRRP, and such methods achieve excellent recognition performance by performing feature extraction through data driving, and have attracted much attention in radar target recognition research.
The existing feature extraction method based on deep learning can roughly adopt the following schemes:
1) self-encoder
The self-coding model AE is a neural network which reproduces input signals as much as possible, belongs to an unsupervised learning algorithm and consists of an encoder and a decoder, wherein the encoder and the decoder can be regarded as a feature extraction process, and the output of the encoder can be used as data features for subsequent tasks such as classification or identification. Such methods have limited ability to extract structured information of HRRP samples.
2) Convolutional neural network
The convolutional neural network CNN extracts local features and deep steady feature information in an original signal through a series of convolution and pooling operations, and has good generalization capability, but the CNN cannot represent the precedence relationship among data segments.
3) Recurrent neural networks
When modeling data with a time sequence relation, the recurrent neural network RNN can exert advantages, the internal structure of the RNN comprises recurrent connection of neurons to the RNN, and the RNN can realize memory and transmission of historical information, so that the RNN is widely applied to classification, prediction and generation tasks of sequence data. However, the method ignores the characteristic that the target region and the noise region in the HRRP sample have different influence degrees on the recognition result, and the recognition capability still needs to be improved.
4) Attention mechanism
Attention is a complex cognitive function in the human brain, meaning the ability of a person to focus on some information while choosing to ignore others. When a neural network is used for processing a large amount of input information, the attention mechanism of the human brain can be used for reference, and some key input information is focused to improve the efficiency of the neural network.
Xu, b.chen, j.wan, h.liu, l.jin, a paper "Target-aware associative network for radial HRRP Target recognition" (Signal Processing, 2019) published therein proposes an attention model based on RNN, the specific steps of the method are: firstly, each data segment of a single HRRP sample is coded by using RNN, time sequence correlation among distance units is extracted, then different weight coefficients are given to output at each time step according to the contribution degree of each data segment to recognition by using an attention mechanism, the output at each time step is weighted and summed by using the coefficients to serve as a separability characteristic, and finally the separability characteristic is connected with a softmax classifier to output a recognition result. The HRRP is modeled by the RNN, so that the long-distance dependence information cannot be extracted, and each data segment in the HRRP cannot be processed in parallel.
In a paper published in the radar academic report of 2019 by Liujia Qi, Chen Bohai and Ji xi, the attention machine system and the bidirectional GRU model-based radar HRRP target identification, the method combines the bidirectional GRU with the attention machine system, and comprises the following specific steps: the method comprises the steps of dividing an HRRP sample of a time domain into a positive sequence and a negative sequence through a sliding window, extracting features of the positive sequence and the negative sequence through two mutually independent GRU networks respectively, and splicing the extracted features at the same time, so that the target identification and classification are carried out by utilizing the bidirectional time sequence information of the HRRP and obtaining the hidden layer features after weighted summation through an attention mechanism. Although this method uses HRRP bidirectional timing information, it still has a problem that long-distance dependent information cannot be extracted.
Disclosure of Invention
The invention aims to provide a radar high-resolution range profile denoising and identifying method based on generation of an antagonistic network GAN and self-attention to overcome the defects of the prior art, so as to reconstruct a high signal-to-noise ratio HRRP and improve the identification performance of a target in a low signal-to-noise ratio environment by extracting remote dependence information in the HRRP.
The technical idea of the invention is as follows: generating a high signal-to-noise ratio HRRP through GAN, processing a one-dimensional high signal-to-noise ratio HRRP sample by using a sliding window method to generate a sequence sample, and obtaining a target identification result of the HRRP sample by constructing a self-attention-based deep neural network, wherein the implementation scheme comprises the following steps:
(1) dividing and processing radar echoes of three types of airplane targets to generate a training sample set S1And a test sample set S2;
(2) Constructing a radar high-resolution range profile noise reduction and identification network:
establishing a generator which is formed by sequentially connecting five convolution layers and five deconvolution layers, wherein each convolution layer is provided with a LeakyReLU operation, and each deconvolution layer is provided with a LeakyReLU and a jump connection operation and is used for denoising a noisy HRRP to generate a denoised high signal-to-noise ratio HRRP;
establishing a discriminator which is formed by sequentially connecting five convolution layers, one flattened layer and two full-connection layers, wherein a LeakyReLU operation is carried out after each convolution layer and is used for assisting a generator to improve the noise reduction performance;
establishing an identifier which comprises a sliding window layer, a position coding layer, three full-connection layers, five groups of feature extractors, a flatten layer and a softmax classifier, wherein the sliding window layer, the position coding layer, the third full-connection layer, the five groups of feature extractors, the flatten layer, the fourth full-connection layer, the fifth full-connection layer and the softmax classifier are sequentially connected and used for performing feature extraction on the high signal-to-noise ratio HRRP and giving a target identification result;
the generator is respectively connected with the discriminator and the recognizer to form a radar high-resolution range profile noise reduction recognition network based on a generated countermeasure network GAN and self attention;
(3) will train the sample set S1Inputting the data into the noise reduction recognition network constructed in the step (2), and performing countermeasure training on the data through a back propagation algorithm to obtain a trained radar high-resolution range profile noise reduction recognition network based on GAN and self attention;
(4) set of test samples S2And inputting the trained radar high-resolution range profile noise reduction recognition network based on the GAN and the self-attention to test to obtain a target recognition result output by the network.
Compared with the prior art, the invention has the following advantages:
firstly, the recognizer of the invention adopts position coding to represent the precedence relationship among HRRP data segments in a position coding layer, can fully utilize time sequence information in a sequence, and adopts a self-attention mechanism in a self-attention layer of a feature extractor to enable a model to pay more attention to a target area with strong distinguishability and fully extract the dependency relationship among the data segments, thereby overcoming the problem of losing remote dependency information in the prior art and improving the feature extraction capability and the recognition performance of the target.
Secondly, the HRRP with the low noise ratio is converted into the HRRP with the high noise ratio after noise reduction through the generator, and the HRRP with the high noise ratio is directly input into the recognizer to obtain the target category, so that the problem that end-to-end training cannot be realized due to step-by-step noise reduction and recognition in the prior art is solved.
Thirdly, the loss function of the recognizer is coupled with the loss function of the generator during training, so that the generator can keep the separability characteristic useful for the recognition process in the noise reduction process, the problem of recognition information loss caused by the noise reduction process in the prior art is effectively solved, and the target recognition performance in the low signal-to-noise ratio environment can be improved.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention.
Fig. 2 is a network framework diagram in the present invention.
Detailed Description
Embodiments and effects of the present invention will be further described below with reference to the accompanying drawings.
Referring to fig. 1, the implementation steps for this embodiment are as follows.
Step 1, generating a training sample set and a testing sample set.
1.1) respectively acquiring original radar echoes of three types of airplane targets, namely an An26 airplane, a trophy airplane and a Jack 42 airplane, sampling the echoes at intervals of 20 times from the first echo of each section of echo data, and respectively adding Gaussian white noise with a fixed signal-to-noise ratio to the sampled original radar echoes to obtain radar echoes containing noise of each type of targets;
1.2) respectively carrying out Fourier transform on each type of sampled target original radar echo and radar echo containing noise to obtain a radar high-resolution range profile HRRP without noise and a radar high-resolution range profile HRRP containing noise;
1.3) respectively carrying out 2 norm normalization and center-of-gravity alignment pretreatment on the HRRP without noise and the HRRP with noise obtained in the step 1.2);
1.4) selecting the fifth and sixth sections of radar echoes of an 26 airplane, the sixth and seventh sections of radar echoes of a trophy airplane, the second and fifth sections of radar echoes of a Jack 42 airplane and 1.3) corresponding to the six sections of radar echo data sections in total to serve as a training sample set S1(ii) a Taking the HRRP without noise and the HRRP with noise after 1.3) pretreatment corresponding to the rest radar echo data segments as a test sample set S2。
And 2, constructing a radar high-resolution range profile noise reduction identification network based on the GAN and the self attention.
2.1) building a generator: the method is characterized by comprising five convolution layers and five deconvolution layers which are sequentially connected, wherein each convolution layer is provided with a LeakyReLU operation, and each deconvolution layer is provided with a LeakyReLU and a jump connection operation and is used for denoising a noise-containing HRRP to generate a denoised high signal-to-noise ratio HRRP;
the generator has the following layer parameter settings:
the sizes of convolution kernels of the five convolution layers and the five deconvolution layers are all 15;
the number of convolution kernels of the first convolution layer and the fourth convolution layer is 16;
the number of convolution kernels of the second convolution layer, the third convolution layer, the second deconvolution layer and the third deconvolution layer is 32;
the number of convolution kernels of the fourth convolution layer, the fifth convolution layer and the first deconvolution layer is 64;
the number of convolution kernels of the fifth deconvolution layer is 1;
2.2) establishing a discriminator: the generator consists of five convolution layers, one flattened layer and two full-connection layers which are connected in sequence, and LeakyReLU operation is carried out after each convolution layer and is used for assisting the generator to improve the noise reduction performance;
the parameters of each layer of the discriminator are set as follows:
the sizes of convolution kernels of the five convolution layers are all 15;
the number of convolution kernels of the sixth convolution layer is 16;
the number of convolution kernels of the seventh convolution layer and the eighth convolution layer is 32;
the number of convolution kernels of the ninth convolution layer and the tenth convolution layer is 64;
the number of neurons in the first fully-connected layer is 8;
the number of neurons in the second fully-connected layer is 1;
2.3) establishing a recognizer: the device comprises a sliding window layer, a position coding layer, three full-connection layers, five groups of feature extractors, a flatten layer and a softmax classifier, wherein the sliding window layer, the position coding layer, the third full-connection layer, the five groups of feature extractors, the flatten layer, the fourth full-connection layer, the fifth full-connection layer and the softmax classifier are sequentially connected and are used for performing feature extraction on the high signal-to-noise ratio HRRP and giving a target identification result;
the function and structure parameters of each layer of the recognizer are as follows:
the sliding window layer is used for generating high signal-to-noise ratio HRRP output by the generator according to window length d equal to 6 and step lengthPerforming sliding window, and converting the sequence into a sequence form HRRP with the sequence length of T-84 as output;
the position coding layer is used for generating T-84 position codes by using a sine function, and the sine function is expressed as follows:
PE(pos,j)=cos(pos/102j/T)
wherein pos belongs to [1, T ] represents the sequence number of the current calculated position code in the sequence, j belongs to [1, d ] represents the j element of the current calculated position code, and the result of adding the T position codes and the output of the sliding window layer according to the corresponding items is the output of the position code layer;
the five groups of feature extractors are used for extracting the dependency relationship among sequence data, and each group of feature extractors consists of a self-attention layer, a residual connecting layer, a normalization layer, a feedforward layer, a residual connecting layer and a normalization layer which are sequentially cascaded;
the input from the attention layer isFor the first set of feature extractors, X represents the output of the position-coding layer, and for the other four sets of feature extractors, X represents the output of the previous set of feature extractors, the output from the attention layer beingIn the formula: wherein, are all learnable parameter matrices, set dmodel=dk=dv=128;
The feedforward layer comprises two full-connection layers, and the number of the neurons of the feedforward layer is 512 and 128 respectively;
the number of the neurons of the third, fourth and fifth full connecting layers is 128, 3072 and 3 respectively;
and 2.4) connecting the generator with the discriminator and the recognizer respectively to form a radar high-resolution range profile noise reduction recognition network based on the GAN and the self-attention.
And 3, training a radar high-resolution distance image noise reduction identification network based on the GAN and the self attention.
Referring to fig. 2, the specific implementation of this step is as follows:
3.1), initializing all learnable parameters in the high-resolution range profile noise reduction identification network based on the GAN and the self-attention radar;
3.2) will train the sample set S1Medium-noise high-resolution range profile x1Inputting the data into a generator, calculating layer by layer along the network structure, and generating a high-resolution range profile x after noise reduction2;
3.3) will train the sample set S1Medium-noise-free high-resolution range profile x3And a noisy high resolution range profile x1Connecting to obtain a true sample r ═ x3,x1]The noise-reduced high-resolution range profile x generated by the generator in 3.2)2And noise-containing high-resolution range profile x1Connecting to obtain a false sample f ═ x2,x1]Inputting the true sample r and the false sample f into a discriminator, and calculating layer by layer along the network structure to obtain the output of the discriminator, namely the true and false samples;
3.4) de-noising high resolution produced by the generator of 3.2)Distance image x2Inputting a recognizer, and calculating layer by layer along a network structure to obtain the output of the recognizer, namely the target category;
3.5) setting the loss function L of the generatorG(D,G):
Wherein,representing a desired operation, D representing a discriminator, G representing a generator, G (x)1) Representing a de-noised high-resolution range profile x2And λ is the coefficient of the regularization term,represents L1A term of regularization is used to normalize the,represents L2Regularization term, α being L1Regularization term and L2The scale factor of the regularization term, beta is the recognition penalty factor,representing recognizer loss function, tkThe kth element, y, in the true class vector representing the targetkRepresenting the kth element in the output vector of the recognizer, wherein K represents the number of target types;
3.6) setting the loss function L of the discriminatorD(D,G):
Wherein L isGPDenotes a gradient penalty term, λGPA coefficient representing a penalty term for the gradient,it is indicated that the operation of the gradient,is a point on a straight line, and ε is a number from [0,1 ]]Randomly chosen numbers in the uniform distribution of (a);
3.7) pairs of learnable parameters w in the arbiterDUpdating, wherein the updating formula is as follows:
wherein,is a parameter obtained by the discriminator after the current update,is a parameter before updating of the discriminator, ηDIs the learning rate of the discriminator,is a loss function LD(D, G) pairA gradient of (a);
3.8) parameters updated with the arbiterThe calculation process of 3.2) to 3.5) is carried out again and the parameter w learnable in the generator is generatedGAnd a parameter w learnable in the recognizerRUpdating, wherein the updating formula is as follows:
wherein,is a parameter that the generator gets after it is currently updated,is the parameter that the recognizer gets after it is currently updated,is a parameter of the generator before it is updated,is a parameter, η, before updating the recognizerGIs the learning rate of the generator and recognizer,is a loss function LG(D, G) pairThe gradient of (a) of (b) is,is a loss function LG(D, G) pairA gradient of (a);
3.9) updated parameters Using the GeneratorAnd identifier updated parametersRepeating 3.2) to 3.7) again, and performing multiple iterative updates until the loss function L is reachedrecAfter stable convergence, stop iterationObtaining a learnable parameter w in the generatorGLearnable parameter w in discriminatorDLearnable parameter w in recognizerRAnd obtaining a trained radar high-resolution range profile noise reduction recognition network based on the GAN and the self attention.
Step 4, testing the sample set S2And inputting the trained radar high-resolution range profile noise reduction recognition network based on the GAN and the self-attention to test to obtain a recognition result output by the network.
The effects of the present invention can be illustrated by the following simulation experiments.
1. Simulation experiment conditions are as follows:
the data used in the simulation experiment of the invention is the data of the actual measurement plane of a certain research institute radar in China, the central frequency of the radar is about 5.5GHz, the signal bandwidth is 400MHz, the measured data comprises 3 types of target planes which are respectively a medium-sized propeller plane 'an 26', a small-sized jet plane 'prize-shaped' and a medium-sized jet plane 'Jack 42'. The measured data is divided into different data segments according to the flight path of the airplane, wherein the Ann 26 and the prize are divided into 7 segments respectively, and the Jack 42 is divided into 5 segments. The simulation experiment starts from the first echo of each section of data, samples are carried out at intervals of 20 echoes, and finally 2600 echoes of 5 th and 6 th sections of 'ann 26', '2600 echoes of 6 th and 7 th sections of' prize-shaped ',' 2198 th echoes of 2 nd and 5 th sections of 'Yake 42' are selected as training data; 6256 echoes of the 1 st to 4 th and 7 th segments of "An 26", 6500 echoes of the 1 st to 5 th segments of "medal", 3900 echoes of the 1 st and 3 th to 4 th segments of "Jack 42" are used as test data. All echoes are 1 x 256 vectors.
Data set 1: fourier transformation, 2 norm normalization and gravity center alignment are carried out on training data and testing data to obtain noise-free high-resolution range profiles (HRRPs) which are respectively used as training sample sets D1And test sample set D2;
Data set 2: gaussian white noise with the signal-to-noise ratio of 15dB is added into the training data and the test data, Fourier transform, 2 norm normalization and gravity center alignment are carried out on the data after the noise is added, a high-resolution range profile HRRP containing the noise under the environment of 15dB is obtained,respectively with training sample set D1And test sample set D2Taking the HRRP combination without noise as a training sample set D3And test sample set D4;
Data set 3: gaussian white noise with the signal-to-noise ratio of 10dB is added into the training data and the test data, Fourier transform, 2 norm normalization and gravity center alignment are carried out on the data after the noise is added, and a high-resolution range profile HRRP containing the noise under the environment of 10dB is obtained and is respectively compared with a training sample set D1And test sample set D2Taking the HRRP combination without noise as a training sample set D5And test sample set D6;
Data set 4: gaussian white noise with the signal-to-noise ratio of 5dB is added into the training data and the test data, Fourier transform, 2 norm normalization and gravity center alignment are carried out on the data after the noise is added, and a high-resolution range profile HRRP containing the noise in the 5dB environment is obtained and is respectively connected with a training sample set D1And test sample set D2Taking the HRRP combination without noise as a training sample set D7And test sample set D8。
The simulation experiment hardware platform is Intel Xeon [email protected] CPU, 64GB RAM, NVIDIA Geforce GTX1080 Ti GPU, and the simulation experiment software platform is Matlab2016, Python 3.6 and Tensorflow 1.8.
2. Simulation experiment content and result analysis:
simulation experiment 1: using the data set 1, the prior attention-cycle neural network method and the identifier of the present invention are applied to target identification of noise-free HRRP, respectively.
Simulation experiment 2: and respectively applying the existing attention recurrent neural network method, the identifier in the invention and the method of the invention to respectively identify the target of the noisy HRRP under the environments of signal-to-noise ratios of 15dB, 10dB and 5dB by using the data set 2, the data set 3 and the data set 4.
The recognition accuracy of the three methods under four environments with no noise and signal-to-noise ratios of 15dB, 10dB and 5dB is respectively calculated by the following formulas:
wherein, a represents the identification accuracy of the test sample set, N represents the sample number of the test sample set, h (-) represents the identification discrimination function, t(i)Representing the true class, y, of the ith test sample in the set of test samples(i)The network output result corresponding to the ith test sample in the test sample set is shown, when t is(i)And y(i)When equal, h (t)(i),y(i)) Equal to 1, otherwise, h (t)(i),y(i)) Equal to 0.
The specific calculation results are shown in table 1:
TABLE 1 comparison of recognition rates for three methods in 5dB, 10dB, 15dB SNR and no noise environment
From the simulation results, the following conclusions can be drawn:
compared with the attention cycle neural network method in the prior art, the recognition accuracy of the recognizer provided by the invention is improved by 2.72% in a noise-free environment, which shows that the recognition accuracy of the radar high-resolution range profile can be improved by using a self-attention mechanism to extract features.
Compared with the method, the identification accuracy of the identifier is respectively improved by 1.77%, 2.12% and 1.88% in the environments of 5dB, 10dB and 15dB signal-to-noise ratios, which shows that the noise reduction module based on the GAN can effectively improve the identification rate of the network in the environment of low signal-to-noise ratio.
Claims (6)
1. The radar high-resolution range profile denoising identification method based on the GAN and the self attention is characterized by comprising the following steps:
(1) dividing and processing radar echoes of three types of airplane targets,generating a training sample set S1And a test sample set S2;
(2) Constructing a radar high-resolution range profile noise reduction and identification network:
establishing a generator which is formed by sequentially connecting five convolution layers and five deconvolution layers, wherein each convolution layer is provided with a LeakyReLU operation, and each deconvolution layer is provided with a LeakyReLU and a jump connection operation and is used for denoising a noisy HRRP to generate a denoised high signal-to-noise ratio HRRP;
establishing a discriminator which is formed by sequentially connecting five convolution layers, one flattened layer and two full-connection layers, wherein a LeakyReLU operation is carried out after each convolution layer and is used for assisting a generator to improve the noise reduction performance;
establishing an identifier which comprises a sliding window layer, a position coding layer, three full-connection layers, five groups of feature extractors, a flatten layer and a softmax classifier, wherein the sliding window layer, the position coding layer, the third full-connection layer, the five groups of feature extractors, the flatten layer, the fourth full-connection layer, the fifth full-connection layer and the softmax classifier are sequentially connected and used for performing feature extraction on the high signal-to-noise ratio HRRP and giving a target identification result;
the generator is respectively connected with the discriminator and the recognizer to form a radar high-resolution range profile noise reduction recognition network based on a generated countermeasure network GAN and self attention;
(3) will train the sample set S1Inputting the data into the noise reduction recognition network constructed in the step (2), and performing countermeasure training on the data through a back propagation algorithm to obtain a trained radar high-resolution range profile noise reduction recognition network based on GAN and self attention;
(4) set of test samples S2And inputting the trained radar high-resolution range profile noise reduction recognition network based on the GAN and the self-attention to test to obtain a target recognition result output by the network.
2. The method of claim 1, wherein the radar returns of three types of aircraft targets are processed and partitioned in (1) to generate a training sample set S1And a test sample set S2The implementation is as follows:
respectively acquiring original radar echoes of three types of airplane targets, namely an An26 airplane, a trophy airplane and an Acke 42 airplane, and respectively adding Gaussian white noise with a fixed signal-to-noise ratio to the original radar echoes to obtain radar echoes of each type containing noise;
respectively carrying out Fourier transform on each type of original radar echo and each type of noise-containing radar echo to obtain a noise-free radar high-resolution range profile HRRP and a noise-containing radar high-resolution range profile HRRP, and respectively carrying out 2-norm normalization and gravity center alignment preprocessing on the noise-free HRRP and the noise-containing HRRP;
according to the three types of airplane tracks, the 5 th and 6 th sections of an '26' airplane, the 6 th and 7 th sections of a 'prize-shaped' airplane and the 2 nd and 5 th sections of an 'Jack 42' airplane are selected, and the HRRP which is not noisy and the HRRP which is noisy after being preprocessed and corresponds to the six sections of radar echoes are used as a training sample set S1(ii) a Taking the preprocessed HRRP without noise and the preprocessed HRRP with noise corresponding to the rest radar echo data segments as a test sample set S2。
3. The method of claim 1, wherein the parameters of each layer of the generator in (2) are set as follows:
the sizes of convolution kernels of the five convolution layers and the five deconvolution layers are all 15;
the number of convolution kernels of the first convolution layer and the fourth convolution layer is 16;
the number of convolution kernels of the second convolution layer, the third convolution layer, the second deconvolution layer and the third deconvolution layer is 32;
the number of convolution kernels of the fourth convolution layer, the fifth convolution layer and the first deconvolution layer is 64;
the number of convolution kernels of the fifth deconvolution layer is 1.
4. The method according to claim 1, wherein the parameters of each layer of the discriminator in (2) are set as follows:
the sizes of convolution kernels of the five convolution layers are all 15;
the number of convolution kernels of the sixth convolution layer is 16;
the number of convolution kernels of the seventh convolution layer and the eighth convolution layer is 32;
the number of convolution kernels of the ninth convolution layer and the tenth convolution layer is 64;
the number of neurons in the first fully-connected layer is 8;
the number of neurons in the second fully-connected layer was 1.
5. The method of claim 1, wherein the role and structure parameters of each layer of the identifier in (2) are as follows:
the sliding window layer is used for generating high signal-to-noise ratio HRRP output by the generator according to window length d equal to 6 and step lengthPerforming sliding window, and converting the sequence into a sequence form HRRP with the sequence length of T-84 as output;
the position coding layer is used for generating 84 position codes by using a sine function, wherein the sine function is expressed as follows:
PE(pos,j)=cos(pos/102j/T)
wherein pos belongs to [1, T ] represents the sequence number of the current calculated position code in the sequence, j belongs to [1, d ] represents the j element of the current calculated position code, and the result of adding the T position codes and the output of the sliding window layer according to the corresponding items is the output of the position code layer;
the five groups of feature extractors are used for extracting the dependency relationship among the sequence data, and each group of feature extractor consists of a self-attention layer, a residual connecting layer, a normalization layer, a feedforward layer, a residual connecting layer and a normalization layer which are sequentially cascaded;
the input from the attention layer isFor the first set of feature extractors, X represents the output of the position-coding layer, and for the other four sets of feature extractors, X represents the output of the previous set of feature extractors, the output from the attention layer beingIn the formula: wherein, are all learnable parameter matrices, set dmodel=dk=dv=128;
The feedforward layer comprises two full-connection layers, and the number of the neurons of the feedforward layer is 512 and 128 respectively;
the number of the neurons of the third, fourth and fifth full connecting layers is 128, 3072 and 3 respectively.
6. The method according to claim 1, wherein in (3), the GAN and self-attention based radar high-resolution range profile noise reduction recognition network is subjected to countermeasure training through a back propagation algorithm, and the following is realized:
(6a) initializing all learnable parameters in the radar high-resolution range profile noise reduction identification network based on the GAN and the self attention;
(6b) will train the sample set S1Medium-noise high-resolution range profile x1Inputting the data into a generator, calculating layer by layer along the network structure, and generating a high-resolution range profile x after noise reduction2;
(6c) Will train the sample set S1Medium-noise-free high-resolution range profile x3And noise-containing high-resolution range profile x1Connecting to obtain a true sample r ═ x3,x1]The high-resolution range profile x after noise reduction in (6b)2And noise-containing high-resolution range profile x1Connecting to obtain a false sample f ═ x2,x1]Inputting the true samples r and the false samples f into a discriminator, and gradually inputting the true samples r and the false samples f into the discriminator along the network structureLayer calculation to obtain the output of the discriminator;
(6d) de-noising the high-resolution range profile x of (6b)2Inputting a recognizer, and calculating layer by layer along a network structure to obtain the output of the recognizer, namely the target category;
(6e) setting the loss function L of the generatorG(D,G):
Wherein,representing a desired operation, D representing a discriminator, G representing a generator, a is a regularization term coefficient,represents L1A term of regularization is used to normalize the,represents L2Regularization term, α being L1Regularization term and L2The scale factor of the regularization term, beta is the recognition penalty factor,representing recognizer loss function, tkThe kth element, y, in the true class vector representing the targetkRepresenting the kth element in the output vector of the recognizer, wherein K represents the number of target types;
(6f) setting a loss function L of a discriminatorD(D,G):
Wherein L isGPDenotes a gradient penalty term, λGPA coefficient representing a penalty term for the gradient,is a point on a straight line, and ε is a number from [0,1 ]]Randomly chosen numbers in the uniform distribution of (a);
(6g) for learnable parameter w in the discriminatorDUpdating, wherein the updating formula is as follows:
wherein,is a parameter obtained by the discriminator after the current update,is a parameter before updating of the discriminator, ηDIs the learning rate of the discriminator,is a loss function LD(D, G) pairA gradient of (a);
(6h) updated parameters using discriminatorsThe calculation processes from (6b) to (6e) are carried out again, and the parameter w which can be learned in the generator is regeneratedGAnd a parameter w learnable in the recognizerRUpdating, wherein the updating formula is as follows:
wherein,is a parameter that the generator gets after it is currently updated,is the parameter that the recognizer gets after it is currently updated,is a parameter of the generator before it is updated,is a parameter, η, before updating the recognizerGIs the learning rate of the generator and recognizer,is a loss function LG(D, G) pairThe gradient of (a) of (b) is,is a loss function LG(D, G) pairA gradient of (a);
(6i) using updated parameters of a generatorAnd identifier updated parametersRepeating (6b) to (6g) for multiple times of iterative updating until the loss function L is reachedregAfter stable convergence, stopping iteration to obtain a learnable parameter w in the generatorGLearnable parameter w in discriminatorDLearnable parameter w in recognizerRAnd obtaining a trained radar high-resolution range profile noise reduction recognition network based on the GAN and the self attention.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111592204.5A CN114428234A (en) | 2021-12-23 | 2021-12-23 | Radar high-resolution range profile noise reduction identification method based on GAN and self-attention |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111592204.5A CN114428234A (en) | 2021-12-23 | 2021-12-23 | Radar high-resolution range profile noise reduction identification method based on GAN and self-attention |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114428234A true CN114428234A (en) | 2022-05-03 |
Family
ID=81311029
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111592204.5A Pending CN114428234A (en) | 2021-12-23 | 2021-12-23 | Radar high-resolution range profile noise reduction identification method based on GAN and self-attention |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114428234A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114943286A (en) * | 2022-05-20 | 2022-08-26 | 电子科技大学 | Unknown target discrimination method based on fusion of time domain features and space domain features |
CN115130750A (en) * | 2022-06-28 | 2022-09-30 | 北京信息科技大学 | Flight path prediction method and device, electronic equipment and readable storage medium |
CN115201768A (en) * | 2022-06-10 | 2022-10-18 | 西安电子科技大学 | Active deception jamming method for generating countermeasure network based on cycle consistency |
CN116250844A (en) * | 2023-03-03 | 2023-06-13 | 山东大学 | Electrocardiosignal noise reduction optimization method and system based on condition generation countermeasure network |
CN117272138A (en) * | 2023-09-15 | 2023-12-22 | 东华理工大学 | Geomagnetic data denoising method and system based on reference channel data constraint and deep learning |
-
2021
- 2021-12-23 CN CN202111592204.5A patent/CN114428234A/en active Pending
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114943286A (en) * | 2022-05-20 | 2022-08-26 | 电子科技大学 | Unknown target discrimination method based on fusion of time domain features and space domain features |
CN114943286B (en) * | 2022-05-20 | 2023-04-07 | 电子科技大学 | Unknown target discrimination method based on fusion of time domain features and space domain features |
CN115201768A (en) * | 2022-06-10 | 2022-10-18 | 西安电子科技大学 | Active deception jamming method for generating countermeasure network based on cycle consistency |
CN115130750A (en) * | 2022-06-28 | 2022-09-30 | 北京信息科技大学 | Flight path prediction method and device, electronic equipment and readable storage medium |
CN116250844A (en) * | 2023-03-03 | 2023-06-13 | 山东大学 | Electrocardiosignal noise reduction optimization method and system based on condition generation countermeasure network |
CN116250844B (en) * | 2023-03-03 | 2024-04-26 | 山东大学 | Electrocardiosignal noise reduction optimization method and system based on condition generation countermeasure network |
CN117272138A (en) * | 2023-09-15 | 2023-12-22 | 东华理工大学 | Geomagnetic data denoising method and system based on reference channel data constraint and deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109522857B (en) | People number estimation method based on generation type confrontation network model | |
CN114428234A (en) | Radar high-resolution range profile noise reduction identification method based on GAN and self-attention | |
CN109683161B (en) | Inverse synthetic aperture radar imaging method based on depth ADMM network | |
CN107220606B (en) | Radar radiation source signal identification method based on one-dimensional convolutional neural network | |
CN111626961B (en) | Radar image clutter suppression method and system based on generation countermeasure network | |
CN112001306A (en) | Electroencephalogram signal decoding method for generating neural network based on deep convolution countermeasure | |
CN112364779A (en) | Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion | |
CN109870729B (en) | Deep neural network magnetic resonance signal denoising method based on discrete cosine transform | |
CN105488466B (en) | A kind of deep-neural-network and Acoustic Object vocal print feature extracting method | |
CN109559736A (en) | A kind of film performer's automatic dubbing method based on confrontation network | |
CN111723701B (en) | Underwater target identification method | |
CN110175560A (en) | A kind of radar signal intra-pulse modulation recognition methods | |
CN113376600B (en) | Pedestrian radar echo denoising method based on RSDNet | |
CN104459668A (en) | Radar target recognition method based on deep learning network | |
CN111160176A (en) | Fusion feature-based ground radar target classification method for one-dimensional convolutional neural network | |
CN112287784B (en) | Radar signal classification method based on deep convolutional neural network and feature fusion | |
Alnujaim et al. | Generative adversarial networks to augment micro-doppler signatures for the classification of human activity | |
Qin et al. | Radar waveform recognition based on deep residual network | |
CN114863938A (en) | Bird language identification method and system based on attention residual error and feature fusion | |
CN113111786A (en) | Underwater target identification method based on small sample training image convolutional network | |
CN116471154A (en) | Modulation signal identification method based on multi-domain mixed attention | |
CN109871907B (en) | Radar target high-resolution range profile identification method based on SAE-HMM model | |
Yang et al. | GAN-based sample expansion for underwater acoustic signal | |
CN112835008B (en) | High-resolution range profile target identification method based on attitude self-adaptive convolutional network | |
Jing et al. | The optimisation of speech recognition based on convolutional neural network |
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 |