Learnable low-frequency wideband radar target parameter estimation method, device and program product

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 program product. The method comprises the following steps: acquiring data collected by a low-frequency broadband radar; inputting the data collected 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 with a plurality of training data pairs with random phases and noises as training samples. The neural network is trained by a network layer structure based on cyclic convolution and training samples based on data enhancement, at least one loss function including a target-background ratio loss function is used, learnable parameters in the neural network are continuously optimized, the parameter estimation network is used for processing low-frequency broadband radar signals, the calculation complexity is reduced, the algorithm efficiency and the generalization capability are improved, and the target is more prominent in a low signal-to-noise ratio scene.

Description

Technical Field

[0001] This application relates to the field of radar signal processing, and more specifically, to a learnable low-frequency broadband radar target parameter estimation method, device, and program product. Background Technology

[0002] Low-frequency broadband radar signals possess the ability to counter stealth and penetrate walls and soil, making them important in both defense and civilian applications. Radar target parameter estimation based on the geometric theory of diffraction (GTD) model is a crucial step in the range-oriented signal processing of low-frequency broadband radar.

[0003] However, existing GTD parameter estimation methods suffer from high computational complexity and complex hyperparameter tuning, resulting in low algorithm efficiency and weak generalization ability for low-frequency broadband radar data in different scenarios. In addition, the signal-to-noise ratio has a significant impact on algorithm performance in existing GTD parameter estimation methods, leading to lower target saliency and a significant decrease in parameter estimation performance in low signal-to-noise ratio scenarios. Summary of the Invention

[0004] This application provides a learnable low-frequency broadband radar target parameter estimation method, device, and program product, aiming to solve the problems of low algorithm efficiency, complex hyperparameter adjustment, weak generalization ability, and low target saliency in low signal-to-noise ratio scenarios in the prior art.

[0005] The first aspect of this application provides a learnable low-frequency broadband radar target parameter estimation method, including:

[0006] Acquire data from low-frequency broadband radar;

[0007] The data collected by the low-frequency broadband radar is input into a pre-trained parameter estimation network to obtain the target parameter estimates output by the parameter estimation network. The parameter estimation network is trained on a neural network based on a recurrent convolutional network layer structure using multiple training data with random phase and noise as training samples.

[0008] Optionally, the neural network includes:

[0009] The input module is used to convert the complex numerical observation vectors in the training samples input to the neural network into real numerical observation vectors, and to calculate the initial input vector of the neural network using the real numerical observation vectors and the real numerical GTD dictionary matrix.

[0010] An iterative optimization module is used to perform multiple iterative optimizations on the initial input vector of the neural network to obtain the real-valued output vector of the neural network.

[0011] The output module is used to convert the real-valued output vector of the neural network into a complex-valued output vector, which is then output as the predicted value of the target parameter.

[0012] Optionally, the iterative optimization module includes:

[0013] Learnable Transformation Structure The learnable transformation structure It is a network layer structure based on recurrent convolution, which iteratively optimizes the initial input vector of the neural network and outputs a real-valued output vector of the neural network, wherein the learnable transformation structure Includes transformation operators Recurrent convolutional layer (CC), Rectified linear unit (ReLU);

[0014] in,

[0015] Transformation Operator This is used to transform the vector x generated by the neural network during the iterative optimization process into the matrix z generated by the neural network during the iterative optimization process, wherein z is used to perform the cyclic convolution operation;

[0016] Inverse transformation operator This is used to transform the matrix z generated by the neural network during the iterative optimization process into a vector generated by the neural network during the iterative optimization process. x ;

[0017] The recurrent convolutional layer CC contains the learnable network parameters. Used to Mapped to Where h = CC K,P,Q (z;w,b), n mod L means that n mod L is congruent to n in the interval from 1 to L modulo L. n Integers;

[0018] The linear rectifier unit ReLU is defined as [ReLU(x)]. i =max(0,x i ).

[0019] Optionally, the training data consists of pairs of complex-valued observation vectors and complex-valued parameter truth vectors, and the training data is generated according to the following steps:

[0020] Generate a set of scattering center parameters;

[0021] Calculate the truth vector of the complex numerical parameters according to the scattering center parameters in the set of scattering center parameters;

[0022] Based on the complex numerical parameter truth vector and the complex numerical GTD dictionary matrix, a complex numerical observation vector is calculated. The complex numerical observation vector and the complex numerical parameter truth vector are combined to form a complex numerical observation vector and complex numerical parameter truth vector pair, which serves as a set of training data.

[0023] Optionally, the set of scattering center parameters is S = {(l n ,α n ,σ n ):n=1,…,N s},in,

[0024] l n The distance unit is defined as an integer randomly drawn without replacement from 1 to L. n Different; α n As a frequency-dependent factor, α n Randomly drawn with replacement from a predefined set of frequency-dependent factors; σ n σ is the scattering coefficient. n amplitude|σ n | Follows a distribution with a preset amplitude, σ n Argument ∠σ n It follows the distribution of preset angles.

[0025] Optionally, calculating the truth vector of the complex numerical parameters according to the scattering center parameters in the set of scattering center parameters includes:

[0026] calculate

[0027]

[0028] Where l = 1, ..., L, α belongs to the preset frequency-dependent factor set {α1, α2, ..., α}. J};

[0029] The result Combined into the truth vector of the complex numerical parameters Defined as:

[0030]

[0031]

[0032] Based on the complex-valued parameter truth vector and the complex-valued GTD dictionary matrix, the complex-valued observation vector is calculated, including:

[0033] Using the complex numerical GTD dictionary matrix, the complex numerical parameter truth vector is converted into the complex numerical observation vector: Wherein, Φ is the complex numerical GTD dictionary matrix.

[0034] Optionally, training the neural network includes:

[0035] In each different training ep, for each set of training data, different random phases and noises are added to the complex numerical observation vectors in the training data, and different random phases are added to the complex numerical parameter ground value vectors in the training data, as training samples;

[0036] The complex numerical observation vectors in the training samples are input into the neural network. Based on the target parameter prediction values ​​output by the neural network and the complex numerical ground value vectors in the training samples, the value of the loss function corresponding to the training samples is calculated. The value of the loss function includes at least the value of the target-to-background ratio loss function.

[0037] Calculate the gradient of the loss function with respect to the learnable network parameters, and optimize the learnable network parameters based on the gradient of the loss function with respect to the learnable network parameters of the neural network.

[0038] Optionally, the target background ratio loss function is defined as:

[0039]

[0040] in, The value of the target background ratio loss function is given by [value], where TBR is the target background ratio. For the target area, For the background area, N T For A T The number of elements, N B For A B The number of elements, The element in the predicted value of the target parameter.

[0041] Optionally, the loss function further includes one or more of a differential loss function and a symmetric loss function, wherein the symmetric loss function is based on the intermediate variables in the iterative optimization module of the neural network and the learnable transformation structure. Defined;

[0042] When the loss function includes the target 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:

[0043]

[0044] in, This is the total loss function value. The value of the target-to-background ratio loss function. The value of the difference loss function, λ1 and λ2 are the values ​​of the symmetry loss function, and λ1 and λ2 are preset balance coefficients.

[0045] A second aspect of this application provides a parameter estimation electronic device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in the learnable low-frequency broadband radar target parameter estimation method proposed in this application.

[0046] A third aspect of this application provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps in the learnable low-frequency broadband radar target parameter estimation method proposed in this application.

[0047] Beneficial effects:

[0048] This application provides a learnable low-frequency broadband radar target parameter estimation method, device, and program product. By training a neural network based on data-augmented training samples, the learnable parameters in the neural network are continuously optimized to obtain a parameter estimation network for processing low-frequency broadband radar signals. This method offers the following advantages:

[0049] (1) A network layer structure based on recurrent convolution was adopted, which reduced the computational complexity and improved the efficiency and accuracy of parameter estimation.

[0050] (2) Multiple training data with random phase and noise are used as training samples to train the neural network and obtain a parameter estimation network for parameter estimation. The parameter estimation network obtained by training and optimization can handle different low-frequency broadband radar signal data, which improves the accuracy and generalization ability of the parameter estimation algorithm.

[0051] (3) The training data in the training dataset used to train the neural network is introduced into random phase and noise data augmentation as training samples. The neural network is trained with a loss function that includes at least the target-to-background ratio loss function. The parameter estimation network obtained by training optimization can handle low-frequency broadband radar signal data well in low signal-to-noise ratio scenarios, effectively improving the saliency of the target. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of a learnable low-frequency broadband radar target parameter estimation method proposed in an embodiment of this application;

[0054] Figure 2 This is a flowchart of a training method for a learnable low-frequency broadband radar target parameter estimation neural network proposed in an embodiment of this application;

[0055] Figure 3 This is a line graph showing the signal-to-noise ratio versus the target-to-background ratio, as proposed in one embodiment of this application.

[0056] Figure 4 This is a signal-to-noise ratio-mean-square error line graph proposed in one embodiment of this application; Detailed Implementation

[0057] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0058] In related technologies, GTD parameter estimation methods based on the geometric theory of diffraction (GTD) model are employed. However, existing GTD parameter estimation methods suffer from high computational complexity and complex hyperparameter tuning, resulting in low algorithm efficiency and weak generalization ability for low-frequency broadband radar data in different scenarios. Furthermore, in existing GTD parameter estimation methods, the signal-to-noise ratio (SNR) has a significant impact on algorithm performance, leading to lower target saliency and a marked decrease in parameter estimation performance in low SNR scenarios.

[0059] In view of this, this application proposes a learnable low-frequency broadband radar target parameter estimation method, aiming to solve the problems of low algorithm efficiency, weak generalization ability, and low target saliency in low signal-to-noise ratio scenarios in the prior art. The parameter estimation method of this application will be described in detail below.

[0060] Figure 1 This is a flowchart illustrating a learnable low-frequency broadband radar target parameter estimation method according to an embodiment of this application. (Refer to...) Figure 1 The learnable low-frequency broadband radar target parameter estimation method proposed in this application includes the following steps:

[0061] S101. Acquire data from low-frequency broadband radar.

[0062] In practice, the following data is obtained from the data collected by the low-frequency broadband radar:

[0063] Radar observation frequency {f m :m=1,…,M}, frequency interval Δf, satisfying f m =f1+(m-1)Δf; the spectrum of the radar observation signal {E(f m ):m=1,…,M};The number of distance units L of the observation scene, requiring L>M.

[0064] S102. Input the data collected by the low-frequency broadband radar into the pre-trained parameter estimation network.

[0065] In specific implementation, the radar observation frequency, radar observation signal spectrum, and number of range units in the observation scene from the data collected by the aforementioned low-frequency broadband radar are input into a pre-trained parameter estimation network (DNN). The input low-frequency broadband radar data is then used to assign values ​​to the corresponding quantities in the parameter estimation network, and the input radar observation signal spectrum {E(f...} is used to... m The combination of m = 1, ..., M forms the radar spectral signal vector e:

[0066]

[0067] The parameter estimation network (DNN) is trained using multiple training data sets with random phases and noise as training samples. The training method for this neural network is detailed in steps S201-S203 of the parameter estimation method below, and will not be repeated here.

[0068] S103, The parameter estimation network outputs the target parameter estimates.

[0069] In practice, the preset algorithm and optimized learnable network parameters Θ in the parameter estimation network DNN are followed. opt Solve the problem to obtain the target parameter estimation vector. It also outputs the estimated values ​​of the target parameters.

[0070]

[0071] Target parameter vector Its elements are Its definition is the same as the parameter truth vector described below, and will not be repeated here. When When , it indicates that there exists a scattering center with a frequency dependence factor of α in the l-th distance cell, and its estimated scattering coefficient is . when When α is α, it means that there is no scattering center with a frequency dependence factor of α in the l-th distance cell.

[0072] Figure 2 This is a flowchart illustrating the training method for a learnable low-frequency broadband radar target parameter estimation neural network, as shown in the embodiments of this application. Figure 2 The training process of the neural network in the learnable low-frequency broadband radar target parameter estimation method proposed in this application includes the following steps:

[0073] S201, Construct the GTD dictionary and neural network.

[0074] In practice, the first step is to construct a GTD dictionary. The GTD dictionary includes a complex-valued GTD dictionary matrix and a real-valued GTD dictionary matrix.

[0075] Constructing a complex numerical GTD dictionary matrix Defined as:

[0076]

[0077] Where, Φ (α) ,α∈{α1,α2,…,α J} is a matrix of J sub-dictionaries. The element in the m-th row and l-th column (m = 1, ..., M, l = 1, ..., L) is defined as:

[0078]

[0079] j is the imaginary unit, c is the speed of light, and f is the speed of light. C =f1+(M-1)Δf / 2 represents the center frequency, r l = (l / L)×c / (2Δf) represents the l-th distance unit.

[0080] Constructing a real-valued GTD dictionary matrix Defined as

[0081]

[0082] Re(·) and Im(·) represent the operations on the real part and the imaginary part, respectively.

[0083] By constructing a GTD dictionary, the data collected by low-frequency broadband radar in real-world scenarios is transformed into subsequent mathematical models, allowing the subsequent neural network operation mathematical models to be processed separately from the previous actual models, thus reducing the complexity of data processing in the embodiments of this application.

[0084] Construct a neural network (DNN). This DNN is an incompletely optimized parameter estimation network (DNN) as described above, containing learnable network parameters Θ and learnable transformations. This application uses the Python programming language and the PyTorch deep learning framework to construct a neural network (DNN). The DNN can utilize any existing programming language and deep learning framework, and this application does not impose any restrictions on this. The neural network (DNN) includes an input module, an iterative optimization module, and an output module.

[0085] The input module is used to convert the complex numerical observation vectors in the training samples input to the neural network into real numerical observation vectors, and to calculate the initial input vector of the neural network using the real numerical observation vectors and the real numerical GTD dictionary matrix.

[0086] In practice, the complex-valued observation vectors in the training samples are input. These training samples are obtained by introducing random phase and noise into the training data in the training dataset. The specific process of generating this training dataset is described in S202 below.

[0087] The input complex numerical observation vector e (here, e = e during training) i,ep e i,ep To add random phase and noise to the complex-valued observation vector (see the definition below), e is transformed into the corresponding real-valued observation vector b during testing.

[0088]

[0089] The initial values ​​are calculated using the real-valued GTD dictionary matrix and used as the initial input vector for the neural network.

[0090] y (1) =x (0) =Ψ T (ΨΨ T ) -1 b

[0091] The iterative optimization module is used to perform multiple iterative optimizations on the initial input vector of the neural network to obtain the real-valued output vector of the neural network.

[0092] In practical implementation, the real-valued output vector of the above neural network is processed by N... p =15 iterations for k=1,…,N p Repeat the following iterative steps:

[0093] r (k) =y (k) -μ (k) Ψ T (Ψy (k) -b)

[0094]

[0095] y (k+1) =x (k) -ρ (k) (x (k) -x (k-1) )

[0096] Among them, μ (k) ,θ (k) ,ρ (k) , The definition is as follows:

[0097] μ (k) ,θ (k) ,ρ (k) Defined as:

[0098] μ (k) =sp(a1k+c1),θ (k) =sp(a²k+c²), in,

[0099] sp(x) = ln(1 + exp(x)), a1, a2, a3, c1, c2, c3 belong to the learnable network parameter Θ1 (the definition of learnable network parameter Θ in a DNN is given below, the same below), S θ It is a threshold shrinkage function with parameter θ, and is an element-wise function defined as:

[0100]

[0101] For learnable transformation, the learnable transformation structure It is a network layer structure based on recurrent convolution, which iteratively optimizes the initial input vector of the neural network and outputs a real-valued output vector of the neural network, wherein the learnable transformation structure Includes transformation operator Recurrent convolutional layer (CC), Rectified linear unit (ReLU);

[0102] In some implementations, the learning transformation structure can be used. (With 5 convolutional layers) The structure is as follows:

[0103]

[0104] in,

[0105] Transformation Operator Defined as: Transformation Operator vector Transform into a matrix The corresponding inverse transformation operator matrix Transform into a vector Satisfy z d,l =x (d-1)L+l Where d = 1, ..., 2J, l = 1, ..., L;

[0106] Recurrent convolutional layers (CCs) contain learnable network parameters. Used to Mapped to in,

[0107] h = CC K,P,Q (z;w,b),

[0108] n mod L means that n mod L is congruent to n in the interval from 1 to L modulo L. n integers, p = 1, ..., P;

[0109] The ReLU linear rectifier unit is an element-wise function, defined as: [ReLU(x)] i =max(0,x i )

[0110] In a neural network (DNN), the learnable network parameter Θ is defined as: Θ = Θ1 ∪ Θ2, where Θ1 = {a1, a2, a3, c1, c2, c3}, and is used to calculate μ. (k) ,θ (k) ,ρ (k) ; Learnable transformations used in the above definition

[0111] The above-mentioned learnable transformations include the number of convolutional layers, kernel size K, and number of feature channels N. F The number of elements J in the frequency-dependent factor set is preset, and the specific value is not limited in this application. The embodiments of this application use 5 convolutional layers, a kernel size of K=3, and N feature channels. F =32, and the number of elements in the frequency-dependent factor set is J=5.

[0112] This application's embodiments introduce an intermediate vector y. Through operations using this intermediate vector y, the output vector of real values ​​is reduced. The number of iterations is reduced, thus decreasing the computational load and improving the algorithm's efficiency.

[0113] The output module is used to convert the real-valued output vector of the neural network into a complex-valued output vector, which is then output as the predicted value of the target parameter.

[0114] In specific implementation, the real-valued output vector of the above iterative optimization module will be used. Convert the vector to a complex value using the following formula:

[0115]

[0116] Use the complex-valued output vector as the predicted value of the target parameter. Output the results.

[0117] S202. Generate complex numerical observation vector-complex numerical parameter truth vector pairs as training data to form a training dataset.

[0118] In practice, a training dataset D = {(e i ,σ i ): i = 1, ..., N D}, where the training data consists of complex-valued observation vectors and complex-valued parameter truth vector pairs (e i ,σ i The number of training data ND is preset, and this application does not limit the specific value of the number of training samples. The number used in the embodiments of this application is N. D =50000. Each training data set (e i ,σ i It is generated according to the following steps:

[0119] Number of scattering centers N s N s The value is a random integer within a preset range, which is not limited in this application. In this embodiment, N is... s The default value is a random integer between 1 and 15;

[0120] Generate a set of scattering center parameters, the set of scattering center parameters being S={(l n ,α n ,σ n ):n=1,…,N s}, where l n A distance unit is defined as an integer randomly drawn without replacement from 1 to L. n Different; α n As a frequency-dependent factor, α n Randomly drawn with replacement from a predefined set of frequency-dependent factors; σ n σ is the scattering coefficient. n amplitude|σ n | Follows a distribution with a preset amplitude, σ n Argument ∠σ n It follows the distribution of preset angles.

[0121] The specific values ​​of the preset frequency-dependent factor set, the preset amplitude distribution, and the preset phase angle distribution are not limited in this application. In the embodiments of this application, the frequency-dependent factor set is {-1,-1 / 2,0,1 / 2,1}, the number of elements in the frequency-dependent factor set is J=5, the amplitude follows a uniform distribution U(0.5,1.5), and the phase angle follows a uniform distribution U(0,2π).

[0122] Calculate the truth vector of the complex numerical parameters according to the scattering center parameters in the set of scattering center parameters.

[0123]

[0124] Where l = 1, ..., L, α belongs to the preset frequency-dependent factor set {α1, α2, ..., α}. J};

[0125] The result Combined into the truth vector of the complex numerical parameters Defined as:

[0126]

[0127]

[0128] Based on the complex numerical parameter truth vector and the complex numerical GTD dictionary matrix, calculate the complex numerical observation vector, and then use the complex numerical GTD dictionary matrix to convert the complex numerical parameter truth vector into the complex numerical observation vector. Each obtained complex-valued observation vector and complex-valued parameter truth vector form a complex-valued observation vector-complex-valued parameter truth vector pair, which is used as a set of training data. All training data constitute the training dataset.

[0129] S203. Calculate the loss function based on the training samples, and optimize the learnable network parameters based on the value of the loss function to train the neural network and obtain the parameter estimation network DNN.

[0130] In practice, the following steps should be followed:

[0131] Initialize the neural network DNN and the optimizer. The learnable network parameters Θ = Θ1∪Θ2 in the neural network DNN, where Θ1 = {a1,a2,a3,c1,c2,c3} are initialized to initial values. The learnable network parameter Θ2 is initialized using an algorithm. The optimizer is used to set the batch size, initial learning rate, and learning rate decay to preset values.

[0132] The initial values ​​of the learnable network parameter Θ1, the optimizer batch size, the initial learning rate, and the learning rate decay are all preset in advance, and the specific values ​​are not limited in this application. The algorithm used for initializing the learnable network parameter Θ2 is prior art, and its specific content is not limited in this application. In the embodiments of this application, the initial value of the learnable network parameter Θ1 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, the preset batch size is 32, the preset learning rate is 0.001, and the learning rate decays to 0.8 every 5 training ep.

[0133] Data augmentation, for each of the training data (e i ,σ i ), for the complex-valued observation vector e in the training data i Add random phase φ i,ep and noise n i,ep For the truth vector σ of the complex-valued parameters in the training data i Add the random phase φ i,ep , to obtain training samples (e i,ep ,σ i,ep Specifically, follow the formula below:

[0134] e i,ep =e i ·exp(jφ i,ep )+n i,ep

[0135] σ i,ep =σ i ·exp(jφ i,ep )

[0136] Where, φ i,ep Follows a preset distribution, noise n i,ep Types and noise n i,ep The required signal-to-noise ratio is preset in advance, and the specific value is not limited in this application. In the embodiments of this application, φ i,ep The noise follows a uniform distribution U(0,2π) and is n i,ep It is Gaussian additive white noise, satisfying a signal-to-noise ratio (SNR) of 5 dB.

[0137] By introducing different random phases and noise, the data augmentation in the original training dataset is improved. i ,σ iMultiple data augmentations are performed using different phases, which significantly increases the amount of data after augmentation and improves the capacity of training samples. As a result, the parameter estimation network trained and optimized based on the augmented training samples has higher accuracy and is closer to the true value. In addition, the use of noisy data as training samples enables the parameter estimation network trained and optimized based on the augmented training samples to have better processing capabilities for data in low signal-to-noise ratio environments, effectively improving the saliency of the target.

[0138] The training samples (e) i,ep ,σ i,ep e in ) i,ep Input a neural network (DNN), and output predicted values ​​of the target parameters based on a pre-defined algorithm of the DNN. Calculate the value of the loss function corresponding to the training sample, wherein the value of the loss function includes at least the value of the target-to-background ratio loss function.

[0139] The target-to-ground ratio (TBR) loss function is defined as follows:

[0140]

[0141] in, The value of the target background ratio loss function is given by [value], where TBR is the target background ratio. For the target area, For the background area, N T For A T The number of elements, N B For A B The number of elements, The element in the predicted value of the target parameter.

[0142] The loss function also includes one or more of the differential loss function and the symmetric loss function.

[0143] The variance loss function (i.e., mean squared error, MSE) measures the degree of deviation between the estimated value and the true value, and is defined as follows:

[0144]

[0145] Where ||·||2 represents the norm of l2.

[0146] The symmetry loss function, used to measure the impact of symmetry loss during training, is defined as follows:

[0147]

[0148] in, This indicates that the input to the neural network DNN is e.i,ep At that time, the corresponding intermediate variable r (k) .

[0149] When the loss function includes the target 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:

[0150]

[0151] in, This is the total loss function value. The value of the target-to-background ratio loss function. The value of the difference loss function, The values ​​of the symmetry loss function are λ1 and λ2, which are preset balance coefficients. compared to The order of magnitude is much larger, so coefficients λ1 and λ2 are set for balancing. The order of magnitude of the values ​​is not limited in this application. In the embodiments of this application, λ1 = 0.1 and λ2 = 0.001.

[0152] Optimize the learnable network parameters Θ in the DNN, apply the backpropagation algorithm, and calculate the total loss function value. gradient with respect to the learnable network parameter Θ Then, the learnable network parameters Θ are optimized using an optimizer.

[0153] In each ep, the above data augmentation is repeated (for each training sample (e) in each different training ep). i,ep ,σ i,ep Add different random phases and noise, input to the neural network output. The loss function value is calculated, and the learnable network parameters Θ are optimized for multiple rounds to obtain the parameter estimation network DNN. The number of repeated rounds is preset and not limited in this application; in this embodiment, the repeated rounds total 60 rounds.

[0154] Figure 3 This is a line graph showing the signal-to-noise ratio versus target-to-background ratio, as illustrated in an embodiment of this application. (Refer to...) Figure 3 The image shows the target background ratio (TBR) values ​​for different algorithms under various low signal-to-noise ratio (SNR) conditions. Figure 3 As can be seen from the figure, the existing FOCUSS algorithm, SVR algorithm, and FISTA-Net algorithm have lower TBR values ​​at the same signal-to-noise ratio compared with the algorithm in the parameter estimation method proposed in this application (TEFISTA-Net in the figure), indicating that the parameter estimation method proposed in this application has stronger target saliency.

[0155] Figure 4 This is a line graph showing the signal-to-noise ratio versus mean square error (MSE) as illustrated in an embodiment of this application. (Refer to...) Figure 4 The figure shows the mean square error (MSE) values ​​of different algorithms under different low signal-to-noise ratio (SNR) conditions. Figure 4 It can be seen that, compared with the algorithm in the parameter estimation method proposed in this application (TEFISTA-Net in the figure), the existing FOCUSS algorithm, SVR algorithm, and FISTA-Net algorithm have higher MSE values ​​at the same signal-to-noise ratio. This indicates that the parameter estimation method proposed in this application has a smaller difference from the expected value, and the parameter estimation value obtained by applying the parameter estimation method proposed in this application is more accurate.

[0156] also, Figure 3 and Figure 4 In the case of low to medium signal-to-noise ratio (SNR, e.g., SNR ≤ 10dB), the parameter estimation method proposed in this application shows better performance than the existing FOCUSS, SVR, and FISTA-Net algorithms. This indicates that the parameter estimation method proposed in this application enhances the training dataset of the neural network by introducing random phase and Gaussian additive white noise into the basic training samples, and introduces a target-to-background ratio loss function for low SNR scenarios. This enables the parameter estimation network obtained by training and optimization to handle low-frequency broadband radar signal data well in low SNR scenarios.

[0157] This application provides a learnable low-frequency broadband radar target parameter estimation method. By training a neural network based on data-augmented training samples and continuously optimizing the learnable parameters within the neural network, a parameter estimation network is obtained for processing low-frequency broadband radar signals. This method offers the following advantages:

[0158] (1) A network layer structure based on recurrent convolution was adopted, which reduced the computational complexity and improved the efficiency of parameter estimation.

[0159] (2) Multiple training data with random phase and noise are used as training samples to train the neural network and obtain a parameter estimation network for parameter estimation. The parameter estimation network obtained by training and optimization can handle different low-frequency broadband radar signal data, reduce computational complexity, and improve the efficiency and generalization ability of parameter estimation algorithm.

[0160] (3) The training data in the training dataset used to train the neural network is introduced into random phase and noise data augmentation as training samples. The neural network is trained with a loss function that includes at least the target-to-background ratio loss function. The parameter estimation network obtained by training optimization can handle low-frequency broadband radar signal data well in low signal-to-noise ratio scenarios, effectively improving the saliency of the target.

[0161] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in the learnable low-frequency broadband radar target parameter estimation method proposed in this application.

[0162] In another embodiment provided in this application, a computer program product is also provided, including a computer program / instructions that, when executed by a processor, implement the steps in the learnable low-frequency broadband radar target parameter estimation method proposed in the embodiments of this application.

[0163] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially 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 this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0164] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0165] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0166] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A learnable low-frequency broadband radar target parameter estimation method, characterized in that, include: Acquire data from low-frequency broadband radar; The data collected by the low-frequency broadband radar is input into a pre-trained parameter estimation network to obtain the target parameter estimates output by the parameter estimation network. The parameter estimation network is trained on a neural network based on a recurrent convolutional network layer structure using multiple training data with random phase and noise as training samples. Training the neural network includes: In each different training ep, for each set of training data, 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 complex-valued parameter ground vectors in the training data, which are used as training samples; The complex numerical observation vectors in the training samples are input into the neural network. Based on the target parameter prediction values ​​output by the neural network and the complex numerical ground value vectors in the training samples, the value of the loss function corresponding to the training samples is calculated. The value of the loss function includes at least the value of the target-to-background ratio loss function. Calculate the gradient of the loss function with respect to the learnable network parameters, and optimize the learnable network parameters based on the gradient of the loss function with respect to the learnable network parameters of the neural network; The target-to-background ratio loss function is defined as follows: in, The value of the target background ratio loss function is given by [value], where TBR is the target background ratio. For the target area, For background area, for The number of elements, for The number of elements, For the elements in the predicted values ​​of the target parameters, For distance units, defined as 1~ The integers drawn without replacement are random numbers. Belongs to the pre-defined set of frequency-dependent factors .

2. The learnable low-frequency broadband radar target parameter estimation method according to claim 1, characterized in that, The neural network includes: The input module is used to convert the complex numerical observation vectors in the training samples input to the neural network into real numerical observation vectors, and to calculate the initial input vector of the neural network using the real numerical observation vectors and the real numerical GTD dictionary matrix. An iterative optimization module is used to perform multiple iterative optimizations on the initial input vector of the neural network to obtain the 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, which is then output as the predicted value of the target parameter.

3. The learnable low-frequency broadband radar target parameter estimation method according to claim 2, characterized in that, The iterative optimization module includes: Learnable Transformation Structure The learnable transformation structure It is a network layer structure based on recurrent convolution, which iteratively optimizes the initial input vector of the neural network and outputs a real-valued output vector of the neural network, wherein the learnable transformation structure Includes transformation operator , recurrent convolutional layer (CC), linear rectified unit ; in, Transformation Operator Used to convert the vectors generated by the neural network during the iterative optimization process Transformed into the matrix generated by the neural network during the iterative optimization process. ,in Used to perform the circular convolution operation; Inverse transformation operator , used to convert the matrix generated by the neural network during the iterative optimization process Transformed into vectors generated by the neural network during the iterative optimization process. ; The recurrent convolutional layer CC contains the learnable network parameters. , used to Mapped to ,in, , , , , Indicates in arrive Between models Congruent to an integer, n is k is the kernel in the time dimension, K is the size of the kernel in the time dimension, p is the output feature channel, P is the number of output feature channels, q is the input feature channel, and Q is the number of input feature channels. For distance units, defined as 1~ The integer drawn without replacement; Linear rectifier unit Defined as .

4. The learnable low-frequency broadband radar target parameter estimation method according to claim 1, characterized in that, The training data consists of pairs of complex-valued observation vectors and complex-valued parameter truth vectors, and the training data is generated according to the following steps: Generate a set of scattering center parameters; Calculate the truth vector of the complex numerical parameters according to the scattering center parameters in the set of scattering center parameters; Based on the complex numerical parameter truth vector and the complex numerical GTD dictionary matrix, a complex numerical observation vector is calculated. The complex numerical observation vector and the complex numerical parameter truth vector are combined to form a complex numerical observation vector and complex numerical parameter truth vector pair, which serves as a set of training data.

5. The learnable low-frequency broadband radar target parameter estimation method according to claim 4, characterized in that, The set of scattering center parameters is as follows: ,in, For distance units, defined as 1~ The integers drawn without replacement are random numbers. They are all different; It is a frequency-dependent factor. Randomly sampled with replacement from a predefined set of frequency-dependent factors; The scattering coefficient is... amplitude It follows a distribution with a preset amplitude. Argument It follows the distribution of preset angles.

6. The learnable low-frequency broadband radar target parameter estimation method according to claim 5, characterized in that, Calculating the truth vector of the complex numerical parameters according to the scattering center parameters in the set of scattering center parameters includes: calculate : in, , Belonging to the preset frequency-dependent factor set ; The result Combined into the truth vector of the complex numerical parameters Defined as: ， ； Based on the complex-valued parameter truth vector and the complex-valued GTD dictionary matrix, the complex-valued observation vector is calculated, including: Using the complex numerical GTD dictionary matrix, the complex numerical parameter truth vector is converted into the complex numerical observation vector: ,in, Let be the complex numerical GTD dictionary matrix.

7. The learnable low-frequency broadband radar target parameter estimation method according to claim 1, characterized in that, include: The loss function further includes one or more of a differential loss function and a symmetric loss function, wherein the symmetric loss function is based on intermediate variables and learnable transformation structures in the iterative optimization module of the neural network. Defined; When the loss function includes the target 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, This is the total loss function value. The value of the target-to-background ratio loss function. The value of the difference loss function, The value of the symmetry loss function. , This is the preset balance coefficient.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps in the learnable low-frequency broadband radar target parameter estimation method according to any one of claims 1-7.

9. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps in the learnable low-frequency broadband radar target parameter estimation method according to any one of claims 1-7.

Images