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

Abstract

The invention discloses an RCS system simulation test method and system based on a deep neural network, and the method comprises the steps: collecting RCS target echo feature data, and carrying out the segmentation of the data, and obtaining a time sequence feature sequence; a double-branch deep neural network structure is constructed, a time sequence feature sequence is processed, features extracted from main and auxiliary branches are fused, and RCS feature representation is obtained; setting a cross-scale feature correction unit in the double-branch deep neural network structure, calculating a cross correlation coefficient between the main and auxiliary branch feature maps, and performing dynamic correction on the main and auxiliary branch features; based on the output of the correction unit, high-frequency and low-frequency RCS simulation features are generated, and adaptive fusion is carried out to obtain RCS system simulation features; and calculating a simulation error between the RCS system simulation characteristic and the reference characteristic, and optimizing the network parameters. According to the invention, the adaptive capability and stability of the measurement system are significantly improved, and the method is of great significance in improving the overall level of electronic countermeasure system performance evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing and test evaluation, and in particular to a RCS system simulation test method and system based on a deep neural network. Background Art

[0002] Radar cross section (RCS) measurement is an important test method in modern electronic countermeasure systems. Its simulation accuracy and reliability directly affect the evaluation results of weapon equipment performance. With the continuous development of stealth technology of modern combat platforms, the accuracy requirements for RCS measurement systems are getting higher and higher. Existing RCS simulation test systems mainly focus on the acquisition and processing efficiency of measurement data, but there are still obvious deficiencies in the quality evaluation and dynamic optimization of measurement results.

[0003] Current RCS measurement systems usually use a single indicator to evaluate measurement quality, such as signal-to-noise ratio or measurement error, which is difficult to fully reflect the scattering characteristics of complex targets at different postures and frequencies. In particular, when dealing with multiple targets or targets with complex structures, there is a lack of a detailed evaluation mechanism for local scattering characteristics. At the same time, existing systems mostly use static measurement strategies, lacking the ability to conduct real-time evaluation and parameter adjustment during the measurement process, which can easily lead to unstable measurement accuracy or loss of local features.

[0004] More importantly, existing technologies rarely consider the interrelationships between the scattering characteristics of different parts of the target, and lack an effective monitoring mechanism for changes in data correlation during the measurement process. In the RCS measurement of complex targets, the problem of maintaining the organizational relationship between local scattering features is often ignored, which may lead to the measurement results being unable to truly reflect the scattering characteristics of the target. In addition, existing systems mostly use fixed measurement parameters, which makes it difficult to adaptively adjust the measurement strategy according to different target characteristics and measurement environments, which seriously restricts the improvement of measurement accuracy. These technical deficiencies not only affect the accuracy and reliability of RCS measurements, but also limit the accuracy of electronic countermeasure system performance evaluation. Summary of the invention

[0005] In view of the problems existing in the existing RCS system simulation test method based on deep neural network, the present invention proposes an RCS system simulation test method and system based on deep neural network.

[0006] Therefore, the problem to be solved by the present invention is to improve the reliability test of the measurement results of the RCS simulation test system.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, an embodiment of the present invention provides an RCS system simulation test method based on a deep neural network, which includes: collecting RCS target echo feature data, segmenting the RCS target echo feature data according to a preset time window length, and obtaining 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 convolution layer to process the time series feature sequence, and the auxiliary branch sets a spectrum reconstruction module, and the features extracted by the main branch and the auxiliary branch are fused to obtain an RCS feature representation; a cross-scale feature correction unit is set in the dual-branch deep neural network structure, and the cross-scale feature correction unit dynamically corrects the main and auxiliary branch features by calculating the mutual correlation coefficient between the main branch and the auxiliary branch feature graphs; based on the output of the cross-scale feature correction unit, a high-frequency RCS simulation feature and a low-frequency RCS simulation feature are generated by a main branch decoder and an auxiliary branch decoder respectively, and a mutual correlation coefficient matrix is ​​used for adaptive fusion to obtain an RCS system simulation feature; the RCS feature representation is used as a reference feature, and a simulation error between the RCS system simulation feature and the reference feature is calculated, and network parameters are optimized based on the simulation error to improve the RCS system simulation accuracy.

[0008] As a preferred solution of the RCS system simulation test method based on deep neural network described in the present invention, the RCS target echo characteristic data is collected, the target echo raw data is band-pass filtered to obtain a filtered echo signal, and complex sampling is performed to obtain a discrete sampling sequence; a window function is applied to the discrete sampling sequence for segmented windowing processing, the windowed signal sequence is fast Fourier transformed to obtain spectrum characteristics, and amplitude spectrum and phase spectrum information are extracted from the spectrum characteristics to form a time series feature sequence.

[0009] As a preferred solution of the RCS system simulation test method based on deep neural network described in the present invention, wherein: the main branch of the dual-branch deep neural network structure is used to process the time series feature sequence, and the secondary branch of the dual-branch deep neural network structure is used to process the frequency domain components, namely the amplitude spectrum and phase spectrum information; the main branch uses a standard convolution layer to process the time series feature sequence; and a residual connection is added every two convolution 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 to the features of the shortcut connection path at the element level, and the merged features are then subjected to the ReLU activation function to obtain the final residual block output; the feature maps output by each residual block are collected, and the feature maps include shallow feature maps and deep feature maps; the feature maps are feature-fused using an adaptive weight mechanism to obtain a fused feature representation; the fused feature representation is dimensionally compressed to obtain a main branch feature vector of a fixed dimension. .

[0010] As a preferred solution of the RCS system simulation test method based on deep neural network described in the present invention, wherein: the secondary 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 convolution layers, and generates frequency distribution characteristics; introduces a compensation mechanism for the phase spectrum, and based on the compensated phase spectrum and the original amplitude spectrum Reconstructing frequency domain signals : in, is an imaginary unit, is the compensated phase spectrum; the reconstructed frequency domain signal is converted into a feature vector through a feature mapping network consisting of two fully connected layers ; The main branch feature vector With branch feature vector Splice on the feature dimension to form a fused feature vector; perform dimensionality reduction mapping on the fused feature vector to obtain the RCS feature representation .

[0011] As a preferred solution of the RCS system simulation test method based on deep neural network described in the present invention, 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 used to calculate the correlation matrix between two branch features, the weight generation module is used to generate the weight factor of the feature channel, and the feature correction module is used to optimize and correct the features; the calculation of the correlation matrix is ​​to extract the main and auxiliary branch feature maps, and adjust them to the same feature space through projection transformation, and calculate the cross-correlation coefficient matrix , specifically including: the main branch feature map is extracted from each residual block feature map ,in , are the height and width of the feature map, respectively. is the number of feature channels, Indicates The auxiliary branch feature map is obtained from the middle layer of the feature mapping network. ; For the main branch feature map Rearrange and convert to matrix form ; Adjust the dimension and normalize the auxiliary branch features and convert them into a matrix form compatible with the main branch features ; Calculate the correlation coefficient matrix between the main and auxiliary branch features : in, is the mutual correlation coefficient matrix, Indicates the direct correlation between features. Represents the autocorrelation matrix of the main branch features, The autocorrelation matrix representing the auxiliary branch characteristics; the mutual correlation coefficient matrix Normalize to get the main branch feature weight factor ; Weight factor Used to characterize the contribution of each feature channel of the main branch to the RCS feature representation; the auxiliary branch feature weight factor is obtained by column normalization ; Weight factor It is used to characterize the contribution of each feature channel of the auxiliary branch to the RCS feature representation; the feature correction module converts the weight factor Applied to the main branch features to obtain the corrected feature map ; The weight factor Applied to the auxiliary branch features to obtain the corrected feature map .

[0012] As a preferred solution of the RCS system simulation test method based on deep neural network of the present invention, wherein: based on the corrected feature map Reconstruct it into a high-frequency RCS simulation feature through a decoder; based on the corrected feature map , reconstruct it into low-frequency RCS simulation features through the auxiliary branch decoder; unify the dimensions of the high-frequency RCS simulation features and the low-frequency RCS simulation features, and normalize them to generate a high-frequency feature weight matrix and the low-frequency feature weight matrix , perform element-by-element multiplication of the weight matrix and the corresponding feature map, perform element-by-element addition of the weighted high-frequency and low-frequency features, and obtain the RCS system simulation feature .

[0013] As a preferred solution of the RCS system simulation test method based on deep neural network of the present invention, wherein: based on the RCS system simulation feature and the RCS feature representation, the feature reconstruction error of the two is calculated, specifically: the feature reconstruction error between the RCS system simulation feature and the RCS feature representation is calculated ; For high frequency RCS characteristics, calculation and RCS characteristics representation Matching degree of mid- and high-frequency components; Evaluation and RCS feature representation of low-frequency RCS features The consistency of low- and medium-frequency components; analysis of the correlation coefficient matrix during feature fusion The fusion quality is evaluated by combining 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. ; The total error is calculated based on the feature reconstruction error weight and the branch feature error weight, expressed as: in, is the reconstruction error weight, is the branch feature error weight.

[0014] As a preferred solution of the RCS system simulation test method based on deep neural network described in the present invention, wherein: based on the changing trend of various errors, the parameter settings are optimized, including: if the main branch optimization is entered, the residual block convolution layer parameters are optimized first, the adaptive weight mechanism parameters are adjusted, the main branch decoder parameters are updated, and the main feature extraction capability is enhanced; if the main branch optimization is completed, the feature mapping network parameters in the auxiliary branch are optimized, the auxiliary branch decoder parameters are updated, and the auxiliary feature extraction capability is improved; if the total error is lower than the preset threshold, or the continuous The round error change is less than the convergence threshold, or the maximum number of iterations is reached , or the validation set error is continuous If there is no improvement in the first round, the optimization process is terminated and the network parameter optimization is completed.

[0015] In a second aspect, an embodiment of the present invention provides an RCS system simulation test system based on a deep neural network, which includes: a data acquisition module, which is used to collect RCS target echo feature data, and 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, which constructs a dual-branch deep neural network structure based on the time series feature sequence, the main branch uses a standard convolution layer to process the time series feature sequence, and the auxiliary branch sets 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, which sets a cross-scale feature correction unit in the dual-branch deep neural network structure, and dynamically corrects the main and auxiliary branch features by calculating the mutual correlation coefficient between the main branch and the auxiliary branch feature graphs; a simulation optimization module, which is used to generate high-frequency RCS simulation features and low-frequency RCS simulation features respectively, and perform adaptive fusion to obtain RCS system simulation features, use the RCS feature representation as a reference feature, calculate the simulation error between the RCS system simulation feature and the reference feature, optimize the network parameters based on the simulation error, and improve the RCS system simulation accuracy.

[0016] The beneficial effect of the present invention is that the present invention constructs an intelligent measurement quality assessment framework by applying deep neural network technology to the RCS simulation test system. Convolutional neural networks and recurrent neural networks are used to extract and analyze the features of the measurement data, and a multi-dimensional evaluation mechanism including average proximity, fluctuation consistency and structural correlation is established, which realizes a comprehensive evaluation of the quality of the measurement data. Based on the feature extraction capability 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 methods. At the same time, the continuous learning and optimization capability of the deep learning model enables the system to continuously improve the evaluation accuracy and robustness, providing a new technical path for the intelligent upgrade of the RCS test system, significantly improving the adaptability and stability of the measurement system, and is of great significance to improving the overall level of performance evaluation of the electronic countermeasure system. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. 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 creative work. Among them: Figure 1 Flowchart of the RCS system simulation test method based on deep neural network. DETAILED DESCRIPTION

[0018] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0019] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0021] Example 1 Reference Figure 1 , which is the first embodiment of the present invention, and provides an RCS system simulation test method based on a deep neural network, comprising: S1: Collect RCS target echo feature data, segment the RCS target echo feature data according to a preset time window length, perform fast Fourier transform on the windowed signal sequence to obtain spectrum features, and extract amplitude spectrum and phase spectrum information from the spectrum features to form a time series feature sequence.

[0022] The RCS target echo feature data is collected and processed using a multi-angle radar, including: The operating frequency is The phased array radar system is spaced at an angle of Scan the RCS target within the measurement angle range to obtain the target echo raw data, where the angle interval is Less than 1 / 30 of the measuring angle range; Perform bandpass filtering on the target echo raw data. The bandwidth of the bandpass filter is Satisfy 0.01 ≤ ≤0.1 , get the filtered echo signal; The filtered echo signal is complex sampled, and the sampling rate of the complex sampling is not less than 2 times the bandwidth , and obtain a discrete sampling sequence.

[0023] Apply window functions to discrete sampling sequences for piecewise windowing, including: The window function is selected according to the frequency characteristics of the discrete sampling sequence, and 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; The window function can smoothly transition to zero value at the time domain endpoint, suppressing the spectrum leakage effect caused by truncation and improving the accuracy of spectrum analysis; The overlap-addition method is used to segment the windowed sequence. The continuity of the signal is maintained by adjusting the length of the overlapping area. The length of the overlapping area is determined according to the time-varying characteristics of the signal. The overlap rate of adjacent windows is between 30% and 70%, and the windowed signal sequence is obtained.

[0024] Perform fast Fourier transform on the windowed signal sequence to obtain the spectrum characteristics, which are: The windowed signal sequence is padded with zeros to an integer power of 2, N sampling points, where N is greater than the number of sampling points of the windowed signal sequence; Perform butterfly operation on the zero-filled signal sequence and decompose the N-point Fourier transform into Level 2 point Fourier transform; Calculate the rotation factor of the 2-point Fourier transform at each level. The rotation factor is ,in, is the frequency serial number; According to the time extraction method, butterfly operation is performed to obtain the spectrum characteristics, which include non-redundant frequencies.

[0025] Amplitude spectrum and phase spectrum information are extracted from the frequency spectrum features to form a time series feature sequence, and the feature dimension of the time series feature sequence does not exceed half of the number of fast Fourier transform points.

[0026] Among them, the amplitude spectrum reflects the energy distribution intensity of the signal at different frequencies, and the phase spectrum reflects the relative phase relationship of each frequency component; in order to construct a time series feature sequence, the amplitude spectrum and phase spectrum information need to be combined. The combination process is to combine the amplitude value and the corresponding phase angle at each frequency point to form a complete frequency domain representation, and then convert this complete frequency domain representation back to the time domain through inverse Fourier transform. The conversion process can be expressed as: all frequency components are superimposed according to their respective amplitudes and phases, and each frequency component will contribute to the final time domain signal. Frequency components with larger amplitudes will have a more significant impact on the time domain signal, and the phase determines the precise position of these components in time.

[0027] After the conversion is completed, 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.

[0028] S2: A dual-branch deep neural network structure is constructed based on the time series feature sequence and the frequency domain component. 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. The spectrum reconstruction module reconstructs the frequency domain information by introducing a phase compensation mechanism, and the features extracted by the main branch and the auxiliary branch are fused to obtain the RCS feature representation.

[0029] By decomposing the signal in time and frequency domain, we can obtain the time series characteristic sequence and frequency domain components, namely the amplitude spectrum and phase spectrum.

[0030] Based on the decomposed time series feature sequence and frequency domain components, a dual-branch deep neural network structure is constructed to process the time domain and frequency domain information respectively, including: The main branch of building a dual-branch deep neural network structure includes: The time series feature sequence is extracted through N layers of standard one-dimensional convolutional layers, and the convolution kernel size of each convolutional layer is , the number of convolution kernels is , the step length is , filled with ; Increasing convolution kernel sizes are used from layer 1 to layer N / 2 to capture local features of different scales. The convolution kernel size is reduced from the N / 2+1th layer to the Nth layer to achieve layer-by-layer refinement of features; A batch normalization layer is connected in series after the output of each convolutional layer to eliminate data distribution offset. Specifically, the feature map output by the convolutional layer is standardized. First, the mean and variance of each batch of data are calculated, and normalized by subtracting the mean and dividing by the variance, and a learnable scaling factor is introduced. and offset factor , increasing the expressive power of the model.

[0031] Residual connections are added every two convolutional layers to form a residual block structure to alleviate the gradient vanishing problem of deep networks. The residual block adopts a dual-path structure including a main path and a shortcut connection path. On the main path, the input features first pass through the first convolution layer for feature extraction, then pass through the batch normalization layer in turn to eliminate data distribution offset, and then introduce nonlinear features through the ReLU activation function. This processed feature continues to be further extracted through the second convolution layer and passes through the batch normalization layer again. At the same time, the input features are directly passed to the end of the residual block through the shortcut connection path. When the dimensions of the input features do not match the output features of the main path, the dimensions of the shortcut connection features need to be adjusted through a 1×1 convolution layer. The features processed by the main path are added to the features of the shortcut connection path at the element level, and the merged features are then activated by the ReLU function to obtain the final residual block output.

[0032] New feature transformations can be learned through the main path, and the original feature information can be retained through the shortcut connection, so that gradient information can be better transmitted in the deep network, effectively alleviating the gradient disappearance problem. The dual convolution structure on the main path enhances the feature extraction capability, while the shortcut connection ensures the effective transmission of information. The combination of the two improves the learning ability and training stability of the entire network.

[0033] The feature maps output by each residual block are collected. The feature maps contain feature information at different levels. Shallow feature maps mainly contain local detail information, while deep feature maps contain more high-level semantic information. In order to make full use of multi-level features, an adaptive weight mechanism is used for feature fusion: the importance weights are calculated for the feature maps of each level, and the importance weights are automatically adjusted through learnable parameters; the features of different levels are weighted according to the calculated importance weights, so that more important features obtain higher weight values; the weighted feature maps are added element-wise to obtain the fused feature representation.

[0034] In order to obtain a fixed-dimensional feature vector, the fused feature representation is dimensionally compressed. Specifically, the time series dimension is compressed through a global pooling operation, and the entire time series feature is compressed into a single feature value. The global pooling process can perform maximum pooling and average pooling operations at the same time, and adaptively fuse the two pooling results through learnable parameters to retain the most representative feature information. Finally, a fixed-dimensional main branch feature vector is obtained. , It not only retains multi-level feature information, but also has a fixed dimension.

[0035] The auxiliary branches of the dual-branch deep neural network structure are constructed, including: Receive the output amplitude spectrum and phase spectrum, input the amplitude spectrum into the frequency domain feature extraction module composed of L layers of one-dimensional convolutional layers, and generate frequency distribution features; A compensation mechanism is introduced for the phase spectrum. The compensated phase spectrum satisfy: in is the original phase spectrum, is the compensation amplitude coefficient, is the compensation frequency coefficient, To compensate for the phase shift, , , These are all learnable parameters.

[0036] Based on the compensated phase spectrum and original amplitude spectrum Reconstructing frequency domain signals : in, is an imaginary unit, and the reconstructed signal enhances the expression of phase characteristics.

[0037] The reconstructed frequency domain signal is converted into a feature vector through a feature mapping network consisting of two fully connected layers. , the output dimension of the feature mapping network is the same as the main branch feature vector The dimensions are the same, including: The reconstructed frequency domain signal passes through a feature mapping network consisting of two fully connected layers. First, the reconstructed frequency domain signal enters the first fully connected layer as input. The number of neurons in this layer is greater than the input dimension, which plays a role in feature expansion. The first fully connected layer is followed by a BatchNorm normalization layer and a ReLU activation function, which introduces nonlinear transformation capabilities while maintaining the stability of data distribution.

[0038] Then, the output features of the first layer are passed to the second fully connected layer, where the number of neurons is precisely set to the same as the main branch feature vector The second fully connected layer is also equipped with a BatchNorm normalization layer, but the Tanh activation function is used instead of ReLU at the end, because the output range of the Tanh function is between [-1, 1], which helps the normalized representation of the features, making the generated feature vector With the main branch feature vector have similar value distribution ranges.

[0039] The main branch feature vector With branch feature vector Concatenate on the feature dimension to form a fused feature vector.

[0040] The fused feature vector is mapped to a dimension reduction layer through a fully connected layer to obtain the final RCS feature representation. , expressed as: in is the weight matrix, is the bias vector, Represents a feature concatenation operation.

[0041] S3: 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 mutual correlation coefficient between the main branch and auxiliary branch feature maps.

[0042] 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 used to calculate the correlation matrix between two branch features, the weight generation module is used to generate the weight factor of the feature channel, and the feature correction module is used to perform weighted optimization on the features.

[0043] The cross-correlation calculation module calculates the correlation matrix between two branch features, including: Extract the main and auxiliary branch feature maps: Main branch: extract feature maps from each residual block ,in , are the height and width of the feature map, respectively. is the number of feature channels, Indicates The features of each residual block retain the temporal dynamic information of different scales; Auxiliary branch: Get feature maps from the middle layer of the feature mapping network , which contains the frequency domain information after phase compensation.

[0044] Since the main branch features come from the residual block and the auxiliary branch features come from the feature mapping network, the feature representations of the two are different and feature alignment is required; the main branch feature map And auxiliary branch feature map Adjust to the same feature space through projection transformation to ensure feature dimension matching. Specifically: Feature map of the main branch Rearrange and convert to matrix form ; The auxiliary branch features are dimensionally adjusted and normalized to be converted into a matrix form compatible with the main branch features. ; The projection transformation ensures that the features of the two branches are represented in the same feature space.

[0045] Calculate the correlation coefficient matrix between the main and auxiliary branch features : in, is the mutual correlation coefficient matrix, Indicates the direct correlation between features. Represents the autocorrelation matrix of the main branch features, Represents the autocorrelation matrix of auxiliary branch features.

[0046] Cross-correlation matrix Elements in Indicates The main branch feature channel and the The correlation strength of the auxiliary branch feature channels.

[0047] Each row of the matrix represents the correlation distribution between a main branch channel and all auxiliary branch channels; Each column of the matrix represents the correlation distribution between an auxiliary branch channel and all main branch channels; By using the mutual correlation matrix Normalize to get the main branch feature weight factor ; Weight Factor Used to characterize the contribution of each feature channel of the main branch to the RCS feature representation; Obtain auxiliary branch feature weight factors through column normalization ; Weight Factor It is used to characterize the contribution of each feature channel of the auxiliary branch to the RCS feature representation.

[0048] The weight factor Applied to the main branch features to obtain the corrected feature map The correction process is to process the weight factor through the sigmoid function , map it to the [0,1] interval, and then compare it with the feature map The correction is performed by multiplication, and the correction process ensures that the key information in the time domain features is highlighted and the secondary features are suppressed.

[0049] The weight factor Applied to the auxiliary branch features to obtain the corrected feature map The correction process also uses the sigmoid function to process the weight factor After that, with the feature map The correction is performed by multiplication, and the correction process ensures the synergy between frequency domain features and time domain features.

[0050] 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 through the main branch decoder and the auxiliary branch decoder respectively, and the cross-correlation coefficient matrix is ​​used for adaptive fusion to obtain the complete RCS system simulation features.

[0051] Corrected feature map It contains the key time domain information of the target, but due to the previous downsampling and convolution operations, the spatial resolution is low. It needs to be reconstructed into high-frequency RCS simulation features through a decoder.

[0052] The decoding process uses a multi-layer deconvolution structure to gradually improve the spatial resolution of the feature map through transposed convolution operations.

[0053] Specifically, First, the first deconvolution layer is input, and the features are mapped to a higher resolution space through the transposed convolution operation. The transposed convolution expands the size of the feature map by inserting zero values ​​between the feature maps and performing convolution operations. At the same time, each deconvolution layer is followed by a nonlinear activation function to enhance the expressiveness of the features. In order to avoid information loss during feature transfer, residual connections are added between adjacent deconvolution layers to retain the original feature information.

[0054] Since each layer of deconvolution can learn feature patterns of different scales, the shallower layers mainly reconstruct the basic structure, while the deeper layers can restore finer local details. As the features are transmitted layer by layer in the multi-layer deconvolution structure, the spatial resolution is continuously improved, and the detailed information of the features is gradually restored. The high-frequency RCS simulation features of the final output not only maintain The key time domain information in the image also contains rich high-frequency details.

[0055] Similarly, after obtaining the corrected auxiliary branch features Afterwards, it is reconstructed into low-frequency RCS simulation features through the auxiliary branch decoder. Since low-frequency features reflect the overall structural characteristics of the target, the decoding process needs to pay attention to the global consistency of the features.

[0056] The auxiliary branch decoder also uses a multi-layer deconvolution structure, but it is different from the main branch decoder. First, After the first layer of deconvolution, a larger convolution kernel is used for feature mapping, so that each output position can obtain a wider range of input information, which is conducive to maintaining the integrity of the features. At the same time, a global average pooling operation is introduced between the deconvolution layers to extract and fuse global feature information and enhance the overall expression ability of the features.

[0057] In the process of feature transfer, the spatial resolution of the feature map is improved through layer-by-layer deconvolution operations. A normalization layer is added after each layer of deconvolution to stabilize the feature distribution and ensure the stability of the reconstruction process. In addition, in order to better maintain the continuity of low-frequency features, jump connections are used between adjacent layers to fuse the basic features of the low layer with the semantic features of the high layer, and output low-frequency RCS simulation features that maintain the main structural information and overall contour features of the target.

[0058] The high-frequency RCS simulation features and the low-frequency RCS simulation features are dimensionally unified so that they have the same spatial resolution and number of feature channels. The Softmax function is used to transform the correlation coefficient matrix Normalize the elements in and map the correlation coefficient to the [0,1] interval to generate a high-frequency feature weight matrix and the low-frequency feature weight matrix , and ensure that the weight matrix satisfies Constraints of The weight matrix is ​​multiplied element by element with the corresponding feature map, and the weighted high-frequency and low-frequency features are added element by element to obtain the complete RCS system simulation features. , expressed as: in, Represents high frequency features, represents low-frequency features, and ⊙ represents element-by-element multiplication operation.

[0059] S5. Use the RCS feature representation as a reference feature, calculate the simulation error between the complete RCS system simulation feature and the reference feature, optimize the network parameters based on the simulation error, and improve the RCS system simulation accuracy.

[0060] Based on RCS system simulation features and RCS feature representation Calculate the feature reconstruction error of the two, specifically: First calculate the simulation characteristics Represented with RCS features The Euclidean distance error between , evaluate the overall reconstruction effect, including: Set the mean Euclidean distance between samples of different categories in the labeled dataset , is the preset convergence threshold, and < , like < , the reconstructed features have achieved matching with the reference features, and the local feature optimization stage is entered; like ≤ < , then the reconstructed features only obtain the target main features, and continue to optimize the feature extraction network; like ≥ , then there is an essential difference between the reconstructed features and the reference features, and the feature extraction scheme is reconstructed.

[0061] The cosine similarity is further used to analyze the directional consistency of the two feature vectors to ensure feature space alignment, including: If the cosine similarity is close to 1, it means that the reconstructed feature and the original feature have a high degree of orientation consistency, confirming that the feature space is completely aligned; If the cosine similarity is at a medium level, it means that the feature vector directions are basically aligned and need to be fine-tuned and optimized; If the cosine similarity is close to 0, it means that there is a significant deviation in the direction of the feature vector and the reconstruction strategy needs to be readjusted.

[0062] Based on feature structure similarity analysis, the fidelity of the reconstructed features in the local area is evaluated, including: If the difference in regional average values ​​is small, it means that the local feature intensities match, ensuring that the local feature levels are consistent; If the regional fluctuations are similar, it means that the local change patterns are consistent, confirming that the local dynamic characteristics are maintained; If the regional correlation is strong, it means that the local structural relationship remains stable, verifying the integrity of the detailed features.

[0063] The above three errors are normalized to obtain the standardized feature reconstruction error .

[0064] For the high-frequency RCS features generated by the main branch, calculation and RCS feature representation The matching degree of mid-high frequency components, including: High frequency RCS characteristics and RCS characteristics representation Perform frequency domain transformation, extract the amplitude distribution of the high-frequency band, obtain high-frequency phase information, obtain the high-frequency characteristic component to be evaluated and the reference high-frequency component, evaluate the error degree of the high-frequency characteristic component and the reference high-frequency component in amplitude difference, phase offset and spectrum structure, and obtain the matching degree of high-frequency RCS.

[0065] Evaluation and RCS feature representation of low-frequency RCS features generated by auxiliary branches The consistency of mid- and low-frequency components, including: For low-frequency RCS characteristics and RCS characteristics Perform frequency domain transformation, extract the amplitude distribution of the low-frequency band, obtain low-frequency phase information, obtain the low-frequency characteristic component to be evaluated and the reference low-frequency component, evaluate the consistency deviation of the low-frequency characteristic component and the reference low-frequency component in terms of baseline offset, trend change consistency and energy distribution characteristics, and evaluate the consistency of the low-frequency RCS characteristics.

[0066] Analyze the correlation coefficient matrix during feature fusion changes to evaluate the fusion quality, including: Stability analysis of diagonal elements: By comparing the changes in the autocorrelation intensity of the main features at consecutive time steps, when the change amplitude is continuously less than the preset threshold, it indicates that the self-expression of the features in each dimension tends to be stable; Convergence analysis of non-diagonal elements: monitor the degree of interaction between different features. When the rate of change of interaction strength decreases and remains at a low level, it means that the information exchange between features has reached a balance and feature fusion has entered a stable stage. Matrix structure consistency analysis: Comparing the overall correlation structure of adjacent time steps, when the structural difference of the matrix continues to decrease and remains at a low level, it indicates that the feature fusion network has reached a coordinated and consistent state and the fusion quality is in the optimal range.

[0067] The standardized branch feature error is obtained by combining the matching degree of high-frequency RCS features, the consistency of low-frequency RCS features and the fusion quality. .

[0068] Define adaptive weight coefficients to dynamically balance the contribution of feature reconstruction error and branch feature error. The weight coefficients are automatically adjusted according to the current training stage and error distribution, and the feature reconstruction error weight gradually increases as training progresses to ensure the final reconstruction quality. The branch feature error weight is larger in the early stages of training to ensure accurate extraction of basic features.

[0069] The total error is calculated based on the feature reconstruction error weight and the branch feature error weight, expressed as: in, is the feature reconstruction error weight, is the branch feature error weight.

[0070] The changing trends of various errors in each round of iteration are recorded. According to the error analysis results, the focus of the next round of optimization is automatically adjusted. When a certain type of error is significantly higher than other errors, its weight coefficient is increased accordingly, and an error change curve is established to guide the dynamic adjustment of the optimization process.

[0071] Based on the changing trends of various errors, optimize parameter settings, including: If you start network training, set the initial learning rate and the maximum number of iterations , thereby initializing the Adam optimizer and determining the learning rate decay strategy.

[0072] If entering the main branch optimization, the residual block convolution layer parameters are optimized first, the adaptive weight mechanism parameters are adjusted, the main branch decoder parameters are updated, and the main feature extraction capability is enhanced; If the optimization of the main branch is completed, optimize the feature mapping network parameters in the auxiliary branch, update the auxiliary branch decoder parameters, and improve the auxiliary feature extraction capability; If the optimization of the main and auxiliary branches is completed, the feature correction unit is optimized, and the mutual correlation coefficient matrix is ​​calculated and optimized. , dynamically adjust the feature weight factors to achieve precise control of feature fusion.

[0073] If each round of iteration is completed, the current overall 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; If the total error is lower than the preset threshold, or continuous The round error change is less than the convergence threshold, or the maximum number of iterations is reached , or the validation set error is continuous If there is no improvement in the first round, the optimization process is terminated and the network parameter optimization is completed.

[0074] In summary, the present invention constructs an intelligent measurement quality assessment framework by applying deep neural network technology to the RCS simulation test system. Convolutional neural networks and recurrent neural networks are used to extract and analyze the features of the measurement data, and a multi-dimensional evaluation mechanism including average proximity, fluctuation consistency and structural correlation is established, which realizes a comprehensive evaluation of the quality of the measurement data. Based on the feature extraction capability of the deep learning model, the system can accurately identify abnormal patterns in the measurement results and adaptively adjust the measurement strategy through the reinforcement learning method. At the same time, the continuous learning and optimization capability of the deep learning model enables the system to continuously improve the evaluation accuracy and robustness, providing a new technical path for the intelligent upgrade of the RCS test system, significantly improving the adaptability and stability of the measurement system, and is of great significance to improving the overall level of performance evaluation of the electronic countermeasure system.

[0075] Example 2 On the basis of the first embodiment, this embodiment further provides an RCS system simulation test system based on a deep neural network, including: A data acquisition module is used to collect RCS target echo characteristic data, and segment the RCS target echo characteristic data according to a preset time window length to obtain a time series characteristic sequence; The feature representation module builds a dual-branch deep neural network structure based on the time series feature sequence. The main branch uses a standard convolutional layer to process the time series feature sequence, and the auxiliary branch sets 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 a cross-scale feature correction unit in the dual-branch deep neural network structure, and dynamically corrects the main and auxiliary branch features by calculating the mutual correlation coefficient between the main branch and auxiliary branch feature maps; The simulation optimization module is used 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 RCS system simulation features, use the RCS feature representation as a reference feature, calculate the simulation error between the RCS system simulation feature and the reference feature, optimize the network parameters based on the simulation error, and improve the RCS system simulation accuracy.

[0076] This embodiment also provides a computer device, which is suitable for the RCS system simulation test method based on a deep neural network, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to implement the RCS system simulation test method based on a deep neural network proposed in the above embodiment.

[0077] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0078] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the RCS system simulation test method based on a deep neural network proposed in the above embodiment is implemented.

[0079] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0080] Example 3 This embodiment provides an RCS system simulation test method based on a deep neural network. In order to verify the beneficial effects of the present invention, a scientific demonstration is carried out through simulation experiments.

[0081] In order to verify the effectiveness of the RCS measurement quality assessment method based on deep neural network in the present invention, a certain fighter model was selected as the test object for simulation experiments. The model contains multiple typical scattering points such as the nose, wing, and vertical tail, and collects RCS data at intervals of 2 degrees within the azimuth range of 0-360 degrees. The experiment was tested in a standard environment and an interference environment, respectively. The interference environment simulated the environmental interference in the actual measurement by superimposing Gaussian white noise with a signal-to-noise ratio of 5dB.

[0082] In the standard environment test, the RCS measurement results of the nose (0 degree azimuth) were analyzed, and the evaluation accuracy of the method of the present invention reached 97.3%, while the traditional statistical method was 89.1%. In the complex scattering area where the wing and the fuselage are connected (45 degree azimuth), the deviation between the evaluation results of the method of the present invention and the theoretical value is kept within 3.2%, which is significantly better than the 8.7% deviation of the traditional method. When the measurement anomaly is artificially introduced at the 90 degree azimuth, the method of the present invention can detect and locate the problem within 0.5 seconds, and the measurement stability after the adaptive adjustment of the system parameters is improved by 85.4%.

[0083] In a test environment with noise interference, the evaluation results of the method of the present invention were analyzed in the range of 180-270 degrees azimuth, and the consistency reached 94.2%, while the traditional method dropped to 76.8%. In particular, in the range of 225-235 degrees where the target posture changes rapidly, the parameter optimization strategy based on deep learning enables the adaptive adjustment of the signal processing parameters of the system, reducing the impact of measurement noise by 72.5%. Through statistical analysis of 50 repeated tests, the standard deviation of the evaluation accuracy of the method of the present invention in complex environments is 1.8%, which verifies that the method has excellent stability and robustness.

[0084] The above experimental data fully demonstrate the significant advantages of the present invention in terms of measurement quality assessment, anomaly detection and environmental adaptability, and provide reliable technical support for the intelligent upgrade of the RCS test system.

[0085] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A RCS system simulation test method based on deep neural network, characterized in that: include: Collecting RCS target echo characteristic data, segmenting the RCS target echo characteristic data according to a preset time window length to obtain a time series characteristic sequence; A dual-branch deep neural network structure is constructed 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 sets a spectrum reconstruction module to fuse the features extracted by the main branch and the auxiliary branch to obtain an RCS feature representation; A cross-scale feature correction unit is provided in the dual-branch deep neural network structure, and the cross-scale feature correction unit dynamically corrects the main and auxiliary branch features by calculating the mutual correlation coefficient between the main branch and the auxiliary branch feature graphs; Based on the output of the cross-scale feature correction unit, a high-frequency RCS simulation feature and a low-frequency RCS simulation feature are generated through a main branch decoder and an auxiliary branch decoder respectively, and a mutual correlation coefficient matrix is ​​used for adaptive fusion to obtain an RCS system simulation feature; The RCS feature representation is used as a reference feature, a simulation error between an RCS system simulation feature and the reference feature is calculated, and network parameters are optimized based on the simulation error to improve the simulation accuracy of the RCS system.

2. The RCS system simulation test method based on deep neural network according to claim 1, characterized in that: Collecting the RCS target echo characteristic data, performing bandpass filtering on the target echo raw data to obtain a filtered echo signal, and performing complex sampling to obtain a discrete sampling sequence; The discrete sampling sequence is subjected to segmented windowing processing by applying a window function, and the windowed signal sequence is subjected to fast Fourier transform to obtain spectrum features, and amplitude spectrum and phase spectrum information are extracted from the spectrum features to form a time series feature sequence.

3. The RCS system simulation test method based on deep neural network according to claim 2, characterized in that: The main branch of the dual-branch deep neural network structure is used to process the time series feature sequence, and the secondary branch of the dual-branch deep neural network structure is used to process the frequency domain component, which is the amplitude spectrum and phase spectrum information; The main branch uses a standard convolutional layer to process the temporal feature sequence; And a residual connection is 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 to the features of the shortcut connection path at the element level, and the combined features are then activated by the ReLU function to obtain the final residual block output; Collecting feature maps output by each residual block, wherein the feature maps include shallow feature maps and deep feature maps; using an adaptive weight mechanism to perform feature fusion on the feature maps to obtain a fused feature representation; The fused feature representation is dimensionally compressed to obtain a main branch feature vector of fixed dimension. .

4. The RCS system simulation test method based on deep neural network according to claim 3, characterized in that: The secondary branch receives the 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, and generates frequency distribution features; A compensation mechanism is introduced for the phase spectrum, based on the compensated phase spectrum and the original amplitude spectrum. Reconstructing frequency domain signals : in, is an imaginary unit, is the phase spectrum after compensation; The reconstructed frequency domain signal is converted into a feature vector through a feature mapping network consisting of two fully connected layers. ; The main branch feature vector and the secondary branch eigenvector Splice on the feature dimension to form a fused feature vector; Perform dimensionality reduction mapping on the fused feature vector to obtain RCS feature representation .

5. The RCS system simulation test method based on deep neural network according to claim 1, characterized in that: The cross-scale feature correction unit includes a cross-correlation calculation module, a weight generation module and a feature correction module, wherein the cross-correlation calculation module is used to calculate the correlation matrix between two branch features, the weight generation module is used to generate the weight factor of the feature channel, and the feature correction module is used 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 mutual correlation coefficient matrix , specifically including: The main branch feature map is extracted from each residual block. ,in , are the height and width of the feature map, respectively. is the number of feature channels, Indicates residual block level; The auxiliary branch feature map is obtained from the middle layer of the feature mapping network. ; Feature map of the main branch Rearrange and convert to matrix form ; The auxiliary branch features are dimensionally adjusted and normalized to be converted into a matrix form compatible with the main branch features. ; Calculate the correlation coefficient matrix between the main and auxiliary branch features : in, is the mutual correlation coefficient matrix, Indicates the direct correlation between features. Represents the autocorrelation matrix of the main branch features, The autocorrelation matrix representing the auxiliary branch characteristics; The weight generation module is a matrix of the mutual correlation coefficients. Normalize to get the main branch feature weight factor ; Weight Factor Used to characterize the contribution of each feature channel of the main branch to the RCS feature representation; Obtain auxiliary branch feature weight factors through column normalization ; Weight Factor Used to characterize the contribution of each feature channel of the auxiliary branch to the RCS feature representation; The feature correction module converts the weight factor Applied to the main branch features to obtain the corrected feature map ; The weight factor Applied to auxiliary branch features to obtain the corrected feature map .

6. The RCS system simulation test method based on deep neural network according to claim 5, characterized in that: Based on the corrected feature map Reconstruct it into high-frequency RCS simulation characteristics through the decoder; Based on the corrected feature map , reconstructed into low-frequency RCS simulation features through the auxiliary branch decoder; The high-frequency RCS simulation features and the low-frequency RCS simulation features are dimensionally unified and normalized to generate a high-frequency feature weight matrix and the low-frequency feature weight matrix ; The weight matrix is ​​multiplied element by element with the corresponding feature map, and the weighted high-frequency and low-frequency features are added element by element to obtain the RCS system simulation features. .

7. The RCS system simulation test method based on deep neural network according to claim 6, characterized in that: The feature reconstruction errors of the two are calculated based on the RCS system simulation features and the RCS feature representation, specifically including: Calculate the feature reconstruction error between the RCS system simulation features and the RCS feature representation ; For high frequency RCS characteristics, calculation and RCS characteristics representation The matching degree of mid-high frequency components; Evaluation and RCS feature representation of low frequency RCS features The consistency of mid- and low-frequency components; Analyze the correlation coefficient matrix during feature fusion changes, evaluate the fusion quality; The branch feature error is obtained by combining the matching degree of high-frequency RCS features, the consistency of low-frequency RCS features and the fusion quality. ; The total error is calculated based on the feature reconstruction error weight and the branch feature error weight, expressed as: in, is the reconstruction error weight, is the branch feature error weight.

8. The RCS system simulation test method based on deep neural network according to claim 7, characterized in that: Based on the changing trends of various errors, optimize parameter settings, including: If entering the main branch optimization, the residual block convolution layer parameters are optimized first, the adaptive weight mechanism parameters are adjusted, the main branch decoder parameters are updated, and the main feature extraction capability is enhanced; If the optimization of the main branch is completed, the feature mapping network parameters in the auxiliary branch are optimized, the auxiliary branch decoder parameters are updated, and the auxiliary feature extraction capability is improved; If the total error is lower than the preset threshold, or continuous The round error change is less than the convergence threshold, or the maximum number of iterations is reached , or the validation set error is continuous If there is no improvement in the first round, the optimization process is terminated and the network parameter optimization is completed.

9. A RCS system simulation test system based on a deep neural network, based on the RCS system simulation test method based on a deep neural network according to any one of claims 1 to 8, characterized in that: include: A data acquisition module is used to collect RCS target echo characteristic data, and segment the RCS target echo characteristic data according to a preset time window length to obtain a time series characteristic sequence; The feature representation module builds a dual-branch deep neural network structure based on the time series feature sequence. The main branch uses a standard convolutional layer to process the time series feature sequence, and the auxiliary branch sets 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 a cross-scale feature correction unit in the dual-branch deep neural network structure, and dynamically corrects the main and auxiliary branch features by calculating the mutual correlation coefficient between the main branch and auxiliary branch feature maps; The simulation optimization module is used to generate high-frequency RCS simulation features and low-frequency RCS simulation features respectively, and perform adaptive fusion to obtain RCS system simulation features, use the RCS feature representation as a reference feature, calculate the simulation error between the RCS system simulation feature and the reference feature, optimize the network parameters based on the simulation error, and improve the RCS system simulation accuracy.

Images