A sea clutter data generation method based on a generative adversarial network
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTHWESTERN POLYTECHNICAL UNIV
- Filing Date
- 2024-04-16
- Publication Date
- 2026-08-07
AI Technical Summary
[0006]针对现有海杂波数据短缺这一问题，本发明的目的在于提供一种基于生成式对抗网络的海杂波数据生成方法，在海杂波数据有限的情况下，结合时域、频域联合注意力机制，生成大量与真实杂波样本分布特性一致的高质量复杂波样本，在一定程度上解决对海目标探测样本缺失的难题
Smart Images
Figure CN118410839B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar signal processing, and in particular to a generative adversarial network method for generating sea clutter based on a joint attention mechanism in the time and frequency domains. Background Technology
[0002] As an active microwave target detection device, radar plays an irreplaceable role in maritime three-dimensional surveillance systems. Unlike ground and air detection systems, maritime target detection operates in a look-down mode, resulting in radar echoes filled with a large amount of clutter. This clutter has a high power level and exhibits significant non-Gaussian, nonlinear, and non-stationary characteristics, making sea clutter a key limiting factor affecting radar detection performance.
[0003] Acquiring sea clutter data is a fundamental and urgent problem to be solved in marine detection technology. Based on different methods of acquiring clutter data, existing methods can be broadly divided into theoretical modeling and measurement experiments. Theoretical clutter modeling is based on electromagnetic scattering theory and generates clutter data with specific distributions through numerical calculations. Measurement experiments use radar systems to collect actual clutter data under different scenarios and radar parameters. However, both methods have certain limitations: 1) Because sea clutter is a complex function influenced by multiple factors, theoretical modeling methods struggle to accurately reproduce the true characteristics of clutter; 2) Data analysis results obtained from sporadic experiments have significant limitations, lacking targeted, long-term, systematic, continuous, and in-depth sea clutter measurement experiments and characteristic studies, making it difficult to obtain large amounts of high-quality clutter data.
[0004] Generative Adversarial Networks (GANs) offer a solution by generating sample data with a distribution consistent with real data using a small amount of sample data. Currently, GANs are mainly used for generating image and video data. However, the clutter data in this project differs from image data in the following two ways: 1) Clutter data is a complex signal. Existing methods for processing complex signals mainly train the real and imaginary parts separately, but this destroys the phase information of the complex signal; 2) Clutter data is mainly one-dimensional signals in the time or frequency domains, while existing GANs are mainly applicable to two-dimensional image data, resulting in a difference in network model structure. Based on this, this invention mainly designs a GAN network model suitable for complex signals. Furthermore, based on traditional one-dimensional time or frequency domain analysis, a joint attention mechanism in the time and frequency domains is introduced to deeply mine the joint features of clutter in the time and frequency domains, improving the realism of the generated clutter data under small sample sizes. Summary of the Invention
[0005] The technical problem to be solved by this invention is:
[0006] To address the problem of insufficient existing sea clutter data, the present invention aims to provide a sea clutter data generation method based on generative adversarial networks. In the case of limited sea clutter data, by combining time-domain and frequency-domain joint attention mechanisms, a large number of high-quality complex wave samples with distribution characteristics consistent with real clutter samples are generated, which to some extent solves the problem of missing sea target detection samples.
[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0008] A method for generating sea clutter data based on generative adversarial networks, characterized by comprising:
[0009] Acquire sea clutter data, filter the sea clutter data, and use the filtered sea clutter data as a training dataset;
[0010] The training dataset is input into a pre-trained generative adversarial network for training, the generative adversarial network including a generator network and a discriminator network;
[0011] The generator network consists of the following: the first fully connected layer takes n-dimensional random noise as input and expands it with a random number sequence; then deconvolution is performed with a stride of 1, batch normalization (BN) is used to standardize the data, ReLU function is used to accelerate convergence, and the corresponding feature mapping sequence is output; the next two fully connected layers map the feature sequence to the same dimension as the generated data; finally, the Tanh activation function is used to output simulated sea clutter data.
[0012] The discriminator network comprises one convolutional module and one fully connected module; the convolutional module includes one convolutional layer, a batch normalization layer, and a LeakyReLU activation function; the fully connected module includes two fully connected layers and a Sigmoid activation function.
[0013] Sea clutter data is generated using a pre-trained generative adversarial network.
[0014] A further technical solution of the present invention: the filtering of the sea clutter data, and the use of the filtered sea clutter data as a training dataset, specifically involves:
[0015] The measured sea clutter time-domain data were filtered, and the sea clutter data that showed obvious targets after pulse compression were deleted, resulting in sea clutter data without obvious targets, which was used as the network training dataset.
[0016] The selected sea clutter time-domain data is subjected to Fast Fourier Transform to obtain its frequency-domain data. Paired data of real and imaginary parts are extracted from the one-dimensional time / frequency domain data of a single pulse and integrated into two-dimensional data as input to the discrimination network. At the same time, a random array of the same size is generated using random numbers as input to the generation network.
[0017] A further technical solution of the present invention: before the generator network is input, the one-dimensional noise is transformed to n dimensions and then input into a fully connected circuit. After the output, the real and imaginary parts of the generated frequency domain data and time domain data are extracted and merged, and then the time domain and frequency domain data are stored separately.
[0018] A further technical solution of the present invention: the discriminator network straightens the two-dimensional time-domain data into one-dimensional data before input.
[0019] A further technical solution of the present invention: the input and output of the last two fully connected layers of the generator network are set to (1024, 512) and (512, 256), respectively.
[0020] A further technical solution of the present invention: adversarial training is performed on the generator network and the discriminator network. The training objective of the generator network is to minimize the objective function, that is, to make the probability distribution of the generated data approximate the distribution of the real data x. The training objective of the discriminator network is to maximize the objective function, that is, to distinguish the generated data from the real data with the highest probability. The loss function is as follows:
[0021] min G max D V(D,G)=E x～pdata(x) [logD(x)]+E z～p(z) [log(1-D(G(z)))].
[0022] That is, the loss function of the generator:
[0023] Loss G =E z～p ( z) [1-D(G(z))]
[0024] The loss function of the discriminator:
[0025]
[0026] Where z is the input random noise of the generator, G(z) is the output of the generator (the generated fake sample), D(x) is the discrimination probability of the discriminator against the real sample, D(G(z)) is the discrimination probability of the discriminator against the generated sample, and p(z) is the distribution of the random noise z. data (x) is the distribution of the real sample x.
[0027] A further technical solution of the present invention: During the training process, the two networks are trained alternately in a loop. The generator network is used to generate simulated samples, and the discriminator network judges the quality of the samples generated by the generator network by learning from real data. The two networks reach equilibrium in the game until the discriminator network can no longer distinguish the image generated by the generator network.
[0028] A computer system is characterized by comprising: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method described above.
[0029] A computer-readable storage medium is characterized by storing computer-executable instructions, which, when executed, are used to implement the above-described method.
[0030] A computer program product is characterized by including computer-executable instructions, which, when executed, are used to implement the above-described method.
[0031] The beneficial effects of this invention are as follows:
[0032] This invention provides a sea clutter generation method based on a generative adversarial network (GAN) with a joint attention mechanism in the time and frequency domains. Building upon existing GANs, a discrimination and generation model for two-dimensional sea clutter data is established. Fourier transform is performed on the measured sea clutter data to obtain its frequency domain information. The real and imaginary parts of the time / frequency domain signals are extracted separately, and the four one-dimensional data points are merged into a matrix form, essentially flattening the one-dimensional time-domain data into two-dimensional data to achieve joint time / frequency domain training. This method effectively solves the problem of uncorrelated amplitude and phase information in sea clutter data generated by one-dimensional networks. By analyzing the amplitude and spectral distribution characteristics of the sea clutter data generated by the generative model, the reliability of the generated sea clutter data is verified. It also verifies that this method can generate more diverse sea clutter data with distributions similar to real sea clutter data. Attached Figure Description
[0033] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.
[0034] Figure 1 This is a flowchart of the method of the present invention;
[0035] Figure 2 The network architecture diagram of the DCGAN network in this invention is shown below;
[0036] Figure 3 (a) shows the discrimination results of the discriminator at each epoch of the measured data and generated data in the example of the present invention;
[0037] Figure 3 (b) represents the loss values of the discriminator and generator at each epoch in this invention example;
[0038] Figure 4(a) is the time-domain portion of the slow-time dimension data generated by the generator in this embodiment of the invention;
[0039] Figure 4 (b) is the spectrum of the slow-time dimension data generated by the generator in the example of the present invention after performing FFT;
[0040] Figure 4 (c) is the time-domain portion of the slow time dimension in the actual test data used for training in this invention example;
[0041] Figure 4 (d) is the spectrum of the slow time dimension after FFT in the actual test data used for training in the example of this invention;
[0042] Figure 4 (e) is the result of MTD generated from the fake data in the example of this invention;
[0043] Figure 4 (f) is the result of the measured data after MTD in the example of this invention;
[0044] Figure 5 (a) is the result of amplitude distribution characteristic analysis of the generated fake data in the example of the present invention. After calculation of the error, it is found to be more consistent with the k-distribution.
[0045] Figure 5 (b) is the result of amplitude distribution characteristic analysis of measured data in the example of the present invention, and the calculated error is more consistent with the k-distribution;
[0046] Figure 5 (c) is the clutter power spectrum of the generated data in the example of the present invention;
[0047] Figure 5 (d) is the clutter power spectrum of the measured sea clutter data in the example of this invention. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0049] This invention provides a generative adversarial network (GAN) method for generating sea clutter based on a joint attention mechanism in the time and frequency domains. This method inputs a small amount of measured clutter time-domain and frequency-domain data into the network and combines a convolutional neural network (CNN) with the GAN to generate a large amount of simulated clutter data with a distribution consistent with the measured clutter data. The basic idea of the network is to learn the data distribution by having two neural networks compete against each other. One neural network, called the generator (G), takes random complex noise as its input to generate fake data similar to real data; the other neural network, called the discriminator (D), takes both real and generated clutter data as input to distinguish between them. The two networks are trained alternately and compete against each other, continuously adjusting their internal parameters, ultimately generating high-quality and diverse simulated sea clutter data with a distribution similar to real sea clutter data, thus solving the problem of difficult data acquisition.
[0050] like Figure 1 As shown, it includes the following steps:
[0051] Step 1: Sea clutter data acquisition and filtering
[0052] 1) Filter the measured sea clutter (time domain) data and delete the sea clutter data that shows obvious targets after pulse compression, so as to obtain sea clutter data without obvious targets, which can be used as the network training dataset.
[0053] 2) The frequency domain data is obtained by performing a Fast Fourier Transform on the selected sea clutter time domain data. Paired data of the real and imaginary parts are extracted from the one-dimensional time / frequency domain data of a single pulse, and these are integrated into two-dimensional data as input to the discriminant network. At the same time, a random array of the same size is generated using random numbers as input to the generator network. The data structure is shown in the attached figure. Figure 1 .
[0054] Step 2: Model Architecture Construction and Training
[0055] The improved deep generative adversarial network (DCGAN) model of this invention comprises two parts: a generator network G (Generator) and a discriminator network D (Discriminator). The structures of the generator and discriminator networks in the DCGAN model are shown in the appendix. Figure 2 The discriminant network consists of one convolutional layer and three fully connected layers. This model does not add pooling layers; instead, it uses convolutional layers with strides.
[0056] 1) Generator Network: The first fully connected layer takes n-dimensional random noise as input and expands it with a random number sequence. Then, deconvolution is performed with a stride of 1. Batch Normalization (BN) is used to standardize the data, and the ReLU function is used to accelerate convergence, outputting the corresponding feature mapping sequence. The input and output of the next two fully connected layers are set to (1024, 512) and (512, 256) respectively. These two fully connected operations map the feature sequence to the same dimension as the generated data. Finally, the Tanh activation function is used to output simulated sea clutter data. The first modification format transforms the one-dimensional noise to n dimensions before inputting it into the fully connected layer. The last modification format extracts the real and imaginary parts of the generated frequency domain data and time domain data, merges them, and then saves the time and frequency domain data separately.
[0057] 2) Discriminator Network: This consists of one convolutional module and one fully connected module. The convolutional module is similar to the generator, including one convolutional layer, a batch normalization layer, and a LeakyReLU activation function. The fully connected module includes two fully connected layers and a Sigmoid activation function. The "straighten" part at the beginning is correct because the measured data is two-dimensional time-domain data. Therefore, the discriminator network needs to straighten the two-dimensional time-domain data into one dimension before processing.
[0058] 3) Loss Function
[0059] Two networks, G and D, are trained adversarially. The training objective of G is to minimize the objective function, i.e., to make the probability distribution of generated data approximate the distribution of real data x. Conversely, the training objective of D is to maximize the objective function, i.e., to distinguish generated data from real data with the highest probability. The loss function is as follows:
[0060] min G max D V(D,G)=E x～pdata(x) [logD(x)]+E z～p(z) [log(1-D(G(z)))].
[0061] That is, the loss function of the generator:
[0062] Loss G =E z～p ( z) [1-D(G(z))]
[0063] The loss function of the discriminator:
[0064]
[0065] Where z is the input random noise of the generator, G(z) is the output of the generator (the generated fake sample), D(x) is the discrimination probability of the discriminator against the real sample, D(G(z)) is the discrimination probability of the discriminator against the generated sample, and p(z) is the distribution of the random noise z. data (x) is the distribution of the real sample x.
[0066] In an ideal situation, the generator's loss is... G The value should approach 0, indicating that the samples generated by the generator can fool the discriminator, making it unable to distinguish between real and generated samples. The discriminator's loss is... D It should also approach 0.5, indicating that the discriminator's probability of distinguishing between real and generated samples is close to 0.5, meaning it cannot clearly distinguish which are real samples and which are generated samples. Reaching this state means the generator and discriminator are in a state of equilibrium; the generator produces high-quality samples, and the discriminator cannot accurately distinguish between real and generated samples.
[0067] 4) Model training and parameter setting
[0068] The generator and discriminator networks are structured as shown above. During training, the two networks alternately train in a loop. The generator network generates simulated samples, while the discriminator network learns from real data to judge the quality of the samples generated by the generator network. The two networks reach equilibrium through this game, until the discriminator network can no longer distinguish the images generated by the generator network.
[0069] Generative and discriminative networks differ in their architecture and therefore their learning efficiencies. In the training of generative adversarial networks (GANs), the learning rate directly controls the magnitude of gradient updates, influencing the extent to which the loss function updates the network and consequently affecting the network model's effective tolerance. An excessively high learning rate can lead to oscillations in the training process (loss function), potentially preventing convergence to the optimal solution; conversely, an excessively low learning rate slows convergence and makes the network prone to getting trapped in local optima. Therefore, the Adam learning rate optimizer is created during initialization. The Adam algorithm assigns different weights to different gradients, enabling the neural network to converge quickly and stably to the optimal point when the learning rate is stable.
[0070] The experimental hardware environment used for training the DCGAN to generate sea clutter data was: NVIDIA GeForce RTX3090 GPU, CUDA 11.3, and 32GB of memory. The software environment consisted of PyTorch 1.12.0 and Python 3.9.
[0071] In this training model, the input data is standardized to zero mean and unit variance. The generator's latent variable z is sampled from N(0,1), and the MSE function is used as the activation function. During the training of the Generative Adversarial Network (GAN), the learning rate directly controls the magnitude of gradient updates, directly affecting the network model's effective tolerance. An excessively high learning rate can lead to oscillations in the training process (loss function), resulting in failure to converge to the optimal solution; an excessively low learning rate will slow down the network's convergence speed and make it prone to getting trapped in local optima. Therefore, an Adam learning rate optimizer is created during initialization, with β1 and β2 set to beta1, 0.999, and gamma set to 0.9. This means the learning rate is updated every 10 generations to 0.9 times the previous generation's learning rate.
[0072] In this training, the initial learning rate was set to 5 × 10. -4 The training consisted of 150 rounds, with one sample selected per training run and 128 random numbers generated. The training results are attached. Figure 3 , 4 As shown in Figure 5, where Figure 3 The fact that the discrimination results all tend to be 0.5 indicates that the discriminator has difficulty distinguishing between generated data and measured data, and the gradual decrease in the loss value indicates that the training effect of the two networks is gradually improving. Figure 4 Images (a), (b), (c), and (d) show the results after processing a randomly selected frame of data. It can be seen that the time-domain and frequency-domain data distributions of the generated simulated sea clutter data are roughly the same as those of the measured data, but... Figure 4 As shown in (e) and (f), the MTD results indicate that the peak amplitude of the generated data differs from that of the measured data, but the amplitude values at other frequency points are almost identical. Figure 5 As an effect evaluation diagram, analysis shows that the amplitudes of both the generated data and the measured data follow a K distribution, and the difference in their clutter power spectral density distribution is small.
[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.
Claims
1. A method for generating sea clutter data based on generative adversarial networks, characterized in that, include: Acquire sea clutter data, filter the sea clutter data, and use the filtered sea clutter data as a training dataset; Specifically: The measured sea clutter time-domain data were filtered, and the sea clutter data that showed obvious targets after pulse compression were deleted, resulting in sea clutter data without obvious targets, which was used as the network training dataset. The selected sea clutter time-domain data is subjected to fast Fourier transform to obtain its frequency-domain data. Paired data of real and imaginary parts are extracted from the one-dimensional time / frequency domain data of a single pulse. The real and imaginary parts of the time / frequency domain signals are extracted separately, and the four one-dimensional data are merged into a matrix form and integrated into two-dimensional data as input to the discrimination network. At the same time, random arrays of the same size are generated using random numbers as input to the generation network. The training dataset is input into a pre-trained generative adversarial network for training, the generative adversarial network including a generator network and a discriminator network; The generator network: input to the first fully connected layer n 3D random noise, expanding the random number sequence; Then, deconvolution is performed with a stride of 1. Batch normalization (BN) is used to standardize the data, and the ReLU function is used to accelerate convergence, outputting the corresponding feature mapping sequence. The last two fully connected layers map the feature sequence to the same dimension as the generated data. Finally, the Tanh activation function is used to output the simulated sea clutter data. The generator network transforms one-dimensional noise to n-dimensional before inputting it into a fully connected layer. After outputting, it extracts the real and imaginary parts of the generated frequency domain data and time domain data, merges them, and then saves the time domain and frequency domain data separately. The discriminator network comprises one convolutional module and one fully connected module; the convolutional module includes one convolutional layer, a batch normalization layer, and a LeakyReLU activation function; the fully connected module includes two fully connected layers and a Sigmoid activation function. Sea clutter data is generated using a pre-trained generative adversarial network. The sea clutter data is then filtered, and the filtered sea clutter data is used as the training dataset.
2. The sea clutter data generation method based on generative adversarial networks according to claim 1, characterized in that, The discriminator network straightens two-dimensional data into one-dimensional data before input.
3. The sea clutter data generation method based on generative adversarial networks according to claim 1, characterized in that, The inputs and outputs of the last two fully connected layers of the generator network are set to (1024, 512) and (512, 256), respectively.
4. The sea clutter data generation method based on generative adversarial networks according to claim 1, characterized in that, The generator network and the discriminator network are trained adversarially. The training objective of the generator network is to minimize the objective function, that is, to make the probability distribution of the generated data approximate the distribution of the real data x. The training objective of the discriminator network is to maximize the objective function, that is, to distinguish the generated data from the real data with the highest probability. The loss function is as follows: That is, the loss function of the generator: The loss function of the discriminator: in, It is random noise input to the generator. It is the output of the generator. It is the probability that the discriminator classifies a real sample. It is the probability that the discriminator will classify the generated sample. It is random noise. The distribution, It is a real sample The distribution of .
5. The sea clutter data generation method based on generative adversarial networks according to claim 4, characterized in that, During training, the two networks are trained alternately in a loop. The generator network is used to generate simulated samples, while the discriminator network judges the quality of the samples generated by the generator network by learning from real data. The two networks reach equilibrium in the game until the discriminator network can no longer distinguish the images generated by the generator network.
6. A computer system, characterized in that... include: One or more processors, a computer-readable storage medium for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of claim 1.
7. A computer-readable storage medium, characterized in that... The device stores computer-executable instructions, which, when executed, are used to implement the method of claim 1.
8. A computer program product, characterized in that... It includes computer-executable instructions, which, when executed, are used to implement the method of claim 1.