RCS system simulation testing method and system based on deep neural network

Abstract

Disclosed in the present disclosure are an RCS system simulation testing method and system based on a deep neural network. The method comprises: collecting RCS target echo feature data and segmenting same to obtain a temporal feature sequence; constructing a dual-branch deep neural network structure to process the temporal feature sequence, and fusing features extracted by a main branch and an auxiliary branch to obtain an RCS feature representation; providing a cross-scale feature correction unit in the dual-branch deep neural network structure, calculating a cross-correlation coefficient between a feature map from the main branch and a feature map from the auxiliary branch, and dynamically correcting features from the main branch and the auxiliary branch; on the basis of an output from the correction unit, generating high-frequency and low-frequency RCS simulation features, and performing adaptive fusion to obtain RCS system simulation features; and calculating a simulation error between the RCS system simulation features and reference features, and optimizing network parameters. The present disclosure significantly improves the adaptability and stability of the measurement system, and is of great significance for improving the overall performance evaluation level for electronic warfare systems.

Description

A simulation and testing method and system for RCS system based on deep neural networks

[0001] Cross-references to related applications

[0002] This application claims priority to Chinese Patent Application No. 202510363483.X, filed on March 26, 2025, entitled "A Simulation Test Method and System for RCS System Based on Deep Neural Network", the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates to the fields of signal processing and testing evaluation technology, and in particular to a simulation testing method and system for RCS systems based on deep neural networks. Background Technology

[0004] Radar cross section (RCS) measurement is a crucial testing method in modern electronic warfare systems, and its simulation accuracy and reliability directly impact the evaluation results of weapon performance. With the continuous development of stealth technology in modern combat platforms, the accuracy requirements for RCS measurement systems are becoming increasingly stringent. Existing RCS simulation testing systems primarily focus on the efficiency of measurement data acquisition and processing, but they still have significant shortcomings in the quality evaluation and dynamic optimization of measurement results.

[0005] Current RCS measurement systems typically use a single metric to evaluate measurement quality, such as signal-to-noise ratio or measurement error, which fails to comprehensively reflect the scattering characteristics of complex targets at different attitudes and frequencies. This is particularly true when dealing with multi-target or complex structural targets, where a detailed evaluation mechanism for local scattering characteristics is lacking. Furthermore, existing systems often employ static measurement strategies, lacking the ability to perform real-time evaluation and parameter adjustment during the measurement process, which can easily lead to unstable measurement accuracy or loss of local features.

[0006] More importantly, existing technologies rarely consider the interrelationships between the scattering characteristics of different parts of the target, and lack effective monitoring mechanisms for changes in data correlation during the measurement process. In RCS measurements of complex targets, the preservation of the organizational relationships between local scattering features is often overlooked, which may lead to measurement results that do not accurately reflect the target's scattering characteristics. Furthermore, existing systems mostly use fixed measurement parameters, making it difficult to adaptively adjust measurement strategies according to different target characteristics and measurement environments, which severely restricts the improvement of measurement accuracy. These technical shortcomings not only affect the accuracy and reliability of RCS measurements but also limit the accuracy of electronic countermeasures system performance evaluation.

[0007] Application content

[0008] This disclosure proposes a simulation test method and system for RCS systems based on deep neural networks, in order to improve the reliability of measurement results of RCS simulation test systems.

[0009] The embodiments of this disclosure can be implemented as follows:

[0010] This disclosure provides an embodiment of an RCS system simulation testing method based on a deep neural network, comprising: acquiring RCS target echo feature data; segmenting the RCS target echo feature data according to a preset time window length to obtain a time-series feature sequence; constructing a dual-branch deep neural network structure based on the time-series feature sequence, wherein the main branch uses a standard convolutional layer to process the time-series feature sequence, and the auxiliary branch is equipped with a spectrum reconstruction module to fuse the features extracted by the main branch and the auxiliary branch to obtain an RCS feature representation; setting a cross-scale feature correction unit in the dual-branch deep neural network structure, wherein the cross-scale feature correction unit dynamically corrects the features of the main branch and the auxiliary branch by calculating the cross-correlation coefficient between the feature maps of the main branch and the auxiliary branch; based on the output of the cross-scale feature correction unit, generating high-frequency RCS simulation features and low-frequency RCS simulation features through the main branch decoder and the auxiliary branch decoder respectively, and adaptively fusing them using a cross-correlation coefficient matrix to obtain RCS system simulation features; using the RCS feature representation as a reference feature, calculating the simulation error between the RCS system simulation features and the reference feature, and optimizing the network parameters based on the simulation error to improve the RCS system simulation accuracy.

[0011] Optionally, the target echo feature data of the RCS is collected, the original target echo data is bandpass filtered to obtain the filtered echo signal, and then resampled to obtain a discrete sampling sequence; a window function is applied to the discrete sampling sequence for segmented windowing processing, and a fast Fourier transform is performed on the windowed signal sequence to obtain spectral features; amplitude spectrum and phase spectrum information are extracted from the spectral features to form a time-series feature sequence.

[0012] Optionally, the main branch of the dual-branch deep neural network structure is configured to process the temporal feature sequence, and the sub-branch of the dual-branch deep neural network structure is configured to process the frequency domain components, namely the amplitude spectrum and phase spectrum information; the main branch uses standard convolutional layers to process the temporal feature sequence; and residual connections are added every two convolutional layers to form a residual block structure; the residual block structure is a dual-path structure, including a main path and a shortcut connection path; the features processed by the main path and the features of the shortcut connection path are added element-wise, and the merged features are then processed by the ReLU activation function to obtain the final residual block output. The feature maps of each residual block output are collected, and the feature maps include shallow feature maps and deep feature maps; an adaptive weight mechanism is used to fuse the feature maps to obtain a fused feature representation; the fused feature representation is subjected to dimensionality compression processing to obtain a fixed-dimensional main branch feature vector F.

[0013] Optionally, the sub-branch receives amplitude spectrum and phase spectrum information, inputs the amplitude spectrum into a frequency domain feature extraction module composed of L layers of one-dimensional convolutional layers to generate frequency distribution features; and introduces a compensation mechanism for the phase spectrum to reconstruct the frequency domain signal Xr based on the compensated phase spectrum and the original amplitude spectrum A.

[0014] Where j is the imaginary unit, The phase spectrum after compensation is obtained; the reconstructed frequency domain signal is converted into a feature vector R through a feature mapping network consisting of two fully connected layers; the main branch feature vector F and the branch feature vector R are concatenated in the feature dimension to form a fused feature vector; the fused feature vector is subjected to dimensionality reduction mapping to obtain the RCS feature representation Y.

[0015] Optionally, the cross-scale feature correction unit includes a cross-correlation calculation module, a weight generation module, and a feature correction module. The cross-correlation calculation module is configured to calculate the correlation matrix between two branch features, the weight generation module is configured to generate weight factors for feature channels, and the feature correction module is configured to optimize and correct the features. The correlation matrix is ​​calculated by extracting the main and auxiliary branch feature maps, adjusting them to the same feature space through projection transformation, and calculating the cross-correlation coefficient matrix G. Specifically, the main branch feature map is extracted from the feature map of each residual block. Where H and W are the height and width of the feature map, respectively, C is the number of feature channels, and L represents the Lth residual block level; the auxiliary branch feature map FR is obtained from the intermediate layer of the feature mapping network; and the main branch feature map FM... L Rearrange and convert to matrix form. The auxiliary branch features are adjusted in dimension and normalized to be converted into a matrix form compatible with the main branch features. Calculate the cross-correlation matrix G between the features of the primary and secondary branches:

[0016] in, It is a cross-correlation matrix. This indicates the direct correlation between features. The autocorrelation matrix representing the characteristics of the main branch. The autocorrelation matrix representing the auxiliary branch features is used; the cross-correlation matrix G is normalized to obtain the main branch feature weight factor λ1; the weight factor λ1 is configured to characterize the contribution of each feature channel of the main branch to the RCS feature representation; the auxiliary branch feature weight factor λ2 is obtained through column normalization; the weight factor λ2 is configured to characterize the contribution of each feature channel of the auxiliary branch to the RCS feature representation; the feature correction module configures the weight factor λ1 to the main branch features to obtain the corrected feature map FM. L The weighting factor λ2 should be configured as an auxiliary branch feature to obtain the corrected feature map FR.

[0017] Optionally, based on the corrected feature map FM L The high-frequency RCS simulation features are reconstructed using a decoder; based on the corrected feature map FR, it is reconstructed into low-frequency RCS simulation features using an auxiliary branch decoder; the high-frequency and low-frequency RCS simulation features are then dimensionality-unified and normalized to generate a high-frequency feature weight matrix G. h and low-frequency feature weight matrix G l The weight matrix is ​​multiplied element-wise with the corresponding feature map, and the weighted high-frequency and low-frequency features are added element-wise to obtain the simulation feature T of the RCS system.

[0018] Optionally, the feature reconstruction error between the RCS system simulation features and the RCS feature representation is calculated, specifically: calculating the feature reconstruction error E between the RCS system simulation features and the RCS feature representation. r For high-frequency RCS features, calculate the degree of matching with the high-frequency components in the RCS feature representation Y; for low-frequency RCS features, evaluate the consistency with the low-frequency components in the RCS feature representation Y; analyze the changes in the cross-correlation matrix G during feature fusion to evaluate the fusion quality; and combine the matching degree of high-frequency RCS features, the consistency of low-frequency RCS features, and the fusion quality to obtain the standardized branch feature error E. b The total error is calculated based on the feature reconstruction error weights and the branch feature error weights, and is expressed as: E t =μ1·E r +μ2·E b

[0019] Where μ1 is the feature reconstruction error weight and μ2 is the branch feature error weight.

[0020] Optionally, based on the changing trends of various errors, parameter settings are optimized, including: if entering the main branch optimization, then prioritizing the optimization of residual block convolutional layer parameters, adjusting adaptive weight mechanism parameters, updating main branch decoder parameters, and enhancing the main feature extraction capability; if the main branch optimization is complete, then optimizing the feature mapping network parameters in the auxiliary branch, updating auxiliary branch decoder parameters, and improving auxiliary feature extraction capability; if the total error is lower than a preset threshold, or the error change for N consecutive rounds is less than the convergence threshold, or the maximum number of iterations I is reached... max If the validation set error does not improve for K consecutive rounds, the optimization process is terminated, and the network parameter optimization is completed.

[0021] This disclosure also provides an RCS system simulation and testing system based on a deep neural network, comprising: a data acquisition module configured to acquire RCS target echo feature data, segment the RCS target echo feature data according to a preset time window length to obtain a time-series feature sequence; a feature representation module, constructing a dual-branch deep neural network structure based on the time-series feature sequence, wherein the main branch uses a standard convolutional layer to process the time-series feature sequence, and the auxiliary branch is equipped with a spectrum reconstruction module to fuse the features extracted by the main branch and the auxiliary branch to obtain an RCS feature representation; a dynamic correction module, setting a cross-scale feature correction unit in the dual-branch deep neural network structure, dynamically correcting the features of the main branch and the auxiliary branch by calculating the cross-correlation coefficient between the feature maps of the main branch and the auxiliary branch; and a simulation optimization module configured to generate high-frequency RCS simulation features and low-frequency RCS simulation features respectively, and adaptively fuse them to obtain RCS system simulation features, using the RCS feature representation as a reference feature, calculating the simulation error between the RCS system simulation features and the reference feature, and optimizing the network parameters based on the simulation error to improve the simulation accuracy of the RCS system.

[0022] The beneficial effects of this disclosure include: By applying deep neural network technology to the RCS simulation testing system, an intelligent measurement quality assessment framework is constructed. Convolutional neural networks and recurrent neural networks are used to extract and analyze features from measurement data, establishing a multi-dimensional assessment mechanism that includes mean approximation, fluctuation consistency, and structural correlation, achieving a comprehensive assessment of measurement data quality. Based on the feature extraction capabilities of the deep learning model, the system can accurately identify abnormal patterns in measurement results and adaptively adjust measurement strategies through reinforcement learning methods. Simultaneously, the continuous learning and optimization capabilities of the deep learning model enable the system to continuously improve assessment accuracy and robustness, providing a new technical path for the intelligent upgrading of RCS testing systems, significantly enhancing the adaptive capability and stability of the measurement system, and playing a significant role in improving the overall performance evaluation level of electronic countermeasures systems. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0024] Figure 1 is a flowchart of the RCS system simulation test method based on deep neural networks. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of this disclosure more apparent and understandable, the specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this disclosure.

[0026] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure. However, this disclosure may be implemented in other ways different from those described herein, and those skilled in the art may make similar extensions without departing from the spirit of this disclosure. Therefore, this disclosure is not limited to the specific embodiments disclosed below.

[0027] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this disclosure. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0028] Referring to Figure 1, an embodiment of this disclosure provides a simulation and testing method for an RCS system based on a deep neural network, comprising:

[0029] S1: Collect RCS target echo feature data, segment the RCS target echo feature data according to the preset time window length, perform fast Fourier transform on the windowed signal sequence to obtain spectral features, and extract amplitude spectrum and phase spectrum information from the spectral features to form a time-series feature sequence.

[0030] The acquisition and processing of RCS target echo characteristic data is performed using multi-angle radar, including:

[0031] A phased array radar system with an operating frequency of F is used to scan the RCS target within the measurement angle range at an angle interval Δθ to obtain the raw target echo data, wherein the angle interval Δθ is less than 1 / 30 of the measurement angle range;

[0032] The original target echo data is bandpass filtered, and the bandwidth B of the bandpass filter satisfies 0.01F≤B≤0.1F to obtain the filtered echo signal.

[0033] The filtered echo signal is resampled at a sampling rate of not less than twice the bandwidth B to obtain a discrete sampling sequence.

[0034] Applying window functions to discrete sampled sequences for piecewise windowing processing includes:

[0035] The window function is selected based on the frequency characteristics of the discrete sampling sequence. Specifically, the balance between frequency resolution and dynamic range is achieved by adjusting the ratio of the main lobe width to the side lobe amplitude of the window function.

[0036] Window functions can smoothly transition to zero at the time domain endpoints, suppressing spectral leakage caused by truncation and improving the accuracy of spectral analysis;

[0037] The windowed sequence is segmented using an overlap-addition method. The continuity of the signal is maintained by adjusting the length of the overlapping region. The length of the overlapping region is determined according to the time-varying characteristics of the signal. The overlap rate of adjacent windows is between 30% and 70%, thus obtaining the windowed signal sequence.

[0038] Performing a Fast Fourier Transform on the windowed signal sequence yields its spectral characteristics, specifically:

[0039] The windowed signal sequence is padded with zeros to a power of 2 N sampling points, where N is greater than the number of sampling points in the windowed signal sequence;

[0040] The zero-padding signal sequence is decomposed by a butterfly operation, which decomposes the N-point Fourier transform into a log2N-level 2-point Fourier transform.

[0041] Calculate the rotation factor for each level of the 2-point Fourier transform, where the rotation factor is e. -j2πk / N , where k is the frequency index;

[0042] The butterfly operation is performed using the time decimation method to obtain the spectral features, which contain N / 2 non-redundant frequency points.

[0043] Amplitude and phase spectrum information are extracted from the spectral features to form a time-series feature sequence. The feature dimension of the time-series feature sequence does not exceed half the number of points of the Fast Fourier Transform.

[0044] The amplitude spectrum reflects the energy distribution intensity of the signal at different frequencies, while the phase spectrum reflects the relative phase relationships of each frequency component. To construct a time-series feature sequence, the amplitude and phase spectrum information needs to be combined. This combination process involves combining the amplitude value and corresponding phase angle at each frequency point to form a complete frequency domain representation. This complete frequency domain representation is then converted back to the time domain using an inverse Fourier transform. The conversion process can be represented as: superimposing all frequency components according to their respective amplitudes and phases; each frequency component contributes to the final time-domain signal. Frequency components with larger amplitudes have a more significant impact on the time-domain signal, while the phase determines the precise temporal position of these components.

[0045] After the transformation, a new time-series feature sequence is obtained. This sequence maintains the same time length as the original windowed signal, but it contains optimized information after frequency domain analysis and reconstruction.

[0046] S2: A dual-branch deep neural network structure is constructed based on temporal feature sequences and frequency domain components. The main branch uses standard convolutional layers to process temporal feature sequences, while the auxiliary branch is equipped with a spectrum reconstruction module. The spectrum reconstruction module reconstructs the frequency domain information by introducing a phase compensation mechanism, and fuses the features extracted by the main branch and the auxiliary branch to obtain the RCS feature representation.

[0047] By decomposing the signal in the time and frequency domains, we obtain the time-series characteristic sequence and frequency domain components, namely the amplitude spectrum and the phase spectrum.

[0048] Based on the decomposed time-series feature sequences and frequency-domain components, a dual-branch deep neural network structure is constructed to process time-domain and frequency-domain information respectively, specifically including:

[0049] The main branch of the dual-branch deep neural network structure includes:

[0050] The temporal feature sequence is extracted through N standard one-dimensional convolutional layers. The kernel size of each convolutional layer is b, the number of kernels is m, the stride is s, and the padding is p.

[0051] Increasing kernel sizes are used from layer 1 to layer N / 2 to capture local features at different scales;

[0052] Decreasing kernel sizes are used from layer N / 2+1 to layer N to achieve layer-by-layer feature refinement;

[0053] Each convolutional layer output is followed by a batch normalization layer, configured to eliminate data distribution bias. Specifically, the feature maps output by the convolutional layers are standardized. First, the mean and variance of each batch of data are calculated. Normalization is performed by subtracting the mean and dividing by the variance. Learnable scaling factor γ and offset factor β are introduced to increase the expressive power of the model.

[0054] Residual connections are added every two convolutional layers to form residual block structures, which alleviates the gradient vanishing problem in deep networks. The residual blocks adopt a dual-path structure, including the main path and the shortcut connection path.

[0055] In the main path, the input features are first extracted through the first convolutional layer, then pass through batch normalization layers to eliminate data distribution bias, and finally pass through the ReLU activation function to introduce non-linear features. These processed features are then extracted through a second convolutional layer and pass through another batch normalization layer. Simultaneously, the input features are directly passed to the end of the residual block via the shortcut connection path. When the dimensions of the input features and the output features of the main path do not match, a 1×1 convolutional layer is used to adjust the dimensions of the shortcut connection features. The features processed by the main path and the features from the shortcut connection path are then element-wise added, and the merged features are then passed through the ReLU activation function to obtain the final residual block output.

[0056] The main path enables the learning of new feature transformations, while shortcut connections preserve the original feature information, thus better transmitting gradient information in deep networks and effectively mitigating the vanishing gradient problem. The double convolutional structure on the main path enhances feature extraction capabilities, while shortcut connections ensure effective information transmission. The combination of these two aspects improves the overall learning ability and training stability of the network.

[0057] Feature maps output from each residual block are collected. These feature maps contain feature information at different levels. Shallow feature maps mainly contain local detail information, while deeper feature maps contain more high-level semantic information. To fully utilize multi-level features, an adaptive weighting mechanism is used for feature fusion: an importance weight is calculated for each level of feature map, and this importance weight is automatically adjusted through learnable parameters; features at different levels are weighted according to the calculated importance weights, so that more important features receive higher weight values; the weighted feature maps are then summed element-wise to obtain the fused feature representation.

[0058] To obtain a fixed-dimensional feature vector, dimensionality compression is performed on the fused feature representation. Specifically, global pooling is used to compress the temporal dimension, reducing the entire temporal feature to a single feature value. The global pooling process can simultaneously perform max pooling and average pooling operations, and adaptively fuse the two pooling results using learnable parameters to retain the most representative feature information. Ultimately, a fixed-dimensional main branch feature vector F is obtained, which retains multi-level feature information while maintaining a fixed dimension.

[0059] The auxiliary branch for constructing a two-branch deep neural network structure includes:

[0060] The amplitude spectrum and phase spectrum of the received output are input into the frequency domain feature extraction module composed of L layers of one-dimensional convolutional layers to generate frequency distribution features.

[0061] A compensation mechanism is introduced into the phase spectrum, resulting in a compensated phase spectrum. satisfy:

[0062] in The original phase spectrum is represented by α, the compensation amplitude coefficient is represented by β, the compensation frequency coefficient is represented by θ, and the compensation phase shift is represented by θ. α, β, and θ are all learnable parameters.

[0063] Reconstruct the frequency domain signal Xr based on the compensated phase spectrum and the original amplitude spectrum A:

[0064] Here, j is the imaginary unit, and the reconstructed signal enhances the expression of phase characteristics.

[0065] The reconstructed frequency domain signal is converted into a feature vector R through a feature mapping network consisting of two fully connected layers. The output dimension of the feature mapping network is the same as the dimension of the main branch feature vector F. Specifically, it includes:

[0066] The reconstructed frequency domain signal is passed through a feature mapping network consisting of two fully connected layers. First, the reconstructed frequency domain signal is input into the first fully connected layer, which has more neurons than the input dimension, thus expanding the features. The first fully connected layer is followed by a BatchNorm normalization layer and a ReLU activation function, introducing nonlinear transformation capabilities while maintaining the stability of the data distribution.

[0067] Then, the output features from the first layer are fed into the second fully connected layer, where the number of neurons is precisely set to the same dimension as the main branch feature vector F. The second fully connected layer also features a BatchNorm normalization layer, but uses the Tanh activation function instead of ReLU at the end. This is because the Tanh function's output range is between -1 and 1, which helps to normalize the feature representation, ensuring that the generated feature vector R has a similar numerical distribution range to the main branch feature vector F.

[0068] The main branch feature vector F and the branch feature vector R are concatenated along the feature dimension to form a fused feature vector.

[0069] The fused feature vector is dimensionality-reduced using a fully connected layer to obtain the final RCS feature representation Y, which is expressed as: Y = V·[F;R] + g

[0070] Where V is the weight matrix, g is the bias vector, and [F;R] represents the feature concatenation operation.

[0071] S3: In the dual-branch deep neural network structure, a cross-scale feature correction unit is set up. The cross-scale feature correction unit dynamically corrects the features of the main branch and the auxiliary branch by calculating the cross-correlation coefficient between the feature maps of the main branch and the auxiliary branch.

[0072] The cross-scale feature correction unit includes a cross-correlation calculation module, a weight generation module, and a feature correction module. The cross-correlation calculation module is configured to calculate the correlation matrix between two branch features, the weight generation module is configured to generate weight factors for feature channels, and the feature correction module is configured to perform weighted optimization on the features.

[0073] The cross-correlation calculation module calculates the correlation matrix between the features of the two branches, including:

[0074] Extracting the main and auxiliary branch feature maps:

[0075] Main branch: Extract feature maps from each residual block. Where H and W are the height and width of the feature map, respectively, C is the number of feature channels, and L represents the Lth residual block level. The features of each residual block retain temporal dynamic information at different scales;

[0076] Auxiliary branch: Obtains feature map FR from the intermediate layer of the feature mapping network, which contains phase-compensated frequency domain information.

[0077] Since the main branch features come from residual blocks and the auxiliary branch features come from the feature mapping network, their feature representations are different, requiring feature alignment; the main branch feature map is then FM-aligned. L The auxiliary branch feature map (FR) is adjusted to the same feature space through projection transformation to ensure feature dimension matching. Specifically:

[0078] FM of the main branch feature map L Rearrange and convert to matrix form.

[0079] The auxiliary branch features are adjusted in dimension and normalized to be converted into a matrix form compatible with the main branch features.

[0080] Projection transformation ensures that the features of the two branches are represented in the same feature space.

[0081] Calculate the cross-correlation matrix G between the features of the primary and secondary branches:

[0082] in, It is a cross-correlation matrix. This indicates the direct correlation between features. The autocorrelation matrix representing the characteristics of the main branch. The autocorrelation matrix represents the characteristics of the secondary branches.

[0083] Elements g in the cross-correlation matrix G ij ∈[-1, 1] represents the correlation strength between the i-th main branch feature channel and the j-th auxiliary branch feature channel.

[0084] Each row of the G matrix represents the correlation distribution between a main branch channel and all auxiliary branch channels;

[0085] Each column of the G matrix represents the correlation distribution between a secondary branch channel and all primary branch channels;

[0086] By normalizing the cross-correlation matrix G, the main branch feature weight factor λ1 is obtained;

[0087] The weighting factor λ1 is configured to characterize the contribution of each feature channel of the main branch to the RCS feature representation;

[0088] The auxiliary branch feature weight factor λ2 is obtained by column normalization;

[0089] The weighting factor λ2 is configured to characterize the contribution of each feature channel of the auxiliary branch to the RCS feature representation.

[0090] The main branch features are processed using a weighting factor λ1 to obtain the corrected feature map FM. L The correction process involves processing the weight factor λ1 using the sigmoid function, mapping it to the [0, 1] interval, and then comparing it with the feature map FM. L Multiplication is used for correction, and the correction process ensures that key information in the time domain features is highlighted while suppressing secondary features.

[0091] The auxiliary branch features are processed using a weighting factor to obtain the corrected feature map FR'. The correction process also uses the sigmoid function to process the weighting factor λ2, and then multiplies it with the feature map FR for correction. The correction process ensures the synergy between the frequency domain features and the time domain features.

[0092] S4. Based on the output of the cross-scale feature correction unit, high-frequency RCS simulation features and low-frequency RCS simulation features are generated by the main branch decoder and the auxiliary branch decoder, respectively. The cross-correlation coefficient matrix is ​​used for adaptive fusion to obtain the complete RCS system simulation features.

[0093] Corrected feature map FM L It contains key temporal information about the target, but due to the initial downsampling and convolution operations, the spatial resolution is low. It needs to be reconstructed into high-frequency RCS simulation features through a decoder.

[0094] The decoding process employs a multi-layer deconvolution structure, which gradually improves the spatial resolution of the feature map through transpose convolution operations.

[0095] Specifically, FM L The input is first fed into the first deconvolutional layer, where the features are mapped to a higher resolution space through transposed convolution. Transposed convolution expands the feature map size by inserting zeros between the feature maps and performing convolution operations. Simultaneously, a non-linear activation function is appended to each deconvolutional layer to enhance the expressive power of the features. To avoid information loss during feature propagation, residual connections are added between adjacent deconvolutional layers to preserve the original feature information.

[0096] Since each deconvolutional layer can learn feature patterns at different scales, shallower layers mainly reconstruct the basic structure, while deeper layers can recover finer local details. As features are passed layer by layer through the multi-layer deconvolutional structure, the spatial resolution continuously improves, and the detailed information of the features is gradually recovered. The final output high-frequency RCS simulation features maintain FM... L The key time-domain information contained in it also includes rich high-frequency details.

[0097] Similarly, after obtaining the corrected auxiliary branch feature FR', it is reconstructed into a low-frequency RCS simulation feature using the auxiliary branch decoder. Since the low-frequency feature reflects the overall structural characteristics of the target, the decoding process needs to pay attention to the global consistency of the feature.

[0098] The auxiliary branch decoder also employs a multi-layer deconvolution structure, but it differs from the main branch decoder. First, after the first deconvolution layer, FR' uses a larger-sized convolutional kernel for feature mapping, allowing each output position to capture a wider range of input information, thus preserving the integrity of the features. Simultaneously, global average pooling is introduced between the deconvolutional layers to extract and fuse global feature information, enhancing the overall expressive power of the features.

[0099] During feature propagation, the spatial resolution of the feature maps is improved through layer-by-layer deconvolution. A normalization layer is added after each deconvolution layer to stabilize the feature distribution and ensure the stability of the reconstruction process. In addition, to better maintain the continuity of low-frequency features, skip connections are used between adjacent layers to fuse the basic features of the lower layers with the semantic features of the higher layers, outputting low-frequency RCS simulation features that retain the main structural information and overall contour features of the target.

[0100] The high-frequency RCS simulation features and low-frequency RCS simulation features are dimensionally unified to have the same spatial resolution and number of feature channels. The elements in the cross-correlation matrix G are normalized using the Softmax function, mapping the correlation coefficients to the [0, 1] interval to generate the high-frequency feature weight matrix G. h and low-frequency feature weight matrix G l And ensure that the weight matrix satisfies G h +G lThe constraint condition is 1;

[0101] The weight matrix is ​​multiplied element-wise with the corresponding feature map, and the weighted high-frequency and low-frequency features are added element-wise to obtain the complete RCS system simulation feature T, expressed as: T = G h ⊙F h +G l ⊙F l

[0102] Among them, F h F represents high-frequency characteristics. l ⊙ indicates low-frequency characteristics, and ⊙ indicates element-wise multiplication.

[0103] S5. Using the RCS feature representation as a reference feature, calculate the simulation error between the complete RCS system simulation feature and the reference feature. Based on the simulation error, optimize the network parameters to improve the simulation accuracy of the RCS system.

[0104] The feature reconstruction error of the RCS system is calculated based on the simulation feature T and the RCS feature representation Y. Specifically:

[0105] First, calculate the Euclidean distance error E between the simulation feature T and the RCS feature representation Y. d To assess the overall reconstruction effect, including:

[0106] The mean Euclidean distances κ1 and κ2 between samples of different classes in the labeled dataset are set as preset convergence thresholds, and κ1 < κ2.

[0107] If E d If <κ1, then the reconstructed features have achieved matching with the reference features, and the process moves to the local feature optimization stage;

[0108] If κ1≤E d If <κ2, then the reconstructed features only obtain the main features of the target, and the feature extraction network optimization continues;

[0109] If E d If the value is ≥κ2, then the reconstructed features are fundamentally different from the reference features, and the feature extraction scheme needs to be reconstructed.

[0110] Optionally, cosine similarity analysis can be used to ensure the directional consistency of two feature vectors, thereby aligning the feature spaces. This includes:

[0111] If the cosine similarity is close to 1, it means that the reconstructed features are highly consistent with the original features in terms of direction, confirming that the feature space is completely aligned.

[0112] If the cosine similarity is at a moderate level, it means that the feature vectors are basically aligned and need to be fine-tuned and optimized.

[0113] If the cosine similarity is close to 0, it indicates that there is a significant deviation in the direction of the feature vector, and the reconstruction strategy needs to be readjusted.

[0114] Based on feature structure similarity analysis, the fidelity of reconstructed features in local regions is evaluated, including:

[0115] If the regional average values ​​are small, it indicates that the local feature strengths are matched, ensuring that the local feature levels are consistent.

[0116] If the degree of fluctuation in different regions is similar, it indicates that the local change patterns are consistent, confirming that the local dynamic characteristics are maintained.

[0117] If the regional correlation is strong, it indicates that the local structural relationship remains stable, verifying the integrity of the detailed features.

[0118] The above three errors are normalized to obtain the standardized feature reconstruction error E. r .

[0119] For the high-frequency RCS features generated by the main branch, calculate the degree of matching with the high-frequency components in the RCS feature representation Y, including:

[0120] Frequency domain transformation is performed on the high-frequency RCS features and RCS feature representation Y to extract the amplitude distribution of the high-frequency band, obtain high-frequency phase information, obtain the high-frequency feature components to be evaluated and the reference high-frequency components, evaluate the degree of error between the high-frequency feature components and the reference high-frequency components in amplitude difference, phase shift and spectral structure, and obtain the degree of matching of high-frequency RCS.

[0121] For the low-frequency RCS features generated by the auxiliary branch, evaluate the consistency with the RCS feature representation of the low-frequency components in Y, including:

[0122] A frequency domain transformation is performed on the low-frequency RCS features and RCS feature representation Y to extract the amplitude distribution of the low-frequency band, obtain low-frequency phase information, and obtain the low-frequency feature components to be evaluated and the reference low-frequency components. The consistency deviation of the low-frequency feature components and the reference low-frequency components in terms of baseline offset, trend change consistency and energy distribution characteristics is evaluated, and the consistency of the low-frequency RCS features is evaluated.

[0123] Analyze the changes in the cross-correlation matrix G during feature fusion to evaluate the fusion quality, including:

[0124] Diagonal element stability analysis: By comparing the changes in the autocorrelation intensity of the main features at consecutive time steps, when the change amplitude is consistently less than the preset threshold, it indicates that the self-expression of each dimension feature tends to be stable.

[0125] Off-diagonal element convergence analysis: Monitor the degree of interaction between different features. When the rate of change of interaction strength decreases and remains at a low level, it indicates that the information exchange between features has reached a balance and feature fusion has entered a stable stage.

[0126] Matrix structure consistency analysis: By comparing the overall correlation structure of adjacent time steps, when the structural differences of the matrix continue to decrease and remain at a low level, it indicates that the feature fusion network has reached a state of coordination and consistency, and the fusion quality is in the optimal range.

[0127] By combining the matching degree of high-frequency RCS features, the consistency of low-frequency RCS features, and the fusion quality, the standardized branch feature error E is obtained. b .

[0128] An adaptive weight coefficient is defined to dynamically balance the contributions of feature reconstruction error and branch feature error. The weight coefficient is automatically adjusted according to the current training stage and error distribution. The weight of feature reconstruction error gradually increases as training progresses to ensure the final reconstruction quality. The weight of branch feature error is larger in the early stage of training to ensure the accurate extraction of basic features.

[0129] The total error is calculated based on the feature reconstruction error weights and the branch feature error weights, and is expressed as: E t =μ1·E r +μ2·E b

[0130] Where μ1 is the feature reconstruction error weight and μ2 is the branch feature error weight.

[0131] Record the changing trends of various errors in each iteration, and automatically adjust the focus of the next round of optimization based on the error analysis results. When a certain type of error is significantly higher than other errors, its weight coefficient is increased accordingly, an error change curve is established, and it is configured to guide the dynamic adjustment of the optimization process.

[0132] Based on the changing trends of various errors, parameter settings are optimized, including:

[0133] If network training begins, set the initial learning rate lr and the maximum number of iterations I. max This initializes the Adam optimizer and determines the learning rate decay strategy.

[0134] If the main branch optimization is entered, the parameters of the residual block convolutional layer are optimized first, the parameters of the adaptive weight mechanism are adjusted, the parameters of the main branch decoder are updated, and the main feature extraction capability is enhanced.

[0135] If the main branch optimization is complete, optimize the feature mapping network parameters in the auxiliary branch, update the decoder parameters of the auxiliary branch, and improve the auxiliary feature extraction capability.

[0136] If the optimization of the main and auxiliary branches is completed, the feature correction unit is optimized, the cross-correlation matrix G is calculated and optimized, and the feature weight factors are dynamically adjusted to achieve precise control of feature fusion.

[0137] If each iteration is completed, the current total error is calculated, the parameters are updated using the Adam optimizer, and the learning rate is dynamically adjusted according to the error changes to ensure the efficiency of the optimization process.

[0138] If the total error is lower than a preset threshold, or the error change is less than the convergence threshold for N consecutive iterations, or the maximum number of iterations I is reached... max If the validation set error does not improve for K consecutive rounds, the optimization process is terminated, and the network parameter optimization is completed.

[0139] In summary, this disclosure constructs an intelligent measurement quality assessment framework by applying deep neural network technology to the RCS simulation testing system. Convolutional neural networks and recurrent neural networks are used to extract and analyze features from the measurement data, establishing a multi-dimensional assessment mechanism that includes mean approximation, fluctuation consistency, and structural correlation, thus achieving a comprehensive assessment of measurement data quality. Based on the feature extraction capabilities of the deep learning model, the system can accurately identify abnormal patterns in the measurement results and adaptively adjust the measurement strategy through reinforcement learning. Simultaneously, the continuous learning and optimization capabilities of the deep learning model enable the system to continuously improve its assessment accuracy and robustness, providing a new technical path for the intelligent upgrade of the RCS testing system. This significantly enhances the adaptive capability and stability of the measurement system and is of great significance for improving the overall performance evaluation level of electronic countermeasures systems.

[0140] This disclosure provides an RCS system simulation and testing system based on a deep neural network, comprising:

[0141] The data acquisition module is configured to acquire RCS target echo feature data, and to segment the RCS target echo feature data according to a preset time window length to obtain a time-series feature sequence.

[0142] The feature representation module constructs a dual-branch deep neural network structure based on temporal feature sequences. The main branch uses standard convolutional layers to process the temporal feature sequences, while the auxiliary branch is equipped with a spectrum reconstruction module to fuse the features extracted by the main branch and the auxiliary branch to obtain the RCS feature representation.

[0143] The dynamic correction module sets up a cross-scale feature correction unit in the dual-branch deep neural network structure, and performs dynamic correction on the main and auxiliary branch features by calculating the cross-correlation coefficient between the feature maps of the main branch and the auxiliary branch.

[0144] The simulation optimization module is configured to generate high-frequency RCS simulation features and low-frequency RCS simulation features through the main branch decoder and the auxiliary branch decoder, respectively, and perform adaptive fusion to obtain the RCS system simulation features. The RCS feature representation is used as the reference feature, the simulation error between the RCS system simulation features and the reference feature is calculated, and the network parameters are optimized based on the simulation error to improve the RCS system simulation accuracy.

[0145] This embodiment also provides a computer device that meets the requirements of the RCS system simulation test method based on deep neural networks, including a memory and a processor; the memory is configured to store computer-executable instructions, and the processor is configured to execute computer-executable instructions to implement the RCS system simulation test method based on deep neural networks as proposed in the above embodiment.

[0146] The computer device can be a terminal, and includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor is configured to provide computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is configured to communicate with external terminals via wired or wireless means. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0147] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the simulation and testing method for RCS system based on deep neural networks as proposed in the above embodiments.

[0148] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0149] To verify the beneficial effects of the embodiments of this disclosure, scientific demonstration is carried out through simulation experiments.

[0150] To verify the effectiveness of the RCS measurement quality assessment method based on deep neural networks provided in the embodiments of this disclosure, a fighter jet model was selected as the test object for simulation experiments. This model includes multiple typical scattering points such as the nose, wings, and vertical tail. RCS data was collected at 2-degree intervals within the 0-360 degree azimuth range. Experiments were conducted in both standard and interference environments. The interference environment was simulated by superimposing Gaussian white noise with a signal-to-noise ratio of 5dB to simulate environmental interference in actual measurements.

[0151] In standard environmental testing, analysis of RCS measurements at the nose (0 azimuth angle) showed that the method provided by the embodiments of this disclosure achieved an evaluation accuracy of 97.3%, compared to 89.1% for traditional statistical methods. In the complex scattering region at the wing-fuselage junction (45 azimuth angle), the deviation between the evaluation results of the method provided by the embodiments of this disclosure and the theoretical value remained within 3.2%, significantly better than the 8.7% deviation of traditional methods. When a measurement anomaly was artificially introduced at the 90 azimuth angle, the method provided by the embodiments of this disclosure could detect and locate the problem within 0.5 seconds, and the measurement stability after adaptive adjustment of system parameters was improved by 85.4%.

[0152] In a test environment with introduced noise interference, analysis of continuous measurement data within the 180-270 degree azimuth range showed that the evaluation results of the method provided in this disclosure achieved a consistency of 94.2%, while the traditional method dropped to 76.8%. Particularly in the 225-235 degree range where the target attitude changes rapidly, the deep learning-based parameter optimization strategy adaptively adjusts the system's signal processing parameters, reducing the impact of measurement noise by 72.5%. Statistical analysis of 50 repeated tests showed that the evaluation accuracy standard deviation of the method provided in this disclosure under complex environments was 1.8%, verifying the excellent stability and robustness of the method.

[0153] The above experimental data fully demonstrate the significant advantages of the method provided by the embodiments of this disclosure in terms of measurement quality assessment, anomaly detection, and environmental adaptability, and provide reliable technical support for the intelligent upgrade of the RCS testing system.

[0154] It should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and are not intended to limit it. Although this disclosure has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this disclosure without departing from the spirit and scope of the technical solutions of this disclosure, and all such modifications and substitutions should be covered within the scope of the claims of this disclosure. Industrial applicability

[0155] In summary, this disclosure provides a method and system for simulating and testing RCS systems based on deep neural networks, which improves the reliability of measurement results from RCS simulation testing systems.

Claims

1. A simulation and testing method for an RCS system based on a deep neural network, characterized in that, include: Collect RCS target echo feature data, and divide the RCS target echo feature data into segments according to a preset time window length to obtain a time-series feature sequence; A dual-branch deep neural network structure is constructed based on the aforementioned temporal feature sequence. The main branch uses a standard convolutional layer to process the temporal feature sequence, while the auxiliary branch is equipped with a spectrum reconstruction module. The features extracted by the main branch and the auxiliary branch are fused to obtain the RCS feature representation. A cross-scale feature correction unit is set in the dual-branch deep neural network structure. The cross-scale feature correction unit dynamically corrects the main and auxiliary branch features by calculating the cross-correlation coefficient between the feature maps of the main branch and the auxiliary branch. Based on the output of the cross-scale feature correction unit, high-frequency RCS simulation features and low-frequency RCS simulation features are generated by the main branch decoder and the auxiliary branch decoder, respectively. The cross-correlation matrix is ​​used for adaptive fusion to obtain the RCS system simulation features. The RCS feature is used as a reference feature. The simulation error between the RCS system simulation feature and the reference feature is calculated. Based on the simulation error, the network parameters are optimized to improve the simulation accuracy of the RCS system.

2. The RCS system simulation and testing method based on deep neural networks as described in claim 1, characterized in that: The target echo feature data of the RCS is collected, the original target echo data is bandpass filtered to obtain the filtered echo signal, and then resampled to obtain a discrete sampling sequence; The discrete sampling sequence is segmented and windowed using a window function. The windowed signal sequence is then subjected to a Fast Fourier Transform to obtain spectral features. Amplitude and phase spectrum information are extracted from the spectral features to form a time-series feature sequence.

3. The RCS system simulation and testing method based on deep neural networks as described in claim 1 or 2, characterized in that: The main branch of the dual-branch deep neural network structure is configured to process the temporal feature sequence, and the sub-branch of the dual-branch deep neural network structure is configured to process the frequency domain components, namely the amplitude spectrum and phase spectrum information; The main branch uses standard convolutional layers to process temporal feature sequences; Furthermore, residual connections are added every two convolutional layers to form a residual block structure; The residual block structure is a dual-path structure, including a main path and a shortcut connection path; The features processed by the main path are added element-wise with the features of the shortcut path, and the merged features are then processed by the ReLU activation function to obtain the final residual block output. Collect feature maps output from each residual block, the feature maps including shallow feature maps and deep feature maps; use an adaptive weighting mechanism to fuse the feature maps to obtain the fused feature representation; The fused feature representations are subjected to dimensionality compression to obtain a fixed-dimensional main branch feature vector F.

4. The simulation and testing method for an RCS system based on a deep neural network as described in any one of claims 1-3, characterized in that: The sub-branch receives amplitude spectrum and phase spectrum information, inputs the amplitude spectrum into the frequency domain feature extraction module composed of L layers of one-dimensional convolutional layers, and generates frequency distribution features; A compensation mechanism is introduced to the phase spectrum, and the frequency domain signal Xr is reconstructed based on the compensated phase spectrum and the original amplitude spectrum A: Where j is the imaginary unit, The phase spectrum after compensation; The reconstructed frequency domain signal is converted into a feature vector R through a feature mapping network consisting of two fully connected layers; The main branch feature vector F and the secondary branch feature vector R are concatenated along the feature dimension to form a fused feature vector; The fused feature vector is subjected to dimensionality reduction mapping to obtain the RCS feature representation Y.

5. The RCS system simulation and testing method based on deep neural networks as described in any one of claims 1-4, characterized in that: The cross-scale feature correction unit includes a cross-correlation calculation module, a weight generation module, and a feature correction module. The cross-correlation calculation module is configured to calculate the correlation matrix between two branch features, the weight generation module is configured to generate weight factors for feature channels, and the feature correction module is configured to optimize and correct the features. The correlation matrix is ​​calculated by extracting the main and auxiliary branch feature maps, adjusting them to the same feature space through projection transformation, and then calculating the cross-correlation matrix G, specifically including: The main branch feature map is the feature map extracted from each residual block. Where H and W are the height and width of the feature map, respectively, C is the number of feature channels, and L represents the Lth residual block level; The auxiliary branch feature map is obtained from the intermediate layer of the feature mapping network; FM of the main branch feature map L Rearrange and convert to matrix form. The auxiliary branch features are adjusted in dimension and normalized to be converted into a matrix form compatible with the main branch features. Calculate the cross-correlation matrix G between the features of the primary and secondary branches: in, It is a cross-correlation matrix. This indicates the direct correlation between features. The autocorrelation matrix representing the characteristics of the main branch. The autocorrelation matrix representing the characteristics of the secondary branches; The weight generation module normalizes the cross-correlation matrix G to obtain the main branch feature weight factor λ1. The weighting factor λ1 is configured to characterize the contribution of each feature channel of the main branch to the RCS feature representation; The auxiliary branch feature weight factor λ2 is obtained by column normalization; The weighting factor λ2 is configured to characterize the contribution of each feature channel of the auxiliary branch to the RCS feature representation; The feature correction module processes the main branch features using a weighting factor λ1 to obtain the corrected feature map FM. L '; The auxiliary branch features are processed using a weighting factor λ2 to obtain the corrected feature map FR'.

6. The RCS system simulation and testing method based on deep neural networks as described in any one of claims 1-5, characterized in that: Based on the corrected feature map FM L 'Reconstruct it into high-frequency RCS simulation features using a decoder;' Based on the corrected feature map FR', it is reconstructed into low-frequency RCS simulation features by the auxiliary branch decoder; The high-frequency RCS simulation features and low-frequency RCS simulation features are subjected to dimensionality unification and normalization to generate the high-frequency feature weight matrix G. h and low-frequency feature weight matrix G l ; The weight matrix is ​​multiplied element-wise with the corresponding feature map, and the weighted high-frequency and low-frequency features are added element-wise to obtain the simulation feature T of the RCS system.

7. The RCS system simulation and testing method based on deep neural networks as described in any one of claims 1-6, characterized in that: The feature reconstruction error is calculated based on the RCS system simulation features and RCS feature representation, specifically including: Calculate the feature reconstruction error E between the simulated features of the RCS system and the feature representation of the RCS system. r ; For high-frequency RCS features, calculate the degree of matching with the high-frequency components in Y represented by the RCS features; For low-frequency RCS features, evaluate the consistency with the RCS feature representation of low-frequency components in Y; Analyze the changes in the cross-correlation matrix G during feature fusion to evaluate the fusion quality; The branch feature error E is obtained by combining the matching degree of high-frequency RCS features, the consistency of low-frequency RCS features, and the fusion quality. b ; The total error is calculated based on the feature reconstruction error weights and the branch feature error weights, and is expressed as follows: AND t =μ1·E r +μ2·E b Where μ1 is the feature reconstruction error weight and μ2 is the branch feature error weight.

8. The RCS system simulation and testing method based on deep neural networks as described in any one of claims 1-7, characterized in that: Based on the changing trends of various errors, parameter settings are optimized, including: If the main branch optimization is entered, the parameters of the residual block convolutional layer are optimized first, the parameters of the adaptive weight mechanism are adjusted, the parameters of the main branch decoder are updated, and the main feature extraction capability is enhanced. If the main branch optimization is complete, then optimize the feature mapping network parameters in the auxiliary branch, update the decoder parameters of the auxiliary branch, and improve the auxiliary feature extraction capability. If the total error is lower than a preset threshold, or the error change is less than the convergence threshold for N consecutive iterations, or the maximum number of iterations I is reached... max If the validation set error does not improve for K consecutive rounds, the optimization process is terminated and the network parameter optimization is completed.

9. A simulation and testing system for an RCS system based on a deep neural network, comprising the simulation and testing method for an RCS system based on a deep neural network as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is configured to acquire RCS target echo feature data, and to segment the RCS target echo feature data according to a preset time window length to obtain a time-series feature sequence. The feature representation module constructs a dual-branch deep neural network structure based on temporal feature sequences. The main branch uses standard convolutional layers to process the temporal feature sequences, while the auxiliary branch is equipped with a spectrum reconstruction module to fuse the features extracted by the main branch and the auxiliary branch to obtain the RCS feature representation. The dynamic correction module sets up a cross-scale feature correction unit in the dual-branch deep neural network structure, and performs dynamic correction on the main and auxiliary branch features by calculating the cross-correlation coefficient between the feature maps of the main branch and the auxiliary branch. The simulation optimization module is configured to generate high-frequency RCS simulation features and low-frequency RCS simulation features respectively, and perform adaptive fusion to obtain RCS system simulation features. The RCS feature representation is used as a reference feature, the simulation error between the RCS system simulation features and the reference feature is calculated, and the network parameters are optimized based on the simulation error to improve the simulation accuracy of the RCS system.

Images