CN115015869A - Can learn low frequency broadband radar target parameter estimation method, equipment and program product - Google Patents

Abstract

The application relates to the field of radar signal processing, and relates to a learnable low-frequency broadband radar target parameter estimation method, equipment and a program product. The method comprises the following steps: acquiring data collected by a low-frequency broadband radar; and inputting the data acquired by the low-frequency broadband radar into a pre-trained parameter estimation network to obtain a target parameter estimation value output by the parameter estimation network, wherein the parameter estimation network is obtained by training a neural network by taking a plurality of training data pairs with random phases and noise as training samples. The neural network constructed by training the training sample based on the network layer structure and the data enhancement of the cyclic convolution adopts the loss function at least comprising the target background ratio loss function, and continuously optimizes learnable parameters in the neural network to obtain the parameter estimation network for processing the low-frequency broadband radar signal, so that the calculation complexity is reduced, the algorithm efficiency and the generalization capability are improved, and the target significance is higher in the low signal-to-noise ratio scene.

Description

Can learn low frequency broadband radar target parameter estimation method, equipment and program product

technical field

The present application relates to the field of radar signal processing, and in particular, to a method, device and program product for learning low-frequency broadband radar target parameter estimation.

Background technique

Low-frequency broadband radar signals have the ability to resist stealth, penetrate walls and soil, and have important applications in the defense and civilian fields. The estimation of radar target parameters based on the geometric theory of diffraction (GTD) model is a key step in the processing of low-frequency broadband radar range signals.

However, the existing GTD parameter estimation methods have problems such as high computational complexity and complex hyperparameter adjustment, resulting in low algorithm efficiency of the existing GTD parameter estimation methods, and weak generalization ability for low-frequency broadband radar data in different scenarios; In addition, in the existing GTD parameter estimation methods, the signal-to-noise ratio has a serious impact on the performance of the algorithm, resulting in a low target saliency in low signal-to-noise ratio scenarios and a significant drop in parameter estimation performance.

SUMMARY OF THE INVENTION

The embodiments of the present application provide a method, device, and program product for estimating target parameters of a learnable low-frequency broadband radar, aiming to solve the problems of low algorithm efficiency, complex hyperparameter adjustment, weak generalization ability, and low performance in the prior art. The problem of low target saliency in signal-to-noise ratio scenarios.

A first aspect of the embodiments of the present application provides a method for estimating target parameters of a learnable low-frequency broadband radar, including:

Obtain data collected by low-frequency broadband radar;

Input the data collected by the low-frequency broadband radar into a pre-trained parameter estimation network to obtain the target parameter estimation value output by the parameter estimation network, wherein the parameter estimation network uses a plurality of training data with random phases and noise as the The training samples are obtained by training the neural network based on the recurrent convolutional network layer structure.

Optionally, the neural network includes:

The input module is used to convert the complex-valued observation vector in the training sample input into the neural network into a real-valued observation vector, and use the real-valued observation vector and the real-valued GTD dictionary matrix to calculate the initial value of the neural network. input vector;

an iterative optimization module for performing multiple iterative optimizations on the initial input vector of the neural network to obtain a real-valued output vector of the neural network;

The output module is used to convert the real-valued output vector of the neural network into a complex-valued output vector, and output it as the predicted value of the target parameter.

Optionally, the iterative optimization module includes:

所述可学习变换结构

是基于循环卷积的网络层结构，对所述神经网络的初始输入向量进行迭代优化，输出所述神经网络的实数值输出向量，其中，所述可学习变换结构

包含变形算子

循环卷积层CC、线性整流单元ReLU；learnable transformation structure

The learnable transform structure

is a network layer structure based on circular convolution, iteratively optimizes the initial input vector of the neural network, and outputs the real-valued output vector of the neural network, wherein the learnable transformation structure

Contains deformation operators

Circular convolution layer CC, linear rectification unit ReLU;

in,

用于将所述神经网络在迭代优化过程中产生的向量x变形为所述神经网络在迭代优化过程中产生的矩阵z，其中，z用于进行所述循环卷积运算；deformation operator

for transforming the vector x generated by the neural network in the iterative optimization process into a matrix z generated by the neural network in the iterative optimization process, where z is used to perform the circular convolution operation;

用于将所述神经网络在迭代优化过程中产生的矩阵z变形为所述神经网络在迭代优化过程中产生的向量_x；Inverse shape operator

for transforming the matrix z generated by the neural network in the iterative optimization process into a vector _x generated by the neural network in the iterative optimization process;

用于将

映射为

其中，h＝CC_K,P,Q(z；w,b)，

n mod L表示在1到L之间模L同余于_n的整数；A recurrent convolutional layer CC containing the learnable network parameters

for the

maps to

where, h=CC _K,P,Q (z; w,b),

n mod L represents an integer between 1 and L that modulo L is congruent to _n ;

Linear rectification unit ReLU, defined as [ReLU(x)] _i =max(0,x _i ).

Optionally, the training data is a pair of a complex-valued observation vector and a complex-valued parameter true value vector, and the training data is generated according to the following steps:

Generate a set of scattering center parameters;

Calculate the true value vector of the complex-valued parameter according to the scattering center parameter in the scattering center parameter set;

Calculate a complex-valued observation vector according to the complex-valued parameter truth value vector and the complex-valued GTD dictionary matrix, and combine the complex-valued observation vector and the complex-valued parameter truth value vector into a complex-valued observation vector and a complex-valued parameter truth value vector pair, as a set of the training data.

Optionally, the set of scattering center parameters is S={(l _n ,α _n ,σ _n ):n=1,...,N _s }, where,

l _n is the distance unit, which is defined as an _integer from ₁ to L that _is randomly selected without replacement; _n is the scattering coefficient, the amplitude of σ _n |σ _n | obeys the distribution of the preset amplitude, and the argument ∠σ _n of σ _n obeys the distribution of the preset argument.

Optionally, calculating the true value vector of the complex-valued parameter according to the scattering center parameter in the scattering center parameter set, including:

calculate

Wherein, l=1,...,L, α belongs to the preset frequency-dependent factor set {α ₁ ,α ₂ ,...,α _J };

组合成所述复数值参数真值向量

定义为：will get

combined into the complex-valued parameter truth vector

defined as:

Calculate the complex-valued observation vector according to the complex-valued parameter truth vector and the complex-valued GTD dictionary matrix, including:

其中，Φ为所述复数值GTD字典矩阵。Using the complex-valued GTD dictionary matrix, convert the complex-valued parameter truth vector into the complex-valued observation vector:

Wherein, Φ is the complex-valued GTD dictionary matrix.

Optionally, training the neural network includes:

In each different training round ep, for each set of the training data, different random phases and noises are added to the complex-valued observation vectors in the training data, and different random phases and noises are added to the complex-valued observation vectors in the training data Different random phases are added to the true value vector of the complex-valued parameter as the training sample;

Input the complex-valued observation vector in the training sample into the neural network, and calculate the loss function corresponding to the training sample according to the predicted value of the target parameter output by the neural network and the complex-valued true value vector in the training sample. value, wherein the value of the loss function includes at least the value of the target-background ratio loss function;

The gradient of the loss function to the learnable network parameter is calculated, and the learnable network parameter is optimized based on the gradient of the loss function to the learnable network parameter of the neural network.

Optionally, the target-background ratio loss function is defined as:

为所述目标背景比损失函数的值，TBR为所述目标背景比，

为目标区域，

为背景区域，N_T为A_T的元素个数，N_B为A_B的元素个数，

为所述目标参数预测值中的元素。in,

is the value of the target-background ratio loss function, TBR is the target-background ratio,

is the target area,

is the background area, N _T is the number of elements of A _T , N _B is the number of elements of A _B ,

element in the predicted value for the target parameter.

定义的；Optionally, the loss function further includes one or more of a dissimilarity loss function and a symmetry loss function, wherein the symmetry loss function is an intermediate variable sum in an iterative optimization module based on the neural network. The learnable transform structure

Defined;

When the loss function includes the target-to-background ratio loss function, the difference loss function and the symmetry loss function, the total loss function value is calculated according to the following formula:

为总损失函数值，

为所述目标背景比损失函数的值，

所述差异性损失函数的值，

为所述对称性损失函数的值，λ₁、λ₂为预设平衡系数。in,

is the total loss function value,

is the value of the target-background ratio loss function,

the value of the dissimilarity loss function,

are the values of the symmetry loss function, and λ ₁ and λ ₂ are preset balance coefficients.

A second aspect of the embodiments of the present application provides an electronic device for parameter estimation, including a memory, a processor, and a computer program stored in the memory, where the processor executes the computer program to implement the learnable learning method proposed by the embodiments of the present application Steps in a low-frequency broadband radar target parameter estimation method.

A third aspect of an embodiment of the present application provides a computer program product, including a computer program/instruction, when the computer program/instruction is executed by a processor to implement the method for estimating target parameters of a learnable low-frequency broadband radar proposed by the embodiment of the present application A step of.

Beneficial effects:

The present application provides a method, equipment and program product for estimating target parameters of a learnable low-frequency broadband radar. By training a neural network constructed based on data augmentation training samples, the learnable parameters in the neural network are continuously optimized to obtain parameters for the estimation network. For processing low frequency broadband radar signals, it has the following advantages:

(1) The network layer structure based on circular convolution is adopted, which reduces the computational complexity and improves the efficiency and accuracy of parameter estimation.

(2) Using multiple training data with random phase and noise as training samples, the parameter estimation network obtained by training the neural network is used for parameter estimation, and the parameter estimation network obtained by training optimization can process different low-frequency broadband radar signal data , which improves the accuracy and generalization ability of the parameter estimation algorithm.

(3) The training data in the training data set used to train the neural network is introduced into data augmentation of random phase and noise as training samples, and the neural network is trained with a loss function including at least the target-background ratio loss function, so that the parameter estimates obtained by the training optimization are used. The network can process low-frequency broadband radar signal data well in low signal-to-noise ratio scenarios, effectively improving the saliency of the target.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present application more clearly, the following briefly introduces the drawings that are used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application. , for those of ordinary skill in the art, other drawings can also be obtained from these drawings without creative labor.

1 is a flowchart of a method for estimating target parameters of a learnable low-frequency broadband radar proposed by an embodiment of the present application;

FIG. 2 is a flowchart of a neural network training method for learning low-frequency broadband radar target parameter estimation proposed by an embodiment of the present application;

3 is a line graph of the signal-to-noise ratio-target background ratio proposed by an embodiment of the present application;

FIG. 4 is a broken line graph of signal-to-noise ratio-mean square error proposed by an embodiment of the present application;

Detailed ways

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

In the related art, the GTD parameter estimation method based on the geometric theory of diffraction (GTD) model is adopted. However, the existing GTD parameter estimation method has problems such as high computational complexity and complex hyperparameter adjustment, which leads to the existing The algorithm efficiency of the GTD parameter estimation method is low, and the generalization ability of low-frequency broadband radar data in different scenarios is weak; in addition, in the existing GTD parameter estimation methods, the signal-to-noise ratio has a serious impact on the performance of the algorithm, resulting in low In the scenario of signal-to-noise ratio, the target saliency is low, and the parameter estimation performance drops significantly.

In view of this, an embodiment of the present application proposes a method for estimating target parameters of a learnable low-frequency broadband radar, which aims to solve the problems of low efficiency, weak generalization ability, and target salience in low signal-to-noise ratio scenarios existing in the prior art. lower problem. The parameter estimation method of the present application will be described in detail below.

FIG. 1 is a flowchart of a method for estimating target parameters of a learnable low-frequency broadband radar according to an embodiment of the present application. Referring to FIG. 1 , the method for estimating target parameters of a learnable low-frequency broadband radar proposed by an embodiment of the present application includes the following steps:

S101. Acquire data collected by a low-frequency broadband radar.

In the specific implementation, the following data are obtained from the data collected by the low-frequency broadband radar:

Radar observation frequency {f _m : m=1,...,M}, frequency interval Δf, satisfying f _m =f ₁ +(m-1)Δf; spectrum of radar observation signal {E(f _m ):m=1, ..., M}; the number of distance units L in the observation scene, L>M is required.

S102. Input the data collected by the low-frequency broadband radar into a pre-trained parameter estimation network.

During specific implementation, the radar observation frequency, the frequency spectrum of the radar observation signal, and the number of distance units of the observation scene in the data collected by the low-frequency broadband radar are input into the pre-trained parameter estimation network DNN. The data collected by the input low-frequency broadband radar is assigned to the corresponding quantity in the parameter estimation network, and the spectrum {E(f _m ):m=1,...,M} of the input radar observation signal is combined into the radar spectrum signal vector e :

The parameter estimation network DNN is obtained by training the neural network with multiple training data with random phase and noise as training samples. For the training method of the neural network, please refer to the steps S201-S203 of training the neural network in the following parameter estimation method, which will not be repeated here.

S103, the parameter estimation network outputs the target parameter estimation value.

并输出目标参数估计值

During specific implementation, the solution is performed according to the preset algorithm in the parameter estimation network DNN and the optimized learnable network parameter Θ _opt , and the target parameter estimation vector is obtained.

and output the target parameter estimates

其元素为

其定义与下述参数真值向量相同，此处不再赘述。当

时，表示第l个距离单元中存在频率依赖因子为α的散射中心，其散射系数估计值为

当

时，表示第l个距离单元中不存在频率依赖因子为α的散射中心。target parameter vector

Its elements are

Its definition is the same as the following parameter truth vector, so it will not be repeated here. when

When , it means that there is a scattering center with a frequency-dependent factor α in the l-th distance unit, and its estimated scattering coefficient is

when

, it means that there is no scattering center with a frequency-dependent factor α in the l-th distance unit.

FIG. 2 is a flowchart of a neural network training method for estimating target parameters of a learnable low-frequency broadband radar shown in an embodiment of the present application. Referring to FIG. 2 , the training process of a neural network in the method for estimating target parameters of a learnable low-frequency wideband radar proposed by an embodiment of the present application includes: Follow the steps below:

S201. Construct GTD dictionary and neural network.

In specific implementation, the GTD dictionary is first constructed. The GTD dictionary includes a complex-valued GTD dictionary matrix and a real-valued GTD dictionary matrix.

定义为：Build a matrix of complex-valued GTD dictionaries

defined as:

的第m行、第l列元素(m＝1,…,M,l＝1,…,L)定义为：Among them, Φ ^(α) ,α∈{α ₁ ,α ₂ ,…,α _J } are J sub-dictionary matrices,

The m-th row and l-th column elements (m=1,...,M,l=1,...,L) are defined as:

j is the imaginary unit, c is the speed of light, f _C =f ₁ +(M-1)Δf/2 represents the center frequency, and r _l =(l/L)×c/(2Δf) represents the lth distance unit.

定义为Build a matrix of real-valued GTD dictionaries

defined as

where Re(·) and Im(·) represent the real and imaginary part operations, respectively.

By constructing a GTD dictionary, the data collected by the low-frequency broadband radar in the actual scene is converted into a subsequent mathematical model, so that the subsequent neural network operation mathematical model is processed separately from the previous actual model, thereby reducing the complexity of the data processing in the embodiment of the present application.

本申请实施例中采用的是Python编程语言和PyTorch深度学习框架构建神经网络DNN，神经网络DNN可以采用现有技术中的编程语言和深度学习框架，对此本申请不做限制。神经网络DNN包含输入模块、迭代优化模块、和输出模块。Build a neural network DNN. The neural network DNN is the above-mentioned parameter estimation network DNN that has not been optimized, and includes a learnable network parameter Θ, a learnable transformation

In the embodiments of this application, the Python programming language and the PyTorch deep learning framework are used to construct the neural network DNN, and the neural network DNN may use the programming language and deep learning framework in the prior art, which is not limited in this application. The neural network DNN includes an input module, an iterative optimization module, and an output module.

The input module is used to convert the complex-valued observation vector in the training sample input into the neural network into a real-valued observation vector, and use the real-valued observation vector and the real-valued GTD dictionary matrix to calculate the initial value of the neural network. input vector.

During specific implementation, the complex-valued observation vector in the training sample is input. The training sample is obtained by introducing random phase and noise into the training data in the training data set, and the specific process of generating the training data set is shown in the following S202.

The input complex-valued observation vector e (here e=e _i,ep during training, e _i,ep is the complex-valued observation vector with random phase and noise added, see the following for the specific definition, and e is the input during testing The complex-valued observation vector of ) is converted into the corresponding real-valued observation vector b:

And use the real-valued GTD dictionary matrix to calculate the initial value as the initial input vector of the neural network:

y ⁽¹⁾ = x ⁽⁰⁾ = Ψ ^T (ΨΨ ^T ) ^-1 b

The iterative optimization module is configured to perform multiple iterative optimizations on the initial input vector of the neural network to obtain a real-valued output vector of the neural network.

During specific implementation, the real-valued output vector of the neural network is subjected to N _p =15 iterations, and the following iterative steps are repeated for k=1, . . . , N _p :

r ^(k) = y ^{(k) -} μ ^(k) Ψ ^T (Ψy ^(k) - b)

y ^(k+1) = x ^{(k) -} ρ ^(k) (x ^{(k) -} x ^(k-1) )

定义如下：Among them, including μ ^(k) , θ ^(k) , ρ ^(k) ,

Defined as follows:

μ ^(k) , θ ^(k) , ρ ^(k) are defined as:

其中，μ ^(k) =sp(a ₁ k+c ₁ ), θ ^(k) =sp(a ₂ k+c ₂ ),

in,

sp(x)=ln(1+exp(x)), a ₁ , a ₂ , a ₃ , c ₁ , c ₂ , c ₃ belong to the learnable network parameter Θ ₁ (the learnable network parameter Θ in the neural network DNN See the definition below, the same below), S _θ is a threshold shrinkage function with parameter θ, which is an element-wise function, defined as:

为可学习变换，所述可学习变换结构

是基于循环卷积的网络层结构，对所述神经网络的初始输入向量进行迭代优化，输出所述神经网络的实数值输出向量，其中，所述可学习变换结构

包含变形算子

循环卷积层CC、线性整流单元ReLU；

is a learnable transform, the learnable transform structure

is a network layer structure based on circular convolution, iteratively optimizes the initial input vector of the neural network, and outputs the real-valued output vector of the neural network, wherein the learnable transformation structure

Contains deformation operators

Circular convolution layer CC, linear rectification unit ReLU;

(卷积层层数为5)结构如下：In some embodiments, the transform structure can be learned

(The number of convolutional layers is 5) The structure is as follows:

in,

定义为：变形算子

将向量

变形为矩阵

相应的逆变形算子

将矩阵

变形为向量

满足z_d,l＝x_(d-1)L+l，其中d＝1,…,2J,l＝1,…,L；deformation operator

Defined as: deformation operator

the vector

morph into a matrix

Corresponding Inverse Transform Operator

put the matrix

morph to vector

Satisfy z _d,l =x _(d-1)L+l , where d=1,...,2J,l=1,...,L;

用于将

映射为

其中，The recurrent convolutional layer CC contains learnable network parameters

for the

maps to

in,

h=CC _K,P,Q (z;w,b),

n mod L represents an integer between 1 and L whose modulus L is congruent to _n , p=1,...,P;

The linear rectification unit ReLU is an element-wise function, which is defined as: [ReLU(x)] _i =max(0,x _i )

用于上述定义的可学习变换

The learnable network parameter Θ in the neural network DNN is defined as: Θ=Θ ₁ ∪Θ ₂ , where Θ ₁ ={a ₁ ,a ₂ ,a ₃ ,c ₁ ,c ₂ ,c ₃ }, for the above Calculate μ ^(k) , θ ^(k) , ρ ^(k) ;

A learnable transform for the above definition

The above learnable transformable convolution layer layers, convolution kernel size K, feature channel number _NF , and element number J of the frequency-dependent factor set are all preset in advance, and the specific values are not limited in this application. The number of convolution layers used in the embodiment of the present application is 5, the size of the convolution kernel is K=3, the number of feature channels is _NF =32, and the number of elements of the frequency-dependent factor set is J=5.

的迭代次数，减小了计算量，提升了算法的效率。The embodiment of the present application introduces an intermediate vector y, and through the operation of the intermediate vector y, the real-valued output vector is reduced

The number of iterations is reduced, the amount of calculation is reduced, and the efficiency of the algorithm is improved.

The output module is used to convert the real-valued output vector of the neural network into a complex-valued output vector, and output it as the predicted value of the target parameter.

按照下式转化为复数值输出向量：In specific implementation, the real value output vector output by the above iterative optimization module is

Convert to a complex-valued output vector as follows:

进行输出。use the complex-valued output vector as the target parameter predictor

to output.

S202. Generate a complex-valued observation vector-complex-valued parameter true value vector pair as training data to form a training data set.

During specific implementation, a training data set D={(e _i ,σ _i ):i=1,...,N _D } is generated, wherein the training data is a complex-valued observation vector-complex-valued parameter true value vector pair (e _i ,σ _i ), the number ND of the training data is preset in advance, and the specific value of the number of training samples is not limited in the present application, and the number used in the embodiment of the present application is _ND =50000. Each set of training data ( _{ei ,σ i )} is generated according to the following steps:

generating the number of scattering centers N _s , where N _s is a random integer within a preset range, which is not limited in this application, and in this embodiment, N _s is preset as a random integer ranging from 1 to 15;

A set of scattering center parameters is generated, and the set of scattering center parameters is S={(l _n ,α _n ,σ _n ):n=1,...,N _s }, where _ln is a distance unit, which is defined as no in 1～L Integers drawn randomly by replacement, l _n are different; α _n is the frequency-dependent factor, α _n is randomly selected from the preset frequency-dependent factor set; σ _n is the scattering coefficient, the amplitude of σ _n |σ _n | It obeys the distribution of preset amplitude, and the argument ∠σ _n of σ _n obeys the distribution of preset argument.

The specific values of the preset frequency-dependent factor set, the distribution of the preset amplitude, and the distribution of the preset argument are not limited in this application. The frequency-dependent factor set used in the embodiments of the present application is {-1, -1/2, 0, 1/2, 1}, the number of elements in the frequency-dependent factor set is J=5, and the amplitude obeys a uniform distribution U(0.5, 1.5), the argument obeys the uniform distribution U(0,2π).

Calculate the true value vector of the complex-valued parameter according to the scattering center parameter in the scattering center parameter set, and calculate

Wherein, l=1,...,L, α belongs to the preset frequency-dependent factor set {α ₁ ,α ₂ ,...,α _J };

组合成所述复数值参数真值向量

定义为：will get

combined into the complex-valued parameter truth vector

defined as:

将每个得到的复数值观测向量和复数值参数真值向量组成的一个复数值观测向量-复数值参数真值向量对，作为一组训练数据，所有训练数据组成训练数据集。Calculate a complex-valued observation vector according to the complex-valued parameter truth vector and complex-valued GTD dictionary matrix, and use the complex-valued GTD dictionary matrix to convert the complex-valued parameter truth vector into the complex-valued observation vector:

A complex-valued observation vector-complex-valued parameter true-value vector pair composed of each obtained complex-valued observation vector and a complex-valued parameter true value vector is used as a set of training data, and all training data constitute a training data set.

S203: Calculate a loss function based on the training samples, optimize the learnable network parameters based on the value of the loss function, and train the neural network to obtain a parameter estimation network DNN.

When implementing, follow the steps below:

Initialize the neural network DNN and optimizer, the learnable network parameters in the neural network DNN Θ=Θ ₁ ∪Θ ₂ , where Θ ₁ ={a ₁ ,a ₂ ,a ₃ ,c ₁ ,c ₂ ,c ₃ } are initialized respectively As the initial value, an algorithm is used to initialize the learnable network parameter Θ ₂ ; an optimizer is used to set the batch size, initial learning rate, and learning rate decay to preset values.

The initial value of the above-mentioned learnable network parameter Θ ₁ and the values of the optimizer batch size, initial learning rate, and learning rate decay are all preset in advance, and the specific values are not limited in this application; the algorithm for the initialization of the learnable network parameter Θ ₂ It is the prior art, and the specific content is not limited in this application. The initial value of the learnable network parameter Θ1 adopted in the embodiment of the present application is {0.5, 0.2, 1, ₂ , ₁ , 0}, the Xavier algorithm is used to initialize the learnable network parameter Θ2, the Adam optimizer is used, and the preset The value is the batch size of 32, the preset learning rate is 0.001, and the ep learning rate decays to 0.8 every 5 training epochs.

Data augmentation, for each of the training data (ei ,σ _i ), adding a random phase φ i,ep and noise ni,ep to the complex-valued observation vector _ei in the training data, and adding a random phase φ _{i ,ep} and noise _ni,ep to the training data Add the random phase φ _i,ep to the true value vector σ _i of the complex-valued parameter in , to obtain the training sample ( _ei,ep ,σ _i,ep ), which is specifically carried out according to the following formula:

e _i,ep = e _i ·exp(jφ _i,ep )+n _i,ep

σ _i,ep =σ _i ·exp(jφ _i,ep )

Among them, φ _i,ep obeys a preset distribution, the type of noise _{ni,ep and} the signal-to-noise ratio satisfied by noise ni,ep are preset in advance, and the specific values are not limited in this application. In the embodiment of the present application, φ _i,ep obeys the uniform distribution U(0,2π), and the noise _ni,ep is Gaussian additive white noise, which satisfies the signal-to-noise ratio SNR=5dB.

By introducing data enhancement with different random phases and noises, the (e _i ,σ _i ) in the original training data set can be enhanced multiple times with different phases. Therefore, the parameter estimation network obtained by training and optimization based on the data-enhanced training sample has higher accuracy and is closer to the true value; in addition, the noise-introduced data is used as the training sample, so that the optimized parameters are trained based on the data-enhanced training sample. Estimation network can have better processing ability for data in low signal-to-noise ratio environment and effectively improve target saliency.

计算所述训练样本对应的损失函数的值，其中，所述损失函数的值至少包括目标背景比损失函数的值。Input the e _i,ep in the training samples (e _i,ep ,σ _i,ep ) into the neural network DNN, and output the predicted value of the target parameter based on the preset algorithm of the neural network DNN

Calculate the value of the loss function corresponding to the training sample, wherein the value of the loss function at least includes the value of the target-background ratio loss function.

The target-to-ground ratio (TBR) loss function is defined as:

为所述目标背景比损失函数的值，TBR为所述目标背景比，

为目标区域，

为背景区域，N_T为A_T的元素个数，N_B为A_B的元素个数，

为所述目标参数预测值中的元素。in,

is the value of the target-background ratio loss function, TBR is the target-background ratio,

is the target area,

is the background area, N _T is the number of elements of A _T , N _B is the number of elements of A _B ,

element in the predicted value for the target parameter.

The loss function also includes one or more of a dissimilarity loss function and a symmetric loss function.

The dissimilarity loss function (that is, the mean squared error MSE), used to measure the degree of deviation of the estimated value from the true value, is defined as:

where ||·|| ₂ represents the norm of l ₂ .

The symmetric loss function, which measures the impact of symmetric loss during training, is defined as:

表示神经网络DNN输入为e_i,ep时，对应的中间变量r^(k)。in,

Represents the corresponding intermediate variable r ^(k) when the input of the neural network DNN is e _i,ep .

When the loss function includes the target-to-background ratio loss function, the difference loss function and the symmetry loss function, the total loss function value is calculated according to the following formula:

为总损失函数值，

为所述目标背景比损失函数的值，

所述差异性损失函数的值，

为所述对称性损失函数的值，λ₁、λ₂为预设平衡系数，由于

相比

的数量级更大，因此设置系数λ₁、λ₂用于平衡

间的数量级，具体数值本申请不作限制，本申请实施例中λ₁＝0.1,λ₂＝0.001。in,

is the total loss function value,

is the value of the target-background ratio loss function,

the value of the dissimilarity loss function,

are the values of the symmetry loss function, λ ₁ and λ ₂ are preset balance coefficients, since

compared to

is an order of magnitude larger, so set the coefficients λ ₁ , λ ₂ for balance

The specific value is not limited in the present application, and in the embodiment of the present application, λ ₁ =0.1, λ ₂ =0.001.

对可学习网络参数Θ的梯度

然后用优化器优化可学习网络参数Θ。Optimize the learnable network parameters Θ in the neural network DNN, apply the backpropagation algorithm, and calculate the total loss function value

Gradient to learnable network parameter Θ

The learnable network parameters Θ are then optimized with an optimizer.

计算损失函数值、优化可学习网络参数Θ操作多个轮次，得到参数估计网络DNN。其中，所述重复轮次的数量为提前预设，本申请不作限制，本申请实施例重复共计60轮次。In each epoch ep, repeat the above data augmentation (add a different random phase and noise to each training sample ( _ei,ep ,σi _,ep ) in each different training epoch), input neural network output

Calculate the loss function value, optimize the learnable network parameter Θ for multiple rounds, and obtain the parameter estimation network DNN. Wherein, the number of the repeated rounds is preset in advance, which is not limited in the present application, and the embodiment of the present application is repeated for a total of 60 rounds.

FIG. 3 is a line graph of the signal-to-noise ratio-target background ratio shown in the embodiment of the present application. Referring to FIG. 3 , the target-to-background ratio TBR values of different algorithms under different low signal-to-noise ratio SNR conditions are shown. It can be seen from FIG. 3 that, compared with the FOCUSS algorithm, SVR algorithm, and FISTA-Net algorithm in the prior art, compared with the algorithm in the parameter estimation method (TEFISTA-Net in the figure) proposed in the embodiment of the present application, the signal-to-noise ratio is the same. The TBR values are lower than those of the time, indicating that the parameter estimation method proposed in the embodiment of the present application has stronger target saliency.

FIG. 4 is a broken line graph of signal-to-noise ratio-mean square error shown in an embodiment of the present application. Referring to FIG. 4 , the mean square error MSE values of different algorithms under different low signal-to-noise ratio SNR conditions are shown. It can be seen from FIG. 4 that, compared with the FOCUSS algorithm, SVR algorithm, and FISTA-Net algorithm in the prior art, compared with the algorithm in the parameter estimation method (TEFISTA-Net in the figure) proposed in the embodiment of the present application, the same signal-to-noise ratio is obtained. When the MSE value is higher, it means that the parameter estimation method proposed in the embodiment of the present application is less different from the expected value, and the parameter estimation value obtained by applying the parameter estimation method proposed in the embodiment of the present application is more accurate.

In addition, the corresponding performance evaluation indicators in the case of low signal-to-noise ratio (for example, SNR≤10dB) in FIGS. 3 and 4, the parameter estimation methods proposed in the embodiments of the present application all show better performance than the FOCUSS algorithm and SVR in the prior art. The better performance of the algorithm and the FISTA-Net algorithm indicates that the parameter estimation method proposed in the embodiment of the present application introduces random phase and Gaussian additive white noise to the basic training sample by performing data enhancement on the training data set for training the neural network, and A target-to-background ratio loss function is introduced for low SNR scenarios, so that the parameter estimation network obtained by training optimization can process low-frequency broadband radar signal data well in low SNR scenarios.

The present application provides a method for estimating target parameters of a learnable low-frequency wideband radar. By training a neural network constructed based on data augmentation training samples, the learnable parameters in the neural network are continuously optimized to obtain a parameter estimation network for processing low-frequency wideband radar signals. Has the following advantages:

(1) The network layer structure based on circular convolution is adopted, which reduces the computational complexity and improves the efficiency of parameter estimation.

(2) Using multiple training data with random phase and noise as training samples, the parameter estimation network obtained by training the neural network is used for parameter estimation, and the parameter estimation network obtained by training optimization can process different low-frequency broadband radar signal data , which reduces the computational complexity and improves the efficiency and generalization ability of the parameter estimation algorithm.

(3) The training data in the training data set used to train the neural network is introduced into data augmentation of random phase and noise as training samples, and the neural network is trained with a loss function including at least the target-background ratio loss function, so that the parameter estimates obtained by the training optimization are used. The network can process low-frequency broadband radar signal data well in low signal-to-noise ratio scenarios, effectively improving the saliency of the target.

Embodiments of the present application further provide an electronic device, including a memory, a processor, and a computer program stored in the memory, where the processor executes the computer program to realize the learnable low-frequency broadband radar target proposed by the embodiments of the present application Steps in a parameter estimation method.

In another embodiment provided by the present application, a computer program product is also provided, including a computer program/instruction, when the computer program/instruction is executed by a processor, to realize the learnable low-frequency broadband radar target proposed by the embodiment of the present application Steps in a parameter estimation method.

In the above-mentioned embodiments, it may be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented in software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer may be a general purpose computer, special purpose computer, computer network, or other programmable device. The computer instructions may be stored in or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions may be downloaded from a website site, computer, server or data center Transmission to another website site, computer, server, or data center is by wire (eg, coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that includes an integration of one or more available media. The usable media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, DVD), or semiconductor media (eg, Solid State Disk (SSD)), among others.

It should be noted that, in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any relationship between these entities or operations. any such actual relationship or sequence exists. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass a non-exclusive inclusion such that a process, method, article or device that includes a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

Each embodiment in this specification is described in a related manner, and the same and similar parts between the various embodiments may be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, as for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and for related parts, please refer to the partial descriptions of the method embodiments.

The above descriptions are only preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application are included in the protection scope of this application.

Claims (11)

1. A low-frequency broadband radar target parameter estimation method that can be learned, is characterized in that, comprises: Obtain data collected by low-frequency broadband radar; Input the data collected by the low-frequency broadband radar into a pre-trained parameter estimation network to obtain the target parameter estimation value output by the parameter estimation network, wherein the parameter estimation network uses a plurality of training data with random phases and noise as the The training samples are obtained by training the neural network based on the recurrent convolutional network layer structure. 2. The method for estimating target parameters of a learnable low-frequency broadband radar according to claim 1, wherein the neural network comprises: The input module is used to convert the complex-valued observation vector in the training sample input into the neural network into a real-valued observation vector, and use the real-valued observation vector and the real-valued GTD dictionary matrix to calculate the initial value of the neural network. input vector; an iterative optimization module for performing multiple iterative optimizations on the initial input vector of the neural network to obtain a real-valued output vector of the neural network; The output module is used to convert the real-valued output vector of the neural network into a complex-valued output vector, and output it as the predicted value of the target parameter. 3. The learnable low-frequency broadband radar target parameter estimation method according to claim 2, wherein the iterative optimization module comprises: learnable transformation structure

The learnable transform structure

is a network layer structure based on circular convolution, iteratively optimizes the initial input vector of the neural network, and outputs the real-valued output vector of the neural network, wherein the learnable transformation structure

Contains deformation operators

Circular convolution layer CC, linear rectification unit ReLU; in, deformation operator

for transforming the vector _x generated by the neural network in the iterative optimization process into a matrix z generated by the neural network in the iterative optimization process, where z is used to perform the circular convolution operation; Inverse shape operator

for transforming the matrix z generated by the neural network in the iterative optimization process into a vector x generated by the neural network in the iterative optimization process; A recurrent convolutional layer CC containing the learnable network parameters

for the

maps to

where, h=CC _K,P,Q (z; w,b),

nmodL represents an integer between 1 and L that modulo L is congruent to n; Linear rectification unit ReLU, defined as [ReLU(x)] _i =max(0,x _i ). 4. The learnable low-frequency broadband radar target parameter estimation method according to claim 1, wherein the training data is a pair of a complex-valued observation vector and a complex-valued parameter true value vector, and the training data is generated according to the following steps of: Generate a set of scattering center parameters; Calculate the true value vector of the complex-valued parameter according to the scattering center parameter in the scattering center parameter set; Calculate a complex-valued observation vector according to the complex-valued parameter truth value vector and the complex-valued GTD dictionary matrix, and combine the complex-valued observation vector and the complex-valued parameter truth value vector into a complex-valued observation vector and a complex-valued parameter truth value vector pair, as a set of the training data. 5. The method for estimating target parameters of a learnable low-frequency broadband radar according to claim 4, wherein the set of scattering center parameters is S={(l _n ,α _n ,σ _n ):n=1,..., N _s }, where, l _n is the distance unit, which is defined as an _integer from ₁ to L that _is randomly selected without replacement; _n is the scattering coefficient, the amplitude of σ _n |σ _n | obeys the distribution of the preset amplitude, and the argument ∠σ _n of σ n obeys the distribution of the preset argument. 6. The method for estimating target parameters of a learnable low-frequency broadband radar according to claim 5, wherein calculating the true value vector of the complex-valued parameter according to the scattering center parameter in the scattering center parameter set, comprising: calculate

Wherein, l=1,...,L, α belongs to the preset frequency-dependent factor set {α ₁ ,α ₂ ,...,α _J }; will get

combined into the complex-valued parameter truth vector

defined as:

Calculate the complex-valued observation vector according to the complex-valued parameter truth vector and the complex-valued GTD dictionary matrix, including: Using the complex-valued GTD dictionary matrix, convert the complex-valued parameter truth vector into the complex-valued observation vector:

Wherein, Φ is the complex-valued GTD dictionary matrix. 7. The method for estimating target parameters of a learnable low-frequency broadband radar according to claim 1, wherein the training of the neural network comprises: In each different training round ep, for each set of the training data, different random phases and noises are added to the complex-valued observation vectors in the training data, and different random phases and noises are added to the complex-valued observation vectors in the training data, and Different random phases are added to the true value vector of the complex-valued parameter as the training sample; Input the complex-valued observation vector in the training sample into the neural network, and calculate the loss function corresponding to the training sample according to the predicted value of the target parameter output by the neural network and the complex-valued truth vector in the training sample. value, wherein the value of the loss function includes at least the value of the target-background ratio loss function; The gradient of the loss function to the learnable network parameter is calculated, and the learnable network parameter is optimized based on the gradient of the loss function to the learnable network parameter of the neural network. 8. The learnable low-frequency broadband radar target parameter estimation method according to claim 7, wherein the target-background ratio loss function is defined as:

in,

is the value of the target-background ratio loss function, TBR is the target-background ratio,

is the target area,

is the background area, N _T is the number of elements of A _T , N _B is the number of elements of A _B ,

element in the predicted value for the target parameter. 9. The learnable low-frequency broadband radar target parameter estimation method according to claim 7, characterized in that, comprising: The loss function further includes one or more of a difference loss function and a symmetry loss function, wherein the symmetry loss function is based on the intermediate variables in the iterative optimization module of the neural network and the learnable loss function. Transform structure

Defined; When the loss function includes the target-to-background ratio loss function, the difference loss function and the symmetry loss function, the total loss function value is calculated according to the following formula:

in,

is the total loss function value,

is the value of the target-background ratio loss function,

the value of the dissimilarity loss function,

are the values of the symmetry loss function, and λ ₁ and λ ₂ are preset balance coefficients. 10. An electronic device, characterized by comprising a memory, a processor and a computer program stored on the memory, the processor executing the computer program to realize the learnable low frequency according to any one of claims 1-9 Steps in a wideband radar target parameter estimation method. 11. A computer program product, comprising a computer program/instruction, characterized in that, when the computer program/instruction is executed by a processor, the method for estimating the target parameter of a learnable low-frequency broadband radar according to any one of claims 1-9 is implemented. A step of.

Images