A sea clutter modeling and suppression method driven by both data and models

Abstract

The present invention discloses a sea clutter modeling and suppression method under dual drive of data model, which belongs to the fields of radar and signal processing technology and artificial intelligence technology. The method comprises the following steps: building a generative adversarial network GAN as a driving model of a sea clutter modeling algorithm, and using Gaussian white noise data and sea clutter data as driving data of the sea clutter modeling algorithm; training the model of the generator and the discriminator in the GAN; building a sea clutter suppression model based on a convolutional neural network CNN, and using the data generated by the GAN model as the input data set of the CNN to train the CNN model; measuring the performance of the model, using the mean square error MSD test to measure the sea clutter modeling simulation effect of the GAN model, and testing the sea clutter suppression effect of the CNN model by comparing the time-frequency spectrum of the radar receiving signal before and after using the CNN model. The method provided by the present invention can more effectively simulate and model sea clutter, and can suppress the sea clutter component in the radar receiving signal.

Description

Technical Field

[0001] The present invention belongs to the field of radar and signal processing technology and artificial intelligence technology, and in particular relates to a sea clutter modeling and suppression method under dual drive of data models. Background Art

[0002] The working environment faced by ocean radar is complex and changeable. In addition to the required target echo signal, the radar echo signal it receives also contains noise and sea clutter components. Since sea clutter has a high power level, the sea clutter component has become one of the main factors affecting the working performance of ocean radar. Sea clutter has non-Gaussian and non-stationary characteristics, so modeling and simulation of sea clutter is an important basis for further studying the suppression of sea clutter components in radar received signals.

[0003] At present, there are two main analysis ideas for the research and analysis of sea clutter, namely, based on statistical distribution and physical characteristics. The methods based on statistical distribution are mainly based on models, such as lognormal distribution, Weibull distribution and K distribution. Since the K distribution model can simultaneously consider the amplitude distribution characteristics and pulse-to-pulse correlation performance of sea clutter, it can reflect the statistical characteristics of sea clutter more accurately than other models. Conte and Marier simulated sea clutter using spherical invariant random process SIRP and zero memory nonlinear transformation method ZMNL respectively. The above two methods are currently the most widely used sea clutter simulation methods. Ritchie obtained the false alarm probability by calculating the cumulative amplitude distribution of sea clutter to evaluate the statistics of sea clutter; Watts used the measurement results of Doppler spectrum records to analyze the main characteristics such as average power spectrum density and Doppler spectrum extreme amplitude, and proposed a sea clutter amplitude statistical model based on composite K distribution, which was verified by a large amount of sea clutter test data; Weinberg used the weighted product of independent Gaussian vectors to construct in-phase orthogonal components, and proposed a KK distribution sea clutter model with simpler calculation. Methods based on physical properties usually analyze the chaotic or fractal characteristics of sea clutter. Haykin et al. studied and analyzed the embedding dimension and Lyapunov exponent of measured sea clutter, and proposed a small target detection algorithm based on the combination of chaotic phase space reconstruction and back propagation BP neural network; subsequently, Leung used a nonlinear prediction method based on correlation dimension and memory bank to detect target signals; Cui Wanzhao et al. predicted chaotic time series through support vector machine SVM, realized target detection and was widely promoted. Unswoorth et al. found that sea clutter does not strictly have chaotic characteristics; Hu et al. used fractional Brownian motion to verify that sea clutter has fractal characteristics, and used Hurst exponent for marine target detection, which was recognized by most scholars and became a hot topic in subsequent research. In China, Xing Hongyan et al. used the chaotic and fractal characteristics of sea clutter to propose multi-fractal target detection algorithms based on SVM and attenuated wave analysis, respectively, and continuously improved and perfected them; Li Zhengzhou et al. used radial basis function neural network and spatial and temporal chaotic reconstruction for small and weak target detection; Liu Ningbo et al. conducted a lot of analysis and research on frequency domain fractal characteristics and combination characteristics. The above literature analyzes and studies sea clutter from the perspectives of statistical distribution and physical characteristics, and both have great reference value. However, the sea clutter modeling based on statistical distribution does not take into account the physical characteristics of sea clutter, and it is difficult to reveal its inherent dynamic characteristics; the method based on physical characteristics uses the difference in fractal parameters between sea clutter with small targets and pure sea clutter to achieve small target detection under the sea clutter background, but fails to establish a sea clutter model under different sea conditions, which is inconvenient for simulation evaluation.

[0004] The data-model dual-driven modeling method is generally used to deal with situations where the object mechanism is unclear or relatively complex. Linear regression methods, SVM methods, and neural network methods can be used. With the introduction of generative adversarial networks, more and more generative tasks are beginning to use GAN. At present, the accumulation of a large amount of measured sea clutter data provides a reliable driving data foundation for the use of GAN for sea clutter modeling. Summary of the invention

[0005] Aiming at the problem that the existing methods cannot fully characterize the characteristics of sea clutter, thus seriously affecting the effect of sea clutter simulation modeling, the present invention provides a sea clutter modeling and suppression method driven by both data and model, aiming to solve the problem of making full use of the statistical distribution characteristics of sea clutter theoretical data and the physical characteristics of sea clutter measured data to model and simulate sea clutter, and generate a large number of scalable sea clutter data sets.

[0006] To solve the above problems, the present invention adopts the following technical solutions.

[0007] A sea clutter modeling and suppression method driven by both data and models comprises the following steps:

[0008] Step 1: Build a GAN network as the driving model for sea clutter modeling, and use Gaussian white noise data, sea clutter simulation data that meets K distribution generated by the spherically invariant random process method SIRP method, and measured sea clutter data as driving data for sea clutter modeling;

[0009] Step 2: Perform model training of the GAN generator and discriminator;

[0010] Step 3: Build a CNN-based sea clutter suppression model and use the sea clutter data generated by the GAN model as the input data set of CNN;

[0011] Step 4: Conduct CNN model training;

[0012] Step 5: Measure the model performance. Use the MSD test to measure the sea clutter modeling and simulation effect of the GAN model. By comparing the time-frequency spectrum of the radar received signal before and after using the CNN model, the sea clutter suppression effect of the CNN model is tested.

[0013] Furthermore, the GAN network constructed in step 1 is used as a driving model for sea clutter modeling, and Gaussian white noise data, sea clutter simulation data satisfying K distribution generated by spherically invariant random process method SIRP method, and measured sea clutter data are used as driving data for sea clutter modeling;

[0014] The overall network structure of the driving model GAN ​​is as follows Figure 2The driving data needs to be preprocessed using short-time Fourier transform (STFT). The specific calculation formula of STFT is:

[0015]

[0016] Where y(u) is the initial signal, g(u) is the window function, the superscript * is the complex conjugate, t is the center position of the window function, u is the time domain time, and f is the frequency; STFT y (t, f) is the STFT of the initial signal y(u). Through the calculation formula, the above three data are respectively STFTed and their respective time-frequency spectrum data sets are obtained. These three data sets are used as the driving data of this article.

[0017] Furthermore, the model training of the GAN generator and discriminator is carried out as described in step 2. The specific structures of the discriminator and generator are as follows Figure 3 As shown in the figure, the specific training method of the GAN model is divided into the following two parts:

[0018] Training 1: Train the discriminator alone

[0019] 1) 500 samples of white noise satisfying Gaussian distribution and 500 samples of sea clutter simulation satisfying K distribution are used to obtain time-frequency spectrum through STFT;

[0020] 2) The above 1000 time-frequency spectrum image samples are mixed and constructed into a training set and a test set in a ratio of 8:2, the image label of Gaussian white noise is set to 0, and the image label of sea clutter simulation data is set to 1;

[0021] 3) The training set is used as the input R of the discriminator, and the discriminator is trained until the discriminator converges. Then the test set is used as the input R of the discriminator to test the convergence effect of the discriminator.

[0022] Training 2: Training the generator and discriminator simultaneously

[0023] 1) Obtain the time-frequency spectrum of 500 measured sea clutter data through STFT;

[0024] 2) The above 500 time-frequency spectrum image samples are constructed into a data set, and the image label is set to 1;

[0025] 3) The data set is used as the input R of the discriminator, and the random sequence that satisfies the Gaussian distribution is used as the input Z of the generator;

[0026] 4) The training iterations of the discriminator and the generator are performed alternately in the order of first training the discriminator once and then training the generator once. After the training is completed, the time-frequency spectrum image of the sea clutter generated by the generator is subsequently compared with the measured time-frequency spectrum image of the sea clutter for simulation.

[0027] Furthermore, the CNN-based sea clutter suppression model described in step 3 is constructed, and the sea clutter data generated by the GAN model is used as the input data set of the CNN. The overall network structure of the CNN is as follows: Figure 4 shown.

[0028] Further, the CNN model training is performed as described in step 4. The specific structure of the CNN network is as follows: Figure 5 The specific training method is as follows:

[0029] 1) The 500 time-frequency spectra of sea clutter generated by the GAN generator are added to the time-frequency spectra of target echoes at three different Doppler frequencies to obtain the data sets of radar received signals at three different Doppler frequencies. The training set and test set are constructed for each data set in a ratio of 8:2, and the label is set to the time-frequency spectra of the target echo at the Doppler frequency corresponding to each data set;

[0030] 2) The training set is input into CNN for iterative training. After the training is completed, the test set is input to obtain the time-frequency spectrum of the radar received signal after the sea clutter component is suppressed.

[0031] Furthermore, the model performance is measured as described in step 5, and the MSD test is used to measure the sea clutter modeling simulation effect of the GAN model. By comparing the time-frequency spectrum of the radar receiving signal before and after using the CNN model, the sea clutter suppression effect of the CNN model is tested;

[0032] The formula for the MSD test is defined as

[0033]

[0034] Among them, p e (x k ) is the probability distribution function PDF of the measured sea clutter data, p t (x k ) is the PDF of the sea clutter model generated by the algorithm, and n represents the length of the selected data sequence. The smaller the MSD value, the greater the similarity between the two distributions, and the better the prediction effect of the algorithm model.

[0035] Beneficial Effects

[0036] The sea clutter modeling and suppression method under dual data model driving of the present invention has the following advantages:

[0037] 1. The driving data used in the sea clutter modeling simulation of the present invention fully utilizes the statistical distribution characteristics of the sea clutter theoretical data and the physical characteristics of the sea clutter measured data;

[0038] 2. The sea clutter simulation modeling method of the present invention has better fitting degree than the sea clutter simulation algorithm based on measured data, the sea clutter simulation algorithm based on back propagation neural network and the sea clutter simulation algorithm based on gated recurrent neural network in terms of fitting measured data, indicating that the sea clutter data generated by the present invention has better scalability;

[0039] 3. The present invention can generate a large number of scalable sea clutter data sets, providing a reliable and massive sea clutter basic data set for subsequent research directions such as sea clutter suppression and target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the sea clutter modeling and suppression method under the dual drive of the data model of the present invention;

[0041] Figure 2 The overall network structure of the driving model GAN ​​of the present invention;

[0042] Figure 3 The specific structures of the discriminator and the generator of the present invention;

[0043] Figure 4 The overall network structure of the CNN of the present invention;

[0044] Figure 5 The specific network structure of the CNN of the present invention;

[0045] Figure 6 is the loss function of the discriminator of the present invention;

[0046] Figure 7 The output comparison of the generator under different training rounds of the present invention;

[0047] Figure 8 is the loss function of the CNN of the present invention;

[0048] Fig. 9 The figure is a time-frequency spectrum comparison before and after the sea clutter suppression of the present invention. DETAILED DESCRIPTION

[0049] In order to better understand the purpose, structure and function of the present invention, a sea clutter modeling and suppression method under dual-drive of a data model of the present invention is further described in detail below with reference to the accompanying drawings.

[0050] Example 1

[0051] This embodiment provides a method for modeling and suppressing sea clutter under dual drive of data and model. Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5As shown, the following steps are included:

[0052] Step 1: Build a GAN network as the driving model for sea clutter modeling, and use Gaussian white noise data, sea clutter simulation data that meets K distribution generated by the spherically invariant random process method SIRP method, and measured sea clutter data as driving data for sea clutter modeling;

[0053] The overall network structure of the driving model GAN ​​is as follows Figure 2 The driving data needs to be preprocessed using short-time Fourier transform (STFT). The specific calculation formula of STFT is:

[0054]

[0055] Where y(u) is the initial signal, g(u) is the window function, the superscript * is the complex conjugate, t is the center position of the window function, u is the time domain time, and f is the frequency; STFT y (t, f) is the STFT of the initial signal y(u). Through the calculation formula, the above three data are respectively STFTed and their respective time-frequency spectrum data sets are obtained. These three data sets are used as the driving data of this article.

[0056] Step 2: Perform model training of the GAN generator and discriminator;

[0057] The specific structures of the discriminator and generator are as follows Figure 3 As shown. In each layer of the network, I represents the input data dimension, O represents the output data dimension, k represents the size of the convolution kernel, and s represents the moving step of the convolution kernel. The specific training method of the GAN model is divided into the following two parts:

[0058] Training 1: Train the discriminator alone

[0059] 1) 500 samples of white noise satisfying Gaussian distribution and 500 samples of sea clutter simulation satisfying K distribution are used to obtain time-frequency spectrum through STFT;

[0060] 2) The above 1000 time-frequency spectrum image samples are mixed and constructed into a training set and a test set in a ratio of 8:2, the image label of Gaussian white noise is set to 0, and the image label of sea clutter simulation data is set to 1;

[0061] 3) The training set is used as the input R of the discriminator, and the discriminator is trained until the discriminator converges. Then the test set is used as the input R of the discriminator to test the convergence effect of the discriminator.

[0062] Training 2: Training the generator and discriminator simultaneously

[0063] 1) Obtain the time-frequency spectrum of 500 measured sea clutter data through STFT;

[0064] 2) The above 500 time-frequency spectrum image samples are constructed into a data set, and the image label is set to 1;

[0065] 3) The data set is used as the input R of the discriminator, and the random sequence that satisfies the Gaussian distribution is used as the input Z of the generator;

[0066] 4) The training iterations of the discriminator and the generator are performed alternately in the order of first training the discriminator once and then training the generator once. After the training is completed, the time-frequency spectrum image of the sea clutter generated by the generator is subsequently compared with the measured time-frequency spectrum image of the sea clutter for simulation.

[0067] Step 3: Build a sea clutter suppression model based on CNN, and use the sea clutter data generated by the GAN model as the input data set of CNN; the overall network structure of CNN is as follows: Figure 4 shown.

[0068] Step 4: Conduct CNN model training;

[0069] The specific structure of CNN network is as follows Figure 5 The specific training method is as follows:

[0070] 1) The 500 time-frequency spectra of sea clutter generated by the GAN generator are added to the time-frequency spectra of target echoes at three different Doppler frequencies to obtain the data sets of radar received signals at three different Doppler frequencies. The training set and test set are constructed for each data set in a ratio of 8:2, and the label is set to the time-frequency spectra of the target echo at the Doppler frequency corresponding to each data set;

[0071] 2) The training set is input into CNN for iterative training. After the training is completed, the test set is input to obtain the time-frequency spectrum of the radar received signal after the sea clutter component is suppressed.

[0072] Step 5: Measure the model performance. Use the MSD test to measure the sea clutter modeling and simulation effect of the GAN model. By comparing the time-frequency spectrum of the radar received signal before and after using the CNN model, the sea clutter suppression effect of the CNN model is tested.

[0073] The formula for the MSD test is defined as

[0074]

[0075] Among them, p e (x k ) is the probability distribution function PDF of the measured sea clutter data, p t (x k) is the PDF of the sea clutter model generated by the algorithm, and n represents the length of the selected data sequence. The smaller the MSD value, the greater the similarity between the two distributions, and the better the prediction effect of the algorithm model.

[0076] First, the GAN model is trained with simulation training 1. The convergence of the discriminator loss function is as follows: Figure 6 As shown. It can be found that the discriminator has converged.

[0077] After the discriminator converges, the GAN model is trained with simulation training 2. When the values ​​of the training round Epoch are {1, 10, 20, 30, 40, 50}, the time-frequency diagram of the sea clutter data generated by the generator is as follows: Figure 7 As shown in Figure 2, it can be found that the generator's ability to learn and express the characteristics of sea clutter is constantly improving.

[0078] The 500 sets of measured data used in this paper to train GAN are also used as input data for three sea clutter simulation algorithms: a sea clutter simulation algorithm based on measured data, a sea clutter simulation algorithm based on a back propagation neural network BPNN, and a sea clutter simulation algorithm based on a gated recurrent neural network GRNN. 500 sets of sea clutter simulation data are generated for each algorithm.

[0079] 500 sets of measured data were taken as the control group to compare and analyze the MSD test results of the proposed algorithm and the above three algorithms. The average results of the MSD test of the four algorithms and the control group are shown in Table 1. It can be found that the MSD average value of the proposed algorithm is the smallest, indicating that the sea clutter simulation data generated by the proposed algorithm has the highest fit with the measured data. Since the sea clutter data generated by these four algorithms are not generated by the control group, it also shows that the sea clutter simulation algorithm in this paper can fit more measured sea clutter data and has the best scalability.

[0080] Table 1 MSD comparison between sea clutter simulation algorithm and control group

[0081] algorithm This article's algorithm Measured data algorithm BPNN algorithm GRNN algorithm MSD <![CDATA[1.28×10 -5 ]]> <![CDATA[5.35×10 -4 ]]> <![CDATA[4.49×10 -5 ]]> <![CDATA[1.62×10 -5 ]]>

[0082] The sea clutter data generated by the algorithm in this paper is used to construct a radar receiving signal data set and simulate the CNN training. Figure 8 This is the convergence of the loss function during CNN training at three different Doppler frequencies.

[0083] After completing the CNN network training, the input test set can suppress the sea clutter component in the spectrum when the radar receives the signal. Fig. 9 The following is a comparison of the sea clutter components before and after suppression at three different Doppler frequencies. It can be found that the CNN model in this paper can effectively suppress the sea clutter components in the radar receiving signal.

[0084] The examples described in the present invention are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should all fall within the protection scope of the present invention.

Claims

1. A sea clutter modeling and suppression method driven by both data and model, characterized in that: The following steps are involved: Step 1: Build a GAN network as the driving model for sea clutter modeling, and use Gaussian white noise data, sea clutter simulation data that meets K distribution generated by the spherically invariant random process method SIRP method, and measured sea clutter data as driving data for sea clutter modeling; Step 2: Perform model training of the GAN generator and discriminator; Step 3: Build a CNN-based sea clutter suppression model and use the sea clutter data generated by the GAN model as the input data set of CNN; Step 4: Conduct CNN model training; Step 5: Measure the model performance. Use the MSD test to measure the sea clutter modeling and simulation effect of the GAN model. Compare the time-frequency spectrum of the radar receiving signal before and after using the CNN model to test the sea clutter suppression effect of the CNN model. The overall structure of the driving model GAN ​​network in step 1 includes a generator G and a discriminator D. The input of the generator G is Z. After Z is input into the generator G, sea clutter data G(Z) that does not exist in the real world will be generated, and then the discriminator D will judge the authenticity of the sea clutter data. Through continuous network iterative training, the generator G and the discriminator D will update their network parameters respectively to minimize the loss function, and finally reach a Nash equilibrium state. At this time, the GAN model reaches the optimal state, and the sea clutter data generated by the generator G is closest to the real sea clutter data. The driving data needs to be preprocessed using the short-time Fourier transform STFT. The calculation formula of STFT is as follows: Where y(u) is the initial signal, g(u) is the window function, the superscript * is the complex conjugate, t is the center position of the window function, u is the time domain time, and f is the frequency; STFT y (t,f) is the STFT of the initial signal y(u); through the calculation formula, STFT is performed on the three driving data respectively and their respective time-frequency spectrum data sets are obtained. These three driving data sets are used as driving data for sea clutter modeling.

2. The method for sea clutter modeling and suppression under dual-drive of data and model according to claim 1 is characterized in that: The discriminator in step 2 includes a convolutional layer and a fully connected layer, in which a Dropout layer is added to prevent overfitting. The generator includes a fully connected layer and a deconvolution layer. The specific training method of the GAN model is divided into the following two parts: Training 1: Train the discriminator alone; 1) 500 samples of white noise satisfying Gaussian distribution and 500 samples of sea clutter simulation satisfying K distribution are used to obtain time-frequency spectrum through STFT; 2) 1000 time-frequency spectrum image samples were mixed and constructed into training set and test set in a ratio of 8:2, the image label of Gaussian white noise was set to 0, and the image label of sea clutter simulation data was set to 1; 3) The training set is used as the input R of the discriminator, and the discriminator is trained until the discriminator converges. Then the test set is used as the input R of the discriminator to test the convergence effect of the discriminator. Training 2: Train the generator and discriminator simultaneously; 1) Obtain the time-frequency spectrum of 500 measured sea clutter data through STFT; 2) The above 500 time-frequency spectrum image samples are constructed into a data set, and the image label is set to 1; 3) The data set is used as the input R of the discriminator, and the random sequence that satisfies the Gaussian distribution is used as the input Z of the generator; 4) The training iterations of the discriminator and the generator are performed alternately in the order of first training the discriminator once and then training the generator once. After the training is completed, the time-frequency spectrum image of the sea clutter generated by the generator is subsequently compared with the measured time-frequency spectrum image of the sea clutter for simulation.

3. The method for sea clutter modeling and suppression under dual-drive of data and model according to claim 1 is characterized in that: The CNN network described in step three includes a data input layer, a convolution layer and a deconvolution layer; the input data set passes through the convolution layer to extract the data features in the time-frequency spectrum, and then passes through the deconvolution layer to reconstruct the features of the time-frequency spectrum.

4. The method for sea clutter modeling and suppression under dual-drive of data and model according to claim 1 is characterized in that: The CNN model training described in step 4 is performed. CNN includes two convolutional layers and three deconvolutional layers. In order to speed up the network training, a batch normalization layer is added after the convolutional layer and the deconvolutional layer. The specific training method is as follows: 1) The 500 time-frequency spectra of sea clutter generated by the GAN generator are added to the time-frequency spectra of target echoes at three different Doppler frequencies to obtain the data sets of radar received signals at three different Doppler frequencies. The training set and test set are constructed for each data set in a ratio of 8:2, and the label is set to the time-frequency spectra of the target echo at the Doppler frequency corresponding to each data set; 2) The training set is input into CNN for iterative training. After the training is completed, the test set is input to obtain the time-frequency spectrum of the radar received signal after the sea clutter component is suppressed.

5. The method for sea clutter modeling and suppression under dual-drive of data and model according to claim 1 is characterized in that: The formula for the MSD test described in step 5 is defined as Among them, p e (x k ) is the probability distribution function of the measured sea clutter data, p t (x k ) is the probability distribution function of the sea clutter model generated by the algorithm, and n represents the length of the selected data sequence; the smaller the MSD value, the greater the similarity between the two distributions, and the better the prediction effect of the algorithm model.

Images