Intelligent sea clutter data sample generation method based on wide area feature extraction

Abstract

The invention discloses a sea clutter data sample intelligent generation method based on wide-area feature extraction, and the method comprises the steps: processing sea clutter data through a sliding window mechanism based on the space-time correlation characteristics of sea clutter, and carrying out the three-channel RGB image sample construction of one-dimensional single-range gate multi-pulse sea clutter data; extracting refined features for wide-area energy distribution of sea clutters, constructing a sea clutter feature extraction module, and embedding the sea clutter feature extraction module into a sea clutter modeling simulation network; a ViT-based generator is improved, a sea clutter transient global information and local detail fusion module and a space-frequency dual-path information extraction module are proposed and introduced, extracted sea clutter refined features are input into the generator as prior information, and a joint loss function constraint is designed to guide network training. And taking the trained generator as a sea clutter generation model until both the generator and the discriminator reach a convergence state. The data finally generated by the method provided by the invention meets the requirement of a given threshold index in the aspect of a multi-domain fidelity index.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning, and specifically is a method for intelligently generating sea clutter data samples based on wide-area feature extraction. Background Art

[0002] Sea clutter is the radar echo reflected by the sea surface, and its intensity is affected by multiple environmental factors such as climatic conditions and wind and waves. In a complex ocean environment, the characteristics of sea clutter are unpredictable, which poses a great challenge to the detection of sea surface targets. High-quality sea clutter data is crucial to studying sea clutter characteristics and improving target detection algorithms. Traditional sea clutter echo data simulation usually uses classical statistical models to model the amplitude statistical characteristics of sea clutter, and then generates specific distributions through zero memory nonlinear transformation method (ZMNL) and spherically invariant random process method (SIRP). However, these traditional models are too simple to fully simulate the non-Gaussian and non-stationary characteristics of sea clutter, resulting in low accuracy and poor applicability of the generated data, which limits the application effect of the model in real scenarios. Therefore, obtaining high-quality and diversified sea clutter data has become an important task in current research.

[0003] In view of the limitations of traditional methods, researchers have gradually turned their attention to generative adversarial networks (GANs) to generate radar data. Through adversarial training, GANs have achieved remarkable results in the fields of image and audio generation, and have shown certain potential in the generation of radar data. Although GANs have improved the defects of traditional methods to a certain extent, their structure is usually based on convolutional neural networks (CNNs). This type of architecture still has certain limitations in processing the spatiotemporal characteristics of sea clutter data and cannot fully meet the needs of generating high-precision sea clutter sample data. In the patent "A method for generating sea clutter data samples based on spatiotemporal joint perception", although the sea clutter generation accuracy is improved by enhancing the spatiotemporal joint perception of sea clutter, the evaluation criteria for sea clutter are still insufficient. It is necessary to test the multi-dimensional and multi-angle characteristics of simulated sea clutter to illustrate the authenticity of the generated sea clutter data. Summary of the invention

[0004] The present invention aims to propose an intelligent generation method for sea clutter data samples based on wide-area feature extraction. The method utilizes the spatiotemporal characteristics of sea clutter, removes target interference, and performs amplitude characteristic analysis, spectrum analysis, and time-frequency analysis. The method combines a generative adversarial network (GAN) to simulate and generate sea clutter data, so as to solve the problem that the traditional method does not accurately grasp the multi-domain characteristics of sea clutter and the data accuracy is insufficient when generating sea clutter data.

[0005] The technical solution to achieve the purpose of the present invention is: a method for intelligently generating sea clutter data samples based on wide-area feature extraction, which specifically includes the following steps:

[0006] Step 1: Obtain the original sea clutter data, extract the pure sea clutter data from the original sea clutter data, and normalize the echoes within the same distance unit;

[0007] Step 2: Obtain independent echo samples of sea clutter using a sliding window mechanism based on temporal and spatial characteristics;

[0008] The amplitude characteristics of the independent echo samples of sea clutter are analyzed, the probability density distribution parameters of the sea clutter data are estimated, and the optimal statistical distribution model is fitted to extract the amplitude distribution type and amplitude distribution parameters;

[0009] Perform power spectrum and time-frequency domain analysis on the independent echo samples of sea clutter to obtain the Doppler shift and Doppler broadening in the power spectrum of sea clutter;

[0010] Step 3: Divide the multi-pulse multi-range gate sea clutter data into multi-pulse one-dimensional samples according to single-range units, and use the multi-pulse one-dimensional samples to construct three-channel RGB image samples;

[0011] Step 4: Use the three-channel RGB image samples and the corresponding amplitude distribution type and amplitude distribution parameter label information to train a refined feature extraction network for wide-area energy distribution of sea clutter. After the model converges and the accuracy is improved, a pre-training model is formed. The three-channel RGB image samples are input into the pre-training model to obtain wide-area refined features.

[0012] Step 5: Building a generative adversarial network (GAN) model based on the improved ViT, wherein the generative adversarial network model based on the improved ViT includes a generator and a discriminator, wherein the generator is used to generate simulated sea clutter data, and the discriminator is used to judge the authenticity, amplitude characteristics, and energy spectral density of the generated data;

[0013] Step 6: The wide-area refined features obtained in step 4 are used as the prior information input of the generator, and the independent echo samples of sea clutter in step 2 are processed into real numbers according to the real and imaginary parts respectively, and multi-pulse multi-range gate real and imaginary dual-channel samples are constructed and used as training sets, which are input into the generator and discriminator for collaborative training;

[0014] Step 7: Generate sea clutter data samples using the trained generator.

[0015] Compared with the prior art, the main advantages of the present invention are: in view of the problems that the characteristics of sea clutter data are complex, the sample size is limited, and the multi-domain characteristics are difficult to accurately capture, a generative adversarial model based on improved ViT is proposed. Through this method, sea clutter data with high fidelity can be efficiently simulated. In addition, this technology only requires a small amount of real data for large-scale generation, and has the characteristics of efficient training, high generation accuracy and strong model transferability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0017] Figure 2 This is the data sample construction method of the present invention.

[0018] Figure 3 This is the feature extraction data sample construction method of the present invention.

[0019] Figure 4 It is a structural diagram of the sea clutter feature extraction module of the present invention.

[0020] Figure 5 It is the structural diagram of the generator and discriminator of the present invention.

[0021] Figure 6 It is a structural diagram of the transient global-local information fusion module of the present invention.

[0022] Figure 7 It is a structural diagram of the space-frequency dual-path information extraction module of the present invention.

[0023] Figure 8 It is a training process diagram of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0025] The present invention provides a method for intelligently generating sea clutter data samples based on wide-area feature extraction, and the specific steps include:

[0026] Step 1: Obtain original sea clutter data, and extract pure sea clutter data from the original sea clutter data; the specific method of extracting pure sea clutter data from the original sea clutter data is:

[0027] The nearshore, farshore, target units (such as sea peaks and specific targets on the sea surface) and other abnormal unit data are removed to obtain pure sea clutter data. The abnormal units include nearshore area distance units, target distance units, and farshore distance units that are seriously affected by the marine terrain environment.

[0028] Some radar echoes in the existing data sets contain not only clutter but also target echoes, such as channel buoys, non-cooperative ships, and distant islands. In order to reduce the interference of other targets on the training sea clutter, non-sea clutter echoes must be screened out. Since the coastal terrain near the coast has a greater impact on sea clutter, the received sea clutter echoes are irregularly distributed, and the sea clutter distance unit collected at the far coast is too far away, and the impact of sea clutter on the echo is too small, resulting in the sea clutter signal being submerged by other noise signals, and the frequency domain appears as white noise. In this way, the echo signal unit of the distance unit also needs to be eliminated. The amplitude intensity of sea clutter data is also affected by distance. In order to eliminate the influence of different amplitude sizes of different distance units, different pulse echoes in the same distance unit must be de-meaned and normalized.

[0029] Step 2: Analyze the time-space correlation characteristics of the normalized sea clutter data, extract the decorrelation time p and decorrelation distance r, and determine the sliding window size as [p, r], and use the sliding window mechanism to obtain independent echo samples of sea clutter. The specific method is:

[0030] The sample extraction method is as follows Figure 2 As shown, the space-time sliding window is used to extract sea clutter independent echo samples from the pure sea clutter data, and each time a matrix of size (P, R) is taken out as the sea clutter independent echo sample. Where P is the number of continuous pulses and P = m 2 ,in Here n is 2, ceil is rounded down, R is the number of distance units and R = lr, where l is 1. The sliding step size of the sliding window in the pulse dimension Where floor is rounded up. The sliding step length Step in the distance dimension R =R / / 2.

[0031] The amplitude characteristics of the independent echo samples of sea clutter with size (P, R) are analyzed according to a single range unit, the probability density distribution parameters of the sea clutter data are estimated, and the optimal statistical distribution model is fitted to extract the amplitude distribution type and amplitude distribution parameters;

[0032] The power spectrum and time-frequency domain analysis of the sea clutter independent echo samples of size (P, R) are performed according to a single distance unit, the spectrum characteristics are analyzed, and the Doppler shift and Doppler broadening are extracted. The specific method is as follows:

[0033] The periodogram method is used to analyze the spectral characteristics of pure sea clutter data and obtain Doppler broadening and Doppler shift. First, the autocorrelation function is calculated: for the complex signal x[n] = x I [n]+jx Q [n], where x I [n] is the real part, x Q[n] is the imaginary part, and the complex autocorrelation function R[k] is defined as:

[0034] R[k]=∑ n x[n]·x * [n+k]

[0035] Here x * The conjugate complex number represented by [n], that is, x * [n] = x I [n]-jx Q [n]. The conjugate is used when computing the autocorrelation to ensure that the result is still a representation of the energy.

[0036] Secondly, applying the discrete Fourier transform (DFT) to the autocorrelation function R[k], we can get the power spectrum of the complex signal:

[0037]

[0038] Where N is the total number of samples and f is the frequency index.

[0039] Finally, the frequency shift and spread in the power spectrum of sea clutter can be obtained through the power spectrum. Frequency shift is the center frequency position in the power spectrum; spread refers to the phenomenon that the spectrum range of the signal in the frequency domain is widened. It describes the width of the signal distribution on the spectrum and reflects the distribution of the frequency components of the signal within a certain range.

[0040] Step 3: Separate the sea clutter echo samples According to the single distance unit, it is divided into one-dimensional samples of multiple pulses, and the three-channel RGB image samples are constructed using the one-dimensional samples of multiple pulses. The specific method is as follows:

[0041] like Figure 3 As shown in the figure, the multi-pulse multi-range gate sea clutter data is divided into multi-pulse one-dimensional samples R according to the single range unit. p ∈R P×1 , the one-dimensional sample R p Convert to time domain, frequency domain, and time-frequency domain respectively, denoted as h 1 、h 2 、h 3 , and h i (i=1,2,3) is converted into a grayscale image of m×m×1, where m is the observation time of sea clutter, expressed as the number of pulses. p The grayscale pixel matrix SH obtained in different analysis domains i (j,k), as shown in the following formula:

[0042]

[0043] Where: j = 1, ..., m; k = 1, ..., m; i = 1, 2, 3; h i (·) represents the intensity value of sea clutter echo data in different analysis domains.

[0044] Finally, the two-dimensional grayscale images SH under different analysis domains are i (j,k)(i=1,2,3) are combined in the channel dimension to generate a three-channel RGB image sample MRGB, the size of which is m×m×3. The generated three-channel RGB image sample uses the multi-domain RGB pixel matrix MRGB(j,k,P rgb ) means, specifically:

[0045] MRGB(j,k,P rgb i)=(SH i (j,k),P rgb i)

[0046] Where: P rgb i=1,2,3 represents the i-th channel of the RGB image.

[0047] Step 4: Use the three-channel RGB image samples and the corresponding amplitude distribution type and amplitude distribution parameter label information to train a refined feature extraction network for wide-area energy distribution of sea clutter. After the model converges and the accuracy is improved, a pre-training model is formed. The three-channel RGB image samples are input into the pre-training model to obtain wide-area refined features as prior information.

[0048] The refined feature extraction network for wide-area energy distribution of sea clutter includes: a lightweight attention module receives three-channel image data of sea clutter as input MRGB and outputs space-frequency features The bidirectional long short-term memory module receives the input MRGB(j,k,1) and outputs the bidirectional historical amplitude feature The gated attention mechanism feature fusion module receives the input f RCBAM and h t And output

[0049] The lightweight attention module includes multiple residual modules and channel spatial attention modules connected in series. The residual module learns the difference between the output and the input, i.e., the residual mapping: H(MRGB)=Res(MRGB)+MRGB, where MRGB is the data input to the residual module, H(·) is the output of the residual module, and Res(·) is the residual module. The output H(MRGB) of the residual module is input into the channel spatial attention module (RCBAM), and the extracted spatial-frequency features f are output after passing through the channel attention module and the spatial attention module respectively. RCBAM .

[0050] The data processing process of the channel attention module (CAM) in the lightweight attention module is specifically as follows:

[0051] First, enter It is divided into multiple parts along the channel, expressed as H(MRGB)=[X 0 ,X 1 ,...,X n-1 ],in C can be divided by n. For each part, the corresponding 2D DCT frequency component is assigned, and the 2D DCT result can be used as the preprocessing result of channel attention. So we have:

[0052]

[0053] where [u,v] is X i The corresponding frequency component 2D index, and is the C'-dimensional vector after preprocessing. The overall preprocessed vector can be obtained by concatenation:

[0054] Freq=cat([Freq 0 ,Freq 1 ,…,Freq n-1 ])

[0055] Freq is the vector of different frequency components obtained.

[0056] The above channel attention can be written as:

[0057] M CAM (H(MRGB))=AFF(Freq(H(MRGB))+GMP(H(MRGB)))

[0058] The channel attention output is:

[0059] CAM(MRGB)=M CAM (H(MRGB))⊙H(MRGB)

[0060] The spatial attention (SAM) in the lightweight attention module processes data as follows:

[0061] Λ(·)=[Conv( 1,1 )(·),Conv( 3,1 )(·),Conv( 3,2 )(·),Conv( 3,3 )(·)]

[0062] M SAM (CAM(MRGB))

[0063] =σ(f1×1 (concat(Λ(GAP(CAM(MRGB));GMP(CAM(MRGB))))))

[0064] The output of the spatial attention module is:

[0065] SAM(CAM(MRGB))=M SAM (CAM(MRGB))⊙CA<(MRGB)

[0066] Where: Λ is the multi-scale convolution path, Conv (i,j) represents the convolution operation using a dilated convolution kernel with a kernel size of i and a dilation factor of j. 1×1 is a convolution operation with a kernel size of 1×1, and σ is the sigmoid activation function.

[0067] In summary, the features extracted by RCBAM output are:

[0068] f RCBAM =SAM(CAM(MRGB))

[0069] Bidirectional long short-term memory module:

[0070] The bidirectional LSTM network includes a long short-term memory network (LSTM) combining two directions, namely forward and backward, which respectively include m memory units and n memory units.

[0071] Input sea clutter time series The process of forward LSTM processing input time series is as follows:

[0072] The forget gate is used to determine the hidden state h of the previous time step t-1 The amount of information that needs to be discarded in . The formula is as follows:

[0073] f t =σ(W f ·x t +U f ·h t-1 +b f )

[0074] Where: : The input vector of the current time step, the dimension is the number of features T. : The hidden state vector of the previous time step, with the dimension being the number of hidden layer units h. : Weight matrix of forget gate. : The weight matrix from the previous hidden state to the forget gate. : Bias vector of forget gate. Output: : The output vector of the forget gate indicates the forgetting ratio, and its dimension is the number of hidden layer units h.

[0075] The input gate is used to determine the input x at the current time step t t What information needs to be added to the cell state?

[0076] i t =σ(W i ·x t +U i ·h t-1 +b i )

[0077]

[0078] Where: Input : Weight matrix of input gate. : The weight matrix from the previous hidden state to the input gate. : Bias vector for input gate. : Weight matrix used to generate candidate cell states. : The hidden state weight matrix of the candidate cell state. : Bias vector of candidate cell states. Output The output vector of the input gate represents the information retention ratio, and its dimension is the number of hidden layer units. : Candidate cell state, used to update the cell state.

[0079] Combined with forget gate f t and input gate i t Update the cell state C t :

[0080]

[0081] Where: Input : The cell state at the previous time step. t ⊙C t-1 : The forgotten proportion of the cell state at the previous time step, dimension h. : Update ratio of candidate cell states, dimension h. Output: : The cell state after the update at the current time step.

[0082] The output gate determines the output h at the current time step t t , that is, the hidden state:

[0083] o t =σ(W o ·x t +U o ·h t-1+b o )

[0084] h t =o t ⊙tanh(C t )

[0085] enter: : Weight matrix of the output gate. : The weight matrix from the previous hidden state to the output gate. : Bias vector of the output gate. Output: : Output vector of the output gate. : Hidden state, representing the output of the current time step t.

[0086] The processing of the backward LSTM is the same as that of the forward LSTM, but the traversal direction of the input sequence is opposite to that of the forward direction.

[0087] BiLSTM hides the historical amplitude state output by the forward LSTM And the historical amplitude hidden state of the backward LSTM output After concatenation, the output of BiLSTM is:

[0088]

[0089] The gated attention feature fusion module is used to extract the spatial-frequency feature f RCBAM and historical amplitude hidden feature h t ,Through the gated attention scoring mechanism, the feature importance scoring is completed, thereby highlighting important features and suppressing unimportant features.

[0090] ms att =sigmod(concat(f RCBAM ⊕h t ))∈R L×1

[0091] Finally, the extracted wide-area features are:

[0092] f ms =ms att f RCBAM +(1-ms att )h t

[0093] Step 5: Build a generative adversarial network (GAN) model based on the improved Vision Transformer (ViT), and the generative adversarial network model based on the improved ViT includes a generator and a discriminator. The generator structure includes a feature connection layer, a linear layer, a deconvolution layer, a position encoding layer, and multiple Transformer encoding blocks, which are mainly used to generate simulated sea clutter data. The discriminator consists of a position encoding layer, multiple Transformer encoding blocks and a true and false classification head, an amplitude distribution type classification head, an amplitude distribution parameter prediction head, and a spectral density function prediction head, which are used to judge the authenticity, amplitude characteristics, and energy spectral density of the generated data.

[0094] The generator first receives the input prior information f ms ∈R L×1 and sea surface parameters Radar parameters , Scene Parameters The feature encoding layer first stacks the prior information and three types of parameters in the channel dimension and outputs the stacked features Linear layer input Output fixed dimension after transformation The deconvolution layer receives , output , where C = 2 represents the real channel and the imaginary channel, H = P, W = R; the position encoding layer further adds position information to the input features, so that the output dimension remains After that, the features enter the encoder composed of multiple Transformer encoding blocks, each of which contains a transient global-local information fusion module and a space-frequency dual-path information extraction module.

[0095] In the transient global-local information fusion module, the input feature First, it is divided into multiple windows, each window size is S×S, and the feature dimension is These windows are flattened into one-dimensional vectors to represent local information, and a self-attention mechanism is applied within each window to obtain the local window self-attention matrix To supplement the long-distance information, this module introduces a global vector , which is expanded by broadcasting Then it is spliced ​​with local features to form fusion features The output of the module is the fused feature , and are sent to the space-frequency dual-path information extraction module for further processing.

[0096] The space-frequency dual-path information extraction module receives the fusion feature And divided into n parts along the channel direction, each part , apply 2D discrete cosine transform (DCT) to each block to obtain the frequency feature By concatenating all frequency features, an overall multi-frequency weight vector is formed The channel attention map is obtained through the adaptive feature fusion (AFF) operation , and finally get the channel enhancement feature The channel enhancement features are further processed through the multi-scale dilated convolution path Λ(·). After the dilated convolution results with different convolution kernel sizes and dilation factors are concatenated, the spatial attention map is obtained through 1×1 convolution. The output spatial features are Finally, the output features of the space-frequency dual-path information extraction module are The final convolutional layer receives the output of multiple encoding blocks. , output the generated dual-channel multi-pulse multi-range gate data

[0097] The discriminator first receives the input of real dual-channel multi-pulse multi-range gate data and generated , where C = 2 represents the real channel and the imaginary channel, H = P, W = R, and the output after multiple Transformer encoding layers is f TR The true and false classification results are output through the true and false classification heads respectively , Amplitude distribution type classification head output , Amplitude distribution parameter prediction head output , spectral density function prediction head output , where FS is the sampling frequency.

[0098] Step 6: Follow the above steps to extract the refined features f ms ∈R L×1 As the prior information input of the generator, the independent echo samples of sea clutter in step 2 are , real number processing is performed according to the real and imaginary parts respectively, and multi-pulse multi-range gate real and imaginary dual-channel samples are constructed and used as training sets, which are input into the generator and discriminator for collaborative training; a joint loss function is designed based on the Wasserstein GAN loss function with a gradient penalty term, and the model training is guided from the time domain and frequency domain constraints respectively. The loss function design is as follows:

[0099]

[0100] Among them, D adv(x) is the discriminator probability distribution, x is the real sample, To generate samples, gp is the joint distribution sample gradient penalty term, is a pair of samples obtained by linear sampling of real sea clutter data and generated sea clutter data, E x To obtain the true sample distribution To generate the sample distribution, To obtain the sampling distribution, is Laplace regularization, CE(·) is the cross entropy loss function of the amplitude distribution type, MSEmask(·) is the mean square error of the amplitude distribution parameters, where mask means that the loss value is calculated only for the parameters of the optimal amplitude distribution type; CSL(·) is the energy spectral density loss function based on cosine similarity, and the calculation formula of CSL(·) is as follows:

[0101]

[0102] Among them G PSD and R PSD are the PSD of generated samples and real samples, respectively, i and r i They represent the same frequency components of the real samples of the generated samples.

[0103] Alternately train the generator and discriminator until a Nash equilibrium state is reached;

[0104] It is determined whether the fidelity index between the sea clutter data output by the generator and the real sea clutter data reaches the threshold. If it does not reach the set threshold, the training process is returned to adjust the model parameters until the fidelity index exceeds the threshold.

[0105] The fidelity index completes the statistical acquisition and calculation of sea clutter in the time domain and frequency domain, as follows:

[0106] Assume that after feature extraction, the generated samples and the real samples form an M×N (M is the number of samples, N is the number of features, here M is 100, N is 6) dimensional feature matrix F R and F G , respectively calculate the similarity between each feature and the overall similarity:

[0107]

[0108] Among them, F i is the similarity of the i-th dimension feature of the sample, and the overall similarity is as follows:

[0109]

[0110] The above formula is the overall feature similarity F between the generated sea clutter and the real sea clutter. overall , whose value range is between 0 and 1. In order to facilitate observation and comparison, let F overall Multiply by 100 to expand its value range to between 0-100.

[0111] Based on the above definition of fidelity index, assuming that the model generates data Y g , the comparison data is Y r , fidelity is defined as:

[0112] S=cos(V g ,V r )

[0113] Where V g Y g The clutter echo fidelity statistics vector includes the mean φ mean , variance φ var , kurtosis φ kur , skewness φ skew , Doppler broadening φ fd , Doppler shift φ dfd That is V g =[φ mean (Y g ),φ var (Y g ),φ kur (Y g ),φ skew (Y g ),φ fd (Y g ),φ dfd (Y g )]. r Y r The clutter echo fidelity statistical vector V r =[φ mean (Y r ),φ var (Y r ),φ kur (Y r ),φ skew (Y r ),φ fd (Y r ),φ dfd (Y r )], S is the fidelity evaluation index of the present invention.

[0114] Step 7: Generate sea clutter data samples using the trained generator.

[0115] The generated data is further verified by quantitatively and qualitatively comparing the sea clutter data output by the trained generator with the real sea clutter data. The main tasks are to compare the time-distance intensity graph, the amplitude probability density distribution function, and the power spectrum to verify the authenticity of the model-generated data and the measured comparison data.

[0116] Example

[0117] In order to verify the effectiveness of the present invention, a series of experiments are conducted to analyze the performance of the generated model.

[0118] First, a data set is constructed and the existing sea clutter data is preprocessed to remove abnormal data units to ensure data continuity. After excluding abnormal units, the amplitude distribution characteristics, spectral characteristics and spatiotemporal correlation characteristics of pure sea clutter data are analyzed in detail. For the spatiotemporal correlation of pure sea clutter data, the temporal correlation length of the sea clutter in the data set is calculated to be p pulses, and the spatial correlation length is r distance units. The sliding window size (P, R) is determined based on the decorrelation distance and decorrelation time, where P is the number of consecutive pulses and P = m 2 ,in , where n is 2, ceil is rounded down, R is the number of distance units and R = lr, where l is 1. The sliding step length of the sliding window in the pulse dimension , where floor is rounded up. The sliding step length Step in the distance dimension R =R / / 2. The sliding window is used to slide on the two-dimensional sea clutter data to obtain sea clutter samples to ensure the temporal and spatial correlation of the samples.

[0119] In one embodiment, the generator is composed of a linear layer, a position encoding layer, multiple Transformer encoding blocks and a convolutional layer, while the discriminator includes a position encoding layer, multiple Transformer encoding blocks and a classification head and a prediction head. Each Transformer encoding block is composed of two sub-blocks: a transient global local information fusion module and a space-frequency dual-path information extraction module. The attention module fuses and extracts the temporal and spatial information of the input data, and passes the features into the feedforward multi-layer perceptron layer. Each sub-block is preceded by a batch normalization layer and followed by a random dropout layer. In this embodiment, the random dropout rate is set to 0.5. The Transformer encoder of the generator consists of m encoding blocks, and the Transformer encoder of the discriminator consists of n encoding blocks. The input of the generator includes prior information f ms ∈R L×1 and sea surface parameters Radar parameters Scene Parameters The output of the generator is the generated dual-channel multi-pulse multi-range gate data The input of the discriminator is a real sample and generate samples The discriminator output consists of four parts: true and false classification results Amplitude distribution type classification head output , Amplitude distribution parameter prediction head output , spectral density function prediction head output , where FS is the sampling frequency. According to the above steps, the adversarial generative network is constructed and the network is iteratively trained. The AdamW optimizer is used with a learning rate of 1×10 -5 , and apply the learning rate increment strategy: when the verification loss fluctuates frequently, the training effect of the model is improved by dynamically increasing the learning rate. In order to balance the training differences between the generator and the discriminator, the training times of the generator and the discriminator in each round of training are set to 1:M, and the model at the end of the current round of training is saved. By setting the fidelity indicator threshold S t , when the similarity score is lower than S t When , return to step 3 to further fine-tune the network model; the fidelity score is higher than S t When , the training stops. At this point, the generator is able to generate high-quality data samples with distributions similar to those of real sea clutter samples.

[0120] The above embodiments are intended to describe the implementation of the present invention rather than to limit its scope. The scope of the present invention is defined by the appended claims, and all equivalent substitutions and modifications are within the scope of protection of the present invention without departing from the spirit and basic principles of the present invention.

Claims

1. A method for intelligently generating sea clutter data samples based on wide-area feature extraction, characterized in that: The specific steps include: Step 1: Obtain the original sea clutter data, extract the pure sea clutter data from the original sea clutter data, and perform normalization processing; Step 2: Obtain independent echo samples of sea clutter using a sliding window mechanism based on temporal and spatial characteristics; The amplitude characteristics of the independent echo samples of sea clutter are analyzed, the probability density distribution parameters of the sea clutter data are estimated, and the optimal statistical distribution model is fitted to extract the amplitude distribution type and amplitude distribution parameters; Perform power spectrum and time-frequency domain analysis on the independent echo samples of sea clutter to obtain the Doppler shift and Doppler broadening in the power spectrum of sea clutter; Step 3: Divide the multi-pulse multi-range gate sea clutter data into multi-pulse one-dimensional samples according to single-range units, and use the multi-pulse one-dimensional samples to construct three-channel RGB image samples; Step 4: Use the three-channel RGB image samples and the corresponding amplitude distribution type and amplitude distribution parameter label information to train a refined feature extraction network for wide-area energy distribution of sea clutter. After the model converges and the accuracy is improved, a pre-training model is formed. The three-channel RGB image samples are input into the pre-training model to obtain wide-area refined features. Step 5: Building a generative adversarial network (GAN) model based on the improved ViT, wherein the generative adversarial network model based on the improved ViT includes a generator and a discriminator, wherein the generator is used to generate simulated sea clutter data, and the discriminator is used to judge the authenticity, amplitude characteristics, and energy spectral density of the generated data; Step 6: The wide-area refined features obtained in step 4 are used as the prior information input of the generator, and the independent echo samples of sea clutter in step 2 are processed into real numbers according to the real and imaginary parts respectively, and multi-pulse multi-range gate real and imaginary dual-channel samples are constructed and used as training sets, which are input into the generator and discriminator for collaborative training; Step 7: Generate sea clutter data samples using the trained generator.

2. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that: The specific method of constructing multi-pulse multi-range gate sea clutter data using a sliding window mechanism based on spatiotemporal characteristics is as follows: Determine the sliding window size, use the time-space sliding window to extract sea clutter independent echo samples from the pure sea clutter data, and take out a matrix of size (P, R) as the sea clutter independent echo sample each time, where P is the number of continuous pulses and P = m 2 ,in n is the set value, r is the decorrelation distance, p is the decorrelation time, ceil is rounded down, R is the number of distance units and R=lr, l is the set value; the sliding step of the sliding window in the pulse dimension Where floor is rounded up, and the sliding step length Step in the distance dimension R =R / / 2.

3. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that: The specific method of analyzing the power spectrum and time-frequency domain of the sea clutter independent echo samples and obtaining the Doppler shift and Doppler broadening in the sea clutter power spectrum is as follows: Calculate the autocorrelation function: For a complex signal x[n] = x I [n]+jx Q [n], where x I [n] is the real part, x Q [n] is the imaginary part, and the complex autocorrelation function R[k] is defined as: R[k]=∑ n x[n]·x * [n+k] x * [n] represents the conjugate complex number of x[n], that is, x * [n] = x I [n]-jx Q [n]; Applying the discrete Fourier transform (DFT) to the autocorrelation function R[k] yields the power spectrum of the complex signal: Where N is the total number of samples and f is the frequency index; The Doppler shift and Doppler broadening in the sea clutter power spectrum are obtained through the power spectrum.

4. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that: The multi-pulse multi-range gate sea clutter data is divided into multi-pulse one-dimensional samples according to single-range units. The specific method of using the multi-pulse one-dimensional samples to construct three-channel RGB image samples is as follows: The multi-pulse multi-range gate sea clutter data is divided into multi-pulse one-dimensional samples Rp according to single-range units. p Convert to time domain, frequency domain, and time-frequency domain respectively, denoted as h 1 、h 2 、h 3 , and h i Converted into a grayscale image of m×m×1, i=1,2,3, m is the observation time of sea clutter; One-dimensional sample R p The grayscale pixel matrix SH obtained in different analysis domains i (j, k), as shown in the following formula: Where, j = 1, ..., m; k = 1, ..., m; i = 1, 2, 3; h i (·) represents the intensity value of sea clutter echo data in different analysis domains; The two-dimensional grayscale images SH under different analysis domains i (j, k) are combined in the channel dimension to generate a three-channel RGB image sample MRGB, the size of which is m×m×3. The generated three-channel RGB image sample uses the multi-domain RGB pixel matrix MRGB(j, k, P rgb ) means, specifically: MRGB(j,k,P rgb i)=(SH i (j,k),P rgb i) Where: P rgb i=1,2,3 represents the i-th channel of the RGB image.

5. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that: The refined feature extraction network for wide-area energy distribution of sea clutter includes: Lightweight attention module, bidirectional long short-term memory module, and gated attention mechanism feature fusion module; The lightweight attention module includes multiple residual modules connected in series and a channel space attention module. The residual module learns the difference between the output and the input, i.e., the residual mapping: H(MRGB)=Res(MRGB)+MRGB, where MRGB is the data input to the residual module, H(·) is the output of the residual module, and Res(·) is the residual module. The output H(MRGB) of the residual module is input into the channel-spatial attention module, and the extracted spatial-frequency features f are output after passing through the channel-spatial attention module and the spatial-attention module respectively. RCBAM , specifically expressed as: f RCBAM =SAM(CAM(MRGB)) The bidirectional long short-term memory module adopts a bidirectional long short-term machine network, and the bidirectional long short-term machine network includes a long short-term memory network combined with two directions, namely forward and backward, and the two networks include m memory units and n memory units respectively; Input sea clutter time series X = {x1, x2, ..., xT}, the process of forward LSTM processing the input time series is as follows: The forget gate is used to determine the hidden state h of the previous time step t-1 The amount of information that needs to be discarded is as follows: f t =σ(W f ·x t +U f ·h t-1 +b f ) Where: x t The input vector of the current time step, the dimension is the number of features T, h t-1 is the hidden state vector of the previous time step, and its dimension is the number of hidden layer units h; W f is the weight matrix of the forget gate, U f is the weight matrix from the previous hidden state to the forget gate; b f is the bias vector of the forget gate, and the output is f t is the output vector of the forget gate, indicating the forgetting ratio, and its dimension is the number of hidden layer units h; The input gate is used to determine the input x at the current time step t t The information that needs to be added to the cell state is: i t =σ(W i ·x t +U i ·h t-1 +b i ) Where: Input W i is the weight matrix of the input gate; U i is the weight matrix from the previous hidden state to the input gate; b i is the bias vector of the input gate; W C is the weight matrix used to generate candidate cell states; U C is the hidden state weight matrix of the candidate cell state; b C is the bias vector of the candidate cell state; i t The output vector of the input gate represents the information retention ratio, and its dimension is the number of hidden layer units h; is the candidate cell state, used to update the cell state; Combined with forget gate f t and input gate i t Update the cell state C t : Where: C t-1 is the cell state at the previous time step; f t ⊙C t-1 is the forgetting ratio of the cell state at the previous time step; is the update ratio of candidate cell states, C t is the cell state updated at the current time step; The output gate determines the output h at the current time step t t , that is, the hidden state: the t =σ(W o ·x t +U o ·h t-1 +b o ) h t =o t ⊙tanh(C t ) Where W o is the weight matrix of the output gate; U o is the weight matrix from the previous hidden state to the output gate; b o is the bias direction of the output gate; o t is the output vector of the output gate; h t is the hidden state, which represents the output of the current time step t; The processing of backward LSTM is the same as that of forward LSTM, and the traversal direction of the input sequence is opposite to that of the forward direction; BiLSTM hides the historical amplitude state output by the forward LSTM And the historical amplitude hidden state of the backward LSTM output After concatenation, the output of BiLSTM is: The gated attention feature fusion module is used to extract the spatial-frequency feature f RCBAM and historical amplitude hidden feature h t , through the gated attention scoring mechanism, the feature importance scoring is completed, which is specifically expressed as: In the formula, ms att is the gated attention output, sigmod(·) is the nonlinear activation function, and concat(·) is the concatenation operation; The extracted wide-area features are: f ms =ms att f RCBAM +(1-ms att )h t 。 6. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 5, characterized in that: The data processing process of the channel attention module is specifically as follows: The input H(MRGB) is divided into multiple parts along the channel, expressed as H(MRGB)=[X 0 , X 1 , ..., X n-1 ], where X i , i∈{0, 1, ..., n-1}, C can be divided by n, for each part, the corresponding 2D DCT frequency component is assigned, and the 2D DCT result is used as the preprocessing result of channel attention, so: where [u, v] is X i The corresponding 2D index of the frequency component, is the C′-dimensional vector after preprocessing; the overall preprocessed vector is obtained by concatenation: Freq=cat([Freq 0 ,Freq 1 ,…,Freq n-1 ]) Freq is the vector of different frequency components obtained; The above channel attention is written as: M CAM (H(MRGB))=AFF(Freq(H(MRGB))+GMP(H(MRGB))) The output of the channel attention module is: CAM(MRGB)=M CAM (H(MRGB))⊙H(MRGB) The specific process of data processing by the spatial attention module is as follows: Λ(·)=[Conv (1,1) (·),Conv (3,1) (·),Conv (3,2) (·),Conv (3,3) (·)] M SAM (CAM(MRGB)) =σ(f 1×1 (concat(Λ(GAP(CAM(MRGB));GMP(CAM(MRGB)))))) The output of the spatial attention module is: SAM(CAM(MRGB))=M SAM (CAM(MRGB))⊙CAM(MRGB) Where: A is the multi-scale convolution path, Conv (i，j) represents the convolution operation using a dilated convolution kernel with a kernel size of i and a dilation factor of j. 1×1 is a convolution operation with a kernel size of 1×1, and σ is the sigmoid activation function.

7. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that: The generator includes a feature connection layer, a linear layer, a deconvolution layer, a position encoding layer, and multiple Transformer encoding blocks; the discriminator includes a position encoding layer, multiple Transformer encoding blocks and a true and false classification head, an amplitude distribution type classification head, an amplitude distribution parameter prediction head, and a spectral density function prediction head; The generator first receives the input prior information f ms and sea surface parameter P sea 、Radar parameters P radar , scene parameter P scene The feature encoding layer first stacks the prior information and three types of parameters in the channel dimension and outputs the stacked features Linear layer input Output fixed dimension after transformation Features; The deconvolution layer receives Output The position encoding layer further adds position information to the input features, keeping the output dimension as The features enter the encoder composed of multiple Transformer encoding blocks, each of which contains a transient global-local information fusion module and a space-frequency dual-path information extraction module; In the transient global-local information fusion module, the input feature First, it is divided into multiple windows, each window size is S×S, and the feature dimension is These windows are flattened into one-dimensional vectors to represent local information, and a self-attention mechanism is applied within each window to obtain the local window self-attention matrix f local , introduce a global vector f global , which is expanded by broadcasting to Then it is spliced ​​with local features to form fusion features The output of the module is the fused features and is sent to the space-frequency dual-path information extraction module; The space-frequency dual-path information extraction module receives the fusion feature And divided into n parts along the channel direction, each part X i , Apply 2D discrete cosine transform (DCT) to each block to obtain the frequency feature Freq i ; By splicing all frequency features, an overall multi-frequency weight vector is formed Channel attention mapping is obtained through adaptive feature fusion (AFF) operation Finally, the channel enhancement feature is obtained The multi-scale atrous convolution path Λ(·) is used to process channel enhancement features. After the atrous convolution results with different convolution kernel sizes and dilation factors are concatenated, the spatial attention map M is obtained through 1×1 convolution. SAM (F), the output spatial feature is Finally, the output features of the space-frequency dual-path information extraction module are The final convolutional layer receives the output of multiple encoding blocks Output generated dual channel multi-pulse multi-range gate data HC gen ; The discriminator first receives the input of the real dual-channel multi-pulse multi-range gate data HC and the generated HC gen , after passing through multiple Transformer encoding layers, the output is f TR , f TR The true and false classification results are output through the true and false classification heads respectively Amplitude distribution type classification head output Amplitude distribution parameter prediction head output Spectral density function prediction head output 8. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that: The specific process of training the generator and discriminator is: Input the real and imaginary dual-channel samples of multi-pulse and multi-range gates into the adversarial network GAN model, and train the generator and discriminator alternately until the Nash equilibrium state is reached; It is determined whether the fidelity index between the sea clutter data output by the generator and the real sea clutter data reaches the threshold. If it does not reach the set threshold, the training process is returned to adjust the model parameters until the fidelity index exceeds the threshold.

9. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that: The loss function for co-training the generator and discriminator is: Among them, D adv (x) is the discriminator probability distribution, x is the real sample, To generate samples, λ gp is the joint distribution sample gradient penalty term, is a pair of samples obtained by linear sampling of real sea clutter data and generated sea clutter data, E x To obtain the true sample distribution, To generate the sample distribution, To obtain the sampling distribution, is Laplace regularization, CE(·) is the amplitude distribution type cross entropy loss function, MSE mask (·) is the mean square error of the amplitude distribution parameters, where mask means that the loss value is calculated only for the parameters of the best amplitude distribution type; CSL(·) is the energy spectral density loss function based on cosine similarity, and the calculation formula of CSL(·) is as follows: Among them G PSD and R PSD are the PSD of generated samples and real samples, respectively, i and r i They represent the same frequency components of the real samples of the generated samples.

10. The method for intelligently generating sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that: The calculation method of the fidelity index is: Assume that the generated samples form an M×N dimensional feature matrix F after feature extraction with the real samples R and F G , M is the number of samples, N is the number of features, and the similarity between each feature and the overall similarity are calculated separately: Among them, F i is the similarity of the i-th dimension feature of the sample, and the overall similarity is as follows: F overall is the overall characteristic similarity between sea clutter and real sea clutter; let F overall Multiply by 100 to expand its value range to between 0 and 100; Assume that the model generates data Y g , the comparison data is Y r , fidelity is defined as: S=cos(V g ,In r ) Where V g Y g The clutter echo fidelity statistics vector includes the mean φ mean , variance φ var , kurtosis φ kur , skewness φ skew , Doppler broadening φ fd , Doppler shift φ dfd That is V g =[φ mean (Y g ), φ var (Y g ), φ kur (Y g ), φ skew (Y g ), φ fd (Y g ), φ dfd (Y g )]; V r Y r The clutter echo fidelity statistics vector V r =[φ mean (Y r ), φ var (Y r ), φ kur (Y r ), φ skew (Y r ), φ fd (Y r ), φ dfd (Y r )], S is the fidelity evaluation index.

Images