A method for generating sea clutter echo data samples based on spatio-temporal joint perception

Abstract

The application discloses a sea clutter data sample generation method based on space-time joint perception, comprising the following steps: obtaining pulse noise close to the real sea clutter distribution according to the amplitude distribution statistical characteristics of the sea clutter data; adopting a sliding window to extract data in the pure sea clutter data to form independent samples, and converting the samples into double-channel real samples; building a sea clutter generator and discriminator based on a transformer; inputting the pulse noise as prior information into the generator, and cooperatively training the generator and the discriminator based on the double-channel real samples until the generator and the discriminator are trained to reach a Nash equilibrium state; taking the sea clutter data generated by the trained generator as a substitute or expansion of the real data, and performing characteristic analysis on the generated echo data and the real echo data. The application can generate a large amount of data from a small amount of sea clutter data, and has the advantages of high precision, easy training, easy migration and the like.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically provides a method for generating sea clutter echo data samples based on spatiotemporal joint perception. Background Art

[0002] Sea clutter is radar backscatter echoes from the sea surface, and its intensity is affected by multiple factors, including the climate, wind, and wave variations. The unpredictable characteristics of sea clutter in complex ocean environments pose significant challenges to surface target detection. High-quality sea clutter data is crucial for understanding sea clutter characteristics and developing target detection algorithms. Traditional sea clutter echo data simulation methods use classic statistical models to fit and model the statistical characteristics of sea clutter amplitude, then generate distributions using the zero-memory nonlinearity (ZMNL) and spherically invariant random processes (SIRP) methods. Traditional models are overly simplistic and struggle to simulate the non-Gaussian and non-stationary characteristics of sea clutter. Consequently, the resulting simulated data suffers from low accuracy and a narrow scope of application. Acquiring high-quality, diverse sea clutter data is a significant undertaking.

[0003] Due to the aforementioned issues with traditional methods, researchers have begun exploring the use of generative adversarial networks (GANs) to generate radar data. Through adversarial training, GANs have achieved remarkable success in fields such as image generation and music generation, and have also demonstrated promising performance in radar data generation. While these approaches address some of the challenges of traditional methods, their network architecture, primarily based on convolutional neural networks (CNNs), lacks the ability to understand sea clutter data, which exhibits spatiotemporal characteristics, making it difficult to meet the accuracy requirements for generating sample data. Summary of the Invention

[0004] The purpose of the present invention is to propose a sea clutter echo data sample generation method based on spatiotemporal joint perception to solve the problems of weak spatiotemporal characteristic perception and insufficient data accuracy when generating sea clutter data.

[0005] The technical solution for achieving the purpose of the present invention is: a method for generating sea clutter data samples based on spatiotemporal joint perception, comprising the following steps:

[0006] Step 1: Obtain the original sea clutter data set and remove abnormal units; obtain the amplitude distribution statistical characteristics and correlation length of the pure sea clutter data after removing the abnormal units; obtain impulse noise close to the actual sea clutter distribution based on the amplitude distribution statistical characteristics; use a sliding window to extract data from the pure sea clutter data to form independent samples, where the sample size is selected according to the correlation length; and convert the samples into dual-channel real samples;

[0007] Step 2: Build a Transformer-based sea clutter generator and discriminator; the generator and discriminator are both composed of multiple Transformer encoding blocks stacked together;

[0008] Step 3: Input the impulse noise obtained in step 1 as prior information into the generator, and collaboratively train the generator and discriminator based on the dual-channel real samples obtained in step 1 until the generator and discriminator reach a Nash equilibrium state.

[0009] Step 4: Compare the sea clutter data generated by the trained generator with the real sea clutter data, calculate the statistical feature similarity, and compare the statistical feature similarity with the threshold. If the statistical feature similarity is less than the threshold, return to step 3 to adjust the network parameters until the statistical feature similarity is greater than the threshold.

[0010] Compared with the existing technology, the present invention has the following significant advantages: in order to solve the problems of complex sea clutter data features, single samples, and difficulty in capturing spatiotemporal features, the present invention models the real distribution of sea clutter, constructs a generative adversarial model based on the Transformer encoder, and establishes a mapping relationship from random noise to real data; the present invention can use a small amount of sea clutter data for mass generation, and has the advantages of high accuracy, easy training, and convenient migration. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0012] Figure 2 It is the data sampling method of the present invention;

[0013] Figure 3 It is the structural diagram of the generator and discriminator of the present invention;

[0014] Figure 4 It is a diagram of the spatiotemporal axial attention structure of the present invention;

[0015] Figure 5 It is a training process diagram of the present invention; DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0017] The present invention provides a method for generating sea clutter data samples based on spatiotemporal joint perception, which specifically includes the following steps:

[0018] Step 1: Obtain the original sea clutter data set and remove abnormal units; obtain the amplitude distribution statistical characteristics and correlation length of the pure sea clutter data after removing the abnormal units; obtain impulse noise close to the actual sea clutter distribution based on the amplitude distribution statistical characteristics; use a sliding window to extract data from the pure sea clutter data to form independent samples, and select the sample size based on the correlation length; and convert the samples into dual-channel real samples;

[0019] Sea clutter datasets often contain small or collaborative targets on the sea surface, as well as possible sea spikes. These cells exhibit significant amplitude differences from surrounding cells and are considered anomalous. In sea clutter research, these anomalous cells must be excluded to avoid affecting experimental results and obtain pure sea clutter data.

[0020] The amplitude characteristics of pure sea clutter data are analyzed to obtain the probability density distribution of the data. The probability density distribution is then fitted into different statistical distribution models. The distribution model with the best fitting degree is selected, and the parameters of the model are obtained. According to the parameters, the impulse noise is obtained through the inverse function transformation method.

[0021] The spatial correlation analysis of sea clutter data is performed, assuming that the sea clutter is spatially uniform, and the correlation length is estimated using the correlation coefficient formula:

[0022]

[0023] Where x(k) is the sea clutter sequence in the kth range unit, N is the total number of range units, * is the complex conjugate, are the squared modulus and mean square of the mean of the clutter sequence, respectively. p takes a value between -1 and +1. The correlation coefficient is large in the neighboring units of x(k), followed by a rapid decay process. When p(l) decays from 1 to 1 / e (i.e., 0.368), the sea clutter distance is considered no longer correlated. The number of range units at this point is the correlation length L of the sea clutter data.

[0024] A sliding window is used to extract data from the pure sea clutter data after removing abnormal units to form independent samples. Each time, a matrix of size (p, r) is taken as an independent sea clutter data sample, where r is the number of distance units, r is selected as n times the correlation length L, and p is the number of pulses, which is a flexibly settable parameter. The sliding window step size is S p The specific method of sample extraction is as follows: Figure 2 .

[0025] After the above operations, several sea clutter samples are obtained. At this time, these samples are in the form of complex numbers. The real and imaginary parts of the samples are separated and fused as two channels. After this operation, the form of the sea clutter samples becomes a dual-channel real number.

[0026] Step 2: Build a Transformer-based sea clutter generator and discriminator; the generator and discriminator are both composed of multiple Transformer encoding blocks stacked together;

[0027] In a further embodiment, the Transformer encoding block includes an axial attention module and a feedforward neural network module, the axial attention module performs attention calculations on the pulse and distance respectively to capture different features in the two dimensions of pulse-distance, the feedforward neural network module includes a GELU activation function and a feedforward multilayer perceptron, and a residual connection is used between the axial attention module and the feedforward neural network module;

[0028] In a further embodiment, the generator includes a linear layer, a position encoding layer, a Transformer encoder, and a convolutional layer with a convolution kernel size of (1, 1), which are connected in sequence. The Transformer encoder is composed of a plurality of stacked Transformer encoding blocks. Each Transformer encoding block includes an axial attention module and a multi-layer perceptron module.

[0029] The axial attention module divides attention into two parts: time and space, and performs attention calculations in the two dimensions of pulse and distance respectively. Figure 4 The implementation of the axial attention designed by the present invention is shown. That is, the attention of the two dimensions of pulse and distance are parallelized and fused in the final output. For a specific position (i, j), the output y of the axial attention is i,j The calculation is as follows:

[0030]

[0031]

[0032]

[0033] Among them, K1, V1 are the corresponding key value matrices along the pulse dimension, and K2, V2 are the corresponding key value matrices along the spatial unit dimension. 1 is the attention output in pulse dimension, y 2is the attention output of the distance dimension. Through axial attention, relevant information is efficiently captured in two dimensions, while reducing the computational complexity burden brought by calculating global attention. In terms of computation, attention is performed independently in two dimensions, reducing the computational complexity required for each position in the spatiotemporal vector from O(P×R) to O(P+R). In addition, the feature vector after the attention operation still retains global information and does not change the size of the input tensor. In order to enable the network to learn features of different dimensions, the two module parameters are not shared. The axial attention designed in this way allows the network to learn and capture feature information separately in the pulse dimension and the spatial dimension, with a certain degree of discrimination, thereby modeling and analyzing different dimensions of sea clutter data.

[0034] In a further embodiment, the discriminator comprises a positional encoding layer, a Transformer encoder, and a classification head, connected in sequence. The positional encoding layer and Transformer encoder are configured similarly to the generator structure. The classification head, comprised of a fully connected layer, performs a binary classification on the final output, mapping it to a true sample label and a generated sample label, respectively.

[0035] To improve network stability, a normalization layer is placed before each submodule to normalize the data. To reduce overfitting, a random elimination layer is added after each submodule. In addition, residual connections are used between submodules.

[0036] Step 3: Input the impulse noise obtained in step 1 as prior information into the generator, and collaboratively train the generator and discriminator based on the dual-channel real samples obtained in step 1; use the Wasserstein distance with a penalty gradient term as the loss function; train the generator and discriminator until they reach a Nash equilibrium state, and save the trained network model;

[0037] The training process is as follows Figure 5 As shown. The impulse noise obtained in step 1 is input into the generator with length N, and is mapped to an H through the linear layer. M After the dimensional space, a multidimensional vector is obtained, where H M is a hyperparameter that can be flexibly set; the multidimensional vector is input to the position encoding layer. The purpose of the position encoding layer is to retain the relative position information of different positions in the data through position encoding. Therefore, the position encoding layer first divides the vector into P N segments, and add corresponding position codes to each segment, where P N is a hyperparameter that can be flexibly configured. The vector is then input into the Transformer encoder, and after the encoder's operation, it is output to a convolutional layer. This convolutional layer resizes the input vector to the same size as the real sample, ultimately generating the sea clutter data sample.

[0038] The sea clutter data samples and the dual-channel real samples obtained in step 1 are used as the training set of the discriminator. Samples are taken from the training set and input into the discriminator. After position encoding through the position encoding layer, feature capture is performed through the Transformer encoder, and finally, the true / false classification result is obtained through the classification head.

[0039] Subsequently, the discriminator updates its gradient according to the loss between this classification result and the true result. The loss function used is the Wasserstein GAN with a penalty gradient term, which is defined as:

[0040]

[0041] where D adv (x) is the discriminator probability distribution, x is the true sample, is the generated sample, λ gp is the joint distribution sample gradient penalty term, is a sample linearly sampled from a pair of real sea clutter data and generated sea clutter data, E x is to take the true sample distribution, is to take the generated sample distribution, is to take the sampled sample distribution, is regularization. The purpose of the discriminator is to reduce this loss to distinguish between generated samples and true samples. Subsequently, the generator will also update its gradient according to the loss between this result and the label. The purpose of the generator is to increase this loss to confuse generated samples and true samples.

[0042] By continuously repeating the above operations, the discriminator and the generator are alternately trained until Nash equilibrium, and then the trained model is saved.

[0043] Step 4: Take the sea clutter data generated by the trained generator for characteristic analysis, compare it with the real sea clutter data, calculate the statistical feature similarity S, set the statistical feature accuracy threshold St. If S < St, return to step 3 to fine-tune the network structure parameters until the condition S > St is satisfied.

[0044] Use the trained generator to generate sea clutter samples and extract the statistical features of the samples. Subsequently, extract the statistical features of the real data. Then, use the cosine distance to calculate the cosine similarity between the two feature vectors. The higher this score, the more similar the generated data and the real data. Set the threshold S t , when the similarity score is less than S t , return to step 3 to continue fine-tuning the network model. When the similarity score is greater than S t , stop training.

[0045] Specifically, the feature vectors of the generated data and the real data are extracted through the pre-trained network model. The feature vectors of the real data obey a distribution A, and the feature vectors of the generated data also obey a distribution B. The distance between the two distributions reflects the difference between the two data. The cosine similarity between the two distributions is calculated using the statistical feature cosine distance. The higher the score, the closer the two distributions are. The present invention takes the threshold St as 99.5. When the similarity score is less than 99.5, return to step 3 to continue fine-tuning the network model. When the similarity score is greater than 99.5, stop training.

[0046] The pulse noise is fed into the trained generator to generate sea clutter data samples.

[0047] The encoder in the present invention consists of two composite blocks. The first block is the attention module, which is used to capture the internal correlation in the input sequence. In order to match the spatiotemporal characteristics of sea clutter data, the present invention uses axial attention instead of the traditional attention mechanism. The second block consists of an activation function and a feedforward MLP, which are used to extract features and enhance representation capabilities. In order to improve the stability of the network, the present invention pre-places a batchnorm layer in each block to perform batch normalization on the data. In order to reduce overfitting, a dropout layer is added after each block. At the same time, residual connections are used between each block to allow information to flow better.

[0048] This paper splits the multi-head attention mechanism into two modules, allowing the network to learn the correlations between each dimension and unit separately, while avoiding the difficulty of computing global attention. The first module performs self-attention on the impulse dimension, and the second module performs self-attention on the spatial dimension. By adopting axial attention, relevant information can be efficiently captured in two dimensions while reducing the computational complexity burden of computing global attention. This approach enables the network to focus on important relationships within each dimension, improving the model's ability to handle sea clutter data.

[0049] Example

[0050] In order to verify the effectiveness of the present invention, the following analysis is conducted through experiments.

[0051] First, a data set is constructed. Abnormal units are removed from the existing sea clutter data set. To maintain data continuity, the amplitude distribution characteristics of the pure clutter data are analyzed after eliminating the abnormal units. The best fitting model is found to be the Weibull distribution. The parameters of the Weibull distribution model are also obtained. Based on these parameters, random noise with a distribution that conforms to the Weibull distribution is obtained by inputting them into the inverse cumulative distribution function of the Weibull model. This random noise is an impulse noise that is more consistent with the non-Gaussian characteristics of real sea clutter data.

[0052] A spatial correlation analysis was then performed on the clean sea clutter data, and the correlation length of the sea clutter in the dataset was calculated to be L. Sea clutter samples were then generated. To fully ensure spatial correlation, the data distance unit dimension was 5L. The pulse dimension was P, and the sliding window length was P / 2. This operation yielded sum true sea clutter samples.

[0053] In one embodiment, the generator is composed of a linear layer, a position encoding layer, a Transformer encoder and a convolutional layer, and the discriminator is composed of a position encoding layer, a Transformer encoder and a classification head. The Transformer encoder is composed of a plurality of stacked encoding blocks, each encoding block is composed of two sub-blocks, namely an axial attention block and a multi-layer perceptron. The attention module extracts and fuses the temporal information and spatial information of the data and then passes it to the feedforward multi-layer perceptron layer, wherein each sub-block is pre-batch normalized and post-randomly eliminated. In one embodiment, the random elimination rate is set to 0.5. The Transformer encoder in the generator is composed of the m encoding blocks mentioned above, and the Transformer encoder in the discriminator is composed of the n encoding blocks mentioned above. The final generator input is N-dimensional impulse noise, and the output is generated data of size (b, c, h, w). The discriminator input is a real sample and a generated sample of size (b, c, h, w), and the output is a 1-dimensional identification result.

[0054] According to the above steps, an adversarial generative network is constructed and the network is iteratively trained. In one embodiment, the Adam optimizer is selected and the learning rate is set to 0.0001. In order to balance the difference between the generator and the discriminator, the discriminator is first trained t times, and then the generator and the discriminator are trained alternately, and finally the trained model is saved. The 1-dimensional impulse noise is used as the input of the generator, and the generator will generate sea clutter data of size (b, c, h, w). In one embodiment, the similarity index threshold is set to S t , when the similarity score is less than S t When the similarity score is greater than S, return to step 3 to continue fine-tuning the network model. t When , the training stops. At this point, the generator can generate samples with a distribution similar to that of real sea clutter samples.

[0055] The above embodiments are provided for the purpose of describing the present invention only and are not intended to limit the scope of the present invention. The scope of the present invention is defined by the appended claims. Various equivalent substitutions and modifications made without departing from the spirit and principles of the present invention are intended to be within the scope of the present invention.

Claims

1. A method for generating sea clutter data samples based on spatiotemporal joint perception, characterized in that: The steps include: Step 1: Obtain an original sea clutter data set and remove abnormal units; obtain the amplitude distribution statistical characteristics and correlation length of the pure sea clutter data after removing the abnormal units. The specific method for obtaining the correlation length of the pure sea clutter data after removing the abnormal units is: Calculate the correlation coefficient estimate of the clean sea clutter data after removing the abnormal units. The specific formula is: Where x(k) is the sea clutter sequence of the kth range unit, x(k+l) is the sea clutter sequence of the k+lth range unit, l is a natural number, N is the total number of range units, * is the complex conjugate, are the square modulus and mean square value of the noise sequence mean respectively; The number of range cells when p(l) decays from 1 to 1 / e is the correlation length L of the sea clutter data; The impulse noise close to the real sea clutter distribution is obtained based on the statistical characteristics of the amplitude distribution; the data is extracted from the pure sea clutter data using a sliding window to form independent samples, where the sample size is selected according to the correlation length; and the samples are converted into dual-channel real samples; Step 2: Build a Transformer-based sea clutter generator and discriminator. The generator and discriminator are both composed of multiple stacked Transformer coding blocks. The generator includes a linear layer, a position coding layer, several stacked Transformer coding blocks, and a convolutional layer connected in sequence. The discriminator includes a position encoding layer, several stacked Transformer encoding blocks and a classification head connected in sequence; The Transformer encoding block includes an axial attention module and a feedforward neural network module. The axial attention module performs attention calculations on the pulse and distance respectively to capture different features in the two dimensions of pulse and distance. The feedforward neural network module includes a GELU activation function and a feedforward multilayer perceptron. A residual connection is used between the axial attention module and the feedforward neural network module. Step 3: Input the impulse noise obtained in step 1 as prior information into the generator. Based on the dual-channel real samples obtained in step 1, the generator and discriminator are trained collaboratively until the generator and discriminator reach the Nash equilibrium state. The specific method is as follows: The impulse noise obtained in step 1 is input into the generator with length N, and is mapped to an H through the linear layer. M Finally, a multi-dimensional vector is obtained in the dimensional space; The multidimensional vector is input to the position encoding layer, which divides the vector into P N segments and add corresponding position codes to each segment; The position-encoded vector is input into the Transformer encoder, and after the encoder operation, it is output to a convolutional layer. The convolutional layer changes the size of the input vector to the same size as the real sample to generate sea clutter data samples. The sea clutter data samples and the dual-channel real samples obtained in step 1 are used as the training set of the discriminator. Samples are taken from the training set and input into the discriminator. After position encoding by the position encoding layer, the features are captured by the Transformer encoder, and the true and false classification results are obtained by the classification head. Perform gradient updates on the discriminator based on the loss between the classification result and the true result; By repeating the above operations, the discriminator and generator are alternately trained to Nash equilibrium, and then the trained model is saved; Step 4: Compare the sea clutter data generated by the trained generator with the real sea clutter data, calculate the statistical feature similarity, and compare the statistical feature similarity with the threshold. If the statistical feature similarity is less than the threshold, return to step 3 to adjust the network parameters until the statistical feature similarity is greater than the threshold.

2. The method for generating sea clutter data samples based on spatiotemporal joint perception according to claim 1, characterized in that: The abnormal units include target units and sea peaks.

3. The method for generating sea clutter data samples based on spatiotemporal joint perception according to claim 1, characterized in that: The specific method for obtaining the amplitude distribution statistical characteristics of the pure sea clutter data after removing the abnormal units and obtaining the impulse noise close to the real sea clutter distribution based on the amplitude distribution statistical characteristics is as follows: The amplitude characteristics of the pure sea clutter data after removing abnormal units are analyzed to obtain the probability density distribution of the data. The probability density distribution is then fitted into different statistical distribution models. The distribution model with the best fitting degree is selected, and the parameters of the optimal distribution model are obtained. According to the parameters of the optimal distribution model, the impulse noise is obtained through the inverse function transformation method.

4. The method for generating sea clutter data samples based on spatiotemporal joint perception according to claim 1, characterized in that: The output y of the axial attention module for position (i, j) i,j Specifically: Among them, K1, V1 are the corresponding key-value matrices along the pulse dimension, K2, V2 are the corresponding key-value matrices along the spatial unit dimension, is the attention output in pulse dimension, is the attention output of the distance dimension.

5. The method for generating sea clutter data samples based on spatiotemporal joint perception according to claim 1, characterized in that: The loss function used is Wasserstein GAN with a penalty gradient term, specifically: Among them, D adv (x) is the probability distribution of the discriminator, x is the real sample, To generate samples, λ gp Joint distribution sample gradient penalty, 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, For regularization.

6. The method for generating sea clutter data samples based on spatiotemporal joint perception according to claim 1, characterized in that: The cosine distance is used to calculate the statistical feature similarity between two feature vectors.

Images