阅读说明：本技术 一种雷达杂波模拟方法及电子设备 (Radar clutter simulation method and electronic equipment ) 是由 张伟 施祖帅 马鑫 于 2021-08-27 设计创作，主要内容包括：本发明公开了一种雷达杂波模拟方法及电子设备,其包括采集原始杂波数据,进行预处理,并获取和拆分预处理后的杂波数据的距离单元；构建包含生成器、判别器和Q网络的info-WaveGAN训练模型；将拆分后的每个距离单元作为一个训练样本输入info-WaveGAN训练模型,依次经过生成器、判别器和Q网络进行模型训练；创建损失函数,通过循环执行最小化损失函数操作对训练模型的参数进行优化处理,得到优化后的杂波数据。本发明提高了模型拟合能力,以生成更加逼真的杂波,且忽略了输入雷达杂波的分布类型,能够将其应用于任意分布类型的雷达杂波生成,增加了本方法的通用性。(The invention discloses a radar clutter simulation method and electronic equipment, which comprises a distance unit for acquiring original clutter data, preprocessing the original clutter data, and acquiring and splitting the preprocessed clutter data; constructing an info-WaveGAN training model comprising a generator, a discriminator and a Q network; inputting each split distance unit serving as a training sample into an info-WaveGAN training model, and sequentially performing model training through a generator, a discriminator and a Q network; and creating a loss function, and performing optimization processing on parameters of the training model by circularly executing the operation of minimizing the loss function to obtain optimized clutter data. The method improves the model fitting capability to generate more vivid clutter, ignores the distribution type of the input radar clutter, can be applied to the radar clutter generation of any distribution type, and increases the universality of the method.)

1. A radar clutter simulation method is characterized by comprising the following sub-steps:

s1, collecting original clutter data, preprocessing the original clutter data, and acquiring and splitting a distance unit of the preprocessed clutter data;

s2, constructing an info-WaveGAN training model containing a generator, a discriminator and a Q network;

s3, inputting each split distance unit as a training sample into an info-WaveGAN training model, and performing model training through a generator, a discriminator and a Q network in sequence;

s4, creating a loss function, and performing optimization processing on the parameters of the info-WaveGAN training model through circularly executing the operation of the minimum loss function to obtain optimized clutter data and complete radar clutter simulation.

2. The method for simulating radar clutter according to claim 1, wherein the specific method of step S1 is:

s1-1, collecting original clutter data, extracting the distance of the original clutter data, and multiplying the distance by the power of three-thirds of the distance on the basis of the original clutter data to obtain clutter data with offset distance attenuation;

s1-2, judging whether the ground-scraping angle secant value is close to 1, if so, entering the step S1-3; otherwise, dividing the clutter data with the offset distance attenuation by one half of the secant value of the ground wiping angle to obtain clutter data with the influence of the ground wiping angle eliminated;

s1-3, judging whether a target echo exists in the current clutter data, if so, entering a step S1-4; otherwise, entering step S1-5;

s1-4, judging whether the position of the target echo is obtained or not, if so, eliminating the distance unit echo of the target echo from the clutter data, and entering the step S1-5; otherwise, eliminating the peak value of the target echo by a windowing filtering method, and entering the step S1-5;

s1-5, acquiring a distance unit and a pulse unit of the current clutter data, performing global normalization, completing preprocessing, and entering the step S1-6;

and S1-6, splitting each distance unit of the preprocessed clutter data.

3. The radar clutter simulation method according to claim 1, wherein the generator constructed in step S2-1 comprises six deconvolution modules connected in series; the first deconvolution module comprises a full-connection operation layer, a reshape matrix conversion operation layer, a batch normalization operation layer and a relu activation function layer which are sequentially connected; the second deconvolution module comprises a one-dimensional deconvolution operation layer, a batch normalization operation layer and a relu activation function layer which are sequentially connected, wherein the second deconvolution module, the third deconvolution module, the fourth deconvolution module and the fifth deconvolution module have the same structure; the sixth deconvolution module comprises a one-dimensional deconvolution operation layer and a Tanh activation function layer which are sequentially connected.

4. The radar clutter simulation method according to claim 1, wherein the discriminator constructed in step S2-2 comprises six convolution modules connected in sequence; the first convolution module comprises a one-dimensional convolution operation layer and a leakage relu function activation layer which are sequentially connected; the second convolution module comprises a one-dimensional convolution operation layer and a leakage relu function activation layer which are sequentially connected, wherein the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module are identical in structure; the sixth convolution module comprises a reshape matrix conversion operation layer and a full-connection operation layer which are sequentially connected.

5. The radar clutter simulation method according to claim 1, wherein the Q network constructed in step S2-2 comprises seven convolution modules connected in sequence; wherein the first five convolution modules are shared with the first five convolution modules of the discriminator; the sixth convolution module comprises a reshape matrix conversion operation layer, a full-connection operation layer, a batchnorm normalization layer and a leakage relu function activation layer which are sequentially connected; the seventh convolution module includes a fully connected operational layer.

6. The method for simulating radar clutter according to claim 1, wherein the method for creating the loss function in step S4 comprises:

according to the formula:

obtaining a loss functionWherein V (G, D) is an objective function of WaveGAN, G is a generator, D is a discriminator, lambda is an adjustable hyper-parameter, and L_l(G, Q) is a mutual information lower bound, and Q is a Q network; x-Pdata (x) represents the distribution of the input data x sampled in the real data, E [ ·]For the expected value, ln is a logarithmic function based on natural logarithm, D (-) is the probability that x comes from the real data, z-P_z(z) represents that a distance unit vector z is sampled in prior distribution, G (-) is generated data, H (-) is entropy of variable, c-P (c) is hidden variable c according with hidden variable distribution, x-G (z, c) is input data according with condition to generate data distribution, log is a logarithmic function with 10 as a base, and Q (c | x) is variable distribution.

7. A radar clutter simulation electronic device, characterized in that the device comprises:

a memory storing executable instructions; and

a processor configured to execute the executable instructions in the memory to implement the method of any of claims 1-6.

Technical Field

The invention relates to the field of radar signal processing, in particular to a radar clutter simulation method and electronic equipment.

Background

The radar echo signals comprise target echoes, clutter echoes, noise and the like, and ground objects, cloud rain, sea waves, foil strips and the like can reflect electromagnetic waves to form clutter and cover or interfere the detection of the radar on the target. In order to deeply research the influence of the clutter on the radar detection performance, radar designers need to analyze and model radar clutter characteristics so as to formulate a reasonable radar design scheme, select radar parameters and take various anti-clutter measures.

The testing environment of the external field test is unstable, is easy to be limited and influenced, will slow down the development progress, and each field of test consumes a large amount of resources. The method has the advantages of flexibility, convenience, economy and the like by taking the vivid simulation clutter as the input of the radar design, and is an indispensable step for estimating and verifying the performance of the radar system, so that the scheme of combining the radar clutter simulation technology and the external field test becomes a main means for developing the current radar system, the cost is reduced, the development period is shortened, and the stability and the accuracy of the test are improved. Therefore, the radar clutter simulation has important significance and great value for radar research.

However, errors are introduced in the radar clutter simulation modeling method in the prior art, secondary errors are introduced in the process of fitting the estimation model, and the fidelity of the simulated clutter is reduced; and the general statistical model has the applicable condition to radar clutter, the statistical models of the radar clutter under different environments are different, and the statistical characteristics met by the radar clutter under the complex environment may not have a specific statistical model to be applicable, so the universality is lower.

Disclosure of Invention

Aiming at the defects in the prior art, the radar clutter simulation method and the electronic equipment provided by the invention solve the problems of poor fidelity and low universality in the prior art.

In order to achieve the purpose of the invention, the invention adopts the technical scheme that:

the radar clutter simulation method comprises the following steps:

s1, collecting original clutter data, preprocessing the original clutter data, and acquiring and splitting a distance unit of the preprocessed clutter data;

s2, constructing an info-WaveGAN training model containing a generator, a discriminator and a Q network;

s3, inputting each split distance unit as a training sample into an info-WaveGAN training model, and performing model training through a generator, a discriminator and a Q network in sequence;

s4, creating a loss function, and performing optimization processing on the parameters of the info-WaveGAN training model through circularly executing the operation of the minimum loss function to obtain optimized clutter data and complete radar clutter simulation.

Further, the specific method of step S1 is:

s1-1, collecting original clutter data, extracting the distance of the original clutter data, and multiplying the distance by the power of three-thirds of the distance on the basis of the original clutter data to obtain clutter data with offset distance attenuation;

s1-2, judging whether the ground-scraping angle secant value is close to 1, if so, entering the step S1-3; otherwise, dividing the clutter data with the offset distance attenuation by one half of the secant value of the ground wiping angle to obtain clutter data with the influence of the ground wiping angle eliminated;

s1-3, judging whether a target echo exists in the current clutter data, if so, entering a step S1-4; otherwise, entering step S1-5;

s1-4, judging whether the position of the target echo is obtained or not, if so, eliminating the distance unit echo of the target echo from the clutter data, and entering the step S1-5; otherwise, eliminating the peak value of the target echo by a windowing filtering method, and entering the step S1-5;

s1-5, acquiring a distance unit and a pulse unit of the current clutter data, performing global normalization, completing preprocessing, and entering the step S1-6;

and S1-6, splitting each distance unit of the preprocessed clutter data.

Further, the generator constructed in step S2-1 includes six deconvolution modules connected in sequence; the first deconvolution module comprises a full-connection operation layer, a reshape matrix conversion operation layer, a batch normalization operation layer and a relu activation function layer which are sequentially connected; the second deconvolution module comprises a one-dimensional deconvolution operation layer, a batch normalization operation layer and a relu activation function layer which are sequentially connected, wherein the second deconvolution module, the third deconvolution module, the fourth deconvolution module and the fifth deconvolution module have the same structure; the sixth deconvolution module comprises a one-dimensional deconvolution operation layer and a Tanh activation function layer which are sequentially connected.

Further, the discriminator constructed in step S2-2 includes six convolution modules connected in sequence; the first convolution module comprises a one-dimensional convolution operation layer and a leakage relu function activation layer which are sequentially connected; the second convolution module comprises a one-dimensional convolution operation layer and a leakage relu function activation layer which are sequentially connected, wherein the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module are identical in structure; the sixth convolution module comprises a reshape matrix conversion operation layer and a full-connection operation layer which are sequentially connected.

Further, the Q network constructed in step S2-2 includes seven convolution modules connected in sequence; wherein the first five convolution modules are shared with the first five convolution modules of the discriminator; the sixth convolution module comprises a reshape matrix conversion operation layer, a full-connection operation layer, a batchnorm normalization layer and a leakage relu function activation layer which are sequentially connected; the seventh convolution module includes a fully connected operational layer.

Further, the specific method for creating the loss function in step S4 is as follows:

according to the formula:

L_l(G,Q)＝H(c)+E_{c～P(c),x～G(z,c)}[logQ(c|x)]

obtaining a loss functionWhere V (G, D) is the objective function of WaveGAN, G is the generator, D is the discriminator, and λ isAdjustable hyperparameter, L_l(G, Q) is a mutual information lower bound, and Q is a Q network; x-Pdata (x) represents the distribution of the input data x sampled in the real data, E [ ·]For the expected value, ln is a logarithmic function based on natural logarithm, D (-) is the probability that x comes from the real data, z-P_z(z) represents that a distance unit vector z is sampled in prior distribution, G (-) is generated data, H (-) is entropy of variable, c-P (c) is hidden variable c according with hidden variable distribution, x-G (z, c) is input data according with condition to generate data distribution, log is a logarithmic function with 10 as a base, and Q (c | x) is variable distribution.

Provided is radar clutter simulation electronic equipment, the equipment comprising:

a memory storing executable instructions; and

a processor configured to execute the executable instructions in the memory to implement a radar clutter simulation method.

The invention has the beneficial effects that:

1. the method is an end-to-end countermeasure generation method, overcomes the defect that the traditional clutter modeling simulation method introduces intermediate process errors, has stronger distribution model fitting capacity, and generates clutter more vividly.

2. The method ignores the distribution type of the input radar clutter, can be applied to the generation of the radar clutter with any distribution type, and increases the universality of the method.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a schematic top view of a radar surface clutter mechanism;

FIG. 3 is a side view of a radar surface clutter mechanism;

FIG. 4 is a general block diagram of an info-WaveGAN training model;

FIG. 5 is a diagram of a generator structure of the info-WaveGAN training model;

FIG. 6 is a diagram of a discriminator structure of an info-WaveGAN training model;

FIG. 7 is a Q network structure diagram of the info-WaveGAN training model;

FIG. 8 is a comparison graph of the clutter MMD index under three sea conditions of 2-4 levels selected in the experiment.

Detailed Description

The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.

As shown in fig. 1, the radar clutter simulation method includes the following steps:

s1, collecting original clutter data, preprocessing the original clutter data, and acquiring and splitting a distance unit of the preprocessed clutter data;

s2, constructing an info-WaveGAN training model containing a generator, a discriminator and a Q network;

s3, inputting each split distance unit as a training sample into an info-WaveGAN training model, and performing model training through a generator, a discriminator and a Q network in sequence;

s4, creating a loss function, and performing optimization processing on the parameters of the info-WaveGAN training model through circularly executing the operation of the minimum loss function to obtain optimized clutter data and complete radar clutter simulation.

The specific method of step S1 is:

s1-1, collecting original clutter data, extracting the distance of the original clutter data, and multiplying the distance by the power of three-thirds of the distance on the basis of the original clutter data to obtain clutter data with offset distance attenuation;

s1-2, judging whether the ground-scraping angle secant value is close to 1, if so, entering the step S1-3; otherwise, dividing the clutter data with the offset distance attenuation by one half of the secant value of the ground wiping angle to obtain clutter data with the influence of the ground wiping angle eliminated;

s1-3, judging whether a target echo exists in the current clutter data, if so, entering a step S1-4; otherwise, entering step S1-5;

s1-4, judging whether the position of the target echo is obtained or not, if so, eliminating the distance unit echo of the target echo from the clutter data, and entering the step S1-5; otherwise, eliminating the peak value of the target echo by a windowing filtering method, and entering the step S1-5;

s1-5, acquiring a distance unit and a pulse unit of the current clutter data, performing global normalization, completing preprocessing, and entering the step S1-6;

and S1-6, splitting each distance unit of the preprocessed clutter data.

As shown in fig. 2 and 3, the preprocessing principle can be analyzed according to the collected data:

according to the formula:

A_c＝Rθ_B(cτ/2)sec(ψ)

obtaining the clutter power P of the radar surface_c(ii) a Wherein, P_tIs the transmitted wave power of the radar, G is the gain of the radar, A_cIs an irradiation area, L_rFor combined losses, σ, of radar systems⁰As scattering coefficient, theta_BThe azimuth resolution of the radar, c is the propagation speed of radar signals in the air, tau is the pulse width emitted by the radar, sec is a secant function, psi is a grazing angle, pi is a circumferential rate, and R is the detection distance of the radar;

from the above formula, the surface clutter power is mainly affected by the distance, antenna parameters, scattering coefficient and ground rubbing angle.

As shown in fig. 4, D is a discriminator for determining whether an input is true or false, x is an input true sample, G is a generator, and a c vector (hidden variable) and a z' vector (random vector) are spliced into a z vector (distance unit vector) and input into a G network for generating a false sample G (z, c), i.e., a condition generating data distribution. In order to reduce the parameter amount of the network, the Q network shares all network parameters of the discriminator D except the last full connection layer, and the full connection layer output of the Q network is c' (hidden variable recovered by the Q network), and the size of the full connection layer output of the Q network is consistent with that of the hidden variable c.

As shown in fig. 5, the generator constructed in step S2-1 includes six deconvolution modules connected in sequence; the first deconvolution module comprises a full-connection operation layer, a reshape matrix conversion operation layer, a batch normalization operation layer and a relu activation function layer which are sequentially connected; the second deconvolution module comprises a one-dimensional deconvolution operation layer, a batch normalization operation layer and a relu activation function layer which are sequentially connected, wherein the second deconvolution module, the third deconvolution module, the fourth deconvolution module and the fifth deconvolution module have the same structure; the sixth deconvolution module comprises a one-dimensional deconvolution operation layer and a Tanh activation function layer which are sequentially connected.

The convolution kernel structure parameters in each deconvolution module are shown in table 1, the size of each convolution kernel is [ M, N, L ], M represents the length of a one-dimensional convolution kernel, except that the length of a full-connection layer convolution kernel is 1, the length of other deconvolution kernels is 25; n represents the depth of convolution kernel, and the size of the convolution kernel is consistent with the depth of the deconvolved feature map matrix; l represents the number of convolution kernels, and each convolution kernel should output a feature map with the size consistent with the depth of the output feature map matrix. The magnitude of the last offset is consistent with the magnitude of L, and an offset is added to the characteristic diagram representing each output. The generator model has an input data size of [64,100] and an output size of [64,4096,2 ].

TABLE 1

Hierarchy level	In the layer	Convolution kernel	Biasing
				1	Dense layer	[1,100,4096]	[4096]
2	conv1d _ transpose layer	[25,1024,512]	[512]
				3	conv1d _ transpose layer	[25,512,256]	[256]
4	conv1d _ transpose layer	[25,256,128]	[128]
				5	conv1d _ transpose layer	[25,128,64]	[64]
6	conv1d _ transpose layer	[25,64,2]	[2]

As shown in fig. 6, the discriminator constructed in step S2-2 includes six convolution modules connected in sequence; the first convolution module comprises a one-dimensional convolution operation layer and a leakage relu function activation layer which are sequentially connected; the second convolution module comprises a one-dimensional convolution operation layer and a leakage relu function activation layer which are sequentially connected, wherein the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module are identical in structure; the sixth convolution module comprises a reshape matrix conversion operation layer and a full-connection operation layer which are sequentially connected.

The convolution structure parameters of the convolution module are shown in table 2, the input data size of the discriminator model is [64,4096,2], and the output size is [64,1 ].

TABLE 2

Hierarchy level	In the layer	Convolution kernel	Biasing
				1	Layer of conv1d	[25,2,64]	[64]
2	Layer of conv1d	[25,64,128]	[128]
				3	Layer of conv1d	[25,128,256]	[256]
4	Layer of conv1d	[25,256,512]	[512]
				5	Layer of conv1d	[25,512,1024]	[1024]
6	Dense layer	[1,4096,1]	[1]

As shown in fig. 7, the Q network constructed in step S2-2 includes seven convolution modules connected in sequence; wherein the first five convolution modules are shared with the first five convolution modules of the discriminator; the sixth convolution module comprises a reshape matrix conversion operation layer, a full-connection operation layer, a batchnorm normalization layer and a leakage relu function activation layer which are sequentially connected; the seventh convolution module includes a fully connected operational layer.

Wherein, the input data size of the Q network is [64,4096,2], and the output size is [64,3 ]. And the second dimension of the Q network output data is consistent with the hidden variable c of the model input noise vector. The hidden variable c can be divided into two parts, one part is a discrete variable which can classify data, generally, a one-hot vector is used here, each element in the one-hot vector represents a category, only one element in the one-hot vector is 1, and the rest elements are all 0, so that the hidden variable c can also be called a one-hot vector. The other part is a continuous variable that can represent other continuously varying features, typically a random vector in the range of [ -1,1 ]. In the training of the scheme, because the clutter data of the training set are the clutter of 2-4 levels of sea states, three points, namely one-hot vectors capable of representing three label states, are used in the training, and the other 2 points are continuous variables and represent continuous change information of the sea clutter. After the output of the Q network is obtained, further processing is required to be used as the recovered hidden variable c'. For a discrete variable part, the value of the discrete variable part needs to be mapped to [0,1] through a softmax function; for the continuous variable part, the mean and variance thereof need to be normalized.

The specific method for creating the loss function in step S4 is as follows:

according to the formula:

L_l(G,Q)＝H(c)+E_{c～P(c),x～G(z,c)}[logQ(c|x)]

obtaining a loss functionWherein V (G, D) is an objective function of WaveGAN, G is a generator, D is a discriminator, lambda is an adjustable hyper-parameter, and L_l(G, Q) is a mutual information lower bound, and Q is a Q network; x-Pdata (x) represents the distribution of the input data x sampled in the real data, E [ ·]For the expected value, ln is a logarithmic function based on natural logarithm, D (-) is the probability that x comes from the real data, z-P_z(z) represents that a distance unit vector z is sampled in prior distribution, G (-) is generated data, H (-) is entropy of variable, c-P (c) is hidden variable c according with hidden variable distribution, x-G (z, c) is input data according with condition to generate data distribution, log is a logarithmic function with 10 as a base, and Q (c | x) is variable distribution.

The radar clutter simulation electronic device includes:

a memory storing executable instructions; and

a processor configured to execute the executable instructions in the memory to implement a radar clutter simulation method.

In one embodiment of the invention, the original data under three sea conditions of 2-4 levels are adopted for simulation:

the experimental simulation test platform is Ubuntu18.04, the system is provided with NVIDA GeForce RTX2080Ti GPU, a tensoflow frame is used for model building, and NVIDA Cuda is used for accelerating calculation. Model optimization selection Using Adam optimization algorithm, Learningrate was set to 1e-5, beta1 was set to 0.5, and beta2 was set to 0.9. The Q network and the generator share an Adam optimizer. During training, the learning strategy of the discriminator, the generator and the Q network is set to be 1:1, namely 1 round of discriminator parameter updating is followed by 1 round of generator and Q network parameter updating.

As shown in fig. 8 and table 3, in the condition-controlled experiment, the present scheme was compared with the MMD index. And respectively and randomly sampling 3 groups of clutter training samples of 2-4 levels of sea conditions, wherein each group comprises 64 sea clutter samples, then estimating, fitting and simulating each real sample by using a traditional method, and then evaluating the difference between the simulated distribution and the real distribution of each group of samples by using an MMD (mass-median-distribution) index. And (3) generating sea clutter of 3 sea condition levels respectively by using the trained condition control model, wherein each sea condition needs 3 groups, each group comprises 64 generated sea clutter samples, and the generated sea clutter samples and the real samples are evaluated respectively by using the MMD index. Table 3 is a MMD quantitative comparison table of the sea clutter generation samples of the 2-4 levels of sea conditions and the real samples, fig. 8 is a MMD index comparison table of the 2-4 levels of sea conditions, and it can be seen from the above-mentioned table and table that the condition control clutter generation method based on the info-WaveGAN training model proposed by the present invention generates clutter at all levels of sea conditions with a smaller MMD value than the clutter generated by using the conventional simulation method, indicating that the generated clutter distribution of the new method proposed by the present invention more conforms to the real clutter distribution.

TABLE 3

The method is an end-to-end countermeasure generation method, overcomes the defect that the traditional clutter modeling simulation method introduces intermediate process errors, has stronger distribution model fitting capacity, and generates clutter more vividly. The method ignores the distribution type of the input radar clutter, can be applied to the generation of the radar clutter with any distribution type, and increases the universality of the method.