A sea clutter data sample intelligent generation method based on wide-area feature extraction

Abstract

The application discloses a sea clutter data sample intelligent generation method based on wide-area feature extraction, comprising the following steps: based on the space-time correlation characteristics of sea clutter, the sliding window mechanism is used to process the sea clutter data, and a three-channel RGB image sample is constructed for one-dimensional single-distance gate multi-pulse sea clutter data; fine features are extracted for wide-area energy distribution of sea clutter, a sea clutter feature extraction module is constructed, and is embedded into a sea clutter modeling and simulation network; a generator based on ViT 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, the extracted fine features of sea clutter are input into the generator as prior information, a joint loss function is designed to constrain and guide network training, until the generator and the discriminator both reach a convergent state, and the trained generator is taken as a sea clutter generation model. The finally generated data of the application reaches the given threshold index requirement in the multi-domain fidelity index.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically a method for intelligent generation of sea clutter data samples based on wide-area feature extraction. Background Technology

[0002] Sea clutter is radar echo reflected from the sea surface, and its intensity is affected by multiple environmental factors such as weather conditions and waves. In complex marine environments, sea clutter characteristics are highly variable, posing a significant challenge to the detection of sea surface targets. High-quality sea clutter data is crucial for studying sea clutter characteristics and improving target detection algorithms. Traditional sea clutter echo data simulation typically uses classical statistical models to model the amplitude statistical characteristics of sea clutter, and then generates specific distributions using the zero-memory nonlinear transform method (ZMNL) and the spherically invariant stochastic process method (SIRP). However, these traditional models are too simplistic and cannot fully simulate the non-Gaussian and non-stationary characteristics of sea clutter, resulting in low accuracy and poor applicability of the generated data, limiting the model's effectiveness in real-world scenarios. Therefore, obtaining high-quality and diverse sea clutter data has become an important research task.

[0003] Given the limitations of traditional methods, researchers have increasingly turned their attention to Generative Adversarial Networks (GANs) for generating radar data. Through adversarial training, GANs have achieved significant results in areas such as image and audio generation, and have shown potential in radar data generation. Although GANs have improved upon the shortcomings of traditional methods to some 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. While the patent "A Method for Generating Sea Clutter Data Samples Based on Spatiotemporal Joint Sensing" improves the accuracy of sea clutter generation by enhancing the spatiotemporal joint sensing of sea clutter, the evaluation criteria for sea clutter are still insufficient. It is necessary to verify the multi-dimensional and multi-angle characteristics of simulated sea clutter to demonstrate the authenticity of the generated sea clutter data. Summary of the Invention

[0004] This invention aims to propose an intelligent generation method for sea clutter data samples based on wide-area feature extraction. By utilizing the spatiotemporal characteristics of sea clutter, removing target interference, and performing amplitude characteristic analysis, spectrum analysis, and time-frequency analysis, combined with generative adversarial networks (GANs), sea clutter data is generated through simulation. This addresses the limitations of traditional methods in that they do not accurately grasp the multi-domain characteristics of sea clutter and have insufficient data accuracy.

[0005] The technical solution to achieve the purpose of this invention is: an intelligent generation method for sea clutter data samples based on wide-area feature extraction, specifically including the following steps:

[0006] Step 1: Acquire raw sea clutter data and extract clean sea clutter data from the raw sea clutter data to normalize the echoes within the same range cell;

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

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

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

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

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

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

[0013] Step 6: Use the wide-area refined features obtained in Step 4 as the prior information input to the generator, and process the independent echo samples of sea clutter in Step 2 into real numbers according to the real and imaginary parts respectively, construct a multi-pulse multi-range gate real and imaginary dual-channel sample and use it as the training set, input it into the generator and discriminator for collaborative training.

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

[0015] Compared with existing technologies, the main advantages of this invention are: to address the problems of complex features, limited sample size, and difficulty in accurately capturing multi-domain characteristics of sea clutter data, a generative adversarial model based on improved ViT is proposed. This method can efficiently simulate sea clutter data with high fidelity. 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. Attached Figure Description

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

[0017] Figure 2 This invention relates to a data sample construction method.

[0018] Figure 3 This invention relates to a method for constructing feature extraction data samples.

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

[0020] Figure 5 This is a structural diagram of the generator and discriminator of this invention.

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

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

[0023] Figure 8 This is a diagram of the training process of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0025] This invention provides an intelligent generation method for sea clutter data samples based on wide-area feature extraction, the specific steps of which include:

[0026] Step 1: Acquire raw sea clutter data and extract clean sea clutter data from it; the specific method for extracting clean sea clutter data from raw sea clutter data is as follows:

[0027] By removing nearshore, offshore, target cells (such as sea crests and specific targets on the sea surface) and other anomalous cell data, clean sea clutter data is obtained. The anomalous cells include nearshore distance cells, target distance cells, and offshore distance cells that are severely affected by the marine topography environment.

[0028] Existing datasets sometimes contain radar echoes that include not only clutter but also target echoes, such as channel buoys, non-cooperative vessels, and distant islands. To reduce interference from other targets on the training sea clutter, non-clutter echoes must be filtered out. Nearshore coastal topography significantly impacts sea clutter, resulting in an irregular distribution of received sea clutter echoes. Furthermore, sea clutter collected from distant shores has too much distance between cells, causing its influence on the echoes to be minimal, resulting in the sea clutter signal being overwhelmed by other noise signals, appearing as white noise in the frequency domain. Therefore, echo signals from these distant cells also need to be removed. The amplitude of sea clutter data is also affected by distance. To eliminate the influence of varying amplitudes across different distance cells, mean-free normalization must be performed on different pulse echoes within the same distance cell.

[0029] Step 2: Perform spatiotemporal correlation analysis on the normalized sea clutter data to extract the decorrelation time p and decorrelation distance r, thereby determining the sliding window size as [p, r]. Use the sliding window mechanism to obtain independent sea clutter echo samples. The specific method is as follows:

[0030] Sample extraction method as follows Figure 2 As shown, a spatiotemporal sliding window is used to extract independent sea clutter echo samples from clean sea clutter data. Each time, a matrix of size (P, R) is extracted as an independent sea clutter echo sample. Here, P is the number of consecutive pulses and P = m. 2 ,in Here, n is set to 2, ceil is the floor function, and R is the number of distance cells, where R = lr, and l is set to 1. The sliding window's step size in the impulse dimension. Where floor represents the floor function, rounded up. Step represents the sliding step size in the distance dimension. R =R / / 2.

[0031] Amplitude characteristics of independent echo samples of sea clutter with size (P,R) are analyzed according to a single distance cell. The probability density distribution parameters of the sea clutter data are estimated and fitted to the optimal statistical distribution model. The amplitude distribution type and amplitude distribution parameters are extracted.

[0032] Power spectrum and time-frequency domain analysis were performed on independent echo samples of sea clutter with size (P,R) at individual distance cells. The spectral characteristics were analyzed, and Doppler frequency shift and Doppler broadening were extracted. The specific method is as follows:

[0033] The spectral characteristics of clean sea clutter data were analyzed using the periodogram method, and Doppler broadening and Doppler frequency shift were obtained. First, the autocorrelation function was calculated: 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:

[0034] R[k]=∑ n x[n]·x * [n+k]

[0035] Here x * [n] represents the conjugate complex number, i.e., x * [n] = x I [n]-jx Q [n]. Conjugate is used when calculating autocorrelation to ensure that the result is still an energy representation.

[0036] Secondly, by applying the Discrete Fourier Transform (DFT) to the autocorrelation function R[k], the power spectrum of the complex signal can be obtained:

[0037]

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

[0039] Finally, the frequency shift and broadening in the power spectrum of sea clutter can be obtained through the power spectrum. The frequency shift refers to the center frequency position in the power spectrum; broadening refers to the phenomenon of the signal's spectral range widening in the frequency domain. It describes the width of the signal's distribution in the spectrum and reflects the distribution of the signal's frequency components within a certain range.

[0040] Step 3: Separate sea clutter echo samples The method involves dividing the data into multi-pulse one-dimensional samples based on single-distance units, and then using these multi-pulse one-dimensional samples to construct three-channel RGB image samples.

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

[0042]

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

[0044] Finally, the two-dimensional grayscale images SH under different analysis domains i (j,k)(i=1,2,3) are merged along the channel dimension to generate three-channel RGB image samples MRGB, each with a size of m×m×3. The generated three-channel RGB image samples are then used to generate a multi-domain RGB pixel matrix MRGB(j,k,P) rgb ) indicates, specifically:

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

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

[0047] Step 4: Train a refined feature extraction network for wide-area energy distribution of sea clutter using three-channel RGB image samples and corresponding amplitude distribution type and amplitude distribution parameter label information. After the model converges and the accuracy is improved, a pre-trained model is formed. Input the three-channel RGB image samples into the pre-trained 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 that receives three-channel sea clutter image data as input MRGB and outputs spatial frequency features. The bidirectional long short-term memory module receives input MRGB(j,k,1) and outputs bidirectional historical amplitude features. The gating attention mechanism feature fusion module receives input f RCBAM and h t and output

[0049] The lightweight attention module comprises multiple cascaded residual modules and a channel spatial attention module. The residual modules learn the difference between their output and 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 itself. The output H(MRGB) of the residual module is input into the channel spatial attention module (RCBAM), passing through both the channel attention module and the spatial attention module, and outputting the extracted spatial frequency feature f. RCBAM .

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

[0051] First, input Dividing the channel into multiple parts, represented as H(MRGB) = [X 0 ,X 1 ,...,X n-1 ],in C is divisible by n. For each part, a corresponding 2D DCT frequency component is assigned, and the 2D DCT result can be used as a preprocessing result for channel attention. Thus, we have:

[0052]

[0053] Where [u,v] is X i The corresponding frequency component 2D exponent, and This is the preprocessed C'-dimensional vector. The entire preprocessed vector can be obtained by concatenation:

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

[0055] Freq represents the vectors of different frequency components obtained.

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

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

[0058] Channel attention output is:

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

[0060] The spatial attention (SAM) process in the lightweight attention module specifically handles the 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] In the formula: Λ represents the multi-scale convolutional pathway, Conv (i,j) This represents a convolution operation performed using a dilated convolution kernel with kernel size i and dilation factor j, f 1×1 This 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 long short-term machine network includes a long short-term memory network (LSTM) that combines two directions, namely forward and backward, each consisting of m and n memory units, respectively.

[0071] Input sea clutter time series The process of processing the input time series using a forward LSTM 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. The formula is as follows:

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

[0074] In the formula: : The input vector at the current time step, with dimension T as the number of features. : The hidden state vector of the previous time step, with dimension h equal to the number of hidden layer units. Weight matrix of the forget gate. : The weight matrix from the previous hidden state to the forget gate. The bias vector for the forget gate. Output: The output vector of the forget gate represents the forgetting ratio, with a dimension equal to 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] In the formula: input : Weight matrix of the input gate. : The weight matrix from the previous hidden state to the input gate. : The bias vector of the input gate. : Weight matrix used to generate candidate cell states. : Hidden state weight matrix of candidate cell states. : Bias vector of candidate cell states. Output The output vector of the input gate represents the information retention ratio, with the dimension being the number of hidden layer units. : Candidate cell state, used to update cell state.

[0079] Combined with the forgetting gate f t and input gate i t Information to update cell state C t :

[0080]

[0081] In the formula: input : Cell state at the previous time step. f t ⊙C t-1 The proportion of the cell state forgotten in the previous time step, dimension h. The update ratio of candidate cell states, in dimension h. Output: : The cell state after the current time step update.

[0082] The output gate determines the output h of 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. : The bias vector of the output gate. Output: : Output vector of the output gate. : Hidden state, representing the output at the current time step t.

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

[0087] BiLSTM hides the historical amplitude output of the forward LSTM. and the historical amplitude hidden state of the backward LSTM output After splicing, the output of BiLSTM is:

[0088]

[0089] The gated attention feature fusion module, for the extracted space-frequency feature f RCBAM and historical amplitude hidden features h t By using a gating attention scoring mechanism, feature importance is scored, 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: Construct a Generative Adversarial Network (GAN) model based on the improved Vision Transformer (ViT). This ViT-based GAN model includes a generator and a discriminator. The generator structure comprises a feature connection layer, a linear layer, a deconvolutional layer, a positional encoding layer, and multiple Transformer encoding blocks, primarily used to generate simulated sea clutter data. The discriminator consists of a positional encoding layer, multiple Transformer encoding blocks, and a true / false classification head, an amplitude distribution type classification head, an amplitude distribution parameter prediction head, and a spectral density function prediction head, used to determine the authenticity, amplitude characteristics, and energy spectral density of the generated data.

[0094] The generator first receives the prior information f as input. ms ∈R L×1 and sea surface parameters Radar parameters Scene parameters The feature encoding layer first stacks the prior information and the three types of parameters along the channel dimension, and outputs the stacked features. Linear layer input Fixed dimension after output transformation Features; deconvolutional layer receives Output Where C=2 represents the real and imaginary channels respectively, H=P, W=R; the positional encoding layer further adds positional information to the input features, keeping the output dimension constant. Subsequently, the features are fed into an encoder consisting of multiple Transformer coding 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 features First, it is divided into multiple windows, each window being S×S in size, with a feature dimension of . These windows are flattened into one-dimensional vectors to represent local information, and a self-attention mechanism is applied within each window to obtain a local window self-attention matrix. To supplement long-range information, this module introduces a global vector. Extended via broadcast Then, it is combined with local features to form a fused feature. The module's output is the fused features. The data is then sent to the space-frequency dual-path information extraction module for further processing.

[0096] Space-frequency dual-path information extraction module receives fusion features It is divided into n parts along the channel direction, each part Applying 2D Discrete Cosine Transform (DCT) to each block yields the frequency characteristics. By concatenating all frequency features, a holistic multi-frequency weight vector is formed. Channel attention maps are obtained through adaptive feature fusion (AFF) operations. Finally, channel enhancement features are obtained. Further channel enhancement features are processed using multi-scale dilated convolutional paths Λ(·). The dilated convolution results with different kernel sizes and dilation factors are concatenated and then spatial attention maps are obtained through 1×1 convolution. The output space features are: Finally, the output features of the space-frequency dual-path information extraction module are: Finally, the convolutional layer receives the output of multiple coded blocks. Output generated dual-channel multi-pulse multi-distance gated data

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

[0098] Step 6: Following the steps above, refine the extracted features f ms ∈R L×1 As prior information input to the generator, and using the independent echo samples of sea clutter from step 2... The real and imaginary parts are respectively converted to real numbers, and a dual-channel sample of real and imaginary parts of a multi-pulse, multi-distance gate is constructed and used as a training set. This sample is then input into the generator and discriminator for collaborative training. Based on the Wasserstein GAN loss function with a gradient penalty term, a joint loss function is designed to guide model training from time-domain and frequency-domain constraints, respectively. The specific design of the loss function is as follows:

[0099]

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

[0101]

[0102] Among them G PSD and R PSD The PSDs of the generated sample and the real sample are g, respectively. i and r i These represent the same frequency components of the real samples that generated the samples.

[0103] The generator and discriminator are trained alternately until a Nash equilibrium is reached.

[0104] Determine whether the fidelity index of the sea clutter data output by the generator reaches the threshold. If it does not reach the set threshold, return to the training process to adjust the model parameters until the fidelity index exceeds the threshold.

[0105] The fidelity index completes the statistical acquisition and calculation of the representation of sea clutter in the time and frequency domains, as detailed below:

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

[0107]

[0108] Where F i Let be the similarity of the i-th dimension feature of the samples. The overall similarity is as follows:

[0109]

[0110] The above formula represents the overall feature similarity F between generated sea clutter and real sea clutter. overall Its value ranges from 0 to 1. For ease of observation and comparison, let F... overall Multiply by 100 to expand its value range to between 0 and 100.

[0111] Based on the above definition of realism metrics, assume that the model generates data Y. g The comparison data is Y. r Realism is defined as:

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

[0113] Where V g For Y g The clutter echo fidelity statistical vector, including the mean φ mean variance φ var kurtosis φ kur skewness φ skew Doppler broadening φ fd Doppler frequency 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 For Y r 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 realism evaluation index of this invention.

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

[0115] Further validation of the generated data involves quantitative and qualitative comparisons between the sea clutter data output by the trained generator and real sea clutter data. This includes comparisons of time-distance intensity maps, amplitude probability density distribution functions, and power spectra to verify the authenticity of the model-generated data and the measured comparison data.

[0116] Example

[0117] To verify the effectiveness of the present invention, the performance of the generated model was analyzed through a series of experiments.

[0118] First, a dataset was constructed and existing sea clutter data was preprocessed to remove outlier cells to ensure data continuity. After removing outliers, the amplitude distribution characteristics, spectral characteristics, and spatiotemporal correlation of the clean sea clutter data were analyzed in detail. For the spatiotemporal correlation of the clean sea clutter data, the temporal correlation length of the sea clutter in the dataset was calculated to be p pulses, and the spatial correlation length was calculated to be r distance cells. The sliding window size (P, R) was determined based on the decorrelation distance and decorrelation time, where P is the number of consecutive pulses and P = m. 2 ,in Here, n is 2, ceil is the floor function, R is the number of distance cells and R = lr, where l is 1. The sliding window's step size in the impulse dimension. Where floor represents the floor function rounded up. The sliding step size in the distance dimension is Step. R =R / / 2. This sliding window is used to obtain sea clutter samples by sliding across the two-dimensional sea clutter data to ensure the spatiotemporal correlation of the samples.

[0119] In one embodiment, the generator consists of a linear layer, a positional encoding layer, multiple Transformer encoding blocks, and a convolutional layer, while the discriminator includes a positional encoding layer, multiple Transformer encoding blocks, 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 spatial-frequency dual-path information extraction module. The attention module fuses and extracts the temporal and spatial information of the input data and feeds the features into a feedforward multilayer perceptron layer. Each sub-block has a pre-batch normalization layer and a post-random dropout layer. In this embodiment, the random dropout rate is set to 0.5. The generator's Transformer encoder consists of m encoding blocks, and the discriminator's Transformer encoder consists of n encoding blocks. The generator's input includes prior information f. ms ∈R L×1 and sea surface parameters Radar parameters Scene parameters The generator output is the generated dual-channel multi-pulse multi-distance gated data. The input to the discriminator is a real sample. and generate samples The discriminator output includes four parts: true / false classification results. Amplitude distribution type classification header output Amplitude distribution parameter prediction head output Spectral density function prediction head output Where FS is the sampling frequency. Following the steps above, an adversarial generative network is constructed and iteratively trained. The AdamW optimizer is used with a learning rate of 1×10⁻⁶. -5 A learning rate increment strategy is applied: when the validation loss fluctuates frequently, the learning rate is dynamically increased to improve the model's training performance. To balance the training differences between the generator and discriminator, the ratio of generator to discriminator training iterations in each round is set to 1:M, and the model at the end of the current round is saved. A fidelity metric threshold S is set. t When the similarity score is lower than S t Then, return to step 3 to further fine-tune the network model; if the fidelity score is higher than S... t If the condition is met, training stops. At this point, the generator can generate high-quality data samples with a distribution similar to real sea clutter samples.

[0120] The above embodiments are intended to describe the implementation of the present invention, and not 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 protection scope of the present invention without departing from the spirit and basic principles thereof.

Claims

1. A method for intelligent generation of sea clutter data samples based on wide-area feature extraction, characterized in that, Specifically, the following steps are included: Step 1: Obtain raw sea clutter data, extract clean sea clutter data from the raw sea clutter data, and perform normalization processing; Step 2: Obtain independent echo samples of sea clutter using a sliding window mechanism based on spatiotemporal characteristics; Amplitude characteristics of independent echo samples of sea clutter are analyzed to estimate the probability density distribution parameters of sea clutter data, and the optimal statistical distribution model is fitted to extract the amplitude distribution type and amplitude distribution parameters. Power spectrum and time-frequency domain analysis were performed on independent echo samples of sea clutter to obtain the Doppler frequency shift and Doppler broadening in the power spectrum of sea clutter. Step 3: Divide the multipulse multi-range gate sea clutter data into one-dimensional samples of multipulse according to single-range units, and use the one-dimensional samples of multipulse to construct three-channel RGB image samples; Step 4: Train a refined feature extraction network for wide-area energy distribution of sea clutter using three-channel RGB image samples and corresponding amplitude distribution type and amplitude distribution parameter label information. After the model converges and the accuracy is improved, a pre-trained model is formed. Input the three-channel RGB image samples into the pre-trained model to obtain wide-area refined features. Step 5: Build a generative adversarial network (GAN) model based on the improved ViT. The improved ViT-based GAN model includes a generator and a discriminator. The generator is used to generate simulated sea clutter data, and the discriminator is used to determine the authenticity, amplitude characteristics, and energy spectral density of the generated data. Step 6: Use the wide-area refined features obtained in Step 4 as the prior information input to the generator, and process the independent echo samples of sea clutter in Step 2 into real numbers according to the real and imaginary parts respectively, construct a multi-pulse multi-range gate real and imaginary dual-channel sample and use it as the training set, input it into the generator and discriminator for collaborative training. Step 7: Generate sea clutter data samples using the trained generator.

2. The intelligent generation method for sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that, The specific method for constructing multi-pulse, multi-range gate sea clutter data based on spatiotemporal characteristics using a sliding window mechanism is as follows: Determine the sliding window size, and use the spatiotemporal sliding window to extract independent sea clutter echo samples from clean sea clutter data. Each time, a matrix of size (P, R) is extracted as an independent sea clutter echo sample, where P is the number of consecutive pulses and P = m. 2 ,in n is a set value, r is the decorrelation distance, p is the decorrelation time, ceil is rounded down, R is the number of distance cells and R = lr, l is a set value; the sliding window's step size in the impulse dimension. Where floor represents the floor function, and Step represents the sliding step size in the distance dimension. R =R / / 2.

3. The intelligent generation method for sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that, The specific method for analyzing the power spectrum and time-frequency domain of independent sea clutter echo samples to obtain the Doppler frequency 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 complex conjugate of x[n], i.e., 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; Doppler shift and Doppler broadening in the power spectrum of sea clutter are obtained by power spectrum analysis.

4. The intelligent generation method for 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 for constructing three-channel RGB image samples using the one-dimensional samples of multi-pulse clutter is as follows: Multipulse, multi-range gate sea clutter data is divided into one-dimensional samples Rp of multipulse according to single-range cells. The one-dimensional samples Rp... p Transform to the time domain, frequency domain, and time-frequency domain respectively, denoted as h. 1 h 2 h 3 and respectively h i Convert to an m×m×1 grayscale image, i=1,2,3, m is the observation duration of sea clutter; One-dimensional sample R p SH of grayscale pixel matrix obtained in different analysis domains i (j, k), as shown in the following formula: In the formula, j = 1, ..., m; k = 1, ..., m; i = 1, 2, 3; h i (·) indicates the intensity value of sea clutter echo data in different analysis domains; Two-dimensional grayscale images SH under different analysis domains i (j, k) are merged along the channel dimension to generate three-channel RGB image samples MRGB, each with a size of m×m×3. The generated three-channel RGB image samples are then processed using a multi-domain RGB pixel matrix MRGB(j, k, P). rgb ) indicates, specifically: MRGB(j,k,P rgb i)=(SH i (j,k),P rgb i) In the formula: P rgb i = 1, 2, 3 represents the i-th channel of the RGB image.

5. The intelligent generation method for 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 cascaded residual modules 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. After passing through the channel attention module and the spatial attention module respectively, the extracted spatial frequency feature f is output. RCBAM Specifically, it is expressed as: f RCBAM =SAM(CAM(MRGB)) The bidirectional long short-term memory module adopts a bidirectional long short-term machine network, which includes a long short-term memory network that combines two directions, namely forward and backward. These two networks include m memory units and n memory units, respectively. Given a sea clutter time series X = {x1, x2, ..., xT}, the process of processing the input time series using a forward LSTM 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 calculated using the following formula: f t =σ(W f ·x t +U f ·h t-1 +b f ) In the formula: x t The input vector at the current time step has dimensions T and h. t-1 W is the hidden state vector from the previous time step, with dimension h equal to the number of hidden layer units. f U is the weight matrix of the forget gate. f b is the weight matrix from the previous hidden state to the forget gate; f Let f be the bias vector of the forget gate, and let f be the output. t The output vector of the forget gate represents 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 as follows: i t =σ(W i ·x t +U i ·h t-1 +b i ) In the formula: input W i U is the weight matrix of the input gate; i b is the weight matrix from the previous hidden state to the input gate; i W is the bias vector of the input gate; C U is the weight matrix used to generate candidate cell states; C b is the hidden state weight matrix for the candidate cell state; C i is the bias vector of the candidate cell state; t The output vector of the input gate represents the information retention ratio, with a dimension equal to the number of hidden layer units h. These are candidate cell states, used to update cell states; Combined with the forgetting gate f t and input gate i t Information to update cell state C t : In the formula: C t-1 The cell state at the previous time step; f t ⊙C t-1 This represents the proportion of the cell state forgotten at the previous time step; C represents the update ratio of candidate cell states. t This represents the cell state updated at the current time step. The output gate determines the output h of 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 ) In the formula, W o U is the weight matrix of the output gate; o b is the weight matrix from the previous hidden state to the output gate; o The bias direction of the output gate; o t h is the output vector of the output gate; t This is a hidden state, representing the output at the current time step t; The processing procedure of backward LSTM is the same as that of forward LSTM, but the traversal direction of the input sequence is opposite to that of forward LSTM. BiLSTM hides the historical amplitude output of the forward LSTM. and the historical amplitude hidden state of the backward LSTM output After splicing, the output of BiLSTM is: The gated attention feature fusion module is used to target the extracted space-frequency features f RCBAM and historical amplitude hidden features h t Feature importance is scored using a gated attention scoring mechanism, specifically as follows: In the formula, ms att For gated attention output, sigmoid(·) is a non-linear activation function, and concat(·) is a connection operation; The extracted wide-area features are: f ms =ms att f RCBAM +(1-ms att )h t 。 6. The intelligent generation method for sea clutter data samples based on wide-area feature extraction according to claim 5, characterized in that, The specific data processing procedure of the channel attention module is as follows: The input H(MRGB) is divided into multiple parts along the channels, represented as H(MRGB) = [X 0 X 1 , ..., X n-1 ], where X i For i ∈ {0, 1, ..., n-1}, C is divisible by n. For each part, a corresponding 2D DCT frequency component is assigned. The 2D DCT result is used as the preprocessing result for channel attention, thus: Where [u, v] is X i The corresponding frequency component 2D exponent, This is the preprocessed C′-dimensional vector; the entire preprocessed vector is obtained by concatenation: Freq=cat([Freq 0 ,Freq 1 ,…,Freq n-1 ]) Freq represents the vectors of different frequency components obtained. The channel attention mentioned above 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 spatial attention module processes data 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) In the formula: A represents the multi-scale convolutional pathway, Conv (i，j) This represents a convolution operation performed using a dilated convolution kernel with kernel size i and dilation factor j, f 1×1 This is a convolution operation with a kernel size of 1×1, and σ is the sigmoid activation function.

7. The intelligent generation method for 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 positional encoding layer, and multiple Transformer encoding blocks; the discriminator includes a positional encoding layer, multiple Transformer encoding blocks, a true / 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 prior information f as input. ms and sea surface parameter P sea Radar parameter P radar Scene parameter P scene The feature encoding layer first stacks the prior information and the three types of parameters along the channel dimension, and outputs the stacked features. Linear layer input Fixed dimension after output transformation Features; Deconvolutional layer reception Output The positional encoding layer further adds positional information to the input features, maintaining the output dimension. Features are fed into an encoder consisting of multiple Transformer coding 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 features First, it is divided into multiple windows, each window being S×S in size, with a feature dimension of . 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 Extended through broadcasting Then, it is combined with local features to form a fused feature. The module output is the fused features. It is then sent to the space-frequency dual-path information extraction module; Space-frequency dual-path information extraction module receives fusion features It is divided into n parts along the channel direction, each part X i Applying a 2D Discrete Cosine Transform (DCT) to each block yields the frequency characteristic Freq. i By concatenating all frequency features, a holistic multi-frequency weight vector is formed. Channel attention mapping is obtained through adaptive feature fusion (AFF) operation. Finally, channel enhancement features are obtained. Channel enhancement features are processed using a multi-scale dilated convolution path Λ(·). The dilated convolution results under different kernel sizes and dilation factors are concatenated and then subjected to a 1×1 convolution to obtain the spatial attention map M. SAM (F), the output space features are Finally, the output features of the space-frequency dual-path information extraction module are: Finally, the convolutional layer receives the output of multiple coded blocks. Output generated dual-channel multi-pulse multi-distance gated data HC gen ; The discriminator first receives the input real dual-channel multi-pulse multi-range gate data HC and the generated HC. gen After passing through multiple Transformer encoding layers, the output f is... TR f TR The true / false classification results are output after passing through the true / false classification head respectively. Amplitude distribution type classification header output Amplitude distribution parameter prediction head output Spectral density function prediction head output 8. The intelligent generation method for 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 as follows: The real and imaginary dual-channel samples of the multi-pulse multi-distance gate are input into the adversarial network GAN model, and the generator and discriminator are trained alternately until the Nash equilibrium is reached. Determine whether the fidelity index of the sea clutter data output by the generator reaches the threshold. If it does not reach the set threshold, return to the training process to adjust the model parameters until the fidelity index exceeds the threshold.

9. The intelligent generation method for 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) represents the probability distribution of the discriminator, where x is the real sample. To generate samples, λ gp For the gradient penalty term of the joint distribution samples, E is a sample obtained by linearly sampling a pair of real sea clutter data and generated sea clutter data. x To obtain the true sample distribution, To obtain the distribution of generated samples, To obtain the distribution of the sampled data, For Laplace regularization, CE(·) is the cross-entropy loss function for amplitude distribution types, and MSE is... mask (·) represents the mean square error of the amplitude distribution parameters, where mask indicates that the loss value is calculated only for parameters of the optimal amplitude distribution type; CSL(·) is the energy spectral density loss function based on cosine similarity, and the calculation formula for CSL(·) is as follows: Among them G PSD and R PSD The PSDs of the generated sample and the real sample are g, respectively. i and r i These represent the same frequency components of the real samples that generated the samples.

10. The intelligent generation method for sea clutter data samples based on wide-area feature extraction according to claim 1, characterized in that, The method for calculating the fidelity index is as follows: Assume that after feature extraction, the generated samples and the real samples form an M×N dimensional feature matrix F. R and F G M is the number of samples, and N is the number of features. Calculate the similarity between each feature and the overall similarity separately: Among them, F i Let be the similarity of the i-th dimension feature of the samples. The overall similarity is as follows: F overall Let F be the overall similarity between sea clutter and real sea clutter; overall Multiply by 100 to expand its value range to between 0 and 100; Assume the model generates data Y g The comparison data is Y. r Realism is defined as: S=cos(V g ,In r ) Where V g For Y g The clutter echo fidelity statistical vector, including the mean φ mean variance φ var kurtosis φ kur skewness φ skew Doppler broadening φ fd Doppler frequency 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 For Y r 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 realism evaluation index.

Images