Modeling Method for Peak Feature of Projectile-Target Encounter Echo Based on Deep Learning

Abstract

The present invention provides a method for modeling the peak characteristics of missile-target encounter echo based on deep learning, which includes the steps of: S1, collecting multiple missile-target encounter echo signals, respectively extracting the corresponding encounter echo state label vectors to establish a training set, a validation set and a test set, and the encounter echo state vector includes missile body attitude, target attitude, miss distance and miss azimuth; S2, establishing a multi-layer deep learning model for predicting the envelope and peak characteristics of missile-target encounter echo signals, and the peak characteristics of missile-target encounter echo include: the number of peaks, peak positions, peak amplitudes, and the multi-layer deep learning model includes a linear module, a residual layer, an upsampling module, a long short-term memory layer, and a convolutional module connected in sequence; S3, training the multi-layer deep learning model through the training set; the validation set is used to adjust the learning rate of the multi-layer deep learning model; S4, inputting the test set into the trained multi-layer deep learning model to evaluate the accuracy of the multi-layer deep learning model.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target near-field characteristic modeling and simulation, and particularly relates to a method for modeling the peak characteristics of the echo of a missile-target encounter based on deep learning. Background Art

[0002] The modeling methods for the echo of a missile-target encounter include the electromagnetic scattering simulation method based on the target geometric model, the transformation method based on the target scattering center, and the transformation method based on the near-field scattering function. The electromagnetic scattering simulation based on the target geometric model has high requirements for the refinement degree of the geometric model, and there are problems such as difficulty in realizing the simulation modeling of the target detailed structure model, material, etc. The transformation method based on the target scattering center requires the ability to provide the scattering center data in all postures, and the amplitude and position errors of the scattering centers will bring large transformation deviations of the missile-target encounter echo. The transformation method based on the near-field scattering function has high requirements for the accuracy of the near-field scanning test data, and factors such as test system errors and background clutter interference will reduce the accuracy of the missile-target encounter echo data. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for modeling the peak characteristics of the echo of a missile-target encounter based on deep learning. First, a multi-layer deep learning model is established, and the corresponding encounter echo state label vectors are extracted from multiple missile-target encounter echo signals to train the model, so as to realize the prediction of the target near-field echo under any missile-target encounter conditions.

[0004] To achieve the above purpose, the present invention provides a method for modeling the peak characteristics of the echo of a missile-target encounter based on deep learning, comprising the steps of:

[0005] S1. Collect multiple missile-target encounter echo signals, and respectively extract the corresponding encounter echo state label vectors; establish a training set, a validation set, and a test set based on the extracted encounter echo state label vectors;

[0006] S2. Establish a multi-layer deep learning model for predicting the envelope of the missile-target encounter echo signal and the peak characteristics of the missile-target encounter echo signal; the multi-layer deep learning model includes a linear module, a residual layer, an upsampling module, a long short-term memory layer, and a convolutional module connected in sequence; the depths of the linear module, the residual layer, the upsampling module, the long short-term memory module, and the convolutional module are H1, H2, H3, H4, and H5 respectively;

[0007] S3. Train the multi-layer deep learning model through the training set; the validation set is used to adjust the learning rate of the multi-layer deep learning model when training the multi-layer deep learning model through the training set;

[0008] S4. Input the test set into the trained multi-layer deep learning model to evaluate the accuracy of the multi-layer deep learning model.

[0009] Optionally, the rendezvous echo state label vector includes the missile body pitch angle, the missile body azimuth angle, the single unit roll angle, the target pitch angle, the target azimuth angle, the target roll angle, as well as the target miss distance and the miss azimuth in the missile-target relative motion coordinate system.

[0010] Optionally, the missile-target rendezvous echo signal envelope includes: the rendezvous coordinate position and the echo amplitude; the missile-target rendezvous echo peak characteristics include: the number of peaks, the peak positions, and the peak amplitudes.

[0011] Optionally, the linear module includes H1 linearly-connected linear layers, and each linear layer includes a fully-connected layer FC, a first regularization layer n1, and a first activation function a1 that are connected in sequence;

[0012] Let L i represent the i-th linear layer, and x i represent the input tensor of the i-th linear layer, where i ∈ [1, H1]; the input tensor of the (i + 1)-th linear layer is the output tensor of the i-th linear layer, and x i+1 = L i (x i );

[0013] The dimension of x1 is N × d, each row of x1 corresponds to a rendezvous echo state label vector, and d is the number of vectors in the rendezvous echo state label vector; the dimension of the output tensor of the first linear layer is N × h; h is a hyperparameter that controls the dimension change of the input tensor during the model propagation process;

[0014] When 1 < i ≤ L1, the dimension of the input tensor x i of the i-th linear layer is N × [(i - 1)·h]; the dimension of the output tensor of the i-th linear layer is N × (i·h); L is the length of the output signal of the multi-layer deep learning model;

[0015] When i > L1, the dimension of the input tensor x i of the i-th linear layer is N × (h·L1), and the dimension of the input tensor of the i-th linear module is N × (h·L1).

[0016] Optionally, the residual layer includes H2 sequentially-connected residual blocks, and each residual block includes a first convolutional layer f1, a second activation function a2, a second regularization layer n2, a second convolutional layer f2, and a third activation function a3 that are connected in sequence;

[0017] The two-dimensional output tensor of the H1-th linear layer is made three-dimensional to form the corresponding three-dimensional matrix, and this three-dimensional matrix serves as the input tensor of the first residual block; the dimensions of the input tensor and the output tensor of each residual block are both N × h × L1;

[0018] Let x be the input tensor of the residual block, and the output tensor of the residual block is a3(x + f2(n2(a1(f1(x)))));

[0019] The number of channels of the input tensors of the first convolutional layer f1 and the second convolutional layer f2 is both C, the number of channels of the output tensors is both C, the convolutional kernel size is both K, the convolutional kernel moving amplitude is both 1, and the number of padding input tensors is both (k - 1) / 2.

[0020] Optionally, the upsampling module includes H3 sequentially connected upsampling units, and each upsampling unit includes a first convolutional block and an upsampling layer g connected in sequence; the first convolutional block includes a third convolutional layer f3, a third regularization layer n3, and a fourth activation function a4; let x' be the input of the first convolutional block, the output of the first convolutional block is a4(n3(f3(x'))), and the output of the corresponding upsampling unit is g(a4(n3(f3(x'))));

[0021] The third convolutional layer f3 does not change the dimension of the input tensor of the convolutional block. The number of channels of the input tensor and the output tensor of the third convolutional layer f3 is both h, the convolutional kernel size is 3, the convolutional kernel moving amplitude is 1, and the number of padding input tensors is 1.

[0022] Optionally, the long short-term memory module includes H4 sequentially connected LSTM layers; the dimensions of the input tensor and the output tensor of the LSTM layer are both N×h×L.

[0023] Optionally, the convolutional module includes H5 - 1 second convolutional blocks and a fifth convolutional layer f5 connected in sequence; the second convolutional block includes a fourth convolutional layer f4, a fourth regularization layer n4, and a fifth activation function a5; the number of channels of the input tensor and the output tensor of the fourth convolutional layer f4 is both h, the convolutional kernel size is 3, the convolutional kernel moving amplitude is 1, and the number of padding input tensors is 1;

[0024] The number of channels of the input tensor of the fifth convolutional layer f5 is h, the number of channels of the output tensor is C, the convolutional kernel size is 1, the convolutional kernel moving amplitude is 1, and the number of padding input tensors is 0.

[0025] Optionally, the loss function of the multi-layer deep learning model is the mean squared error function, and the multi-layer deep learning model uses the Adam optimizer.

[0026] Optionally, in step S1, the extracted intersection echo state label vectors are divided into a training set, a validation set, and a test set according to a quantity ratio of 8:1:1.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] The method for modeling the peak characteristics of the echo during missile-target encounter based on deep learning of the present invention fills the gap in predicting the envelope and peak characteristics of the echo signal during missile-target encounter through the big data deep learning method in this field. The present invention extracts the corresponding encounter echo state label vectors from multiple missile-target encounter echo signals for training a multi-layer deep learning model, and realizes the prediction of the target near-field echo under any missile-target encounter condition (including the envelope of the missile-target encounter echo signal and the peak characteristics of the missile-target encounter echo signal) through the trained model. The prediction of the present invention has high accuracy and low computational complexity, and has good popularization value. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solution of the present invention, the drawings required to be used in the description will be briefly introduced below. Obviously, the drawings in the following description are an embodiment of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts:

[0030] Figure 1 It is a schematic diagram of the multi-layer deep learning model structure of the present invention;

[0031] Figure 2 It is a schematic diagram of the residual block structure of the present invention;

[0032] Figure 3 It is a schematic diagram of the upsampling unit structure of the present invention;

[0033] Figure 4a In an embodiment of the present invention, it is a schematic diagram of the envelope of the predicted missile-target encounter echo signal and the envelope of the corresponding missile-target encounter echo signal in the test set;

[0034] Figure 4b 、 Figure 4c 、 Figure 4d respectively are Figure 4a a schematic diagram of the comparison of the number of peaks, peak amplitude, and peak position between the envelope of the predicted missile-target encounter echo signal and the envelope of the corresponding missile-target encounter echo signal in the test set in

[0035] Figure 5a In another embodiment of the present invention, it is a schematic diagram of the envelope of the predicted missile-target encounter echo signal and the envelope of the corresponding actual missile-target encounter echo signal;

[0036] Figure 5b 、 Figure 5c 、 Figure 5d respectively are Figure 5a a schematic diagram of the comparison of the number of peaks, peak amplitude, and peak position between the envelope of the predicted missile-target encounter echo signal and the envelope of the corresponding missile-target encounter echo signal in the test set in DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] The modeling methods for the projectile-target encounter echo include the electromagnetic scattering simulation method based on the target geometric model, the transformation method based on the target scattering center, and the transformation method based on the near-field scattering function.

[0039] The near-field echo electromagnetic scattering modeling method based on the target geometric model includes the following various schemes: 1) Simulate the scattering fields of each scattering unit on the target geometric surface through high-frequency algorithms, and superimpose the scattering fields of all scattering units to obtain the target echo under the near-field encounter conditions. 2) Calculate the physical optical scattering fields of the surface elements of the target geometric model through the decomposition of the antenna incident beam, and simulate the dynamic echo of the target locally irradiated by the antenna under the projectile-target encounter motion state. 3) Establish a detector antenna model and a projectile-target encounter model, and directly use the electromagnetic algorithm in FEKO software to simulate and model the near-field echo of the projectile-target encounter. The modeling of the projectile-target encounter echo based on the target geometric model needs to calculate the echo scattering field through the solution of the physical domain of the target electromagnetic scattering, and does not involve the content of modeling the near-field echo of the projectile-target encounter by methods such as data mining and deep learning.

[0040] The near-field echo modeling method based on the target scattering center includes the following various schemes: 1) The method of transforming the near-field dynamic characteristics of the target from the scattering center model. First, establish a three-dimensional scattering center model of the target under far-field conditions, and then model the change of the dynamic radar cross-section of the target in the near-field projectile-target encounter section through the method of weighting by the observation distance. 2) Use equivalent scattering points to model the Doppler echo of the near-field volume target to realize the echo simulation of the encounter between the short-range radar and the target. 3) Use equivalent scattering points to simulate the scattering effect of the near-field target body and realize it through digital signal channel modulation. The above modeling of the projectile-target encounter target echo is based on the assumptions of the target scattering center or the equivalent scattering point model, and does not involve the echo modeling method based on data mining and deep learning of the projectile-target encounter echo.

[0041] The near-field echo modeling method based on the near-field scattering function includes the projectile-target encounter echo simulation method based on the weighted transformation of the near-field scattering function. The near-field scattering function is obtained by scanning the cylindrical enclosing surface of the target for dipole antenna incidence and reception. It does not involve the echo modeling method of big data and deep learning for the projectile-target encounter echo.

[0042] The present invention provides a modeling method for the peak characteristics of the projectile-target encounter echo based on deep learning, which includes the steps:

[0043] S1. Collect multiple missile-target intersection echo signals, and extract the corresponding intersection echo state label vectors respectively; establish a training set, a validation set, and a test set based on the extracted intersection echo state label vectors; in this embodiment, collect 10,000 missile-target intersection echo signals, and divide the extracted intersection echo state label vectors into a training set, a validation set, and a test set according to the quantity ratio of 8:1:1.

[0044] For the near-field intersection scenario of a missile and a target, on the premise that the radar incident wavelength and antenna pattern parameters are fixed, change the missile body attitude (pitch angle α m , azimuth angle β m , roll angle γ m ), target attitude (pitch angle α t , azimuth angle β t , roll angle γ t ), as well as the target miss distance ρ and miss azimuth ψ in the missile-target relative motion coordinate system. The change of the target scatterer within the radar antenna beam illumination range causes the change of the target near-field echo.

[0045] In this embodiment, the intersection echo state label vector includes: missile body pitch angle α m , missile body azimuth angle β m , single body roll angle γ m , target pitch angle α t , target azimuth angle β t , target roll angle γ t , as well as the target miss distance ρ and miss azimuth ψ in the missile-target relative motion coordinate system. Among them, α m , β m , γ m , α t , β t , γ t ∈[-180°, 180°], the miss distance ρ∈[0, 50m], and the miss azimuth ψ∈[0, 360°].

[0046] S2. Establish a multi-layer deep learning model for predicting the envelope of the missile-target intersection echo signal and the peak characteristics of the missile-target intersection echo signal.

[0047] The multi-layer deep learning model takes the missile-target intersection state vector as the input of the model, and predicts the envelope of the missile-target intersection echo signal and the peak characteristics through the model. In this embodiment, the envelope of the missile-target intersection echo signal includes: intersection coordinate position and echo amplitude; the peak characteristics of the missile-target intersection echo signal include: the number of peaks, peak positions, and peak amplitudes.

[0048] As Figure 1 shown, the multi-layer deep learning model includes a linear module, a residual layer, an upsampling module, a long short-term memory module, and a convolutional module connected in sequence.

[0049] To better describe the model, the following parameters are defined first:

[0050] H1, H2, H3, H4, and H5 are the depths of the linear module, residual layer, upsampling module, long short-term memory layer, and convolutional module, respectively.

[0051] d is the dimension of the rendezvous echo state label vector (the number of vectors in the rendezvous echo state label vector). In this embodiment, d = 8;

[0052] C is the number of channels of the model output signal;

[0053] L is the length of the model output signal;

[0054] h is a hyperparameter that controls the dimensionality change of the input during model propagation.

[0055] In this embodiment, d = 8, C = 2, L = 2048, h = 24, H1 = 20, H2 = 40, H3 = 10, H4 = 10, and H5 = 10.

[0056] In this embodiment, the number of channels of the model output signal is 2. One channel represents the position of the output signal, and the other channel represents the amplitude of the output signal. The length of the output signal is 2048. That is, an output signal (predicted missile-target rendezvous echo signal) can be represented by 2048 binary tuples (a i , b i ), where i ∈ [1, 2048]. a i corresponds to the channel representing the position, and b i corresponds to the channel representing the amplitude. In this way, the envelope of the predicted missile-target rendezvous echo signal can be obtained, and the envelope peak feature can be extracted from this envelope.

[0057] As Figure 1 shown, the linear module includes H1 linearly connected linear layers. The linear layer includes a fully connected layer FC, a first regularization layer n1, and a first activation function a1 connected in sequence.

[0058] Let L i represent the i-th linear layer, and x i represent the input tensor of the i-th linear layer, where i ∈ [1, H1]; the input tensor of the (i + 1)-th linear layer is the output tensor of the i-th linear layer, and x i+1 = L i (x i ).

[0059] Among them, the dimension of the input tensor x1 of the first linear layer is N × d. Each row of x1 corresponds to a rendezvous echo state label vector, and the dimension of the output tensor of the first linear layer is N × h.

[0060] The fully connected layer FC performs a simple linear operation. Taking the first fully connected layer FC in the linear module as an example:

[0061] Let x represent the input of the first fully connected layer FC, and y represent the output of the first fully connected layer FC. y = Wx + b. x = x1, and x is an (N, d)-dimensional matrix. The network parameters of the first fully connected layer FC include W and b. W is a d×h-dimensional matrix, and b is an h-dimensional vector.

[0062] In this embodiment, the first regularization layer n1 adopts BN (Batch Normalization). The core idea of BN is a normalization operation of learnable parameters, which reduces the sensitivity to the learning rate and improves the training speed and final performance. An intuitive explanation is that BN forcibly converts the input into a state similar to a normal distribution, with an empirical expectation of 0 and an empirical variance of 1 for its output, thereby reducing the sensitivity to the gradient. If BN is not used, the input distribution of each layer of the model will strongly depend on the output of the previous layer. If the output of a certain layer is not ideal, there will be a problem of model collapse, while BN constrains the input distribution to reduce the sensitivity.

[0063] The first activation function a1 is the ReLU (Rectified Linear Unit) activation function. ReLU is one of the most widely used activation functions currently. Its advantage is that there is no problem of gradient disappearance like the activation functions Sigmoid and tanh, and the calculation is simpler.

[0064] When 1 < i ≤ L1, the input tensor x of the i-th linear layer i has a dimension of N×[(i - 1)·h]; the dimension of the output tensor of the i-th linear layer is N×(i·h); L is the length of the output signal of the multi-layer deep learning model;

[0065] When i > L1, the input tensor x of the i-th linear layer i has a dimension of N×(h·L1), and the dimension of the input tensor of the i-th linear module is N×(h·L1).

[0066] As Figure 1 shown, the residual layer contains H2 sequentially connected residual blocks (Res-Block), as Figure 2As shown, each of the residual blocks includes a first convolutional layer f1, a second activation function a2, a second regularization layer n2, a second convolutional layer f2, and a third activation function a3 connected in sequence. The residual block is used to define the information flow mode between layers. Its core lies in the skip connection - allowing the next layer of the network to use the sum of the output and input of this layer as the input, which is equivalent to allowing this layer of the network to learn the changes on the input rather than directly learning the expected output. The skip connection has significantly improved the network performance in multiple tasks.

[0067] The input of the first residual block is the output of the H1-th linear layer. The dimension of the output tensor of the last linear layer is N×(h·L1). The two-dimensional output tensor of the H1-th linear layer is stereoscopically transformed into a corresponding three-dimensional matrix, and the dimension of this three-dimensional matrix is N×h×L1 (this is the prior art). The stereoscopically transformed three-dimensional matrix is used as the input tensor of the first residual block. In the present invention, each residual block does not change the dimensions of the input tensor and the output tensor, and the dimensions of its input tensor and output tensor are both N×h×L1.

[0068] The number of channels of the input tensors of the first convolutional layer f1 and the second convolutional layer f2 is both C, the number of channels of the output tensors is both C, the kernel size is both K, the kernel movement amplitude is both 1, and the number of input tensors to be padded is both (k - 1) / 2. Denote r as the input tensor of the residual block, and the output tensor of this residual block is a3(x + f2(n2(a1(f1(r))))). In this embodiment, the first convolutional layer f1 and the second convolutional layer f2 can be represented by conv(C, C, K, 1, (K - 1) / 2).

[0069] In this embodiment, the second regularization layer n2 uses BN, and the second activation function a2 and the third activation function a3 are ReLU.

[0070] As Figure 1 shown, the upsampling module includes H3 upsampling units (up-Samplers) connected in sequence. Each upsampling unit includes a first convolutional block and an upsampling layer g connected in sequence; as Figure 3 shown, the first convolutional block includes a third convolutional layer f3, a third regularization layer n3, and a fourth activation function a4 connected in sequence; let x′ be the input of the first convolutional block, and the output of the first convolutional block is a4(n3(f3(x′))), and the output of the corresponding upsampling unit is g(a4(n3(f3(x′)))).

[0071] The third convolutional layer f3 does not change the dimension of the input tensor of the convolutional block. The number of channels of the input tensor and the output tensor of the third convolutional layer f3 is both h, the kernel size is 3, the kernel movement amplitude is 1, and the number of input tensors to be padded is 1. The third convolutional layer f3 can be represented by conv(h, h, 3, 1, 1).

[0072] The upsampling layer g is essentially a linear interpolation process with learnable weights. To define an upsampling layer, it is necessary to implement the agreed number of input channels C, which only contains C learnable parameters w1, w2,..., w C . Use an example to describe its calculation process: Upsample the 1×2 tensor x = [x1, x2]. The corresponding upsampling layer requires a parameter w1, and the output tensor is [x1, w1·x1+(1 - w1)·x2, x2, w1·x2]. Similarly, generalize to the case where the output tensor is an arbitrary 1×N tensor, N is an arbitrary constant, and at this time the corresponding upsampling layer still only requires one parameter. For example, for the output tensor of x = [x1, x2,..., x N , the result of its upsampling is:

[0073] [x1, w1·x1+(1 - w1)·x2, x2, w1·x2+(1 - w1)·x3, x3, w1·x3+(1 - w1)·x4,...].

[0074] For the case where the input tensor dimension is C×N, its calculation process is to regard the input tensor as C 1×N tensors stacked together. These tensors respectively pass through an upsampling layer with only one parameter, and then stack the obtained C 1×N output tensors together to form a C×N tensor as the final result.

[0075] In this embodiment, the long short-term memory module includes H4 sequentially connected long short-term memory layers (Long Short-Term Memory Layer); the long short-term memory layer includes multiple standard LSTM units (one unit corresponds to one step executed by the long short-term memory layer). The dimensions of the input tensor and output tensor of the LSTM layer are both N×h×L. The long short-term memory module does not change the dimension of the input tensor and sequentially passes the input tensor into the H4 long and short memory layers. The purpose of this part is to artificially introduce temporal information into the input. Experiments have found that introducing this part will greatly increase the time consumption and needs to be selected according to actual needs. Usually, it is introduced when the number of samples is small (that is, when N is relatively small).

[0076] The long short-term memory layer simulates the human memory mechanism, and its calculation process is divided into three stages: the forgetting stage, the selective memory stage, and the output stage. As Figure 1 shown, the convolutional module includes H5 - 1 sequentially connected second convolutional blocks and the fifth convolutional layer f5; the second convolutional block includes the fourth convolutional layer f4, the fourth regularization layer n4, and the fifth activation function a5 connected in sequence; the number of input and output channels of the fourth convolutional layer f4 is h, the convolutional kernel size is 3, the convolutional kernel movement amplitude is 1, and the number of padding input tensors is 1. The fourth convolutional layer f4 can be expressed as conv(h, h, 3, 1, 1).

[0077] The number of channels of the input tensor of the fifth convolutional layer f5 is h, the number of channels of the output tensor is C, the size of the convolutional kernel is 1, the moving amplitude of the convolutional kernel is 1, and the number of supplementary input tensors is 0. The fifth convolutional layer f5 can be expressed as conv(h, C, 1, 1, 0).

[0078] In this embodiment, the loss function of the multi-layer deep learning model is the least square error function, and the multi-layer deep learning model adopts the Adam optimizer.

[0079] The design idea of the multi-layer deep learning model of the present invention includes two premises: one is that the model should be as deep as possible to make the fitting of the model strong enough; the other is that the model should have at least two parts: ① the part that maps the input parameters to the high-dimensional space, and this part should be relatively deep; ② the part that "stretches" the high-dimensional input, and an LSTM can be introduced in this part to artificially introduce temporal information. The dimension of the input tensor of the model is N×d, and the dimension of the output tensor is N×C×L.

[0080] According to the previous idea, the first part maps the N×d tensor to an N×D tensor, where D = h×L1, that is, the high-dimensional representation of the learning data. Since the model finally needs to obtain a 3D signal, this N×D tensor should be regarded as the result after stretching the N×h×L1 tensor, and this tensor is re-stereotyped into an N×h×L1 tensor. The second part will map the N×h×L1 tensor to an N×C×L tensor. For practical considerations, this part should consider the case where L1 is much smaller than L, that is, upsampling, to save the memory occupied by the model. Since it is already an upsampling operation in terms of length, it is hoped that it is a downsampling operation in terms of channels to reduce the training difficulty, that is, h is much larger than C.

[0081] S3. Train the multi-layer deep learning model through the training set; the validation set is used to adjust the learning rate of the multi-layer deep learning model when training the multi-layer deep learning model through the training set;

[0082] In this embodiment, the training of the model is divided into two processes: the preprocessing stage and the normal stage. In the preprocessing stage, that is, in the first 100 iterations, the learning rate of the optimizer linearly increases from 1e-6 to 1e-4. In the normal stage, that is, after 100 iterations, when the model iterates each time, in addition to calculating the loss and backpropagating the gradient according to the data on the training set, it will additionally evaluate the performance of the model on the validation set. If the performance is worse than the previous iteration, the learning rate is reduced by half and the iteration continues; if the performance has not improved for 50 consecutive iterations, it is considered that the model has converged and the training ends.

[0083] S4. Input the test set into the trained multi-layer deep learning model and evaluate the accuracy of the multi-layer deep learning model.

[0084] In the actual application of the present invention, it is not necessary to predict all the wave peaks of the missile-target encounter echo signal. It is only necessary to predict several wave peaks with the largest peak amplitudes. The wave peaks of the predicted missile-target encounter echo signal envelope and the wave peaks of the missile-target encounter echo signal envelope for training are arranged from largest to smallest, and the amplitudes and positions of the corresponding wave peaks are compared. For each pair of corresponding wave peaks, their relative amplitude errors and relative position errors are calculated, and finally the average of all wave peak amplitudes and position errors is taken.

[0085] On the test set, the peak amplitude error of the encounter echo is 12.15%, and the peak position error is 13.07%. Two groups of prediction results of the missile-target encounter echo peak envelopes are selected, as shown in Figures 4a to 4d 、 Figures 5a to 5d respectively. Among them, Figures 4a to 4d represents the case where the number of predicted peaks is less than the number of test echo peaks, and Figures 5a to 5d represents the case where the number of predicted peaks is more than the number of test echo peaks. The experimental results show that through the deep learning training of the missile-target encounter echo big data, a deep learning multi-layer neural network model for predicting the characteristics of the echo peak envelope can be established to realize the prediction and modeling of the missile-target encounter peak envelope and characteristic parameters.

[0086] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0087] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for modeling the peak feature of the echo during missile-target encounter based on deep learning, characterized in that, Including the steps: S1. Collect multiple projectile-target encounter echo signals, and respectively extract the corresponding encounter echo state label vectors; establish a training set, a validation set, and a test set based on the extracted encounter echo state label vectors; S2. Establish a multi-layer deep learning model for predicting the envelope of the projectile-target encounter echo signal and the peak characteristics of the projectile-target encounter echo signal; the multi-layer deep learning model includes a linear module, a residual layer, an upsampling module, a long short-term memory module, and a convolutional module connected in sequence; the depths of the linear module, the residual layer, the upsampling module, the long short-term memory layer, and the convolutional module are H1, H2, H3, H4, and H5 respectively; The linear module includes H1 linear layers connected in sequence, and each linear layer includes a fully connected layer FC, a first regularization layer n1, and a first activation function a1 connected in sequence; Let L i represent the i-th linear layer, and x i represent the input tensor of the i-th linear layer, where i ∈ [1, H1]; the input tensor of the (i + 1)-th linear layer is the output tensor of the i-th linear layer, and x i+1 = L i (x i ); The dimension of x1 is N×d, each row of x1 corresponds to an encounter echo state label vector, and d is the number of vectors in the encounter echo state label vector; the dimension of the output tensor of the first linear layer is N×h; h is a hyperparameter that controls the dimension change of the input tensor during the propagation of the model; When \(1 < i \leq L1\), the input tensor \(x\) of the \(i\)-th linear layer i has a dimension of \(N\times[(i - 1)\cdot h]\); the output tensor of the \(i\)-th linear layer has a dimension of \(N\times(i\cdot h)\); \(L1 = L / 2\) H3 , where \(L\) is the length of the output signal of the multi-layer deep learning model; When i > L1, the input tensor x of the i-th linear layer i has the dimension of N×(h·L1), and the input tensor of the i-th linear module has the dimension of N×(h·L1); S3. Train the multi-layer deep learning model through the training set; the validation set is used to adjust the learning rate of the multi-layer deep learning model when training the multi-layer deep learning model through the training set; S4. Input the test set into the trained multi-layer deep learning model to evaluate the accuracy of the multi-layer deep learning model.

2. The method for modeling the peak feature of the echo of missile-target encounter based on deep learning according to claim 1, wherein The encounter echo state label vector includes the pitch angle of the projectile, the azimuth angle of the projectile, the roll angle of the projectile, the pitch angle of the target, the azimuth angle of the target, the roll angle of the target, as well as the target miss distance and the miss azimuth in the projectile-target relative motion coordinate system.

3. The method for modeling the peak feature of the echo of the missile-target encounter based on deep learning according to claim 1, wherein The envelope of the projectile-target encounter echo signal includes: the encounter coordinate position and the echo amplitude; the peak characteristics of the projectile-target encounter echo signal include: the number of peaks, the peak positions, and the peak amplitudes.

4. The method for modeling the peak feature of the echo of the missile-target encounter based on deep learning according to claim 1, wherein The residual layer includes H2 residual blocks connected in sequence, and each residual block includes a first convolutional layer f1, a second activation function a2, a second regularization layer n2, a second convolutional layer f2, and a third activation function a3 connected in sequence; Stereoify the two-dimensional output tensor of the H1-th linear layer into a corresponding three-dimensional matrix, and this three-dimensional matrix serves as the input tensor of the first residual block; the dimensions of the input tensor and the output tensor of each residual block are both N×h×L1; Denote x as the input tensor of the residual block, and the output tensor of this residual block is a3(x + f2(n2(a1(f1(x))))); The number of channels of the input tensors of the first convolutional layer f1 and the second convolutional layer f2 is both C, the number of channels of the output tensors is both C, the size of the convolutional kernel is both K, the moving amplitude of the convolutional kernel is both 1, and the number of padding input tensors is both (k - 1) / 2.

5. The method for modeling the peak feature of the echo of missile-target encounter based on deep learning according to claim 4, wherein The upsampling module includes H3 sequentially connected upsampling units, and each upsampling unit includes a first convolutional block and an upsampling layer g connected in sequence; the first convolutional block includes a third convolutional layer f3, a third regularization layer n3, and a fourth activation function a4 connected in sequence; let x' be the input of the first convolutional block, the output of the first convolutional block is a4(n3(f3(x'))), and the output of the corresponding upsampling unit is g(a4(n3(f3(x')))); The third convolutional layer f3 does not change the dimension of the input tensor of the convolutional block. The number of channels of the input tensor and the output tensor of the third convolutional layer f3 are both h, the convolutional kernel size is 3, the convolutional kernel moving amplitude is 1, and the number of padding input tensors is 1.

6. The method for modeling the peak feature of the echo at the missile-target encounter based on deep learning according to claim 5, wherein The long short-term memory module includes H4 sequentially connected LSTM layers; the dimensions of the input tensor and the output tensor of the LSTM layer are both N×h×L.

7. The method for modeling the peak feature of the echo of the missile-target encounter based on deep learning according to claim 6, wherein The convolutional module includes H5-1 sequentially connected second convolutional blocks and a fifth convolutional layer f5; the second convolutional block includes a fourth convolutional layer f4, a fourth regularization layer n4, and a fifth activation function a5 connected in sequence; the number of channels of the input tensor and the output tensor of the fourth convolutional layer f4 are both h, the convolutional kernel size is 3, the convolutional kernel moving amplitude is 1, and the number of padding input tensors is 1; The number of channels of the input tensor of the fifth convolutional layer f5 is h, the number of channels of the output tensor is C, the convolutional kernel size is 1, the convolutional kernel moving amplitude is 1, and the number of padding input tensors is 0.

8. The method for modeling the peak feature of the echo of the missile-target encounter based on deep learning according to claim 1, wherein The loss function of the multi-layer deep learning model is the mean squared error function, and the multi-layer deep learning model uses the Adam optimizer.

9. The method for modeling the peak feature of the echo of missile-target encounter based on deep learning according to claim 1, wherein In step S1, the extracted intersection echo state label vectors are divided into a training set, a validation set, and a test set according to the quantity ratio of 8:1:1.

Images