A method for simulating clutter under continuous conditions

Abstract

The application provides a clutter simulation method under continuous conditions, a trained multilayer perceptron (MLP) is used to map a label value in scalar form to a high-dimensional hidden vector space, and a high-dimensional hidden vector is obtained; compared with a one-hot vector, the high-dimensional hidden vector can carry more label information; a trained multilayer perceptron (MLP) is used to map a label value in scalar form to a high-dimensional hidden vector space, and a high-dimensional hidden vector is obtained. Compared with a one-hot vector, the high-dimensional hidden vector can carry more label information, an end-to-end generative adversarial method is adopted, the intermediate process error introduced by a traditional clutter modeling simulation method is overcome, the generative adversarial method has stronger distribution model fitting capability, clutter generation under continuous conditions is realized, radar clutter data with arbitrary distribution characteristics can be trained, radar clutter data with corresponding distribution is generated, the radar clutter generation can be widely applied to various environments, and the fidelity and universality of the radar clutter simulation technology are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to radar detection technology, in particular to radar clutter suppression technology. BACKGROUND

[0002] Radar can quickly and accurately find targets, so it plays an increasingly important role in many fields, such as military, civil aviation, automatic driving, etc. However, the radar receives clutter signal power level is high and complex, which seriously affects the detection performance of the radar. For example, when the sea detection radar is working, it will receive a large amount of high-power sea clutter signals, which sometimes even drown out the echo signals of slow-moving small targets on the sea. Therefore, during the development of the radar, the clutter suppression performance of the radar needs to be tested in all directions so as to take appropriate clutter suppression measures.

[0003] In the clutter generation task, it is generally desired to generate corresponding clutter data according to certain conditions. Radar clutter data is related to a variety of factors, including radar system parameters such as wavelength, polarization mode, azimuth angle and elevation angle, and environmental parameters such as wind speed, wave height, etc. Most of the parameters related to clutter are scalar with continuous values, which are called regression labels. The current conditional generation model such as CGAN (Conditional Generative Adversarial Network) can work well under discrete labels, but the generation effect of these two models is poor for regression labels.

[0004] The current method for generating clutter based on CGAN with continuous conditions is to first linearly normalize the regression label, X represents the original label, X max and X min represent the maximum and minimum values of the original label respectively; the normalized label value range is [0, 1], which is directly spliced to the input vector of the generator and the output vector of the last fully connected layer of the discriminator to complete the label embedding.

[0005] CGAN generally converts discrete labels into one-hot vectors and embeds them into the generator, but regression labels are scalar with continuous values, and using one-hot vectors will cause a large quantization error. The loss function of CGAN is for discrete labels, and its loss function only includes the label values in the training set, and its training target is to reduce the loss under certain discrete labels. Regression labels are continuous, and the loss under certain label values cannot be reduced. In addition, under regression labels, the distribution of the training set is often uneven, and there are many data for some labels and very few or even no data for some labels, which causes difficulties in training the generator. SUMMARY

[0006] The technical problem to be solved by the present application is that the label embedding method of CGAN is not suitable for regression labels, and a clutter simulation generation method based on regression labels is provided.

[0007] The technical scheme adopted by the present application to solve the above technical problems is a clutter simulation method under continuous conditions, characterized by comprising the following steps:

[0008] 1) Training step:

[0009] The generated clutter data and labels are used to form training data; the clutter data are composed of a clutter data matrix;

[0010] The labels are mapped to a high-dimensional hidden vector h after the training of the trained multilayer perceptron MLP;

[0011] The high-dimensional hidden vector h is spliced with a random noise z to form a combined vector, and the combined vector is input into the generator used for generating clutter data;

[0012] The generator outputs the clutter data to the discriminator, which extracts features from the input clutter data and embeds the high-dimensional hidden vector h in the discriminator through multiplication and addition, and outputs the true or false discrimination result of the current input clutter data;

[0013] 2) Test step:

[0014] The label of the clutter data to be generated is mapped to a high-dimensional hidden vector h using the trained MLP, and the high-dimensional hidden vector h is input into the trained generator, which outputs the generated clutter data.

[0015] The training process of the multilayer perceptron MLP is as follows:

[0016] The MLP pre-training process: a pre-training network is constructed, which includes an x2h part and an h2y part; the x2h part maps the input clutter data matrix to a high-dimensional hidden vector h; the h2y part restores the hidden vector h to a label, and then compares the difference between the restored label and the real label through a loss function to adjust the pre-training network parameters until the training is completed.

[0017] The MLP training process: an MLP training network is constructed, which includes an MLP and a trained h2y part; the MLP network maps the input label to a high-dimensional hidden vector h; the trained h2y part restores the hidden vector h to a label, and then compares the difference between the restored label and the input label through a loss function to adjust the parameters of the MLP until the training is completed.

[0018] In the training step, a noise ε obeying a standard normal distribution is added to the generated label as a new label, and the clutter data and the new label are used as a pair of training data to join the training set.

[0019] The beneficial effects of this invention are that by adopting an end-to-end generative adversarial method, it overcomes the shortcomings of traditional clutter modeling and simulation methods that introduce intermediate process errors. The generative adversarial method has a stronger ability to fit the distribution model, realizes clutter generation under continuous conditions, and can train radar clutter data with arbitrary distribution characteristics to generate radar clutter data with corresponding distribution. This makes the scheme widely applicable to radar clutter generation in various environments, and improves the realism and versatility of radar clutter simulation technology. Attached Figure Description

[0020] Figure 1 PDF curves of the Weibull distribution under different scale parameters (shape parameter p = 10);

[0021] Figure 2 The time-domain waveforms of the Weibull distribution under different scale parameters (shape parameter p = 10);

[0022] Figure 3 For pre-trained networks;

[0023] Figure 4 For the MLP training process;

[0024] Figure 5 The generator structure for the fast-Reg-DCGAN model;

[0025] Figure 6 This is the discriminator structure for the fast-Reg-DCGAN model;

[0026] Figure 7 The loss function curves are shown in (a) for the discriminator and (b) for the generator.

[0027] Figure 8 The time-domain waveforms and amplitude probability density curves of generated clutter under different training rounds: (a) Epoch = 100; (b) Epoch = 800; (c) Epoch = 3000; (d) Epoch = 6000; (e) Epoch = 10000; (f) Epoch = 30000;

[0028] Figure 9 Comparison of the amplitude probability density curve of generated clutter with the theoretical curve;

[0029] Figure 10 Label loss for clutter data generated by the three models;

[0030] Figure 11 Visualize clutter data under different labels;

[0031] Figure 12 Label loss for clutter under different numbers of training labels. Detailed Implementation

[0032] Training dataset construction:

[0033] First, fix the shape parameter p of Weibull distribution as 10, and increase the scale parameter q from 6 to 14.9 with a step of 0.1, generate 50 pieces of simulated clutter data under each set of parameters using the ZMNL method, and then reshape each piece of simulated clutter data into a 64x64 matrix with a length of 4096; finally, the real part and the imaginary part of the clutter data are separated to form two-channel data.

[0034] The scale parameter q of Weibull distribution is used as the label value, and there are a total of 90 labels, each with 50 clutter data matrices, totaling 4500 clutter data matrices. The data-label pairs with the last digit of the label value being odd are used as the training set, and the remaining data-label pairs are used as the test set. The data-label pairs are organized and saved in HDF5 format.

[0035] The shape parameter p of Weibull distribution is generally not more than 2, but in order to have a relatively obvious distinction between clutter data with different label values for later analysis, the shape parameter p is set to be larger. Figure 1 The amplitude probability density of Weibull distribution is given when the shape parameter p is 10 and the scale parameter q is increased from 6.0 to 15.0 with a step of 3.0, Figure 2 The time-domain waveform of the clutter when the shape parameter p is 10 and the scale parameter q is 6, 9, 12, and 15 is given.

[0036] The fast-Reg-DCGAN model structure for implementing the method of the application:

[0037] First, a trained multilayer perceptron (MLP) is used to map the scalar form of the label value to a high-dimensional hidden vector space to obtain a hidden vector h. The dimension of h is much larger than that of the original label y, so it can represent more label feature information. The MLP will map the label y to a high-dimensional hidden vector h. In order to make the hidden vector h correspond to the real picture-label pair, a trained network is needed to restore the hidden vector h to the label y', and then the MLP is trained to make y' approach y. Figure 3 The specific structure of the MLP pre-training network is given, which is based on ResNet34, and after two modules, the x2h part maps the picture x to the hidden vector h. The processing of each module includes full connection Linear operation, batch normalization BN operation and nonlinear activation function ReLU; the h2y part restores the hidden vector h to the label y' through the full connection layer Linear and the activation function ReLU.

[0038] In order to train the MLP network, the real data-label pair is first used to train the network in Figure 3The pre-training network shown is trained, and after training, the h2y module is extracted for training the MLP network. As shown in the figure Figure 4 As shown, the MLP network first converts the label y into a hidden vector h, and the h2y module restores the hidden vector h into a label y', calculates the error between y and y' using the Mean Square Error (MSE) algorithm, and then trains the MLP network using the back propagation algorithm to make y' well approximate y. The h2y module not only makes the training of the MLP network easier, but also makes the hidden vector h generated by the MLP one-to-one corresponding to the real data-label pair.

[0039] Figure 5 The specific way of embedding the label into the generator of the DCGAN is given in the embodiment, which is named fast-Reg-DCGAN. The label y is mapped into a high-dimensional hidden vector h after passing through the MLP, and then h is spliced with a random noise z to form a new vector, which is used as the input of the generator for the generation of clutter data. This is similar to the embedding form of the one-hot vector, but this time the one-hot vector is spliced on the random noise, not the one-hot vector. Compared with the one-hot vector, the high-dimensional hidden vector h can carry more label information, so it can obtain better performance than the traditional CGAN.

[0040] The generator structure can be roughly divided into six modules. The first module includes a full connection Linear operation, a matrix transformation reshape operation, a batch normalization BN operation and a nonlinear activation function ReLU. The second to fifth modules are the same, including a two times upsampling upsample (x2), a residual module, a BN operation and a ReLU nonlinear activation function. A residual block contains two two-dimensional convolution modules conv2d before and after, and also includes a BN operation and a ReLU nonlinear activation function in the middle. The convolution kernel size of the two convolution modules is set to [3, 3], the convolution step size is set to 1, and zero padding is used to keep the size of the feature map unchanged. The two times upsampling upsample is output to the second module on one hand, and is added to the output of the two-dimensional convolution module conv2d after the residual block on the other hand to serve as the input of the sixth module. The step size of the transpose convolution conv2d_transpose in the sixth module is set to 1 to further reduce the "checkerboard effect", and finally the Tanh is used as the nonlinear activation function to obtain the generated data.

[0041] Embedding the label into the discriminator is relatively simple, only the high-dimensional latent vector is embedded into the last full connection layer Linear of the discriminator through multiplication and addition, Figure 6The specific way of embedding high-dimensional latent vectors into the fast-Reg-DCGAN discriminator is given. The discriminator is divided into five modules, the first four modules are the same, including two-dimensional convolution operation conv2d and BN operation, using Leaky ReLU as the nonlinear activation function, using convolution operation with step instead of pooling layer to achieve the effect of reducing training parameters, because the pooling operation will slow down the training speed; The last module contains reshape operation and full connection operation, and uses Sigmoid function to output the discrimination result.

[0042] The parameters of the convolution module in the discriminator are shown in Table 1, the size of each filter is [M, N, L, K], M and N represent the two-dimensional convolution kernel size in the filter; L represents the input channel number, which is consistent with the depth of the feature map to be convolved; K represents the output channel number, which is equal to the number of filters, indicating that each filter corresponds to output a feature map; The number of biases is equal to K, indicating that a bias is added to each output feature map; The input data size of the discriminator model is [64, 2, 128, 128], and the output size is [64, 1], which represents analyzing 64 clutter data matrices in a batch, and discriminating the true and false of the clutter.

[0043] Table 1 Convolution kernel structure parameters of convolution module

[0044]

[0045] Loss function:

[0046] Fast-Reg-DCGANd discriminator loss function L d And the generator loss function L g As shown in equation (1) and equation (2).

[0047]

[0048]

[0049] K is a hyperparameter related to the maximum distance of the labels in the training set, generally set to the maximum value of the adjacent label distance. g N represents the total number of generated samples, represents the i-th generated sample, represents the label of the i-th generated sample, ε j represents a Gaussian random quantity, D represents the generator, K is a hyperparameter related to the maximum distance of the labels in the training set, generally set to the maximum value of the adjacent label distance N r N represents the total number of training samples, represents the i-th training sample, represents the i-th generated sample.

[0050] Training process:

[0051] Generally, the continuity of the label and the continuity of the data are corresponding, and if two labels are close, the corresponding samples will not be far apart. If y is the label of data x, y+ε can also be regarded as the label of data x, and when the noise ε is small enough, the error can be controlled within an acceptable range, so in actual training, the real data-label pair is not directly used for training, but the real label is replaced with a certain range of values, and this range is controlled by ε. Therefore, the data that does not exist in the training set table can be regarded as the "weighted average" of the adjacent data, and the weight of the "weighted average" is obtained by self-learning of the network. Table 2 gives the specific training method, and the core idea is to replace the label y in each real data-label pair with a value within a certain range during training i Add a noise ε obeying the standard normal distribution N(0,σ 2 ), where σ=κ / 3, κ is a hyperparameter related to the maximum distance of the labels in the training set, and is generally set to the maximum value of the distance between adjacent labels. According to the 3σ principle, the new label y i +ε will fall into the adjacent interval on both sides of y i with a high probability, so that labels that do not exist in the training set can also be trained, which will not affect the network structure, so this training method is universal and applicable to all GAN architectures.

[0052] Table 2 fast-Reg-DCGAN training process

[0053]

[0054]

[0055] Training result analysis:

[0056] The experimental simulation test platform is Ubuntu 20.04, the system is equipped with NVIDA GeForce RTX2080Ti GPU, the model is built using pytorch framework, and NVIDA Cuda is used to accelerate the calculation. The model optimization selects the Adam optimization algorithm, the LearningRate is set to 1e-5, beta1 is set to 0.5, and beta2 is set to 0.9. The specific training parameters are shown in Table 3

[0057] Table 3 fast-Reg-DCGAN training parameters

[0058]

[0059]

[0060] The loss function variation curves of the discriminator and the generator during the training process are shown in Figure 7

[0061] Amplitude distribution characteristic analysis:

[0062] In order to show the effect of fast-Reg-DCGAN generation model in conditional control, the amplitude distribution characteristics of generated clutter data under different scale parameters are compared. In order to verify whether the amplitude probability density curve of the clutter data generated by fast-Reg-DCGAN is consistent with the theoretical curve, different scale parameters q = 6.49, 9.16, 11.83, 14.50 are used to generate clutter data at different training rounds Epoch = 100, 800, 3000, 6000, 10000, 30000, as shown in Figure 8 The amplitude probability density curve of the clutter data shows that there are obvious differences in the amplitude probability density curve under different scale parameters with the increase of training rounds, which shows that the conditional control of fast-Reg-DCGAN is successful.

[0063] Figure 9 On the basis of Figure 8 The theoretical probability density curve of Weibull distribution with shape parameter p = 10 and scale parameter q = 6.49, 9.16, 11.83, 14.50 is given in the form of "point line", it can be seen that fast-Reg-DCGAN can well fit the theoretical probability density curve, which shows that fast-Reg-DCGAN model can well learn the distribution characteristics of clutter data under regression label.

[0064] Label consistency analysis:

[0065] Label consistency refers to the difference between the true label of generated data and the expected label. The true label of generated data can be obtained by a trained data-label conversion network, which is based on ResNet34 and can extract features from the input clutter matrix to obtain its corresponding label value; The expected label of generated data is the label value input to the generator when generating the data. The smaller the difference between the true label and the expected label, the better the label consistency.

[0066] ​To quantitatively analyze the label consistency of the clutter data generated by fast-Reg-DCGAN, a label scoring network based on ResNet34 network is built to calculate the label loss of the generated clutter data. The lower the label loss, the higher the label consistency. At the same time, in order to illustrate the performance of fast-Reg-DCGAN model and the traditional method of directly embedding regression labels into the generator and discriminator of DCGAN model (concat-DCGAN).

[0067] Three models generate 100 sets of clutter data under each label, and then calculate the average label loss under each label, Figure 10 The label loss curves under 90 sets of labels are given, where the CGAN with existing label embedding method is called concat-CGAN, concat-DCGAN uses all 90 sets of labels for training, and Reg-DCGAN and fast-Reg-DCGAN use the last 45 sets of labels with odd numbers for training. It can be seen that Reg-DCGAN and fast-Reg-DCGAN have similar performance in label consistency, both of which are much better than concat-CGAN. This shows that compared with the traditional CGAN model, the fast-Reg-DCGAN model can effectively embed the data features of the regression label into the generation network and learn the distribution characteristics of the regression label, so the two models perform better in the clutter generation task based on the regression label.

[0068] In order to intuitively show the label consistency of the data generated by the three models under different labels, Figure 11 The training set data under different labels and the simulated clutter data generated by the three models are visualized, with label values of 6.49, 7.38, 8.27, 9.16, 10.05, 10.94, 11.83, 12.72, 13.61, and 14.50. The three models generate three sets of clutter data under each label. It can be seen that concat-DCGAN has serious label inconsistency, and the generated clutter data cannot change with the change of label value, while Reg-DCGAN and fast-Reg-DCGAN generate data with excellent label consistency, and the transition between different labels is smoother.

[0069] Clutter generation under sparse labels:

[0070] In order to study the performance of the models used in this chapter in clutter generation under sparse labels, different numbers of labels are used for training, with training labels of 45 sets (6.1:0.2:14.9), 22 sets (6.1:0.4:14.9), 14 sets (6.1:0.6:14.9), and 11 sets (6.1:0.8:14.9). Each set of labels selects 25 clutter matrices for training.Figure 12 The label loss comparison curves of the generated clutter under 90 groups of labels are given in different cases, from which it can be seen that Figure 12 It can be seen that although the label loss of the generated clutter will increase with the decrease of the number of training labels, showing higher label inconsistency, the overall label loss is still low, much lower than that of concat-DCGAN. This shows that fast-Reg-DCGAN can maintain a high level of label consistency under sparse labels and has practical application value.

Claims

1. A method of simulating clutter under continuous conditions, characterized in that, The method comprises the following steps: 1) training step: The generated clutter data and the label form training data; the clutter data is composed of a clutter data matrix; The label is mapped to a high-dimensional hidden vector h after the training of the completed multi-layer perceptron MLP; The high-dimensional hidden vector h is spliced with a random noise z to form a combined vector, and the combined vector is input into the generator for clutter data generation; The generator outputs the clutter data to the discriminator, which extracts features from the input clutter data and embeds the high-dimensional hidden vector h into the discriminator through multiplication and addition, and outputs the true or false discrimination result of the current input clutter data; 2) test step: The label of the clutter data to be generated is mapped to a high-dimensional hidden vector h using the trained MLP, and the high-dimensional hidden vector h is input into the trained generator, and the generator outputs the generated clutter data; The training process of the multi-layer perceptron MLP is as follows: The pre-training process of MLP: a pre-training network is constructed, which includes an x2h part and an h2y part; the x2h part maps the input clutter data matrix x to a high-dimensional hidden vector h; the h2y part restores the hidden vector h to the label y, and then compares the difference between the restored label and the real label through a loss function to adjust the pre-training network parameters until the training is completed; The training process of MLP: an MLP training network is constructed, which includes an MLP and a trained h2y part; the MLP network maps the input label to a high-dimensional hidden vector h; the trained h2y part restores the hidden vector h to the label, and then compares the difference between the restored label and the input label through a loss function to adjust the parameters of the MLP until the training is completed.

2. The method of claim 1, wherein, The label of the clutter data is the scale parameter of the Weibull distribution.

3. The method of claim 2, wherein, The x2h part processes the input clutter data matrix through a deep residual network ResNet34, and then processes it twice through a module, each module processing including a full connection Linear operation, a batch normalization BN operation and a nonlinear activation function ReLU; The h2y part restores the hidden vector h to the label through a full connection layer Linear and an activation function ReLU.

4. The method of claim 1, wherein, In the training step, a noise ε obeying a standard normal distribution is added to the generated label as a new label, and the clutter data and the new label are added to the training set as a pair of training data.

5. The method of claim 1, wherein, The generator outputs the clutter data through 6 module processes: The first module processing includes full connection operation, matrix transformation operation, batch normalization BN operation and nonlinear activation function ReLU; The second to fifth module processes are the same, including twice upsampling, residual processing, BN operation and nonlinear activation function ReLU; the residual processing includes a first two-dimensional convolution, a BN operation, a nonlinear activation function ReLU and a second two-dimensional convolution, and the output of the first two-dimensional convolution is added to the output of the second two-dimensional convolution as the output result of the residual processing; The sixth module processing obtains the generated clutter data through transpose convolution and nonlinear activation function Tanh.

6. The method of claim 1, wherein, The discriminator outputs the discrimination result through 5 modules: The first to fourth modules are the same in processing, and feature extraction is performed through one two-dimensional convolution, a batch normalization (BN) operation and a nonlinear activation function Leaky ReLU; The fifth module processing includes a matrix transformation reshape operation, a full connection operation and a Sigmoid function judgment. The features output by the fourth module are embedded into the full connection operation through the multiplication and addition of the embedded high-dimensional hidden vectors h after the reshape operation, and finally the true and false discrimination results of the clutter data are obtained through the Sigmoid function judgment.

Images