RCS prediction method based on deep learning, computer equipment and readable storage medium

Abstract

The invention provides an RCS prediction method based on deep learning, computer equipment and a readable storage medium. The method comprises the following steps: firstly, constructing an electromagnetic simulation model of a target, and resolving to obtain noise-free RCS sample data corresponding to different incident wave angles; and then simulation errors are introduced to form noise sample data. Then, constructing a convolutional auto-encoder network as an RCS denoising model, and training the RCS denoising model by using noise-containing and noise-free sample data; and then a deep neural network model is built based on a grey wolf algorithm to serve as an RCS prediction model, the trained denoising model is used for preprocessing the training data set, then the prediction model is trained, and finally prediction of a target RCS value is achieved. According to the method, the GWO-DNN deep learning model is constructed for prediction, so that the dependence on a physical mechanism and the manual parameter adjustment cost are reduced. Meanwhile, noise reduction preprocessing is carried out on the actually measured RCS result through CDAE, the too high influence of measurement errors on the prediction result is overcome, and the RCS prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a deep learning-based RCS prediction method, a computer device, and a readable storage medium. Background Art

[0002] In the field of electronic countermeasures, when actively canceling a stealth target, high-precision prediction of the protected target's Radar Cross Section (RCS) is required. Existing high-precision RCS prediction methods are often based on physical mechanisms and machine learning algorithms, which can achieve high prediction accuracy. However, the RCS distribution reflected by the physical mechanism often differs significantly from the actual measurement results, and the prediction results are significantly affected by measurement errors. Furthermore, the prediction results provided by existing technologies are significantly affected by measurement errors, making it difficult to truly achieve high-precision RCS prediction.

[0003] Chinese patent application number 202310026682.2 discloses a method for predicting the RCS of conductor targets that combines machine learning and physical mechanisms. This method is implemented as follows: First, the single-station RCS calculation formula for conductor targets under the PO mechanism is used to formally approximate the SVR function by analogy, and the most suitable kernel function is found. Second, a training dataset is obtained through sampling and preprocessing of the experimental process. Third, the IPOI□SVR model proposed in the first step is trained using the dataset sampled in the second step to obtain an approximate function for predicting the RCS of conductor targets. Fourth, based on the trained approximate model, the RCS of conductor targets at different azimuth and zenith angles is predicted. This method addresses the low accuracy problem of existing RCS acquisition technologies that rely solely on physical methods or machine learning. However, a drawback of this method is that, because the physical mechanism cannot truly reflect the actual measurement results, excessive reliance on the physical mechanism may lead to distortion in the prediction of actual measurement data.

[0004] Chinese patent application number 202110469304.2 discloses a method for rapid prediction of target electromagnetic scattering characteristics based on deep learning. The implementation scheme of this method is: S1: Establish a target electromagnetic simulation model; S2: Determine the factors that affect the target electromagnetic scattering characteristics; S3: Simulate the target model under the condition that the influencing factors take different values to obtain its far-field RCS, and establish a training set and a test set; S4: Use the BP neural network algorithm to construct a BP neural network model; S5: Use the training set to train the BP neural network model; S6: Use the test set to test the trained BP neural network model. If the test meets the standard, proceed to step S7. If the test does not meet the standard, return to step S4; S7: Use the BP neural network model that meets the test standard to quickly predict the electromagnetic scattering characteristics. This method effectively solves the problems of large computational complexity and low solution efficiency of existing methods, and meets the needs of rapid prediction of electromagnetic scattering characteristics of high-dynamic targets. However, this method still has some shortcomings. The parameters of the BP neural network model need to be adjusted manually, which requires high labor costs, and the prediction accuracy is affected by subjective experience factors. At the same time, this method does not consider the impact of actual measurement errors on the training results. Using RCS data with errors as samples for prediction will also cause the prediction error to increase. Summary of the Invention

[0005] The present invention aims to solve at least one of the above-mentioned technical problems existing in the prior art.

[0006] To this end, the first aspect of the present invention provides an RCS prediction method based on deep learning.

[0007] A second aspect of the present invention provides a computer device.

[0008] A second aspect of the present invention provides a computer-readable storage medium.

[0009] The present invention provides an RCS prediction method based on deep learning, comprising:

[0010] Construct an electromagnetic simulation model of the target;

[0011] The RCS parameters of the electromagnetic simulation model of the target are solved to obtain the RCS values corresponding to the electromagnetic simulation model at different incident wave azimuth angles and incident wave pitch angles to form noise-free sample data;

[0012] A simulation error is introduced into the calculated RCS value data to form noisy sample data;

[0013] Constructing a convolutional autoencoder network as an RCS denoising model, wherein the input data of the convolutional autoencoder network is noisy sample data, and the output data of the convolutional autoencoder network is the denoised sample data; training the convolutional autoencoder network based on the noisy sample data and the noise-free sample data;

[0014] Generate a training data set and a test data set for RCS prediction, wherein the training data set includes noisy sample data and the test data set includes noise-free sample data;

[0015] A deep neural network model is built based on the Gray Wolf Algorithm as an RCS prediction model. The input data of the RCS prediction model are the azimuth and pitch angle of the incident wave, and the output data of the RCS prediction model are the RCS values corresponding to the azimuth and pitch angle of the incident wave. The Gray Wolf Algorithm is used to optimize the hyperparameters of the deep neural network model.

[0016] Use the trained RCS denoising model to perform denoising preprocessing on the training data set used for RCS prediction, and train the RCS prediction model based on the training data set that has completed denoising preprocessing;

[0017] The trained RCS denoising model and RCS prediction model are used to predict the RCS value of the target.

[0018] The RCS prediction method based on deep learning according to the above technical solution of the present invention may also have the following additional technical features:

[0019] In the above technical solution, in the process of constructing the electromagnetic simulation model of the target, several electromagnetic simulation models with different accuracies are established for subsequent parameter calculation.

[0020] In the above technical solution, the simulation error includes the random error caused by the inherent noise in the measurement system and the environment and the random error generated in the manual measurement process at different directions.

[0021] In the above technical solution, the method of introducing simulation error into the calculated RCS value data includes:

[0022]

[0023] Among them, σ i,j It represents the RCS value calculation result corresponding to the incident wave azimuth angle i and the incident wave pitch angle j; Indicates the RCS value data with simulation error introduced; represents the mean of all RCS data in a single electromagnetic simulation model; randn(i,j) represents the normal distribution function; error represents the introduced error power;

[0024] in, Used to simulate the random error caused by the inherent noise in the measurement system and environment, σ i,j randn(i,j) / 10 error / 10 Used to simulate the random errors generated during manual measurement in different directions.

[0025] In the above technical solution, the constructed convolutional autoencoder network includes an encoder and a decoder;

[0026] The encoder of the convolutional autoencoder network includes a first convolutional layer, a second convolutional layer, a first pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, and a second pooling layer connected in sequence;

[0027] The decoder of the convolutional autoencoder network includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, a fourth deconvolution layer, a fifth deconvolution layer, a sixth deconvolution layer and a fully connected layer connected in sequence;

[0028] The encoder of the convolutional autoencoder network maps the input noisy sample data to a low-dimensional potential representation; the decoder of the convolutional autoencoder network restores the low-dimensional potential representation obtained by the encoder to an output as close as possible to the noise-free sample data, that is, the denoised sample data;

[0029] The mean square error between the denoised sample data and the noise-free sample data is used as the loss function of the convolutional autoencoder network.

[0030] In the above technical solution, CST software is used to solve the RCS parameters of the electromagnetic simulation model of the target.

[0031] In the above technical solution, the deep neural network includes an input layer, an output layer and a hidden layer;

[0032] In the process of optimizing the hyperparameters of the deep neural network model using the gray wolf algorithm, the optimization parameters include the number of neurons in the hidden layer, the learning rate, and the training batch size.

[0033] In the above technical solution, the training process of the RCS prediction model includes:

[0034] Initialize model parameters, randomly generate a population of gray wolves that can identify hyperparameters of the deep neural network model, and set the initial state of the optimization parameters;

[0035] Use the initialized parameters to train the RCS prediction model based on the training dataset that has completed denoising preprocessing;

[0036] According to the wild wolf algorithm, the fitness of each individual in the population is calculated to evaluate the performance of the model when it is used as a hyperparameter;

[0037] Determine whether the maximum number of iterations has been reached. If not, update the individual position according to the fitness, that is, adjust the hyperparameter value, and then return to continue model training. If the maximum number of iterations has been reached, output the hyperparameter combination with the optimal fitness.

[0038] Build a DNN model based on the optimal hyperparameters, make predictions on the test dataset, and evaluate the accuracy and reliability of the prediction results.

[0039] The present invention also provides a computer device comprising a processor and a memory, wherein a computer program is stored in the memory. When the computer program is loaded and executed by the processor, the deep learning-based RCS prediction method as described in any one of the above technical solutions is implemented.

[0040] The present invention further provides a computer-readable storage medium storing a program, which, when loaded by a processor, implements the deep learning-based RCS prediction method as described in any one of the above technical solutions.

[0041] In summary, due to the adoption of the above technical features, the beneficial effects of the present invention are:

[0042] To address the high reliance of existing methods on physical mechanisms, this paper develops a GWO-DNN deep learning model for prediction, eliminating this reliance on physical mechanisms while reducing the cost of manual parameter adjustment. Furthermore, to address the significant impact of measurement errors on prediction results in existing methods, this paper develops a CDAE convolutional denoising autoencoder for noise reduction preprocessing. This allows for high-precision RCS predictions even when test results contain certain errors.

[0043] Specifically, the CDAE convolutional denoising autoencoder constructed by the present invention can perform noise reduction preprocessing on the actual measured RCS results, overcoming the problem that the actual measurement error has an excessive impact on the RCS prediction results. Compared with the existing technology, it has higher versatility and robustness.

[0044] The GWO-DNN model constructed in the present invention can achieve high-precision prediction of RCS without the need for physical prior knowledge, overcoming the problems of existing technologies that rely heavily on physical prior knowledge and require manual model tuning. Compared with existing technologies, it has the advantages of low labor costs, high flexibility, and high versatility.

[0045] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0047] Figure 1 is a flowchart of an RCS prediction method based on deep learning according to an embodiment of the present invention;

[0048] Figure 2 Schematic diagram of the visualization structure of the convolutional autoencoder network in the RCS prediction method based on deep learning according to an embodiment of the present invention;

[0049] Figure 3 This is a flowchart of optimizing deep neural network hyperparameters using the wolf algorithm in one embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0051] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0052] Refer to the following Figures 1 to 3 To describe the deep learning-based RCS prediction method, computer device, and readable storage medium provided according to some embodiments of the present invention.

[0053] Some embodiments of the present application provide an RCS prediction method based on deep learning.

[0054] like Figure 1 As shown, the first embodiment of the present invention proposes a deep learning-based RCS prediction method, including the following steps S1-S8. It should be noted that the order of steps S1-S8 is only a schematic representation of the present disclosure. Those skilled in the art can adjust the order of steps according to actual needs, and steps of different orders can also be performed simultaneously.

[0055] S1. Build an electromagnetic simulation model of the target.

[0056] Specifically, a target electromagnetic simulation model refers to a model established for the predicted target or a device of the same type as the predicted target, capable of performing electromagnetic simulation calculations. This model includes the target geometry, structural materials, simulation frequency, and boundary conditions configured in the simulation software. The methods for constructing electromagnetic simulation models are well known to those skilled in the art and will not be elaborated upon here. In this disclosure, an aircraft is used as the target for electromagnetic simulation modeling.

[0057] In some embodiments, during the construction of an electromagnetic simulation model of a target, several electromagnetic simulation models of varying precision are established for subsequent parameter calculation. Subsequent model training based on the relevant data generated by these models of the same target with varying precision provides a richer data sample and effectively improves model training accuracy. In a specific embodiment of the present disclosure, subsequent data processing is performed using four electromagnetic simulation models of the same target with varying precision.

[0058] S2. Calculate the RCS parameters of the electromagnetic simulation model of the target to obtain the RCS values of the electromagnetic simulation model corresponding to different incident wave azimuth angles and incident wave pitch angles to form noise-free sample data.

[0059] In a specific embodiment, the detailed execution process of step S2 is:

[0060] CST software was used to calculate the RCS parameters of four electromagnetic simulation models of the target with different accuracies, and the RCS values of the electromagnetic simulation models at different incident wave azimuths and incident wave pitch angles were obtained. The simulation intervals of the incident wave azimuth and incident wave pitch angles were both 0.5°.

[0061] S3. Introducing simulation errors into the calculated RCS value data to form noisy sample data; in some embodiments, the simulation errors include random errors caused by inherent noise in the measurement system and environment and random errors generated during manual measurement at different directions.

[0062] In some embodiments, the method of introducing simulation errors into the calculated RCS value data includes:

[0063]

[0064] Among them, σ i,j It represents the RCS value calculation result corresponding to the incident wave azimuth angle i and the incident wave pitch angle j; Indicates the RCS value data with simulation error introduced; represents the mean of all RCS data in a single electromagnetic simulation model; randn(i,j) represents the normal distribution function; error represents the introduced error power, error∈[10,15]dB;

[0065] in, Used to simulate the random error caused by the inherent noise in the measurement system and environment, σ i,j randn(i,j) / 10 error / 10 Used to simulate the random errors generated during manual measurement in different directions.

[0066] In a specific embodiment, based on the above method of introducing simulation errors, 500 noisy sample data are randomly generated for each electromagnetic simulation model of each precision.

[0067] S4. Construct a convolutional autoencoder network as an RCS denoising model, where the input data of the convolutional autoencoder network is noisy sample data, and the output data of the convolutional autoencoder network is the denoised sample data; and train the convolutional autoencoder network based on the noisy sample data and the noise-free sample data.

[0068] Specifically, the training and test datasets required for the convolutional autoencoder (CDAE) network can be divided based on the noisy and noise-free sample data of the four electromagnetic simulation models with different accuracies. For example, the noisy sample data of three of the electromagnetic simulation models is used as the training set for model training, and the noisy sample data of the fourth electromagnetic simulation model is used as the test set to verify the model's performance.

[0069] In some embodiments, as Figure 2 As shown, the constructed convolutional autoencoder network includes an encoder and a decoder;

[0070] The encoder of the convolutional autoencoder network includes a first convolutional layer, a second convolutional layer, a first pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, and a second pooling layer connected in sequence;

[0071] The decoder of the convolutional autoencoder network includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, a fourth deconvolution layer, a fifth deconvolution layer, a sixth deconvolution layer and a fully connected layer connected in sequence;

[0072] The encoder of the convolutional autoencoder network maps the input noisy sample data to a low-dimensional potential representation; the decoder of the convolutional autoencoder network restores the low-dimensional potential representation obtained by the encoder to an output as close as possible to the noise-free sample data, that is, the denoised sample data;

[0073] The mean square error (MSE) between the denoised sample data and the noise-free sample data is used as the loss function of the convolutional autoencoder network.

[0074] In a specific embodiment, the numbers of convolution kernels in the first convolution layer, the second convolution layer, the third convolution layer, the fourth convolution layer, the fifth convolution layer, the first deconvolution layer, the second deconvolution layer, the third deconvolution layer, the fourth deconvolution layer, the fifth deconvolution layer, and the sixth deconvolution layer are 64, 64, 128, 128, 256, 128, 128, 64, 32, 32, and 16 respectively; the size of the convolution kernel is set to 3×3; the pooling layers adopt the maximum pooling method, the size of the pooling kernel is set to 2×2, and the step size is 2.

[0075] S5. Generate a training data set and a test data set for RCS prediction, wherein the training data set includes noisy sample data and the test data set includes noise-free sample data.

[0076] In one specific embodiment, the noisy sample data of a certain electromagnetic simulation model in step S3 can be directly used as a training data set for RCS prediction. The parameters of the electromagnetic simulation model established in step S1 are recalculated according to step S2. When recalculating the parameters, the simulation intervals of the incident wave azimuth angle and the incident wave pitch angle should be set to a smaller value, for example, 0.1° for both. The resulting noise-free sample data is used as a test data set.

[0077] S6. Based on the Grey Wolf Algorithm (GWO), a deep neural network model DNN is constructed as an RCS prediction model (named GWO-DNN model in this disclosure). The input data of the RCS prediction model are the azimuth and pitch angle of the incident wave, and the output data of the RCS prediction model are the RCS values corresponding to the azimuth and pitch angle of the incident wave. The Grey Wolf Algorithm is used to perform hyperparameter optimization on the deep neural network model.

[0078] In some embodiments, the deep neural network includes an input layer, an output layer, and a hidden layer; in a specific embodiment, the number of hidden layers is 6.

[0079] In the process of optimizing the hyperparameters of the deep neural network model using the gray wolf algorithm, the optimization parameters include the number of neurons in the hidden layer, the learning rate, and the training batch size.

[0080] In one embodiment, the optimization parameters include the number of neurons in the four hidden layers, the training batch size, and the learning rate. The population size of the Gray Wolf Algorithm is set to 20, the number of iterations is set to 20, and the upper bounds of the optimization of each hyperparameter are set to [128, 256, 256, 128, 100, 0.02] in the order of the number of neurons in the four hidden layers, the training batch size, and the learning rate, respectively. The lower bounds are set to [16, 32, 32, 16, 20, 0.0005] in the order of the number of neurons in the four hidden layers, the training batch size, and the learning rate.

[0081] S7. Use the trained RCS denoising model to perform denoising preprocessing on the training data set for RCS prediction, and train the RCS prediction model based on the training data set that has completed the denoising preprocessing.

[0082] Figure 3 The training process of the RCS prediction model is shown, including:

[0083] Initialize the model parameters, randomly generate the gray wolf population of the deep neural network model hyperparameters, and set the initial state of the optimization parameters; the number of neurons in each hidden layer is Figure 3 The number of hidden nodes is simply shown in the figure, and the training batch size is Figure 3 The batch size is simply indicated in the figure;

[0084] Use the initialized parameters to train the RCS prediction model based on the training dataset that has completed denoising preprocessing;

[0085] According to the wild wolf algorithm, the fitness of each individual in the population is calculated to evaluate the performance of the model when it is used as a hyperparameter;

[0086] Determine whether the maximum number of iterations has been reached. If not, update the individual position according to the fitness, that is, adjust the hyperparameter value, and then return to continue model training. If the maximum number of iterations has been reached, output the hyperparameter combination with the optimal fitness.

[0087] Build a DNN model based on the optimal hyperparameters, make predictions on the test dataset, and evaluate the accuracy and reliability of the prediction results.

[0088] It should be noted that the specific content of the wild wolf algorithm and DNN model training is well known to those skilled in the art and will not be repeated here.

[0089] S8. Use the trained RCS denoising model and RCS prediction model to predict the RCS value of the target.

[0090] In a specific embodiment, the impact of the various improvements disclosed in this disclosure on the prediction effect was simulated and verified. The results are as follows:

[0091] Different prediction methods are used to conduct simulation experiments on the data set after noise reduction in step S7. The simulation results are shown in Table 1.

[0092] Table 1 Simulation results of the impact of different predictions on prediction accuracy

[0093]

[0094] GWO-GPR refers to a hybrid model that combines the Grey Wolf Optimizer (GWO) and Gaussian Process Regression (GPR). DNN refers to a single deep neural network model. GWO-DNN is the RCS prediction model proposed in this disclosure.

[0095] GWO-DNN simulation experiments were conducted on the datasets before and after denoising, that is, the training datasets before and after CDAE preprocessing. The simulation results are shown in Table 2.

[0096] Table 2 Simulation results of the effect of CDAE preprocessing on prediction accuracy

[0097]

[0098] In Table 2, GWO-DNN indicates that the noisy sample data that has not been processed by the RCS denoising model (CDAE) is directly predicted using the GWO-DNN model. CDAE-GWO-DNN, on the other hand, is processed by the RCS denoising model (CDAE) before using the GWO-DNN model for prediction.

[0099] The square correlation coefficient (r 2 ) and mean absolute error (MAE), whose expressions are

[0100]

[0101] Among them, the mean absolute error (MAE) can accurately reflect the size of the actual prediction error. MAE is used to evaluate the degree of deviation between the true value and the fitted value. The closer the MAE value is to 0, the better the model fit is and the higher the model prediction accuracy is. 2 It represents the correlation between the model prediction value and the actual value, r 2 The closer it is to 1, the higher the prediction accuracy of the model. From Table 1 and Table 2, it can be seen that the changes in each part of the present invention can improve the prediction accuracy to a certain extent, and the present invention can still have a high prediction accuracy in the presence of measurement errors.

[0102] Other embodiments of the present invention provide a computer device comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is loaded and executed by the processor, the deep learning-based RCS prediction method as described in any of the above embodiments is implemented.

[0103] Still other embodiments of the present invention provide a computer-readable storage medium storing a program, which, when loaded by a processor, implements the deep learning-based RCS prediction method as described in any of the above embodiments.

[0104] In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any suitable manner in any one or more embodiments or examples.

[0105] Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A RCS prediction method based on deep learning, characterized in that: include: Construct an electromagnetic simulation model of the target; The RCS parameters of the electromagnetic simulation model of the target are solved to obtain the RCS values corresponding to the electromagnetic simulation model at different incident wave azimuth angles and incident wave pitch angles to form noise-free sample data; A simulation error is introduced into the calculated RCS value data to form noisy sample data; Constructing a convolutional autoencoder network as an RCS denoising model, wherein the input data of the convolutional autoencoder network is noisy sample data, and the output data of the convolutional autoencoder network is the denoised sample data; training the convolutional autoencoder network based on the noisy sample data and the noise-free sample data; Generate a training data set and a test data set for RCS prediction, wherein the training data set includes noisy sample data and the test data set includes noise-free sample data; A deep neural network model is built based on the Gray Wolf Algorithm as an RCS prediction model. The input data of the RCS prediction model are the azimuth and pitch angle of the incident wave, and the output data of the RCS prediction model are the RCS values corresponding to the azimuth and pitch angle of the incident wave. The Gray Wolf Algorithm is used to optimize the hyperparameters of the deep neural network model. Use the trained RCS denoising model to perform denoising preprocessing on the training data set used for RCS prediction, and train the RCS prediction model based on the training data set that has completed denoising preprocessing; The trained RCS denoising model and RCS prediction model are used to predict the RCS value of the target.

2. The RCS prediction method based on deep learning according to claim 1, characterized in that In the process of constructing the target electromagnetic simulation model, several electromagnetic simulation models with different accuracies are established for subsequent parameter calculation.

3. The RCS prediction method based on deep learning according to claim 1, characterized in that The simulation error includes random errors caused by inherent noise in the measurement system and environment and random errors generated during manual measurement at different directions.

4. The RCS prediction method based on deep learning according to claim 3, characterized in that Methods for introducing simulation errors into the calculated RCS value data include: Among them, σ i,j It represents the RCS value calculation result corresponding to the incident wave azimuth angle i and the incident wave pitch angle j; Indicates the RCS value data with simulation error introduced; represents the mean of all RCS data in a single electromagnetic simulation model; randn(i,j) represents the normal distribution function; error represents the introduced error power; in, Used to simulate the random error caused by the inherent noise in the measurement system and environment, σ i,j randn(i,j) / 10 error / 10 Used to simulate the random errors generated during manual measurement in different directions.

5. The RCS prediction method based on deep learning according to claim 1, characterized in that: The constructed convolutional autoencoder network includes an encoder and a decoder; The encoder of the convolutional autoencoder network includes a first convolutional layer, a second convolutional layer, a first pooling layer, a third convolutional layer, a fourth convolutional layer, a fifth convolutional layer, and a second pooling layer connected in sequence; The decoder of the convolutional autoencoder network includes a first deconvolution layer, a second deconvolution layer, a third deconvolution layer, a fourth deconvolution layer, a fifth deconvolution layer, a sixth deconvolution layer and a fully connected layer connected in sequence; The encoder of the convolutional autoencoder network maps the input noisy sample data to a low-dimensional potential representation; the decoder of the convolutional autoencoder network restores the low-dimensional potential representation obtained by the encoder to an output as close as possible to the noise-free sample data, that is, the denoised sample data; The mean square error between the denoised sample data and the noise-free sample data is used as the loss function of the convolutional autoencoder network.

6. The RCS prediction method based on deep learning according to claim 1, characterized in that: CST software is used to solve the RCS parameters of the target's electromagnetic simulation model.

7. The RCS prediction method based on deep learning according to claim 1, characterized in that: The deep neural network includes an input layer, an output layer and a hidden layer; In the process of optimizing the hyperparameters of the deep neural network model using the gray wolf algorithm, the optimization parameters include the number of neurons in the hidden layer, the learning rate, and the training batch size.

8. The RCS prediction method based on deep learning according to claim 7, characterized in that: The training process of the RCS prediction model includes: Initialize model parameters, randomly generate a population of gray wolves that can identify hyperparameters of the deep neural network model, and set the initial state of the optimization parameters; Use the initialized parameters to train the RCS prediction model based on the training dataset that has completed denoising preprocessing; According to the wild wolf algorithm, the fitness of each individual in the population is calculated to evaluate the performance of the model when it is used as a hyperparameter; Determine whether the maximum number of iterations has been reached. If not, update the individual position according to the fitness, that is, adjust the hyperparameter value, and then return to continue model training. If the maximum number of iterations has been reached, output the hyperparameter combination with the optimal fitness. Build a DNN model based on the optimal hyperparameters, make predictions on the test dataset, and evaluate the accuracy and reliability of the prediction results.

9. A computer device, characterized in that: The invention comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded and executed by the processor, the RCS prediction method based on deep learning as described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium, characterized in that A program is stored, and when the program is loaded by a processor, the RCS prediction method based on deep learning as described in any one of claims 1 to 8 is implemented.

Images