Target RCS prediction method and device based on stacked integrated network

Abstract

The invention belongs to the technical field of artificial intelligence and electromagnetic scattering application, and particularly relates to a target RCS prediction method and device based on a stacked integrated network, and the method comprises the specific steps: network construction: constructing the stacked integrated network comprising a base model and a meta model, and taking a predicted value of the base model as an input feature sequence of the meta model; data preprocessing: collecting RCS data of different targets as data sets, performing standardization processing on the data sets of the different targets, and constructing a training data set and a verification data set; model training: training each base model by using the preprocessed training data set, predicting the verification data set by using the trained base model, and outputting and combining the prediction results of the plurality of base models into a new feature as the input of the meta-model to train the meta-model; and performing RCS prediction: performing target RCS prediction by using the trained stacked integrated network.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and electromagnetic scattering applications, and in particular relates to a target RCS prediction method and device based on a stacked integrated network. Background Art

[0002] The study of radar cross section (RCS) has extremely important application value for radar target classification, detection, tracking and scattering imaging. At present, the RCS characteristics of targets are mainly obtained through theoretical simulation and actual measurement. Experimental measurement is complicated and expensive, and theoretical simulation is generally considered to be a feasible alternative. Commonly used RCS simulation methods include high-frequency approximation methods (physical optics method (PO), geometric optics method (GO), and bouncing ray method, etc.) and numerical methods (moment method (MOM), finite-difference time-domain method, parallel multilayer fast multipole method (PMLFMA), etc.).

[0003] In recent years, machine learning and deep learning have been applied to solve complex nonlinear computational electromagnetic problems. Gaussian process regression (GPR) is applied to analyze the relationship between radar frequency and target RCS amplitude. Sub-surrogate models are constructed using GPR, and Bayesian committee machine (BCM) fuses all predictions to analyze electromagnetic scattering. Deep neural network (DNN) is applied to bistatic scattering interpolation of finite dielectric cylinders. Support vector regression (SVR) and GPR are used to predict the RCS of single-station targets. Convolutional neural network (CNN) is combined to analyze background scattering and target scattering.

[0004] The numerical method has high calculation accuracy but limited calculation efficiency. The high-frequency method has fast calculation speed but insufficient accuracy. The above-mentioned machine learning and deep learning studies all show that the corresponding target RCS curve does not have fast-changing oscillations, and some do not involve complex models. In addition, the training time is long, and most of them cannot make real-time predictions. Summary of the invention

[0005] In view of this, the present invention provides a target RCS prediction method and device based on a stacked integrated network, which can ensure accuracy while taking into account computational efficiency.

[0006] The technical solution for implementing the present invention is as follows:

[0007] In the first aspect, a target RCS prediction method based on a stacked integrated network is provided, wherein the specific process is as follows:

[0008] Network construction: construct a stacked integrated network including a base model and a meta-model, wherein the prediction value of the base model is used as an input feature sequence of the meta-model;

[0009] Data preprocessing: Collect RCS data of different targets as data sets, standardize the data sets of different targets, and construct training data sets and verification data sets;

[0010] Model training: Use the preprocessed training data set to train each base model, use the trained base model to predict the validation data set, and merge the prediction results of multiple base models into new features as the input of the meta-model to train the meta-model;

[0011] RCS prediction: Use the trained stacked ensemble network to predict the target RCS.

[0012] Optionally, the base model of the present invention is: XGB Regression, Random Forest, Decision Tree, Polynomial Regression or Gradient Boosting Regression, and the base model uses Grid Search to optimize training hyperparameters.

[0013] Optionally, the meta-model described in the present invention is a bidirectional long short-term memory network (Bi-directional LongShort-Term Memory, Bi-LSTM).

[0014] Optionally, the Bi-LSTM model described in the present invention adopts extrapolation technology to predict the RCS sequence, and sets the window size to 10, that is, in a set of azimuths containing 361 RCS data, the first 10 data points are used to predict the subsequent sequence.

[0015] Optionally, the present invention guides the training of the meta-model by introducing physical equations into the loss function, specifically: the loss function includes a data-driven loss function and a physics-driven loss function.

[0016] Optionally, the loss function of the present invention is:

[0017]

[0018] in, represents the physics-based drive loss function, represents the data-driven loss function, is the meta-model prediction output, y i is the true label value, λ is the regularization parameter, W (l) and denote the parameters of the neural network and physical information mechanism, respectively;

[0019]

[0020] in, represents the L2 norm, N R is the total number of residual terms R;

[0021]

[0022]

[0023] in, Represents the solution RCS of the neural network during the iteration process, that is, the prediction result output by the meta-neural network σ(θ,φ,f) represents the radar cross section calculated based on the physical mechanism equation; E s is the scattered field, E i is the incident field, R is the distance between the radar and the target under far-field conditions,

[0024] Optionally, the targets selected in the present invention include reference body PEC targets, complex PEC targets and complex medium targets.

[0025] Optionally, the standardization process described in the present invention is: standardizing the data using a zero mean and unit variance method.

[0026] Optionally, the training process of the present invention is:

[0027] First, for each base model m j (j=1,2,...,M), set the data set D={D 1 ,D 2 ,…,D 5}50% off;

[0028] Second, for each fold k, use all data except the kth fold as the training set The k-th fold data As a validation set;

[0029] Then, using the training set Training base model m j ,get

[0030] Finally, use the trained model For the k-th fold data Make a prediction and get the predicted value The results predicted by multiple base models The output is merged into a new feature as the input of the meta-model for meta-model training. The meta-model outputs the final prediction result by learning the mapping relationship between the base model output and the true target value.

[0031] In a second aspect, the present invention provides a target RCS prediction device based on a stacked integrated network, comprising a data preprocessing module and a stacked integrated network; wherein:

[0032] Data preprocessing: Collect RCS data of different targets as data sets, and standardize the data sets of different targets;

[0033] Stacked ensemble network: Use the stacked ensemble network trained by the above method to predict the target RCS of the standardized data set.

[0034] Beneficial effects:

[0035] First, this paper proposes a scattering center driven stacking ensemble learning (SCD-stacking) algorithm, which combines the advantages of machine, deep and ensemble learning, and more importantly, introduces a physical information mechanism to accurately and near real-time predict the RCS of perfect electric conductor (PEC) targets.

[0036] Second, the present invention transfers the best performing stacking model to complex targets and complex medium targets via transfer learning, which not only verifies the robustness of the method but also provides a potential solution when data generation is computationally expensive.

[0037] Third, compared with existing machine learning and deep learning, the stacked ensemble learning network proposed in the present invention uses a complex model, the RCS has rapidly changing oscillations, the training time is relatively short, and it reaches a level of near real-time prediction on the trained model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 Integrate network structure for stacking;

[0040] Figure 2 For the slice target I model and scattering center;

[0041] Figure 3 It is the complex PEC target II and its scattering center type distribution;

[0042] Figure 4 It is the complex medium target III model and its scattering center type distribution;

[0043] Figure 5RCS comparison of dataset I, (a) before improvement, (b) after improvement;

[0044] Figure 6 RCS comparison of dataset II, (a) before improvement, (b) after improvement;

[0045] Figure 7 RCS comparison of dataset III, (a) before improvement, (b) after improvement. DETAILED DESCRIPTION

[0046] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] It should be noted that the following embodiments and features in the embodiments may be combined with each other in the absence of conflict; and, based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making any creative work are within the scope of protection of the present disclosure.

[0048] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0049] The design ideas of this application are:

[0050] First, machine learning training is faster, and the model has better stability and efficiency. Deep learning usually performs higher accuracy in automatic feature learning in complex tasks. This application combines the advantages of both and proposes a stacked integration network that uses machine learning and deep learning as the base model and meta-model combination training respectively. A major difference from other types of integration methods such as bagging (such as random forests (RF)) and boosting (such as gradient boosting (GB)) is that stacking can use bagging and boosting techniques at the same time, combined with simpler algorithms, and stacked in different layers. It uses a meta-learner to aggregate the predictions of the last layer and obtain the best performance, which does not exist in other integration methods.

[0051] Secondly, scattering centers are introduced as physical data drivers in the stacking model to guide network training. Scattering centers have inherent physical characteristics of the target, including geometric information, materials, structures, etc. The backscattered echo of the target can be modeled as the sum of the responses of individual attribute scattering centers (ASCs). Many studies have shown the effectiveness of this model in RCS prediction, and these methods that combine scattering features have good interpretability in practical applications.

[0052] The present application embodiment provides a target RCS prediction method based on a stacked integrated network, and the specific process is as follows:

[0053] Network construction: Construct a stacked integrated network containing a base model and a meta-model, where the predicted value of the base model is used as the input feature sequence of the meta-model, such as Figure 1 As shown;

[0054] In the specific implementation of this step, the following representative models were selected based on experience and pre-experiments, such as XGB Regression, Random Forest, Decision Tree, Polynomial Regression, and Gradient Boosting Regression as base models. In this process, grid search was used to optimize their hyperparameters.

[0055] In the specific implementation of this step, according to the characteristics of the input feature data sequence, a bidirectional long short-term memory network (Bi-directional Long Short-Term Memory, Bi-LSTM) is used as the meta-model. The detailed process of the meta-model learner is in Figure 1 .

[0056]

[0057]

[0058] in, represents the loss function, ω * is the optimal parameter of the meta-model Bi-LSTM obtained after training, y represents the target value, ω represents the parameter value of the meta-model, Indicates the final prediction result output by the meta-model.

[0059] Data preprocessing: Collect RCS data of different targets as data sets, standardize the data sets of different targets, and construct training data sets and verification data sets.

[0060] In the specific implementation of this step, the selected targets include the reference PEC target, the complex PEC target and the complex medium target. The standardized processing of this step is:

[0061] The dataset is defined as Where N is the number of samples, x i represents the feature vector, y i The original data were preprocessed and standardized using zero mean and unit variance methods.

[0062]

[0063]

[0064] in, is the i-th sample of data X, μ is the mean value, is zero mean, δ 2 is the variance of the data X, is a data sample normalized by variance, and the preprocessed data is

[0065] Model training: Use the preprocessed training data set to train each base model, use the trained base model to predict the validation data set, and merge the prediction results of multiple base models into new features as the input of the meta-model for meta-model training. The specific process of this step is:

[0066] First, for each base model m j (j=1,2,...,M), set the data set D={D 1 ,D 2 ,…,D 5}50% off;

[0067] Second, for each fold k, use all data except the kth fold as the training set The k-th fold data As a validation set;

[0068] Then, using the training set Training base model m j ,get

[0069] Finally, use the trained model For the k-th fold data Make a prediction and get the predicted value

[0070] In this step, each trained base model m is used j(j=1,2,...,M), and generate the features of the meta-model through 5-fold cross-validation:

[0071]

[0072] Reuse the results predicted by the base model As the input of the meta-model, the meta-model outputs the final prediction result by learning the mapping relationship between the base model output and the true target value.

[0073] RCS prediction: Use the trained stacked ensemble network to predict the target RCS.

[0074] This embodiment first adopts the Bi-LSTM model when designing the stacked integrated network. Unlike the traditional unidirectional LSTM, the Bi-LSTM learns in two directions of the time series, namely forward and backward. The output at the current moment is not only related to the state of the previous time series, but may also be related to the state of the future time series. For each moment, the input is simultaneously provided to two LSTMs in opposite directions, and the output is jointly determined by these two unidirectional LSTMs. This structure can better capture contextual information. Then, during model training, the original data set is preprocessed and divided into several sub-data sets. Each basic model predicts the output by using a 5-fold cross-validation method to reduce the risk of overfitting; the output of each base model is merged (Σ) into a new feature as the input of the meta-model to train the meta-learner Bi-LSTM.

[0075] Furthermore, in another embodiment of the present application, the present application adopts an extrapolation technique in the Bi-LSTM model to predict the RCS sequence, and sets the window size to 10, that is, in a set of azimuths containing 361 RCS data, the first 10 data points are used to predict the subsequent sequence.

[0076] Furthermore, in another embodiment of the present application, physical equations are introduced into the loss function to guide network training, specifically:

[0077] The loss functions of the meta-neural network include: data-driven loss functions and physics-driven loss functions, such as Figure 1 Specifically:

[0078] In traditional neural network training, the loss function usually measures the error between the predicted value and the true value. In order to strengthen physics, physical principles can be embedded into the network architecture as prior information, using physics as a soft penalty to constrain. Therefore, this application designs a loss function that not only includes data-driven loss, but also includes a physics-driven loss function, which guides the network training process by embedding physical data.

[0079] Specifically, in the meta-model, the physical drive loss is used to measure whether the output predicted by the network satisfies the given physical equations, and they are introduced into the loss function, so that the loss function is:

[0080]

[0081] in, represents the physics-based drive loss function, represents the data-driven loss function, is the meta-model prediction output, y i is the true label value, λ is the regularization parameter, W (l) and denote the parameters of the neural network and the physical information mechanism, respectively.

[0082] The radar cross-section (RCS) calculated based on the physical mechanism equation is:

[0083]

[0084] Among them, E s is the scattered field, E i When the incident field is configured as a single station, it is usually normalized to |E i ||=1, R is the distance between the radar and the target under far-field conditions, θ, φ, f are the elevation angle, azimuth angle and working frequency of the radar observation direction respectively. The solution of the neural network in the iterative process (i.e. the prediction result output by the meta-neural network ) into the equation. If the predicted solution does not satisfy the equation, the physical law is not followed and the residual is

[0085]

[0086] Therefore, the physics-driven loss function is defined as the sum of squares of the residuals of the physics equations:

[0087]

[0088] in, represents the L2 norm, N Ris the total number of residual terms R. This term ensures that the predicted solution RCS of the neural network not only conforms to the data, but also satisfies the physical equation. Then the above physical formula evolves into the process of solving the scattered field.

[0089] Considering that the change of target RCS with angle often shows different fluctuation patterns in azimuth and elevation, inspired by the scattering field formula of the ASC method, the change of target RCS with angle can be regarded as the superposition of sinc function and exponential decay function, so it shows obvious fluctuation pattern. Therefore, the physical characteristics of the ASC function are used to solve the scattering field E s (θ, φ, f) introduces the network design.

[0090] As an extension of physical optics, the ASC model is based on the geometric theory of diffraction (GTD) and is currently the most widely used scattering center model. It is used to explain the physical and geometric properties of the target's scattering center. It uses parameters such as position, amplitude, frequency, orientation, and polarization to describe the characteristics of the scattering center. According to the position distribution characteristics of the scattering center, the attribute scattering center of the target can be divided into the localized scattering center the localized SC (LSC), the distributed scattering center the distributed SC (DSC) and the sliding scattering center the sliding SC (SSC). The formula is as follows:

[0091]

[0092]

[0093]

[0094] Wherein, A is an amplitude constant; in the embodiment of the present application, a genetic algorithm (GA) is used for parameter estimation. 1，2 (θ, φ) represents the two surface curvatures in orthogonal directions at the effective reflection point of the hyperboloid under the radar incident viewing angle (θ, φ). F is the azimuth-dependent function, and the subscripts VV and HH represent vertical and horizontal polarizations. ξ=(θ, φ) is the Euler angle of the radar line of sight (LOS); r is the position of the target in the local coordinate system; is the direction of LOS (the direction of the radar position under the local mark); α is the frequency dependence factor, which varies with the formation mechanism of the scattering center and is an integer multiple of 1 / 2; k is the incident wave number; f c is the center frequency; γ(≥0) is the amplitude attenuation factor of the local scattering center; is the main scattering direction of the local scattering center; L represents the length of the distributed scattering center. For the local scattering center, L = 0; represents the observable angle of the scattering center.

[0095] Based on the electromagnetic wave's multiple reflection path and propagation distance, incident direction and geometric structure, the equivalent position of the scattering center corresponding to the number of reflections can be directly obtained. At one incident angle, a multiple reflection scattering center (MSC) will be formed, and its scattering center model expression is as follows

[0096]

[0097] Where I represents the number of MSCs at the corresponding angle, that is, multiple MSCs are formed at the first incident angle. denote the amplitude function, frequency dependence factor and equivalent position of the i-th MSC respectively.

[0098] The above are all based on the scattering center model of metal PEC targets. For the formula of the scattering center model of thin-coated medium targets, it is also necessary to introduce the amplitude coefficient R of the coating medium. c As the attenuation factor of the amplitude we get:

[0099]

[0100] Among them, E C_SSC 、E C_DSC and E C_LSC They are the medium sliding scattering center, medium distribution scattering center and medium local scattering center of the coating material.

[0101] Dielectric amplitude coefficient R c The expression is

[0102]

[0103] Among them, ε r is the relative dielectric constant of the coating medium material, ψ=kd represents the phase difference corresponding to the propagation distance d of the electromagnetic wave in the medium, k and d represent the wave number and coating thickness in the coating medium respectively.

[0104] Therefore, for both the benchmark PEC target and the complex PEC target, E s (θ,φ,f) equals E SSC (θ,φ,f),E DSC (θ,φ,f),E LSC (θ, φ, f) and E MSC (θ, φ, f) superposition; for complex medium targets, E s (θ,φ,f) equals E coated (θ,φ,f).

[0105] Furthermore, in the physical drive data design, for the reference PEC target I, complex PEC target II, and complex medium target III, the embodiment of the present application respectively establishes 35, 45, and 41 scattering center models. The number and properties of the designed scattering centers are shown in Table 1.

[0106] Table 1 Scattering center information

[0107]

[0108]

[0109] SLICY target I has multiple scattering mechanisms and is often used as a benchmark for validating electromagnetic scattering methods. Its model and scattering center type distribution are shown in Figure 2. Figure 2 The target's dimensions are: 0.5625 meters long, 0.5 meters wide, and 0.3436 meters high. Its geometric shape is as follows: Figure 2 As shown. The multilevel fast multipole method (PMLFMA) is used as the label for network training. The frequency is 3GHz-4GHz, and the interval is 100MHz. The pitch angle θ = 0°-90°, the interval is 5°, the azimuth angle φ = 0°-360°, the interval is 1°, and the polarization is VV and HH. The VV and HH polarization datasets have 75449 samples (named as dataset I), among which the division ratio of training, validation and test sets is 6:2:2.

[0110] The geometric structure of complex PEC target II and its scattering center type distribution are shown in Figure 2. Figure 3 As shown, the size is 5.6m×2.6m×1.05m. Real-world test data is used as labels during training. The frequency bands are X and Ku, VV and HH polarizations, the pitch angle θ=75°-105°, and the azimuth angle φ=0°-360°. Dataset II has a total of 43212 samples, among which the division ratio of training, validation and test sets is 6:2:2.

[0111] For complex medium target III, the whole body is covered with 0.01m thick absorbing material, and the medium parameter is 6-1.0j. Figure 4The target III model and its medium scattering center type distribution are shown, with a size of 5.6m×2.6m×1.128m. PMLFMA is used as the label, the frequency bands are L band, S band, C band and X band, VV and HH polarization, the pitch angle θ=30°-90°, the azimuth angle φ=0°-180° (symmetrical about the XOZ plane, the range of angles displayed as 0°-180°), and the dataset III has a total of 43212 samples, among which the division ratio of training, validation and test sets is 6:2:2.

[0112] experiment

[0113] First, the slice target dataset I was trained using machine learning methods XGBoost Regression, Random Forest, Decision Tree, Polynomial Regression, and Gradient Boosting Regression. On the entire slice dataset I, the prediction results of VV and HH at the same pitch angle are not much different; secondly, the display of representative angles of VV polarization was randomly selected (when θ = 0°, the scattering component is relatively simple, almost all contributed by the plane DSC, which is a "quasi-straight line"; and as the angle changes, the scattering component gradually increases, and at 90°, the changes are dense and the RCS fluctuates violently), such as Figure 5 As shown in (a), AVG in the figure represents the average performance of each method under the current corresponding angle, frequency, and polarization. The prediction results of the stacked integrated network after the improvement by introducing the physical information mechanism are shown in Figure 5 (b) shows the results under polarization and frequency. Regardless of whether the RCS fluctuates violently, good results can be achieved.

[0114] Table 2 compares the training and testing time of each method we proposed. Due to the high nonlinearity of electromagnetic scattering, the high polynomial degree of PR leads to long training time. The training time of other methods is relatively similar. The test time of each method is not much different, almost reaching the level of real-time prediction. Taking the well-trained SCD-DT network as an example, it can accurately predict the RCS of the target within 0.078s, which can be considered for application in real-time scenarios.

[0115] Table 2 Comparison of time performance of each network model

[0116]

[0117] On the target I dataset, the SCD-DT method not only takes the least total training and testing time, but also has the best performance and improvement in the prediction evaluation. Considering the comprehensive time performance and prediction index performance, this section uses SCD-DT as a transfer learning method to predict the single-station RCS of complex targets. Specifically, we use the prior information contained in the large dataset I of the PEC benchmark target to reduce the number of datasets by 43%, use the "pre-training and fine-tuning" technology of transfer learning on the SCD-DT network, use the weights of the pre-trained network as the initial weights, and further fine-tune on the measured dataset II of complex PEC targets and the full-wave method dataset III of dielectric-coated scattering targets (adding new feature medium parameters and coating thickness, freezing the underlying network of the pre-trained model, and fine-tuning on the new input layer) to achieve better performance.

[0118] For complex measured PEC targets such as Figure 6 As shown, we randomly selected the Ku-band HH polarization mode θ=90° in Dataset II. For complex medium targets such as Figure 7 As shown, we randomly selected the C-band VV polarization θ = 30° in Dataset III. Figure 6 and 7 The prediction results before and after adding the physical information mechanism are shown respectively. The improved prediction values ​​are in good agreement with the label data. These indicate that the transfer learning method with physical information of scattering centers is an effective method, which not only reduces the amount of training data, but also assists in analyzing the electromagnetic backscattering of measured complex targets and complex medium targets.

[0119] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A target RCS prediction method based on a stacked integrated network, characterized in that: The specific process is: Network construction: construct a stacked integrated network including a base model and a meta-model, wherein the prediction value of the base model is used as an input feature sequence of the meta-model; Data preprocessing: Collect RCS data of different targets as data sets, standardize the data sets of different targets, and construct training data sets and verification data sets; Model training: Use the preprocessed training data set to train each base model, use the trained base model to predict the validation data set, and merge the prediction results of multiple base models into new features as the input of the meta-model to train the meta-model; RCS prediction: Use the trained stacked ensemble network to predict the target RCS.

2. The target RCS prediction method based on stacked integrated network according to claim 1, characterized in that: The meta-model is a bidirectional long short-term memory network Bi-LSTM.

3. The target RCS prediction method based on stacked integrated network according to claim 1 or 2, characterized in that: The training of the meta-model is guided by introducing physical equations into the loss function. Specifically, the loss function includes a data-driven loss function and a physics-driven loss function.

4. The target RCS prediction method based on stacked integrated network according to claim 3, characterized in that: The loss function is: in, represents the physics-based drive loss function, represents the data-driven loss function, is the meta-model prediction output, y i is the true label value, λ is the regularization parameter, W (l) and denote the parameters of the neural network and physical information mechanism, respectively; in, represents the L2 norm, N R is the total number of residual terms R; in, Represents the solution RCS of the neural network during the iteration process, that is, the prediction result output by the meta-neural network σ(θ,φ,f) represents the radar cross section calculated based on the physical mechanism equation; E s is the scattered field, E i is the incident field, and R is the distance between the radar and the target under far-field conditions.

5. The target RCS prediction method based on stacked integrated network according to claim 1, characterized in that: The training process is: First, for each base model m j (j=1,2,...,M), divide the data set D={D1,D2,…,D5} into 5 folds; Second, for each fold k, use all data except the kth fold as the training set The k-th fold data As a validation set; Then, using the training set Training base model m j ,get Finally, use the trained model For the k-th fold data Make a prediction and get the predicted value The results predicted by multiple base models The output is merged into a new feature as the input of the meta-model for meta-model training. The meta-model outputs the final prediction result by learning the mapping relationship between the base model output and the true target value.

6. The target RCS prediction method based on stacked integrated network according to claim 1, characterized in that: The base model is: XGB Regression, random forest, decision tree, polynomial regression or gradient boosting regression, and the base model uses grid search to optimize training hyperparameters.

7. The target RCS prediction method based on stacked integrated network according to claim 3, characterized in that: The Bi-LSTM model uses an extrapolation technique to predict the RCS sequence, and sets the window size to 10, that is, in a set of azimuths containing 361 RCS data, the first 10 data points are used to predict the subsequent sequence.

8. The target RCS prediction method based on stacked integrated network according to claim 1, characterized in that: The selected targets include reference PEC targets, complex PEC targets and complex medium targets.

9. The target RCS prediction method based on stacked integrated network according to claim 1, characterized in that: The standardization process is: using zero mean and unit variance method to standardize the data.

10. A target RCS prediction device based on a stacked integrated network, characterized in that: It includes a data preprocessing module and a stacking integration network; wherein, Data preprocessing: Collect RCS data of different targets as data sets, and standardize the data sets of different targets; Stacked ensemble network: Use the stacked ensemble network trained by the above method to predict the target RCS of the standardized data set.

Images