CN113987674A - Radar HRRP continuous learning method based on generation of countermeasure network - Google Patents

Abstract

The invention discloses a radar HRRP continuous learning method based on a generation countermeasure network, which comprises the following steps: s1, the used actual measurement data comprises three types of airplanes, namely a medium propeller airplane An-26, a small jet airplane Cessna and a large jet airplane Yark-42, wherein the radar works in a C wave band, the signal bandwidth is 400MHz, and the pulse repetition frequency is 400 Hz; s2, processing the intensity sensitivity and the translation sensitivity in the pretreatment process; s3, under the background of radar HRRP automatic target identification, three different HRRP incremental learning task settings are used; s4, training the HRRP data processed by the S2 under three settings according to a continuous learning correlation method, and redesigning a decoder for a pseudo replay DGR method to be a conditional auto-encoder countermeasure network CVAEGAN, wherein the CVAEGAN comprises an encoder, a decoder, a discriminator and a classifier; s5, training the HRRP data processed by S2 by using the DGR method of the CVAEGAN network.

Description

Radar HRRP continuous learning method based on generation of countermeasure network

Technical Field

The invention belongs to the technical field of radars, and particularly relates to a radar HRRP continuous learning method based on a generation countermeasure network.

Background

The Range Resolution of a High-Resolution broadband radar is much smaller than the target size, and the echo is also called a High Resolution Range Profile (HRRP) of the target. The HRRP not only contains the radial size of the target, the distribution of scattering points and other structural information, but also has the advantages of relatively simple acquisition and processing, convenient storage and great benefit to engineering application. Therefore, the target identification method based on the HRRP has attracted much attention in the field of radar automatic target identification.

Past target recognition research on HRRP has mostly focused on static off-line tasks. The tasks firstly collect training HRRP data and establish a sample library, then establish a model training target recognition model based on statistics or deep learning and the like, and finally deploy the model to hardware equipment for real recognition tasks in an off-line manner. The methods have achieved good application results, however, when a new target HRRP sample is obtained, the models cannot be expanded online to obtain the identification capability of the new target, and therefore the models are difficult to apply to a variable actual scene.

Disclosure of Invention

In view of the above technical problems, the present invention provides a method for HRRP continuous learning of radar based on generation of countermeasure network.

In order to solve the technical problems, the invention adopts the following technical scheme:

a radar HRRP continuous learning method based on a generation countermeasure network comprises the following steps:

s1, the used measured data comprises three types of airplanes, namely a medium propeller airplane An-26, a small jet airplane Cessna and a large jet airplane Yark-42, the radar works in a C wave band, the signal bandwidth is 400MHz, the pulse repetition frequency is 400Hz, each HRRP sample in the data set comprises 256 distance units, the 2 nd and 5 th sections of the An-26 airplane, the 6 th and 7 th sections of the Cessna and the 5 th and 6 th sections of the Yark-42 are taken as training samples, the rest sections are taken as test samples, An equal interval sampling method is used for extracting the training samples from the training data, the training samples are as many as the training samples of the simulation data, a certain amount of white Gaussian noise is added to enable the signal to noise ratio to be 25dB, the used data set of the 9 types of airplanes is An electromagnetic simulation HRRP data set generated by FECO software based on a model, each type of airplane data set for training covers the angular domain of 360 degrees in the azimuth dimension, the method comprises 1600 training samples, wherein each HRRP sample comprises 256 distance units, the target type, the azimuth dimension coverage angle domain and the sample number of a simulated airplane data set used for testing are consistent with those of the airplane data set used for training, only the difference of about 10 degrees exists between the pitching angle and the simulated airplane data set, and in order to ensure that the simulated data are more real, a certain amount of white Gaussian noise is added to ensure that the signal-to-noise ratio of the simulated data set is 25dB during simulation;

s2, processing the intensity sensitivity and the translation sensitivity in the pretreatment process;

s3, under the background of radar HRRP automatic target identification, three different HRRP incremental learning task settings are used;

s4, training the HRRP data processed by the S2 under three settings according to a continuous learning correlation method, and redesigning a decoder for a pseudo replay DGR method to be a conditional auto-encoder countermeasure network CVAEGAN, wherein the CVAEGAN comprises an encoder, a decoder, a discriminator and a classifier;

s5, training the HRRP data processed by S2 by using the DGR method of the CVAEGAN network.

Preferably, the S2 further includes:

s201, performing L on original HRRP echo₂Normalization processing, raw HRRP data denoted X_raw＝[x₁，x₂，...，x_M]Wherein M represents the total number of HRRP distance units, x obtained after normalization_{normalization}Expressed as:

wherein x_iRepresenting the intensity of the ith distance cell;

s202, using a gravity center alignment method to eliminate translation sensitivity, wherein in the gravity center alignment method, firstly, the gravity center position of data is calculated, and then the gravity center of the HRRP data is close to the data center position by translating left and right, wherein the calculation process of the gravity center g is represented as:

preferably, the three different HRRP incremental learning tasks in S3 include task incremental learning, domain incremental learning, and class incremental learning.

The invention has the following beneficial effects:

(1) the problem of catastrophic forgetting existing in the past offline training model is solved, so that the radar can update parameters in real time according to new data on a battlefield, and the actual application range is expanded;

(2) the model not only has certain data privacy and fixed model capacity characteristics, but also has better recognition effect compared with a model based on a regularization method.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In an actual battlefield environment, new HRRP data of non-cooperative targets may be collected, and if the model parameters of previous offline deployment are directly updated by using a new sample, the classification knowledge obtained by offline learning in the past will be faced with catastrophic forgetfulness. How to alleviate the catastrophic forgetting problem and enable the model to have expansibility on new samples, that is, how to have the recognition capability on new targets while keeping the recognition capability of targets obtained by past training becomes an engineering problem to be solved urgently at present.

The embodiment of the invention provides a radar HRRP continuous learning method based on a generation countermeasure network, which comprises the following steps:

s1, due to the lack of the measured data, the validity of the method of the embodiment of the invention is verified by adopting the form of combining the measured data and the simulation data. The measured data used includes three types of aircraft. The radar works in the C wave band, the signal bandwidth is 400MHz, and the pulse repetition frequency is 400 Hz. The three types of aircraft are respectively a medium propeller aircraft 'An-26', a small jet aircraft 'Cessna' and a large jet aircraft 'Yark-42'. Each HRRP sample in the data set contains 256 range units. In order to make the azimuth domain of the training data cover the azimuth domain of the test data, the 2 nd and 5 th sections of An-26 airplane, the 6 th and 7 th sections of Cessna, and the 5 th and 6 th sections of Yark-42 are used as training samples, and the rest sections are used as test samples. And in order to avoid the class imbalance phenomenon existing in the samples, an equal interval sampling method is used for extracting the samples from the training data, the number of the samples is as large as that of the training samples of the simulation data, and a certain amount of white Gaussian noise is added to enable the signal-to-noise ratio to be 25 dB. The data set for the type 9 aircraft used was an electromagnetic simulation HRRP data set generated by FECO software based on a turntable model. Each type of airplane data set for training covers 360-degree angular domain in azimuth dimension, and comprises 1600 training samples, and each HRRP sample comprises 256 distance units. The target type, the azimuth dimension coverage angle domain and the sample number of the simulated airplane data set used for the test are consistent with the airplane data set used for training, and only the difference of about 10 degrees exists between the two pitch angles. In order to make the simulation data more realistic, a certain amount of white Gaussian noise is added to make the signal-to-noise ratio of the simulation data set to be 25dB during simulation.

S2, because the acquired original HRRP data contains the intensity sensitivity and the translation sensitivity, the two sensitivities are processed in the preprocessing process to eliminate the instability influence of the two sensitivities on the back-end deep neural network model;

s3, under the background of radar HRRP automatic target identification, in order to enable comparison among various incremental learning methods to reflect actual performance better, three different HRRP incremental learning task settings are used;

s4, training the HRRP data in S2 under three settings according to a continuous learning correlation method, and redesigning a decoder for a fake replay DGR method to generate a counter-reactive network CVAEGAN, wherein the CVAEGAN comprises an encoder, a decoder, a discriminator and a classifier;

s5, training HRRP data in S2 by using DGR method of CVAEGAN network, the final test results are superior to other methods.

In an embodiment of the present invention, S2 further includes:

s201, in order to eliminate the influence caused by the intensity sensitivity, L is carried out on the original HRRP echo₂And (6) normalization processing. If the original HRRP data can be expressed as x_raw＝[x₁，x₂，...，x_M]Wherein M represents the total number of HRRP distance units, x obtained after normalization_{normalization}Can be expressed as:

wherein x_iIndicating the intensity of the ith range bin.

S202, eliminating the translational sensitivity by using a gravity center alignment method. In the barycentric alignment method, the barycentric position of the data is first calculated, and then the barycentric position of the HRRP data is brought close to the barycentric position by translating the HRRP data left and right, wherein the calculation process of the barycentric g can be expressed as:

in an embodiment of the present invention, S3 further includes:

s301, firstly, a data set is divided into a plurality of different tasks, and the Task setting of the HRRP incremental learning is called Task incremental learning and is represented by Task-IL. In Task-IL, the Task number to which the test sample belongs is known, and the model only needs to indicate to which Task number the HRRP sample tested belongs. The second task setting for HRRP incremental learning is called domain incremental learning, denoted by domain-IL. Under this setting, we only focus on what test HRRP samples correspond to is the class ii target in the incremental learning task. The third set of HRRP incremental learning tasks is called class-incremental learning, denoted by class-IL. After a test HRRP sample is given, the model needs to output the target to which the HRRP sample corresponds.

In an embodiment of the present invention, S4 further includes:

s401, the purpose of the encoder is to derive the mean and variance of the random variable z used for decoder generation of HRRP. And inputting a label y corresponding to the HRRP data x into an encoder network together with the data as a condition, namely, obtaining an N-dimensional vector x _ hat by one-hot encoding the condition label y, and then directly splicing the N-dimensional vector x _ hat with the condition label y to obtain an x _ input [ x, x _ hat ] as a new vector to be input. The input generates codes through the combination of two full-link layers and a ReLU activation function, and finally the obtained codes respectively obtain a mean value and a variance which are used as input z of a decoder through different full-link layers. The data can be expressed by the formula for fully connected layers and ReLu as:

y_FC＝f(Wx_input+b)

where W represents the weight matrix of the fully-connected layer, b represents the bias of the fully-connected layer, and the function f (·) represents the ReLU activation function.

f(x)＝max(0，x)

S402, the decoder is to generate HRRP samples corresponding to the category by using the random variable and the condition category label y, and the encoder and the decoder form a codec pair, so the design is relatively similar. Similar to the encoder, this input is characterized by a combination module of two fully-connected layers and a ReLU activation function. At the last combination of HRRP samples generated, we exchanged the activation function for a sigmoid activation function, which keeps the range of the output between 0 and 1.

Where exp (·) represents an exponential operation.

And S403, distinguishing the 'true' sample from the 'false' sample through a discriminator. To reduce the risk of model collapse of the CVAE-GAN model and to make the HRRP samples generated by the decoder part more realistic, we use a more complex discriminator.

S404, the purpose is to make the reconstructed data decoded by the conditional coding condition still obtain the same classified capability as the real data, so that the generated data of the generator can meet its condition (label). The classifier network is also characterized in that two full connection layers are matched with a ReLU activation function to extract features, a full connection layer and softmax are used to obtain a discrimination result, the discrimination result is different from an output discrimination result (only true and false discrimination) in the GAN, and in a final classifier, the discrimination range is determined by three task settings.

We assume that the input before passing softmax is denoted by input. If the total number of targets contained in the training set is C, then output by softmax can be expressed as:

where exp (·) denotes an index operation, and input (i) and output (i) refer to the ith element in the vectors input and output, respectively. Where output (i) is also the maximum a posteriori probability P (i | x), representing the probability that x is classified as label i.

In an embodiment of the present invention, S5 further includes:

s501, in the discriminator, an HRRP sample can pass through SE-Block after being subjected to dimension change, and in an experiment, the number of channels is set to be 256, so that the module can be more concentrated in more important distance units in the HRRP sample. We get the adjusted value c of the weight by passing x through layers composed of two fully connected modules, and finally we get x _ se which can be expressed as:

x_SE(i)＝c(i)×x(i)

where x is the input of SE-block, c is the weight adjustment result, x_SEFor the output of the SE-block, i in parentheses represents the ith element in the corresponding vector.

S502, after the importance of each layer is adjusted, x _ se is sent into a module consisting of two layers of transformers. Where two layers of transformers indicate that there are two identical encoders inside, and two identical decoders.

The overall input is x _ se, and 256 HRRP distance units of x _ se are respectively sent into a self-attention module, and the layer not only uses the information of the layer but also uses the information of other distance units when processing one distance unit. Z after treatment by self-attention is as follows:

where Q, K and V are the inputs x _ se and three weight matrices W^Q，W^KAnd W^VThe result of multiplication represents the matrix of Query, Key, Value, respectively. d_kDenotes the length of the key vector, and the superscript T of K denotes it as a transpose.

After passing through self-attention module, we proceed with Add & normaize operations:

z＝LayerMorm(x_se+z)

where LayerNorm denotes layer normalization and x _ se + z denotes residual concatenation, Add.

Then, residual error normalization is carried out through a feedforward neural network, and the result of passing through the feedforward neural network is as follows:

z＝FFNN(z)＝max(0，zW₁+b₁)W₂+b₂

the feedforward neural network comprises a combination of two layers of fully connected and activation functions, the activation function of the first layer being a ReLU and the second layer being a linear activation function. I.e. W₁，b₁And W₂，b₂The weight matrix and bias of the two fully-connected layers are represented, respectively.

The result of this processing is processed once more by one and the same encoder, and the final result obtained by the last two encoders is a set of attention vectors K and V.

The decoder has most of its components similar to the encoder, with the input again being the x _ se input to self-entry. Except that self-entries in the decoder only allow attention to information preceding the current range cell and not all.

The principle of the Encode-Decoder Attention layer is similar to the multi-head Attention mechanism, except that it uses the output of the previous layer to construct a Query matrix, and the Key matrix and Value matrix come from the output obtained by the Encoder.

Where Q is from the layer above the decoder and K and V are from the encoder.

It is to be understood that the exemplary embodiments described herein are illustrative and not restrictive. While one or more embodiments of the present invention have been described, it will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present invention as defined by the following claims.

Claims (3)

1. A radar HRRP continuous learning method based on a generation countermeasure network is characterized by comprising the following steps:

s1, the used measured data comprises three types of airplanes, namely a medium propeller airplane An-26, a small jet airplane Cessna and a large jet airplane Yark-42, the radar works in a C wave band, the signal bandwidth is 400MHz, the pulse repetition frequency is 400Hz, each HRRP sample in the data set comprises 256 distance units, the 2 nd and 5 th sections of the An-26 airplane, the 6 th and 7 th sections of the Cessna and the 5 th and 6 th sections of the Yark-42 are taken as training samples, the rest sections are taken as test samples, An equal interval sampling method is used for extracting the training samples from the training data, the training samples are as many as the training samples of the simulation data, a certain amount of white Gaussian noise is added to enable the signal to noise ratio to be 25dB, the used data set of the 9 types of airplanes is An electromagnetic simulation HRRP data set generated by FECO software based on a model, each type of airplane data set for training covers the angular domain of 360 degrees in the azimuth dimension, the method comprises 1600 training samples, wherein each HRRP sample comprises 256 distance units, the target type, the azimuth dimension coverage angle domain and the sample number of a simulated airplane data set used for testing are consistent with those of the airplane data set used for training, only the difference of about 10 degrees exists between the pitching angle and the simulated airplane data set, and in order to ensure that the simulated data are more real, a certain amount of white Gaussian noise is added to ensure that the signal-to-noise ratio of the simulated data set is 25dB during simulation;

s2, processing the intensity sensitivity and the translation sensitivity in the pretreatment process;

s3, under the background of radar HRRP automatic target identification, three different HRRP incremental learning task settings are used;

s4, training the HRRP data processed by the S2 under three settings according to a continuous learning correlation method, and redesigning a decoder for a pseudo replay DGR method to be a conditional auto-encoder countermeasure network CVAEGAN, wherein the CVAEGAN comprises an encoder, a decoder, a discriminator and a classifier;

s5, training the HRRP data processed by S2 by using the DGR method of the CVAEGAN network.

2. The HRRP continuous learning method for radar based on generation of countermeasure network as claimed in claim 1, wherein said S2 further comprises:

s201, performing L on original HRRP echo₂Normalization processing, raw HRRP data denoted x_raw＝[x₁，x₂，…，x_M]Wherein M represents the total number of HRRP distance units, x obtained after normalization_{normalization}Expressed as:

wherein x_iIndicating the intensity of the ith distance cell；

S202, using a gravity center alignment method to eliminate translation sensitivity, wherein in the gravity center alignment method, firstly, the gravity center position of data is calculated, and then the gravity center of the HRRP data is close to the data center position by translating left and right, wherein the calculation process of the gravity center g is represented as:

3. the HRRP continuous learning method for radar based on generation countermeasure network as claimed in claim 1, wherein the three different HRRP incremental learning tasks in S3 include task incremental learning, domain incremental learning and class incremental learning.