CN112731330B - Radar carrier frequency parameter change steady target identification method based on transfer learning - Google Patents

Radar carrier frequency parameter change steady target identification method based on transfer learning Download PDF

Info

Publication number
CN112731330B
CN112731330B CN202110032917.XA CN202110032917A CN112731330B CN 112731330 B CN112731330 B CN 112731330B CN 202110032917 A CN202110032917 A CN 202110032917A CN 112731330 B CN112731330 B CN 112731330B
Authority
CN
China
Prior art keywords
neural network
echo signal
layer
radar
signal
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.)
Active
Application number
CN202110032917.XA
Other languages
Chinese (zh)
Other versions
CN112731330A (en
Inventor
王鹏辉
刘宏伟
孙嘉琪
丁军
邓心慰
陈渤
徐一兼
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN202110032917.XA priority Critical patent/CN112731330B/en
Publication of CN112731330A publication Critical patent/CN112731330A/en
Application granted granted Critical
Publication of CN112731330B publication Critical patent/CN112731330B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Probability & Statistics with Applications (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The invention discloses a method for identifying a radar carrier frequency parameter change steady target based on transfer learning, which mainly solves the problem of identifying a moving target in real time under a radar carrier frequency parameter change scene. The method comprises the following implementation steps: (1) generating a training set; (2) generating a fine tuning set; (3) constructing a self-encoder neural network; (4) constructing a feedforward neural network; (5) training a self-encoder neural network; (6) training a feedforward neural network; (7) and (4) target identification. The method and the device utilize the idea of transfer learning to transfer the knowledge learned by radar target identification under specific carrier frequency to the problem of radar target identification learning under new carrier frequency, and utilize a small amount of fine tuning data sets under new carrier frequency to update the model in real time, thereby improving the real-time property and the identification accuracy of target identification.

Description

Radar carrier frequency parameter change steady target identification method based on transfer learning
Technical Field
The invention belongs to the technical field of radars, and further relates to a radar carrier frequency parameter change steady target identification method based on transfer learning in the technical field of radar target classification. The invention can identify moving targets such as air and ground in real time under the scene of radar carrier frequency parameter change.
Background
The radar target identification means that the type of a target is judged by using a radar echo signal of a moving target. At present, the micro doppler feature based on the target echo signal is one of the important ways for the conventional narrow-band radar to realize target identification. The extraction of the micro Doppler features mainly comprises the steps of extracting stable features capable of reflecting the self structure information of the target from the micro Doppler modulation characteristics in radar echoes by analyzing the micro Doppler modulation characteristics, obtaining the estimation of parameters such as the size, the motion state, the attribute and the like of the target by analyzing the features, and providing a basis for target classification or identification. Most of the classification methods based on micro-doppler features proposed in the existing documents have a precondition that the radar operating parameters in the training samples and the test samples are kept unchanged. And under the radar actual work condition, in order to improve the anti-interference performance of the radar system and solve the problems of blind speed, range ambiguity and the like, the radar carrier frequency parameter and the pulse repetition frequency parameter can be changed, which means that the radar working parameters in the training sample and the test sample are not matched.
The university of people's liberation force information engineering in China proposed a target identification feature extraction method in the patent document "target identification feature extraction method based on radar time domain echo" (patent application No. CN201510458912.8, application publication No. CN 105093199A). The method comprises the following specific steps: firstly, preprocessing work such as mean value removal, energy normalization and the like is carried out on a target echo signal; secondly, intra-pulse motion compensation is carried out on the high-speed moving target; filtering out non-structural attributes such as target speed, distance and the like coupled in radar echoes; and thirdly, fully utilizing the amplitude-phase information of the filtered signals, and extracting a product spectrum of the signals as the identification characteristics of the target. The method fully utilizes the amplitude-phase information of the signal, takes the product spectrum of the signal as the identification characteristic of the target, and can realize the individual identification of the target by the narrow-band radar because the echo product spectrum characteristic is not subjected to distance compression and the information quantity is not lost. However, the method still has the following defects: the method adopts the product spectrum characteristic of an extracted signal to identify a target, the product spectrum characteristic is an artificially extracted characteristic, the method for artificially extracting the characteristic relies on artificial design, the product spectrum characteristic of the signal is only stable to parameters such as target speed distance and the like, the problem of model mismatch still occurs when the target is identified under a new carrier frequency, the conventional method for solving the model mismatch is to establish a sample library under various radar carrier frequency parameters and respectively train corresponding classification templates, and select a proper classification template according to the echo signal carrier frequency of a test sample, but because the cost for developing a radar recorded data experiment is higher, the sample data amount under the new carrier frequency is less, and because the method for establishing the template library has poor flexibility, the template library can not be updated in real time according to the change of the carrier frequency parameters, so that the real-time identification of the targets such as air, ground and the like can not be realized.
The university of Harbin engineering Master thesis 2019, Zhang Tianyi et al, in the published paper, "narrow-band radar target classification system design based on micro-Doppler", provides a method for identifying time-frequency image features based on micro-Doppler features. The method comprises the following specific steps: the method comprises the steps of firstly, processing a narrow-band radar echo signal by using a short-time Fourier transform method to obtain a time-frequency image; secondly, constructing and training a CNN-LeNet-5 network; and thirdly, processing the time-frequency image, inputting the processed time-frequency image into a CNN-LeNet-5 network, and reducing the overfitting condition of the network by using a Dropout algorithm. Compared with the traditional artificial feature extraction method, the method avoids artificial subjective intervention, realizes the autonomous learning and identification of the features, and has better stability and higher identification accuracy. However, the method still has the following defects: although the method can automatically select the target characteristics by utilizing the CNN-LeNet-5 network and reduce the construction cost of the template base, the characteristics extracted under different radar carrier frequency parameters are the same, and the change of the radar carrier frequency parameters can influence the micro-Doppler modulation effect of the target, so that the dispersion condition of the original characteristics in a characteristic space is changed, and the identification accuracy of the method is greatly reduced.
Disclosure of Invention
The invention aims to provide a radar carrier frequency parameter change steady target identification method based on transfer learning aiming at the defects in the prior art, which is used for solving the problems that the method for establishing a template library cannot realize real-time identification of targets such as air, ground and the like due to higher experimental cost and poor real-time performance of the method for testing the echo signal carrier frequency parameter change caused model mismatch in the prior art; and the distribution of the original characteristics in the characteristic space is changed due to the change of the carrier frequency parameters, so that the identification rate is reduced.
The idea of realizing the purpose of the invention is that the invention utilizes the idea of transfer learning to transfer the knowledge learned by radar target identification under specific carrier frequency to the problem of radar target identification learning under new carrier frequency, and can utilize a small amount of fine tuning data sets under new carrier frequency to update the model in real time, thereby realizing real-time identification of targets in the air, on the ground and the like. The invention constructs the fine tuning set and the feedforward neural network which are formed by the narrow-band radar echo signals under various carrier frequencies with less sample numbers, can utilize the fine tuning set to perform fine tuning on the feedforward neural network, and can adaptively extract the carrier frequency steady characteristics from the radar echo information under the condition of different carrier frequency parameter changes, so that the characteristics which can be extracted by the invention have separability, and the identification accuracy is improved.
The method comprises the following concrete implementation steps:
(1) generating a training set:
(1a) extracting radar echo signals with data dimension L under a single carrier frequency and containing N category targets as a training data set, wherein each category at least contains 800 radar echo signals, N is more than or equal to 3, and L is more than or equal to 64;
(1b) performing clutter suppression on each echo signal in the training data set by using a regional CLEAN method;
(1c) removing the main component in each echo signal after clutter suppression by using a global CLEAN method;
(1d) carrying out modular two-norm normalization processing on the amplitude of each echo signal with the main component removed to obtain a training set;
(2) generating a fine tuning set:
(2a) radar echo signals with data dimension L under M carrier frequencies containing N types of targets are extracted to form a fine adjustment data set, each type at least contains 200 radar echo signals, wherein N is more than or equal to 3, M is more than or equal to 5, and L is more than or equal to 64;
(2b) performing clutter suppression on each echo signal in the fine tuning data set by using a regional clear method;
(2c) removing the main component in each echo signal after clutter suppression by using a global CLEAN method;
(2d) carrying out modular two-norm normalization processing on the amplitude of each echo signal with the main component removed to obtain a fine tuning set;
(3) constructing a self-encoder neural network:
(3a) a seven-layer self-encoder neural network is built, and the structure of the self-encoder neural network is as follows in sequence: the decoding device comprises an input layer, a first coding layer, a second coding layer, a third coding layer, a first decoding layer, a second decoding layer and a third decoding layer;
(3b) the parameters of each layer are set as follows: setting the numbers of the neurons of the first to third coding layers to 500, 256 and 64, respectively; setting the number of neurons of the first decoding layer, the second decoding layer and the third decoding layer as 64, 256 and T respectively, wherein the number of the neurons T is equal to the data dimension L of the radar echo signal;
(4) constructing a feedforward neural network:
(4a) a six-layer feedforward neural network is built, and the structure of the feedforward neural network is as follows in sequence: the system comprises an input layer, a first coding layer, a second coding layer, a third coding layer, a full connection layer and a SoftMax layer;
(4b) the parameters of each layer are set as follows: setting the numbers of the neurons of the first to third coding layers to 500, 256 and 64, respectively; setting the number of the neurons of the full connection layer as M, wherein the number M of the neurons of the full connection layer is equal to the total number N of the target categories; the SoftMax layer adopts a SoftMax activation function to calculate the probability that the input echo signals are classified into each class;
(5) training a self-encoder neural network:
(5a) randomly extracting P radar echo signals in a training set, inputting the P radar echo signals into a self-encoder neural network, and outputting reconstructed signals of the P radar echo signals, wherein P represents the number of samples input into the self-encoder neural network, and P is more than or equal to 64;
(5b) calculating a loss value between a radar echo signal currently input into the neural network of the self-encoder and an output reconstruction signal by using a reconstruction loss function, and iteratively updating network parameters by using a back propagation algorithm;
(5c) judging whether the reconstruction loss function is converged, if so, executing the step (6) after obtaining the trained self-encoder neural network, otherwise, continuously and randomly extracting P radar echo signals in the training set and inputting the P radar echo signals into the self-encoder neural network and then executing the step (5 b);
(6) training a feedforward neural network:
(6a) assigning the weight parameters of the input layer, the first coding layer, the second coding layer and the third coding layer of the trained self-encoder neural network to the weight parameters of the input layer, the first coding layer, the second coding layer and the third coding layer in the feedforward neural network, and generating the weight parameters of the full-connection layer in the feedforward neural network by adopting a Gaussian initialization method;
(6b) randomly extracting Q radar echo signals in the fine tuning set, inputting the Q radar echo signals into a feedforward neural network, and outputting a prediction classification label of each signal in the Q radar echo signals, wherein Q represents the number of samples input into the feedforward neural network, and Q is more than or equal to 64;
(6c) calculating a loss value between a prediction classification label and a real classification label of each echo signal currently input into the feedforward neural network by using a cross entropy loss function;
(6d) fixing the weight parameters of an input layer, a first coding layer, a second coding layer and a third coding layer in the feedforward neural network to be unchanged, and iteratively updating the weight parameters of the full-connection layer of the feedforward neural network by using a back propagation algorithm;
(6e) judging whether the cross entropy loss function of the current iteration is converged, if so, executing the step (7) after obtaining the trained feedforward neural network, otherwise, continuously randomly extracting Q radar echo signals from the fine tuning set and inputting the Q radar echo signals into the feedforward neural network and then executing the step (6 c);
(7) target identification:
(7a) performing clutter suppression on each echo signal to be identified by using a regional clear method;
(7b) removing the main component of each echo signal subjected to clutter suppression by utilizing a global CLEAN method;
(7c) and inputting each echo signal with the main component removed into a trained feedforward neural network, calculating the probability of dividing the target to be identified into various types through a SoftMax layer, and selecting the type corresponding to the highest probability as an identification result.
Compared with the prior art, the invention has the following advantages:
firstly, the method adopts the idea of transfer learning, firstly, a self-encoder neural network model is trained by utilizing a radar echo signal of a source domain, then, the trained parameters of the self-encoder neural network are used as initial parameters of the front three coding layers of the feedforward neural network, and then, the parameters of the full connection layer of the feedforward neural network are finely adjusted by utilizing the radar echo signal of a target domain, so that the problem that the template library method in the prior art cannot update the template library in real time according to the change of carrier frequency parameters, and thus, the real-time identification of targets such as air, ground and the like cannot be realized is solved. The invention can update the model in real time by using a small amount of fine tuning data sets under new carrier frequency, thereby improving the real-time property of target identification.
Secondly, the invention constructs a fine tuning set formed by narrow-band radar echo signals under multiple carrier frequencies with less sample number, and utilizes the fine tuning set to perform fine tuning on the feedforward neural network, thereby realizing the self-adaptive extraction of carrier frequency steady characteristics from radar echo information under the condition of different carrier frequency parameter changes, and solving the problem that the identification accuracy rate is reduced due to the dispersion condition of the original characteristics in a characteristic space because the radar carrier frequency parameter changes can influence the micro Doppler modulation effect of a target in the prior art, so that the extracted characteristics of the invention have separability, and the accuracy rate of target identification is improved.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a graph of the results of a simulation experiment of the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and specific examples.
The specific steps implemented by the present invention are described in further detail with reference to fig. 1.
Step 1, generating a training set.
Narrow-band radar echo signals with data dimension L under a single carrier frequency and containing N category targets are extracted to serve as a training data set, each category at least contains 800 narrow-band radar echo signals, wherein N is larger than or equal to 3, and L is larger than or equal to 64.
And performing clutter suppression on each echo signal in the training data set by using a region CLEAN method.
The method for area CLEAN comprises the following specific steps:
the method comprises the first step of estimating ground clutter energy in radar echo according to the extracted radar working parameters of the narrow-band radar echo signals.
And secondly, performing discrete Fourier transform on each echo signal to obtain a Doppler spectrum of the echo signal, and taking a region near zero frequency of the Doppler spectrum as a clutter region, wherein the region near zero frequency is determined by the type of clutter.
Thirdly, reconstructing a time domain signal corresponding to the maximum Doppler amplitude in the clutter region in each echo signal according to the following formula:
Figure BDA0002893079850000051
wherein, Ci(t) represents the signal amplitude of the reconstructed time domain signal corresponding to the maximum Doppler amplitude in the clutter region in the ith echo signal at the time t, YiRepresents the maximum Doppler amplitude of a clutter region of the Doppler spectrum of the ith echo signal, K represents the number of points of discrete Fourier transform, exp represents exponential operation with a natural constant e as the base, j represents an imaginary unit symbol, pi represents a circumferential ratio, ξiIs represented by the formulaiThe corresponding doppler frequency of the doppler frequency is,
Figure BDA0002893079850000062
is represented by the formulaiThe corresponding phase.
And fourthly, subtracting the reconstructed time domain signal from each echo signal to obtain the echo processed signal.
And fifthly, calculating the energy of each processed echo signal in the clutter region.
Judging whether the energy of each processed echo signal in a clutter area is smaller than the energy of a ground clutter, if so, obtaining the echo signal after clutter suppression; otherwise, executing the second step of the step.
And removing the main component in each echo signal after clutter suppression by using a global CLEAN method.
The specific steps of the global CLEAN method are as follows:
firstly, performing discrete Fourier transform on each echo signal to obtain the Doppler spectrum of the echo signal.
Secondly, reconstructing a time domain signal corresponding to the main component in each echo signal according to the following formula:
Figure BDA0002893079850000061
wherein, Bi(t) represents the signal amplitude at time t of the main component echo signal reconstructed from the main component echo signal in the ith echo signal, RiIndicating the maximum Doppler amplitude, f, in the Doppler spectrum of the ith echo signaliIs represented by the formulaiCorresponding Doppler frequency, θiIs represented by the formulaiThe corresponding phase.
And thirdly, subtracting the reconstructed time domain signal from each echo signal subjected to clutter suppression to obtain an echo signal from which the main component is removed.
And carrying out the norm normalization processing on the amplitude of each echo signal without the main body component to obtain a training set.
And 2, generating a fine tuning set.
Narrow-band radar echo signals with data dimension L under M carrier frequencies containing N category targets are extracted to form a fine-tuning data set, each category at least contains 200 narrow-band radar echo signals, wherein N is larger than or equal to 3, M is larger than or equal to 5, and L is larger than or equal to 64.
And performing clutter suppression on each echo signal in the fine tuning data set by using a region CLEAN method.
The method for area CLEAN comprises the following specific steps:
the method comprises the first step of estimating ground clutter energy in radar echo according to the extracted radar working parameters of the narrow-band radar echo signals.
And secondly, performing discrete Fourier transform on each echo signal to obtain a Doppler spectrum of the echo signal, and taking a region near zero frequency of the Doppler spectrum as a clutter region, wherein the region near zero frequency is determined by the type of clutter.
Thirdly, reconstructing a time domain signal corresponding to the maximum Doppler amplitude in the clutter region in each echo signal according to the following formula:
Figure BDA0002893079850000071
wherein, Ci(t) represents the signal amplitude of the reconstructed time domain signal corresponding to the maximum Doppler amplitude in the clutter region in the ith echo signal at the time t, YiRepresents the maximum Doppler amplitude of a clutter region of the Doppler spectrum of the ith echo signal, K represents the number of points of discrete Fourier transform, exp represents exponential operation with a natural constant e as the base, j represents an imaginary unit symbol, pi represents a circumferential ratio, ξiIs represented by the formulaiThe corresponding doppler frequency of the doppler frequency is,
Figure BDA0002893079850000073
is represented by the formulaiThe corresponding phase.
And fourthly, subtracting the reconstructed time domain signal from each echo signal to obtain the echo processed signal.
And fifthly, calculating the energy of each processed echo signal in the clutter region.
Judging whether the energy of each processed echo signal in a clutter area is smaller than the energy of a ground clutter, if so, obtaining the echo signal after clutter suppression; otherwise, executing the second step of the step.
And removing the main component in each echo signal after clutter suppression by using a global CLEAN method.
The specific steps of the global CLEAN method are as follows:
firstly, performing discrete Fourier transform on each echo signal to obtain the Doppler spectrum of the echo signal.
Secondly, reconstructing a time domain signal corresponding to the main component in each echo signal according to the following formula:
Figure BDA0002893079850000072
wherein, Bi(t) represents the signal amplitude at time t of the main component echo signal reconstructed from the main component echo signal in the ith echo signal, RiIndicating the maximum Doppler amplitude, f, in the Doppler spectrum of the ith echo signaliIs represented by the formulaiCorresponding Doppler frequency, θiIs represented by the formulaiThe corresponding phase.
And thirdly, subtracting the reconstructed time domain signal from each echo signal subjected to clutter suppression to obtain an echo signal from which the main component is removed.
And carrying out the norm normalization processing on the amplitude of each echo signal without the main component to obtain a fine tuning set.
And 3, constructing a self-encoder neural network.
A seven-layer self-encoder neural network is built, and the structure of the self-encoder neural network is as follows in sequence: the device comprises an input layer, a first coding layer, a second coding layer, a third coding layer, a first decoding layer, a second decoding layer and a third decoding layer.
The parameters of each layer are set as follows: setting the numbers of the neurons of the first to third coding layers to 500, 256 and 64, respectively; the numbers of neurons of the first to third decoding layers are set to 64, 256, and T, respectively, where the number of neurons T is equal to the data dimension L of the narrowband radar echo signal.
And 4, constructing a feedforward neural network.
A six-layer feedforward neural network is built, and the structure of the feedforward neural network is as follows in sequence: the system comprises an input layer, a first coding layer, a second coding layer, a third coding layer, a full connection layer and a SoftMax layer;
the parameters of each layer are set as follows: setting the numbers of the neurons of the first to third coding layers to 500, 256 and 64, respectively; setting the number of the neurons of the full connection layer as M, wherein the number M of the neurons of the full connection layer is equal to the total number N of the target categories; the SoftMax layer adopts a SoftMax activation function to calculate the probability that the input echo signals are classified into each class;
and 5, training a neural network of the self-encoder.
The first step is as follows: randomly extracting P radar echo signals in a training set, inputting the P radar echo signals into a self-encoder neural network, and outputting reconstructed signals of the P radar echo signals, wherein P represents the number of samples input into the self-encoder neural network, and P is more than or equal to 64. Generally, the value of P is related to the hardware setting of the working machine used for target recognition in the invention, and the larger the P is, the larger the GPU memory of the working machine is required to be, and the smaller the P is, the longer the required training time is.
The second step is that: and calculating a loss value between the radar echo signal currently input into the self-encoder neural network and the output reconstruction signal by using a reconstruction loss function, and iteratively updating network parameters by using a back propagation algorithm.
The third step: and (4) judging whether the reconstruction loss function is converged, if so, executing the step (6) after obtaining the trained self-encoder neural network, otherwise, continuously and randomly extracting P radar echo signals in the training set, inputting the P radar echo signals into the self-encoder neural network, and executing the second step of the step.
The reconstruction loss function is as follows:
Figure BDA0002893079850000081
where J denotes a reconstruction loss function, L1, 2.. and L denotes the data dimension of the narrowband radar echo signal, x(l)Representing the signal amplitude value at the l-th position of the radar echo signal input to the self-encoder neural network,
Figure BDA0002893079850000082
denotes x(l)I-represents an absolute value operation.
And 6, training a feedforward neural network.
The first step is as follows: and assigning the weight parameters of the input layer, the first coding layer, the second coding layer and the third coding layer of the trained self-encoder neural network to the weight parameters of the input layer, the first coding layer, the second coding layer and the third coding layer in the feedforward neural network, and generating the weight parameters of the full-connection layer in the feedforward neural network by adopting a Gaussian initialization method.
The second step is that: q radar echo signals in the fine tuning set are randomly extracted and input into the feedforward neural network, a prediction classification label of each signal in the Q radar echo signals is output, Q represents the number of samples input into the feedforward neural network, and Q is larger than or equal to 64.
The third step: and calculating the loss value between the prediction classification label and the real classification label of each echo signal currently input into the feedforward neural network by using a cross entropy loss function.
The cross entropy loss function is as follows:
Figure BDA0002893079850000091
wherein H represents a cross entropy loss function, YprePredictive class label, Y, representing an amplitude-phase two-channel networktrainAnd the true class label represents a target sample in the training set, p is 1,2, …, N, p represents a class serial number of the target sample in the training set, and log represents a logarithm operation with a base 10.
The fourth step: and fixing the weight parameters of the input layer, the first coding layer, the second coding layer and the third coding layer in the feedforward neural network to be unchanged, and iteratively updating the weight parameters of the full-connection layer of the feedforward neural network by using a back propagation algorithm.
The fifth step: and (4) judging whether the cross entropy loss function of the current iteration is converged, if so, executing the step (7) after obtaining the trained feedforward neural network, otherwise, continuously and randomly extracting Q radar echo signals from the fine tuning set and inputting the Q radar echo signals into the feedforward neural network and then executing the third step of the step.
And 7, identifying the target.
And performing clutter suppression on each echo signal to be identified by using a region CLEAN method.
The method for area CLEAN comprises the following specific steps:
the method comprises the first step of estimating ground clutter energy in radar echo according to the extracted radar working parameters of the narrow-band radar echo signals.
And secondly, performing discrete Fourier transform on each echo signal to obtain a Doppler spectrum of the echo signal, and taking a region near zero frequency of the Doppler spectrum as a clutter region, wherein the region near zero frequency is determined by the type of clutter.
Thirdly, reconstructing a time domain signal corresponding to the maximum Doppler amplitude in the clutter region in each echo signal according to the following formula:
Figure BDA0002893079850000092
wherein, Ci(t) represents the signal amplitude of the reconstructed time domain signal corresponding to the maximum Doppler amplitude in the clutter region in the ith echo signal at the time t, YiRepresents the maximum Doppler amplitude of a clutter region of the Doppler spectrum of the ith echo signal, K represents the number of points of discrete Fourier transform, exp represents exponential operation with a natural constant e as the base, j represents an imaginary unit symbol, pi represents a circumferential ratio, ξiIs represented by the formulaiThe corresponding doppler frequency of the doppler frequency is,
Figure BDA0002893079850000102
is represented by the formulaiThe corresponding phase.
And fourthly, subtracting the reconstructed time domain signal from each echo signal to obtain the echo processed signal.
And fifthly, calculating the energy of each processed echo signal in the clutter region.
Judging whether the energy of each processed echo signal in a clutter area is smaller than the energy of a ground clutter, if so, obtaining the echo signal after clutter suppression; otherwise, executing the second step of the step.
And removing the main component of each echo signal subjected to clutter suppression by using a global CLEAN method.
The specific steps of the global CLEAN method are as follows:
firstly, performing discrete Fourier transform on each echo signal to obtain the Doppler spectrum of the echo signal.
Secondly, reconstructing a time domain signal corresponding to the main component in each echo signal according to the following formula:
Figure BDA0002893079850000101
wherein, Bi(t) represents the signal amplitude at time t of the main component echo signal reconstructed from the main component echo signal in the ith echo signal, RiIndicating the maximum Doppler amplitude, f, in the Doppler spectrum of the ith echo signaliIs represented by the formulaiCorresponding Doppler frequency, θiIs represented by the formulaiThe corresponding phase.
And thirdly, subtracting the reconstructed time domain signal from each echo signal subjected to clutter suppression to obtain an echo signal from which the main component is removed.
Inputting the preprocessed test sample into a trained feedforward neural network, calculating the probability of dividing the target to be recognized into various types through a SoftMax layer, and selecting the type corresponding to the highest probability as a recognition result.
1. Simulation conditions
The hardware test platform of the simulation experiment of the invention is as follows: the processor is an Intel Core i7 CPU, the main frequency is 3.4GHz, and the memory is 16 GB.
The software platform of the simulation experiment of the invention is as follows: windows 10 operating system, MatlabR2016a, and python 3.6.
The target echo of the training data set used by the simulation experiment of the invention comes from the simulation data of three types of airplane targets, the structural parameters of the three types of airplane targets: the number of blades, blade angle and blade length are randomly generated in a certain range. The specific parameters of the radar are as follows: the pulse repetition frequency is 5KHz, the dwell time is 110ms, and the training sample carrier frequency is 10 GHZ. The helicopter, the propeller and the jet machine respectively generate 300 groups of echoes randomly, and the total number of the echoes is 900 groups of echoes to form a training data set.
Gaussian white noise with a signal-to-noise ratio of 15dB, which is the ratio of the micromotion component to the noise, is added to the training data set. The helicopter, the propeller and the jet in the simulation experiment all have four different models, and specific model parameters are shown in the following table.
Target model of airplane Number of blades L1(m) L2(m) Rotating speed (rmin)
Helicopter BK17 4 0 5.5 383
Helicopter rice-17 5 0 10.645 185
Helicopter AS350 3 0 5.345 394
Helicopter bell 212 2 0 7.315 324
Propeller SAAB2000 6 0.28 1.905 950
Propeller L-420 5 0.12 1.15 1650
Propeller L-610G 4 0.23 1.675 1150
Propeller F406 3 0.23 1.18 1690
Jet plane A 30 0.3 1.0 3000
Jet B 38 0.38 1.1 3520
Jet C 27 0.18 0.51 8615
Jet D 33 0.2 0.6 5000
The target echoes of the fine tuning data set and the test data set used in the simulation experiment of the invention are also from the simulation data of the three types of airplane targets, the target echoes are divided into 9 groups of data sets according to the difference of the carrier frequencies of the simulation echo samples, and the carrier frequencies from the first group of data sets to the ninth group of data sets are respectively as follows: 8GHZ, 8.5GHZ, 9GHZ, 9.5GHZ, 10GHZ, 10.5GHZ, 11GHZ, 11.5GHZ, 12 GHZ. Setting sample carrier frequency as the following 9 groups of structural parameters of three types of airplane targets: the number of blades, blade angle and blade length are randomly generated in a certain range. The specific parameters of the radar are as follows: the pulse repetition frequency is 5KHz, the dwell time is 110ms, and the training sample carrier frequency is 10 GHZ. 300 samples are generated for each model of airplane, namely 1200 samples for each type of airplane, and 3600 samples are obtained according to the following steps of 1: the scale of 4 is divided into a trim data set and a test data set. Gaussian white noise with a signal-to-noise ratio of 15dB was added to both the fine tuning data set and the test data set during the experiment.
2. Simulation content and result analysis:
the simulation experiment of the invention is to adopt the support vector machine SVM classification method of the invention and the prior art to respectively carry out the simulation experiment on the 3 types of airplane targets to obtain the recognition result.
In a simulation experiment, the classification method of the Support Vector Machine (SVM) in the prior art is as follows: the radar Doppler target classification method, which is called a Support Vector Machine (SVM) classification method for short, is proposed by the university of Western electronic technology in the patent document 'ground target classification method based on robustness time-frequency characteristics' (patent application No. CN201510475477.X, publication No. CN 105044701A).
The method is used for identifying the test data set, firstly, echo signals in a training data set are input into a self-encoder neural network, the self-encoder neural network is trained, then, weight parameters of an input layer, a first encoding layer, a second encoding layer and a third encoding layer of the trained self-encoder neural network are respectively used as weight parameters of the input layer, the first encoding layer, the second encoding layer and the third encoding layer of a feedforward neural network, a fine tuning set is input into the feedforward neural network, the feedforward neural network is trained, target samples in the test data set after preprocessing are input into the trained feedforward neural network, the identification result of the test data set is obtained, and the identification accuracy of the test data under each signal-to-noise ratio condition is counted.
The results of the identification accuracy under different carrier frequencies of the method and the method for extracting the narrowband radar target classification based on the frequency domain features are plotted as shown in figure 2. In fig. 2, the abscissa represents the carrier frequency of the echo sample of the test data set, which is 8GHZ, 8.5GHZ, 9GHZ, 9.5GHZ, 10GHZ, 10.5GHZ, 11GHZ, 11.5GHZ, and 12GHZ, respectively, and the ordinate represents the identification accuracy of the echo signal in the test data set. In the figure, a diamond solid line represents a relation curve between the identification accuracy of the test sample obtained by the method and the carrier frequency, and a fork solid line represents a relation curve between the identification accuracy of the test sample obtained by the SVM classification method and the carrier frequency.
As shown in FIG. 2, the method of the invention has better recognition accuracy than the SVM classification method under any test sample carrier frequency. From the perspective of identification accuracy, under different test carrier frequencies, the average identification accuracy of the method is improved by nearly three percent. From the view point of carrier frequency robustness, when the difference between the test carrier frequency and the training carrier frequency is not large, the identification performance of the two methods has no obvious change; when the test carrier frequency is 8GHz, the identification performance of the SVM classification method is reduced by nearly two percentage points, but the identification performance of the method is hardly reduced; when the test carrier frequency is 11GHz and 11.5GHz, the identification performance of the SVM classification method is reduced by nearly 1.5 percentage points, but the identification performance of the method is almost not reduced; when the test carrier frequency is 12GHz, the identification performance of the SVM classification method is reduced by nearly 1.25 percent, and the identification performance of the method is reduced by nearly one percent. In general, the method has better robustness to carrier frequency change than the SVM classification method.
In conclusion, the simulation experiment results prove that the method can effectively and automatically extract the characteristics by utilizing the nonlinear fitting capability of the neural network, can finely adjust partial network parameters by utilizing a small amount of samples, enables the classification model to have a better fitting effect under the condition of carrier frequency change, obviously improves the carrier frequency robustness and the identification accuracy of radar target identification, and verifies the effectiveness of the method through experiments.

Claims (5)

1. A method for identifying a target with stable radar carrier frequency parameter change based on transfer learning is characterized in that a self-encoder neural network and a feedforward neural network are constructed, the self-encoder neural network is trained by utilizing radar echo signals under a single carrier frequency with a large number of samples, and network parameters of the feedforward neural network are finely adjusted by utilizing radar echo signals under multiple carrier frequencies with a small number of samples, and the method comprises the following steps:
(1) generating a training set:
(1a) extracting radar echo signals with data dimension L under a single carrier frequency and containing N category targets as a training data set, wherein each category at least contains 800 radar echo signals, N is more than or equal to 3, and L is more than or equal to 64;
(1b) performing clutter suppression on each echo signal in the training data set by using a regional CLEAN method;
(1c) removing the main component in each echo signal after clutter suppression by using a global CLEAN method;
(1d) carrying out modular two-norm normalization processing on the amplitude of each echo signal with the main component removed to obtain a training set;
(2) generating a fine tuning set:
(2a) radar echo signals with data dimension L under M carrier frequencies containing N types of targets are extracted to form a fine adjustment data set, each type at least contains 200 radar echo signals, wherein N is more than or equal to 3, M is more than or equal to 5, and L is more than or equal to 64;
(2b) performing clutter suppression on each echo signal in the fine tuning data set by using a regional clear method;
(2c) removing the main component in each echo signal after clutter suppression by using a global CLEAN method;
(2d) carrying out modular two-norm normalization processing on the amplitude of each echo signal with the main component removed to obtain a fine tuning set;
(3) constructing a self-encoder neural network:
(3a) a seven-layer self-encoder neural network is built, and the structure of the self-encoder neural network is as follows in sequence: the decoding device comprises an input layer, a first coding layer, a second coding layer, a third coding layer, a first decoding layer, a second decoding layer and a third decoding layer;
(3b) the parameters of each layer are set as follows: setting the numbers of the neurons of the first to third coding layers to 500, 256 and 64, respectively; setting the number of neurons of the first decoding layer, the second decoding layer and the third decoding layer as 64, 256 and T respectively, wherein the number of the neurons T is equal to the data dimension L of the radar echo signal;
(4) constructing a feedforward neural network:
(4a) a six-layer feedforward neural network is built, and the structure of the feedforward neural network is as follows in sequence: the system comprises an input layer, a first coding layer, a second coding layer, a third coding layer, a full connection layer and a SoftMax layer;
(4b) the parameters of each layer are set as follows: setting the numbers of the neurons of the first to third coding layers to 500, 256 and 64, respectively; setting the number of the neurons of the full connection layer as M, wherein the number M of the neurons of the full connection layer is equal to the total number N of the target categories; the SoftMax layer adopts a SoftMax activation function to calculate the probability that the input echo signals are classified into each class;
(5) training a self-encoder neural network:
(5a) randomly extracting P radar echo signals in a training set, inputting the P radar echo signals into a self-encoder neural network, and outputting reconstructed signals of the P radar echo signals, wherein P represents the number of samples input into the self-encoder neural network, and P is more than or equal to 64;
(5b) calculating a loss value between a radar echo signal currently input into the neural network of the self-encoder and an output reconstruction signal by using a reconstruction loss function, and iteratively updating network parameters by using a back propagation algorithm;
(5c) judging whether the reconstruction loss function is converged, if so, executing the step (6) after obtaining the trained self-encoder neural network, otherwise, continuously and randomly extracting P radar echo signals in the training set and inputting the P radar echo signals into the self-encoder neural network and then executing the step (5 b);
(6) training a feedforward neural network:
(6a) assigning the weight parameters of the input layer, the first coding layer, the second coding layer and the third coding layer of the trained self-encoder neural network to the weight parameters of the input layer, the first coding layer, the second coding layer and the third coding layer in the feedforward neural network, and generating the weight parameters of the full-connection layer in the feedforward neural network by adopting a Gaussian initialization method;
(6b) randomly extracting Q radar echo signals in the fine tuning set, inputting the Q radar echo signals into a feedforward neural network, and outputting a prediction classification label of each signal in the Q radar echo signals, wherein Q represents the number of samples input into the feedforward neural network, and Q is more than or equal to 64;
(6c) calculating a loss value between a prediction classification label and a real classification label of each echo signal currently input into the feedforward neural network by using a cross entropy loss function;
(6d) fixing the weight parameters of an input layer, a first coding layer, a second coding layer and a third coding layer in the feedforward neural network to be unchanged, and iteratively updating the weight parameters of the full-connection layer of the feedforward neural network by using a back propagation algorithm;
(6e) judging whether the cross entropy loss function of the current iteration is converged, if so, executing the step (7) after obtaining the trained feedforward neural network, otherwise, continuously randomly extracting Q radar echo signals from the fine tuning set and inputting the Q radar echo signals into the feedforward neural network and then executing the step (6 c);
(7) target identification:
(7a) performing clutter suppression on each echo signal to be identified by using a regional clear method;
(7b) removing the main component of each echo signal subjected to clutter suppression by utilizing a global CLEAN method;
(7c) and inputting each echo signal with the main component removed into a trained feedforward neural network, calculating the probability of dividing the target to be identified into various types through a SoftMax layer, and selecting the type corresponding to the highest probability as an identification result.
2. The method for identifying radar carrier frequency parameter variation robust targets based on transfer learning according to claim 1, wherein the specific steps of the area clear method in step (1b), step (2b) and step (7a) are as follows:
firstly, estimating ground clutter energy in radar echo according to the extracted radar working parameters of the radar echo signal;
secondly, performing discrete Fourier transform on each echo signal to obtain a Doppler spectrum of the echo signal, and taking a region near zero frequency of the Doppler spectrum as a clutter region, wherein the region near zero frequency is determined by the type of clutter;
thirdly, reconstructing a time domain signal corresponding to the maximum Doppler amplitude in the clutter region in each echo signal according to the following formula:
Figure FDA0002893079840000041
wherein, Ci(t) represents the signal amplitude of the reconstructed time domain signal corresponding to the maximum Doppler amplitude in the clutter region in the ith echo signal at the time t, YiRepresents the maximum Doppler amplitude of a clutter region of the Doppler spectrum of the ith echo signal, K represents the number of points of discrete Fourier transform, exp represents exponential operation with a natural constant e as the base, j represents an imaginary unit symbol, pi represents a circumferential ratio, ξiIs represented by the formulaiThe corresponding doppler frequency of the doppler frequency is,
Figure FDA0002893079840000043
is represented by the formulaiA corresponding phase;
fourthly, subtracting the reconstructed time domain signal from each echo signal to obtain an echo processed signal;
fifthly, calculating the energy of each processed echo signal in the clutter area;
judging whether the energy of each processed echo signal in a clutter area is smaller than the energy of a ground clutter, if so, obtaining the echo signal after clutter suppression; otherwise, the second step is executed.
3. The method for identifying radar carrier frequency parameter variation robust targets based on transfer learning according to claim 2, wherein the specific steps of the global CLEAN method in step (1c), step (2c) and step (7b) are as follows:
firstly, performing discrete Fourier transform on each echo signal to obtain a Doppler spectrum of the echo signal;
secondly, reconstructing a time domain signal corresponding to the main component in each echo signal according to the following formula:
Figure FDA0002893079840000042
wherein, Bi(t) represents the signal amplitude at time t of the main component echo signal reconstructed from the main component echo signal in the ith echo signal, RiIndicating the maximum Doppler amplitude, f, in the Doppler spectrum of the ith echo signaliIs represented by the formulaiCorresponding Doppler frequency, θiIs represented by the formulaiA corresponding phase;
and thirdly, subtracting the reconstructed time domain signal from each echo signal subjected to clutter suppression to obtain an echo signal from which the main component is removed.
4. The method for identifying radar carrier frequency parameter variation robust targets based on transfer learning according to claim 1, wherein the reconstruction loss function in step (5b) is as follows:
Figure FDA0002893079840000051
where J denotes a reconstruction loss function, L1, 2.. and L denotes the data dimension of the radar echo signal, x(l)Representing the signal amplitude value at the l-th position of the radar echo signal input to the self-encoder neural network,
Figure FDA0002893079840000053
denotes x(l)I-represents an absolute value operation.
5. The method for identifying radar carrier frequency parameter variation robust targets based on transfer learning according to claim 1, wherein the cross entropy loss function in step (6c) is as follows:
Figure FDA0002893079840000052
wherein H represents a cross entropy loss function, YprePredictive class label, Y, representing an amplitude-phase two-channel networktrainAnd the true class label represents a target sample in the training set, p is 1,2, …, N, p represents a class serial number of the target sample in the training set, and log represents a logarithm operation with a base 10.
CN202110032917.XA 2021-01-12 2021-01-12 Radar carrier frequency parameter change steady target identification method based on transfer learning Active CN112731330B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110032917.XA CN112731330B (en) 2021-01-12 2021-01-12 Radar carrier frequency parameter change steady target identification method based on transfer learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110032917.XA CN112731330B (en) 2021-01-12 2021-01-12 Radar carrier frequency parameter change steady target identification method based on transfer learning

Publications (2)

Publication Number Publication Date
CN112731330A CN112731330A (en) 2021-04-30
CN112731330B true CN112731330B (en) 2022-04-19

Family

ID=75590391

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110032917.XA Active CN112731330B (en) 2021-01-12 2021-01-12 Radar carrier frequency parameter change steady target identification method based on transfer learning

Country Status (1)

Country Link
CN (1) CN112731330B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113240081B (en) * 2021-05-06 2022-03-22 西安电子科技大学 High-resolution range profile target robust identification method aiming at radar carrier frequency transformation
CN114814776B (en) * 2022-06-24 2022-10-14 中国空气动力研究与发展中心计算空气动力研究所 PD radar target detection method based on graph attention network and transfer learning
CN117129963A (en) * 2023-10-09 2023-11-28 北京大学重庆大数据研究院 Radar echo extrapolation method based on migration prediction and readable storage medium thereof

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107194433B (en) * 2017-06-14 2019-09-13 电子科技大学 A kind of Radar range profile's target identification method based on depth autoencoder network
US11176477B2 (en) * 2018-02-06 2021-11-16 Hrl Laboratories, Llc System and method for unsupervised domain adaptation via sliced-wasserstein distance
CN110275147B (en) * 2018-03-13 2022-01-04 中国人民解放军国防科技大学 Human behavior micro-Doppler classification and identification method based on migration depth neural network
CN108508411B (en) * 2018-03-22 2022-04-08 天津大学 Passive radar external radiation source signal identification method based on transfer learning
CN109086700B (en) * 2018-07-20 2021-08-13 杭州电子科技大学 Radar one-dimensional range profile target identification method based on deep convolutional neural network
CN110472483B (en) * 2019-07-02 2022-11-15 五邑大学 SAR image-oriented small sample semantic feature enhancement method and device
CN110824450B (en) * 2019-10-15 2020-09-29 中国人民解放军国防科技大学 Radar target HRRP robust identification method in noise environment
CN111736127B (en) * 2020-06-11 2022-09-13 北京理工大学 Source domain generation and distribution parameter generalization method for unknown sea clutter

Also Published As

Publication number Publication date
CN112731330A (en) 2021-04-30

Similar Documents

Publication Publication Date Title
CN112731330B (en) Radar carrier frequency parameter change steady target identification method based on transfer learning
CN112882009B (en) Radar micro Doppler target identification method based on amplitude and phase dual-channel network
CN109188414A (en) A kind of gesture motion detection method based on millimetre-wave radar
CN113298846B (en) Interference intelligent detection method based on time-frequency semantic perception
CN112666533B (en) Repetition frequency change steady target identification method based on spatial pyramid pooling network
CN112287784B (en) Radar signal classification method based on deep convolutional neural network and feature fusion
CN112859014A (en) Radar interference suppression method, device and medium based on radar signal sorting
CN114595732B (en) Radar radiation source sorting method based on depth clustering
CN112882012B (en) Radar target noise robust identification method based on signal-to-noise ratio matching and echo enhancement
CN112882011A (en) Radar carrier frequency variation robust target identification method based on frequency domain correlation characteristics
Qin et al. Radar waveform recognition based on deep residual network
CN111680737A (en) Radar radiation source individual identification method under differential signal-to-noise ratio condition
CN114117912A (en) Sea clutter modeling and inhibiting method under data model dual drive
CN112731331B (en) Micro-motion target noise steady identification method based on signal-to-noise ratio adaptive network
CN112835003B (en) Radar repetition frequency variation steady target recognition method based on resampling preprocessing
Xie et al. Data augmentation and deep neural network classification based on ship radiated noise
CN112784916B (en) Air target micro-motion parameter real-time extraction method based on multitask convolutional network
CN115061094A (en) Radar target identification method based on neural network and SVM
CN114676721A (en) Radar active suppression interference identification method and system based on radial basis function neural network
Tian et al. Multiresolution jamming recognition with few-shot learning
Akhtar et al. Detection of helicopters using neural nets
Kang et al. Aircraft classification method based on the kurtosis–skewness feature and wavelet decomposition and linear discriminant analysis
CN114462457B (en) Ship underwater noise deep learning identification method based on intrinsic probability density function
CN112859025A (en) Radar signal modulation type classification method based on hybrid network
CN114994657A (en) Radar repetition frequency change steady target identification method based on repetition frequency self-adaptive 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
GR01 Patent grant
GR01 Patent grant