Physical model driven complex natural resonance frequency correction method and device

Abstract

The present disclosure provides a physical model-driven complex natural resonance frequency correction method and device, which belongs to the field of radar signal processing technology. The method uses a resonant scattering mechanism model to calculate the theoretical value of the complex natural resonance frequency corresponding to the target. The complex natural resonance frequency data enters the C‑VAE network, and the C‑VAE outputs a correction value; the difference with the correction value is used to construct the loss function term of the target resonance scattering mechanism; the input data and the correction value are used to construct the loss function term of the C‑VAE network. And the composition loss function; the C‑VAE network is trained using the loss function, and the trained C‑VAE network is used for complex natural resonance frequency correction. The use of the present invention can provide a more accurate data set for subsequent radar target recognition, which is conducive to improving the accuracy of radar target recognition based on complex natural resonance frequency characteristics.

Description

Technical Field

[0001] The invention belongs to the field of radar signal processing, radar target detection and recognition, and in particular relates to a physical model driven complex natural resonance frequency correction method and device. Background Art

[0002] As one of the important research directions in the development and application of modern radar technology, radar target recognition plays an extremely important role in the fields of sea situation monitoring, topographic mapping, weather forecasting, etc. At present, the research on radar target recognition focuses on the characteristics of the optical region, and the resonance region characteristics of the target are less utilized. However, the optical region characteristics such as RCS and ISAR images are sensitive to the incident direction and target attitude, and a sufficient database needs to be built. The complex natural resonance frequency feature is the most critical feature of the radar target in the resonance region. It is only related to the inherent properties of the target such as size, shape, and material, and is insensitive to factors such as the incident angle, polarization mode, and target attitude. Therefore, radar target recognition based on the complex natural resonance frequency feature has important research value.

[0003] However, the current research on complex natural resonance frequency characteristics mainly has the following problems:

[0004] 1. The existing complex natural resonance frequency feature extraction process will introduce calculation errors, resulting in inaccurate target feature acquisition, the existence of false features and feature data offset, which seriously restricts the improvement of radar target recognition accuracy based on complex natural resonance frequency features.

[0005] 2. In order to solve the problem of inaccurate data caused by errors in the extraction process, existing radar target recognition methods generally use deep learning networks to automatically learn the high-dimensional features of data to restore data accuracy. However, they only rely on data-driven and lack physical interpretability. There is a problem that the network reconstructed data is not consistent with the target resonant scattering mechanism.

[0006] Therefore, a complex natural resonance frequency correction method driven by a physical model and capable of effectively restoring feature accuracy is designed to provide a more accurate data set for subsequent radar target recognition, which is conducive to improving the accuracy of radar target recognition based on complex natural resonance frequency features. Summary of the invention

[0007] In view of this, the present invention provides a physical model driven complex natural resonance frequency correction method and device, which can provide a more accurate data set for subsequent radar target identification, which is conducive to improving the accuracy of radar target identification based on complex natural resonance frequency characteristics.

[0008] The present invention introduces a physical model to drive the network to correct the complex natural resonant frequency data, and the corrected data can be used for radar target recognition based on the complex natural resonant frequency characteristics. The processing method of the present invention can also be applied to tasks such as correction of radar target resonance area echoes.

[0009] In order to solve the above technical problems, the present invention is implemented as follows.

[0010] A physical model driven complex natural resonant frequency correction method, comprising:

[0011] Step 1: Use the resonant scattering mechanism model to calculate the theoretical value of the complex natural resonant frequency corresponding to the target ;

[0012] Step 2: Input the complex natural resonant frequency data Enter the complex domain variational autoencoder network C-VAE, and the C-VAE network outputs the correction value ; The C-VAE network includes an encoder and a decoder. The encoder consists of multiple layers of complex domain one-dimensional convolutional layers and complex domain linear layers, and the decoder consists of multiple layers of complex domain one-dimensional transposed convolutional layers and complex domain linear layers;

[0013] Step 3: Using the theoretical value of the complex natural resonant frequency With correction value The difference between the two constructs the loss function term of the target resonant scattering mechanism. ; Using the input complex natural resonant frequency data and correction value Constructing the loss function term of the C-VAE network ; Loss function term of target resonant scattering mechanism And the loss function term of the C-VAE network The loss function is composed of the following components: When training the C-VAE network, the network training goal is to minimize the value of the loss function, so that the correction value of the complex natural resonant frequency The consistency with the target resonant scattering mechanism is improved, and the correction value The complex natural resonant frequency data of the input The error between them is the smallest. The trained C-VAE network is used for complex natural resonant frequency correction.

[0014] Preferably, in step 1, the resonant scattering mechanism model is used to calculate the theoretical value of the complex natural resonant frequency corresponding to the target. for:

[0015] According to the minor semi-axis of the target a and the major semiaxis b Get the local curvature radius of the surface wave in the propagation direction and creeping arc length The complex natural resonant frequency is introduced Function , substitute into the target resonance equation and solve the theoretical value of the target complex natural resonance frequency :

[0016] The target resonance equation is:

[0017]

[0018] In the formula, is the number of crawling circles, is the complex natural resonant frequency The function of a , b are the minor and major semi-axis of the target, respectively; is the angle between the incident direction of the electromagnetic wave and the major axis of the ellipsoid.

[0019] Preferably, in step 3, the loss function term of the target resonant scattering mechanism is for:

[0020] .

[0021] Preferably, in step 3, the loss function term of the C-VAE network is Includes the reconstruction error term, which is the correction value of the complex natural resonant frequency The complex natural resonant frequency data of the input The mean square error between .

[0022] Preferably, in step 3, the loss function term of the C-VAE is A regularization term is further included, which is the KL divergence between the distribution of the encoder output in the C-VAE network and the standard normal distribution.

[0023] Preferably, in step 2, the complex domain one-dimensional convolution layer, the complex domain one-dimensional transposed convolution layer and the complex domain linear layer in the C-VAE network all adopt a complex domain network structure layer;

[0024] The input data and weights of the complex domain network structure layer are all in complex form; the complex domain network structure layer converts the real part of the complex input data A and the real part of the complex weight C , the imaginary part of the complex input data B and the imaginary part of the complex weight D Multiply them separately and take the difference to get the new real part ; Input the real part of the complex number data A The imaginary part D of the complex weight and the imaginary part of the complex input data B and the real part of the complex weight CMultiply them separately and sum them to get the new imaginary part ; New real and imaginary composite complex numbers Output.

[0025] Preferably, the method further comprises:

[0026] Step 4: Determine the input complex natural resonant frequency data Is the corresponding loss function value less than the set threshold? If so, output the correction value ; Otherwise, determine the complex natural resonant frequency data currently input is abnormal data.

[0027] The present invention also provides a physical model driven complex natural resonance frequency correction device, which includes a complex natural resonance frequency theoretical value calculation module, a C-VAE network, a loss function determination module, a training module and a judgment and reasoning module;

[0028] The complex natural resonant frequency theoretical value calculation module is used to calculate the theoretical value of the complex natural resonant frequency corresponding to the target using the resonant scattering mechanism model , sent to the loss function determination module;

[0029] The C-VAE network includes an encoder and a decoder, wherein the encoder is composed of multiple layers of complex domain one-dimensional convolution and complex domain linear layers, and the decoder is composed of multiple layers of complex domain one-dimensional transposed convolution and complex domain linear layers; the input complex natural resonant frequency data Enter the C-VAE network, and the C-VAE network outputs the correction value ; Input complex natural resonant frequency data and correction value Provided to the loss function determination module;

[0030] The loss function determination module receives the theoretical value , correction value , input complex natural resonant frequency data , substitute into the loss function, and output the loss function value; the loss function includes: using the theoretical value of the complex natural resonant frequency With correction value The loss function of the target resonant scattering mechanism constructed by the difference , and using the complex natural resonant frequency data of the input and correction value The loss function term of the constructed C-VAE network ;

[0031] The training module, when training the C-VAE network, takes the value of the minimization loss function as the network training target, so that the correction value of the complex natural resonant frequency The consistency with the target resonant scattering mechanism is improved, and the correction value The complex natural resonant frequency data of the input The error between them is minimal;

[0032] The judgment and reasoning module is used to judge the input complex natural resonance frequency data Is the corresponding loss function value less than the set threshold? If so, it is considered that the current input complex natural resonant frequency data For normal data, output correction value ; Otherwise, determine the complex natural resonant frequency data currently input is abnormal data.

[0033] Preferably, the loss function used by the loss function determination module is Loss for:

[0034]

[0035]

[0036]

[0037] in, for The reconstruction error term in for The regularization term in ; is the latent variable distribution of the encoder output in the C-VAE network, is the prior distribution and is assumed to be a standard normal distribution, for and KL divergence of two distributions; is the number of crawling circles, is the local radius of curvature of the surface wave in the propagation direction, is the creeping arc length of the surface wave, is the complex natural resonant frequency The function of a , b are the minor and major semi-axis of the target, respectively; is the angle between the incident direction of the electromagnetic wave and the major axis of the ellipsoid; , , is the weighting coefficient; It means mean square error.

[0038] Preferably, the complex domain one-dimensional convolution layer, the complex domain one-dimensional transposed convolution layer and the complex domain linear layer in the C-VAE network all adopt a complex domain network structure layer; the input data and weights of the complex domain network structure layer are in complex form;

[0039] The complex domain network structure layer includes a real part processing module, an imaginary part processing module, a difference finder, a summer, and a fusion module;

[0040] The real part processing module converts complex input data The real part of A and the imaginary part B With complex weights The real part of C Multiply to get and , sent to the difference finder and the summer;

[0041] The imaginary part processing module converts complex input data The real part of A and the imaginary part B With complex weights The real part of D Multiply to get and , sent to the difference finder and the summer;

[0042] The subtractor computes the new real part ;

[0043] The summer calculates the new imaginary part ;

[0044] The fusion module combines the new real part and the new imaginary part into a complex number Output.

[0045] Beneficial effects:

[0046] (1) The present invention incorporates the target resonant scattering mechanism into the loss function of network learning, and drives network learning with the target's resonant scattering mechanism, which can enhance the physical interpretability of the algorithm and effectively restore the accuracy of the target's complex natural resonant frequency data, reduce the impact of computational errors in the extraction process on subsequent recognition, and improve the target recognition accuracy based on the complex natural resonant frequency characteristics.

[0047] (2) The loss function of the complex domain variational autoencoder network of the present invention consists of two parts: the reconstruction error term and the regularization term. The reconstruction error is the corrected data With input data The regularization term is the KL divergence between the distribution of the encoder output and the standard normal distribution. Through network training, the gradients and weights in the network are continuously updated to reduce the loss function value, thereby minimizing the reconstruction error and maximizing the similarity between the encoder output distribution and the standard normal distribution. Compared with traditional autoencoder networks, the variational autoencoder network introduces regularization, which can effectively reduce the overfitting phenomenon caused by the high degree of freedom of the autoencoder network, and constrain the network to extract effective information from the input data instead of blindly copying the input data.

[0048] (3) In a preferred embodiment of the present invention, a loss function threshold is set to calculate the loss function value of the test data after it passes through the network. If the loss function value exceeds the set threshold, it is determined to be abnormal data, thereby preventing the input of abnormal data from affecting the network effect.

[0049] (4) The present invention designs a complex domain variational autoencoder network C-VAE, which uses complex domain one-dimensional convolution, complex domain one-dimensional transposed convolution and complex domain linear layer to replace real-valued one-dimensional convolution, one-dimensional transposed convolution and linear layer, making full use of the correlation between the real part and the imaginary part of the complex natural resonant frequency characteristics to explore richer intrinsic features. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the principle of the complex natural resonance frequency correction method driven by the physical model of the present invention.

[0051] Figure 2 The present invention is a flowchart of the complex natural resonant frequency correction method driven by the physical model.

[0052] Figure 3 This is the reconstruction effect diagram.

[0053] Figure 4 Implement methods for complex networks.

[0054] Figure 5 This is a structural diagram of the complex natural resonance frequency correction device driven by the physical model of the present invention. DETAILED DESCRIPTION

[0055] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0056] The present invention designs a physics model driven complex natural resonant frequency correction method (Physics Guided Complex Natural Resonant Frequency Correction, PG-CNRFC), which can effectively solve the problem of feature inaccuracy caused by calculation errors in the process of complex natural resonant frequency extraction, improve data accuracy while complying with the target resonant scattering mechanism, and realize the correction of complex natural resonant frequency characteristics, thereby effectively improving the accuracy of subsequent radar target recognition.

[0057] The core idea of ​​the present invention is to design a loss function based on the target resonant scattering mechanism, introduce the target's resonant scattering mechanism to drive the learning of the network, obtain the complex natural resonant frequency characteristics that are more consistent with the target resonant scattering mechanism, effectively realize the correction of data, and thus effectively improve the accuracy of subsequent radar target recognition; convert the variational autoencoder network based on real-valued operations into a complex domain variational autoencoder network (Complex Domian Variational AutoEncoder, C-VAE), fully consider the correlation between the real part and the imaginary part of the complex natural resonant frequency characteristics, and allow the input, output and weight of the model to be complex data.

[0058] The specific structure of the complex natural resonant frequency correction method driven by the physical model is as follows: Figure 1 As shown. Figure 1 It can be seen that the overall loss function corresponding to the upper and lower channels includes two parts. One part is the loss function based on the target resonant scattering mechanism. , and the other part is the loss function of the complex domain variational autoencoder network , forming the overall loss function.

[0059] The construction of these two loss functions is described below.

[0060] 1. Loss function based on target resonant scattering mechanism

[0061] The idea of ​​constructing this loss function is to introduce the resonant scattering mechanism of the target and use the resonant scattering mechanism model to calculate the theoretical value of the complex natural resonant frequency corresponding to the target. , the loss function is constructed as the corrected data Theoretical value of complex natural resonant frequency The difference is minimized, and the corrected data is constrained The consistency with the target resonant scattering mechanism.

[0062] Specifically, for an ellipsoid, the creeping wave trajectory on the spherical surface will pass through two caustics, each of which produces The phase jump of the target is therefore the theoretical value of the complex natural resonance With the target minor semi-axis , semi-major axis The resonance equation satisfied between them is shown in formula (1).

[0063] (1)

[0064] in, is the number of crawling circles, is the local radius of curvature of the surface wave in the propagation direction, is the creeping arc length of the surface wave, is the complex natural resonant frequency function. and The expressions of are shown in formulas (2) and (3).

[0065] (2)

[0066] (3)

[0067] For TM waves, The expression is

[0068] (4)

[0069] For TE waves, The expression is

[0070] (5)

[0071] in, , is the complex natural resonant angular frequency, is the speed of light, is the angle between the incident direction of the electromagnetic wave and the major axis of the ellipsoid, , , are the zeros of the Airy integral and its derivative, respectively.

[0072] The corrected complex natural resonant frequency data Substitute the deviation value of the target resonance equation as the loss function based on the target resonance scattering mechanism , integrated into the overall loss function of the algorithm, training the network to reduce the loss function based on the target resonant scattering mechanism , improve the consistency between the corrected data and the target resonant scattering mechanism. Loss function based on the target resonant scattering mechanism As shown in formula (6).

[0073] (6)

[0074] Formula (6) can be equivalent to , so the loss function based on the target resonant scattering mechanism can be expressed as , which is equivalent to calculating the corrected data Theoretical value of complex natural resonant frequency The mean square error (MSE) between the two is used to train the model to reduce the loss function based on the target resonant scattering mechanism, and constrain the consistency between the corrected data and the target resonant scattering mechanism.

[0075] 2. Complex Domain Variational Autoencoder Network

[0076] The complex domain variational autoencoder network C-VAE includes an encoder and a decoder, wherein the encoder is composed of multiple layers of complex domain one-dimensional convolutions and complex domain linear layers, and the decoder is composed of multiple layers of complex domain one-dimensional transposed convolutions and complex domain linear layers. The difference from the traditional real-valued variational autoencoder network is that the present invention uses complex domain one-dimensional convolutions, complex domain one-dimensional transposed convolutions and complex domain linear layers to replace real-valued one-dimensional convolutions, one-dimensional transposed convolutions and linear layers, transforming the existing real-valued operation network into a network structure capable of processing complex data, making full use of the correlation between the real and imaginary parts of the complex natural resonant frequency characteristics, and mining richer intrinsic features.

[0077] The complex domain one-dimensional convolution, complex domain one-dimensional transposed convolution, and complex domain linear layer in the C-VAE network all use the complex domain network structure layer. The input data and weights of the complex domain network structure layer are all in complex form.

[0078] The specific operation process of the complex domain network structure layer is as follows: Assume that the data input to the complex domain network structure layer is , for the first complex domain one-dimensional convolutional layer of the encoder, the input data is the complex natural resonant frequency , for other layers, the input data is the complex features from the previous layer. The complex weights in the network are , the complex convolution result of the data output by the complex domain network structure layer is the Hadamard product of the two, as shown in formula (7).

[0079] (7)

[0080] It can be seen from formula (7) that the real part and imaginary part of the complex natural resonant frequency are related, and they jointly affect the real part and imaginary part of the network output characteristics, which can obtain richer features than real-valued operations. Therefore, in order to obtain richer features, the present invention adopts the real part of the complex natural resonant frequency A and the imaginary part B After splitting, they are sent to the complex domain network structure layer respectively, and then the real part of the complex natural resonant frequency is Aand the imaginary part of the complex weights of the network layer D , the imaginary part of the complex natural resonant frequency B and the real part of the complex weights of the network layer C Multiply them separately and sum them to get the new imaginary part At the same time, the product of the real parts of the two and the product of the imaginary parts of the two are subtracted to obtain the new real part . Combine the two parts into a plural form , which is the complex feature obtained by the complex network layer from the complex natural resonant frequency of the input. The specific implementation of the complex domain network structure layer is as follows Figure 4 shown.

[0081] The main difference between the variational autoencoder (VAE) and the traditional autoencoder (AE) is that the encoder in AE outputs the specific value of the latent variable, while the encoder in VAE outputs the specific distribution of the latent variable. It consists of two parts: reconstruction error term and regularization term. The reconstruction error is the corrected data With input data The regularization term is the KL divergence between the distribution of the encoder output and the standard normal distribution, as shown in formula (8).

[0082] (8)

[0083] in, For input data, is the corrected data; is the weight coefficient corresponding to the reconstruction error term, is the mean square error between the corrected data and the input data; is the weight coefficient corresponding to the regularization term, is the latent variable distribution output by the C-VAE encoder, is the prior distribution and is assumed to be a standard normal distribution, for and The KL divergence of the two distributions indicates the similarity between the latent variable distribution output by the encoder and the prior distribution.

[0084] The process of training C-VAE is to continuously update the gradients and weights in the network to reduce the loss function value, minimize the reconstruction error and maximize the similarity between the encoder output distribution and the standard normal distribution. Compared with the traditional autoencoder network, the variational autoencoder network introduces regularization, which can effectively reduce the overfitting phenomenon caused by the high degree of freedom of the autoencoder network, and constrain the network to extract effective information from the input data instead of blindly copying the input data.

[0085] In summary, the complete loss function expression of the PG-CNRFC network proposed in the present invention is shown in formula (9).

[0086] (9)

[0087] The physical model-driven complex natural resonant frequency correction method proposed in the present invention takes minimizing the above-mentioned loss function value as the training goal of the network. While improving the consistency between the corrected data and the target resonant scattering mechanism, it also takes into account the similarity with the input data and maximizes the similarity between the latent variable distribution generated by the encoder and the standard normal distribution.

[0088] Based on the above description of the C-VAE network structure and the principle of loss function construction, the following is combined with Figure 2 The flowchart of the complex natural resonance frequency correction method driven by the physical model of the present invention is shown to describe the implementation process of the method. Figure 2 As shown, the method comprises the following steps:

[0089] Step 1: Use the resonant scattering mechanism model to calculate the theoretical value of the complex natural resonant frequency corresponding to the target .

[0090] In this step, according to the minor semi-axis of the target a and the major semiaxis b Get the local curvature radius of the creeping wave in the propagation direction and creeping arc length The complex natural resonant frequency is introduced Function , substitute into the target resonance equation and solve the target complex natural resonance frequency theoretical value .

[0091] Step 2: Input the complex natural resonant frequency data Enter the complex domain variational autoencoder network C-VAE, and the C-VAE network outputs the correction value .

[0092] The complex domain variational autoencoder network (C-VAE) consists of two parts: an encoder and a decoder. It is implemented by converting one-dimensional convolution, one-dimensional transposed convolution and linear layers into complex domain one-dimensional convolution, complex domain one-dimensional transposed convolution and complex domain linear layers. It can meet the input of complex features and extract the correlation between the real and imaginary parts of the complex features at each layer.

[0093] Input Data Multiple layers of complex domain one-dimensional convolutional layers and complex domain linear layers are used to encode the probability distribution in the latent space and extract feature information. A sampling point on the probability distribution is further decoded and reconstructed through multiple layers of complex domain one-dimensional transposed convolutional layers and complex domain linear layers.

[0094] Step 3: Substitute the theoretical value obtained in step 1 Substitute the loss function term of the target resonant scattering mechanism into , the input data obtained in step 2 and correction value Substitute the loss function of the complex domain variational autoencoder network The loss function term based on the target resonant scattering mechanism And the loss function of the complex domain variational autoencoder network Composition loss function.

[0095] When training the complex domain variational autoencoder network, the data corrected by C-VAE are calculated in sequence With input data The reconstruction error between the two, the loss function based on the target resonant scattering mechanism, and the similarity between the latent variable distribution generated by the encoder and the standard normal distribution are weighted and summed to obtain the loss function value.

[0096] The goal of network training is to minimize the loss function value, so that the corrected data is more consistent with the target resonant scattering mechanism while taking into account the similarity with the input data and maximizing the similarity between the latent variable distribution generated by the encoder and the standard normal distribution, thereby updating the gradient and weights in the network. When the maximum number of iterations is reached or the loss value of the model on the validation set increases, the training is stopped and the network parameters at this time are saved.

[0097] The trained C-VAE network is used for complex natural resonant frequency correction.

[0098] Step 4: To prevent the input of abnormal data from affecting the network effect, a loss function threshold is set in the PG-CNRFC architecture, and the loss function value of the test data after passing through the PG-CNRFC architecture is calculated. If the loss function value exceeds the set threshold, the test data is judged to be abnormal data, that is, it is not the target data, and a prompt can be given without correcting the data; if the loss function value is less than the set threshold, the corrected result of the test data is output for subsequent radar target identification.

[0099] This is the end of the process.

[0100] The trained C-VAE network is used to correct the input complex natural resonant frequency data. The corrected data can be used to achieve target recognition, classification and other tasks through a classifier. The processing method of the present invention can also be applied to tasks such as correction of radar target resonance area echoes.

[0101] Based on the above method, the present invention also provides a physical model driven complex natural resonance frequency correction device, such as Figure 5 As shown, the device includes a complex natural resonant frequency theoretical value calculation module, a complex domain variational autoencoder network, a loss function determination module, a training module, and a judgment and reasoning module.

[0102] The complex natural resonance frequency theoretical value calculation module is used to calculate the theoretical value of the complex natural resonance frequency corresponding to the target using the resonant scattering mechanism model , sent to the loss function determination module;

[0103] Complex domain variational autoencoder network C-VAE, including encoder and decoder, the encoder consists of multiple layers of complex domain one-dimensional convolution and complex domain linear layer, the decoder consists of multiple layers of complex domain one-dimensional transposed convolution and complex domain linear layer; the input complex natural resonant frequency data Enter the C-VAE network, and the C-VAE network outputs the correction value ; During the training process, the input data and correction value Provided to the loss function determination module;

[0104] The loss function determination module receives the theoretical value , correction value , input data , substitute into the loss function, output the loss function value, the loss function includes: using the theoretical value With correction value The loss function of the target resonant scattering mechanism constructed by the difference , and input data and correction value The loss function of the constructed complex domain variational autoencoder network ;

[0105] The training module, when training the C-VAE network, takes the value of the minimization loss function as the network training goal, so that the correction value of the complex natural resonant frequency The consistency with the target resonant scattering mechanism is improved, and the correction value With input data The error between them is minimal;

[0106] Judgment and reasoning module, used to judge the input data of the loss function determination module output Is the corresponding loss function value less than the set threshold? If so, the current input data is considered normal data and the correction value is output. ; Otherwise, determine the current input data is abnormal data.

[0107] The loss function used by the loss function determination module Loss for:

[0108]

[0109]

[0110]

[0111] in, for The reconstruction error term in for The regularization term in ; is the latent variable distribution of the encoder output in the C-VAE network, is the prior distribution and is assumed to be a standard normal distribution, for and KL divergence of two distributions; is the number of crawling circles, is the local radius of curvature of the surface wave in the propagation direction, is the creeping arc length of the surface wave, is the complex natural resonant frequency The function of a , b are the minor and major semi-axis of the target, respectively; is the angle between the incident direction of the electromagnetic wave and the major axis of the ellipsoid; , , is the weighting coefficient; It means mean square error.

[0112] In a preferred embodiment, the complex domain one-dimensional convolution, the complex domain one-dimensional transposed convolution and the complex domain linear layer in the C-VAE network all adopt a complex domain network structure layer; the input data and weights of the complex domain network structure layer are in complex form.

[0113] like Figure 4 As shown, the complex domain network structure layer includes a real part processing module, an imaginary part processing module, a difference finder, a summer, and a fusion module;

[0114] The real part processing module converts complex input data The real part of A and the imaginary part B With complex weights The real part of C Multiply to get and , sent to the difference finder and the summer;

[0115] The imaginary part processing module converts complex input data The real part of A and the imaginary part B With complex weights The imaginary part D Multiply to get and , sent to the difference finder and the summer;

[0116] The subtractor computes the new real part ;

[0117] The summer calculates the new imaginary part ;

[0118] The fusion module combines the new real part and the new imaginary part into a complex number Output.

[0119] The following experimental results will verify the correction effect of the present invention on the complex natural resonance frequency characteristics. The experiment constructs an ellipsoid model with a major semi-axis of 1m and a minor semi-axis of 0.5m, uses FEKO electromagnetic simulation software to simulate the model scattering data, uses the matrix beam method to extract the complex natural resonance frequency characteristics of the model, divides the training set and the test set into a ratio of 8:2, trains the C-VAE network and saves the weight parameters for subsequent testing. The specific parameters of the experimental setting are shown in Table 1.

[0120] Table 1 Experimental parameters

[0121]

[0122] The experiment used a pre-trained C-VAE network and input the constructed ellipsoid model test data. The average MSE of 89 test data and theoretical values ​​was 30.5736, and the average MSE of the corrected data and theoretical values ​​was 0.5558. The comparison between the test data, corrected data and theoretical values ​​was as follows: Figure 3 The MSE between the corrected data and the theoretical value is reduced by 30.0178, which fully demonstrates that the PG-CNRFC architecture can effectively correct the complex natural resonant frequency characteristics, making the corrected data more consistent with the target resonant scattering mechanism.

[0123] In addition, the PG-CNRFC architecture also sets up the judgment of abnormal data input. When the input is the constructed ellipsoid model data, the average loss function of 89 test data is 5.84, and the loss function value of the model is less than the loss function threshold; when the input is other model data, the average loss function value of 89 test data is 110.08, and the loss function value of the model is greater than the loss function threshold, and the data is judged to be abnormal data. In summary, the PG-CNRFC architecture can not only effectively improve the accuracy of the data and correct the complex natural resonant frequency characteristics, but also effectively distinguish the input of abnormal data to avoid the phenomenon of data miscorrection.

[0124] In order to verify that the present invention can effectively improve the accuracy of radar target recognition based on complex natural resonant frequency characteristics, the experiment was designed to compare the uncorrected data and the data corrected by PG-CNRFC as data sets, and the KNN algorithm with the same parameters was input for classification. The specific experimental results are shown in Table 2. The experimental results show that compared with the uncorrected data as the database, the recognition accuracy of the data corrected by PG-CNRFC as the database is improved by 3.85%. The data corrected by PG-CNRFC is more in line with the resonant scattering mechanism of the target, which increases the data difference of the three types of models, can effectively improve the accuracy of the target complex natural resonant frequency data, and reduce the influence of errors in the extraction process on recognition, thereby improving the target recognition accuracy, which fully reflects the advantages of the present invention in the radar target recognition task based on complex natural resonant frequency characteristics.

[0125] Table 2 Target recognition rate before and after PG-CNRFC correction

[0126]

[0127] The beneficial effects that can be achieved by the present invention are as follows: the accuracy of target complex natural resonance frequency data can be effectively restored, and the correction function of complex natural resonance frequency characteristics can be realized, thereby improving the target recognition accuracy based on the complex natural resonance frequency characteristics; a PG-CNRFC architecture is designed, and a loss function based on the target resonance scattering mechanism is designed, and the network learning is driven by the target resonance scattering mechanism of the target, thereby enhancing the physical interpretability of the algorithm while effectively restoring the accuracy of the data; a complex domain variational autoencoder network C-VAE is designed, and complex domain one-dimensional convolution, complex domain one-dimensional transposed convolution and complex domain linear layer are used instead of real-valued one-dimensional convolution, one-dimensional transposed convolution and linear layer, so as to make full use of the correlation between the real part and the imaginary part of the complex natural resonance frequency characteristics.

[0128] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in the description may be different and are not limited. Therefore, those skilled in the art in the field of the present invention may modify or replace the technical solutions recorded in the above embodiments; and these modifications and replacements do not deviate from the creative purpose and technical solutions of the present invention and should all fall within the protection scope of the present invention.

Claims

1. A physical model driven complex natural resonance frequency correction method, characterized in that: include: Step 1: Use the resonant scattering mechanism model to calculate the theoretical value of the complex natural resonant frequency corresponding to the target ; Step 2: Input the complex natural resonant frequency data Enter the complex domain variational autoencoder network C-VAE, and the C-VAE network outputs the correction value ; The C-VAE network includes an encoder and a decoder. The encoder consists of multiple layers of complex domain one-dimensional convolutional layers and complex domain linear layers, and the decoder consists of multiple layers of complex domain one-dimensional transposed convolutional layers and complex domain linear layers; Step 3: Using the theoretical value of the complex natural resonant frequency With correction value The difference between the two constructs the loss function term of the target resonant scattering mechanism. ; Using the input complex natural resonant frequency data and correction value Constructing the loss function term of the C-VAE network ; Loss function term for target resonant scattering mechanism And the loss function term of the C-VAE network The loss function is composed of the following components: When training the C-VAE network, the network training goal is to minimize the value of the loss function, so that the correction value of the complex natural resonant frequency The consistency with the target resonant scattering mechanism is improved, and the correction value The complex natural resonant frequency data of the input The error between them is the smallest; the trained C-VAE network is used for complex natural resonant frequency correction.

2. The physical model driven complex natural resonance frequency correction method according to claim 1, characterized in that: In step 1, the resonant scattering mechanism model is used to calculate the theoretical value of the complex natural resonant frequency corresponding to the target. for: According to the minor semi-axis of the target a and the major semiaxis b Get the local curvature radius of the surface wave in the propagation direction and creeping arc length The expression of complex natural resonant frequency is introduced Function , substitute into the target resonance equation and solve the theoretical value of the target complex natural resonance frequency : The target resonance equation is: In the formula, is the number of crawling circles, is the complex natural resonant frequency The function of a , b are the minor and major semi-axis of the target, respectively; is the angle between the incident direction of the electromagnetic wave and the major axis of the ellipsoid.

3. The physical model driven complex natural resonance frequency correction method according to claim 2, characterized in that: In step 3, the loss function term of the target resonant scattering mechanism is for: 。 4. The physical model driven complex natural resonance frequency correction method according to claim 1, characterized in that: In step 3, the loss function term of the C-VAE network is Includes the reconstruction error term, which is the correction value of the complex natural resonant frequency The complex natural resonant frequency data of the input The mean square error between .

5. The physical model driven complex natural resonance frequency correction method according to claim 4, characterized in that: In step 3, the loss function term of the C-VAE A regularization term is further included, which is the KL divergence between the distribution of the encoder output in the C-VAE network and the standard normal distribution.

6. The physical model driven complex natural resonance frequency correction method according to claim 1, characterized in that: In step 2, the complex domain one-dimensional convolution layer, the complex domain one-dimensional transposed convolution layer and the complex domain linear layer in the C-VAE network all adopt a complex domain network structure layer; The input data and weights of the complex domain network structure layer are all in complex form; the complex domain network structure layer converts the real part of the complex input data A and the real part of the complex weight C , the imaginary part of the complex input data B and the imaginary part of the complex weight D Multiply them separately and take the difference to get the new real part ; Input the real part of the complex number data A The imaginary part D of the complex weight and the imaginary part of the complex input data B and the real part of the complex weight C Multiply them separately and sum them to get the new imaginary part ; New real and imaginary composite complex numbers Output.

7. The physical model driven complex natural resonance frequency correction method according to claim 1, characterized in that: The method further comprises: Step 4: Determine the input complex natural resonant frequency data Is the corresponding loss function value less than the set threshold? If so, output the correction value ; Otherwise, determine the complex natural resonant frequency data currently input is abnormal data.

8. A physical model driven complex natural resonance frequency correction device, characterized in that: The device includes a complex natural resonance frequency theoretical value calculation module, a C-VAE network, a loss function determination module, a training module and a judgment and reasoning module; The complex natural resonant frequency theoretical value calculation module is used to calculate the theoretical value of the complex natural resonant frequency corresponding to the target using the resonant scattering mechanism model , sent to the loss function determination module; The C-VAE network includes an encoder and a decoder, wherein the encoder is composed of multiple layers of complex domain one-dimensional convolution and complex domain linear layers, and the decoder is composed of multiple layers of complex domain one-dimensional transposed convolution and complex domain linear layers; Input complex natural resonant frequency data Enter the C-VAE network, and the C-VAE network outputs the correction value ; Input complex natural resonant frequency data and correction value Provided to the loss function determination module; The loss function determination module receives the theoretical value , correction value , input complex natural resonant frequency data , substitute into the loss function and output the loss function value; The loss function includes: Using the theoretical value of the complex natural resonant frequency With correction value The loss function of the target resonant scattering mechanism constructed by the difference , and using the complex natural resonant frequency data of the input and correction value The loss function term of the constructed C-VAE network ; The training module, when training the C-VAE network, takes the value of the minimization loss function as the network training target, so that the correction value of the complex natural resonant frequency The consistency with the target resonant scattering mechanism is improved, and the correction value The complex natural resonant frequency data of the input The error between them is minimal; The judgment and reasoning module is used to judge the input complex natural resonance frequency data Is the corresponding loss function value less than the set threshold? If so, it is considered that the current input complex natural resonant frequency data For normal data, output correction value ; Otherwise, determine the complex natural resonant frequency data currently input is abnormal data.

9. The physical model driven complex natural resonance frequency correction device according to claim 8, characterized in that: The loss function used by the loss function determination module Loss for: in, for The reconstruction error term in for The regularization term in ; is the latent variable distribution of the encoder output in the C-VAE network, is the prior distribution and is assumed to be a standard normal distribution, for and KL divergence of two distributions; is the number of crawling circles, is the local radius of curvature of the surface wave in the propagation direction, is the creeping arc length of the surface wave, is the complex natural resonant frequency The function of a , b are the minor and major semi-axis of the target, respectively; is the angle between the incident direction of the electromagnetic wave and the major axis of the ellipsoid; , , is the weighting coefficient; It means mean square error.

10. The physical model driven complex natural resonance frequency correction device according to claim 8, characterized in that: The complex domain one-dimensional convolution layer, the complex domain one-dimensional transposed convolution layer and the complex domain linear layer in the C-VAE network all adopt a complex domain network structure layer; the input data and weights of the complex domain network structure layer are all in complex form; The complex domain network structure layer includes a real part processing module, an imaginary part processing module, a difference finder, a summer, and a fusion module; The real part processing module converts complex input data The real part of A and the imaginary part B With complex weights The real part of C Multiply to get and , sent to the difference finder and the summer; The imaginary part processing module converts complex input data The real part of A and the imaginary part B With complex weights The real part of D Multiply to get and , sent to the difference finder and the summer; The subtractor computes the new real part ; The summer calculates the new imaginary part ; The fusion module combines the new real part and the new imaginary part into a complex number Output.

Images