Direction finding method based on generative adversarial network deep learning under small sample condition

Abstract

The application discloses a small sample condition generation adversarial network deep learning direction finding method, relates to the radio direction finding technical field, and solves the technical problem that in the prior art, it is difficult to construct a complete training sample library, and in the actual application environment condition, the number of samples that can be collected is less than the number of required training samples; the steps of the method are as follows: constructing a generation adversarial network; a generator and a discriminator are mutually opposed to generate generated samples approximating a real distribution; data preprocessing; extracting relative amplitude and covariance characteristics of each array element as input data of the CNN; realizing nonlinear feature mapping of the data through convolution operation; dividing the input feature mapping into multiple non-overlapping regions; small sample CNN direction finding; the small sample condition for the actual application environment is realized, generated samples approximating the real sample distribution are generated by using the generation adversarial network, the training sample data set is effectively expanded, and then the convolution neural network direction finding model is utilized, so that high-precision direction finding is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of radio direction finding, and relates to neural network technology, in particular to a generative adversarial network deep learning direction finding method under a small sample condition. BACKGROUND

[0002] Direction finding of a radiation source target is that an electronic reconnaissance device estimates the direction of arrival of a target signal through signal processing by using the received target signal. The direction finding method for a radiation source target mainly includes amplitude method direction finding, phase method direction finding, and spatial spectrum estimation direction finding. These methods are traditional non-intelligent direction finding methods. At present, intelligent direction finding methods have become a research hotspot. The patent "arbitrary array direction of arrival estimation method based on deep learning" proposes an array signal direction finding method based on a convolutional neural network (CNN). The method is aimed at a given multi-element array. By extracting the phase difference between the array elements and other characteristic information of the sampling data of the array elements, the method realizes fast and high-precision direction finding of a radiation source target on the basis of deep learning.

[0003] However, the CNN direction finding method needs a large number of training samples. First, the method is a direction finding method based on classification and identification. Assuming that the direction finding range is 360° and the angle resolution is 0.1°, and the number of targets is unknown, the number of angle classifications is At this time, the number of training samples is huge. Second, when facing the demand for wide-band direction finding, a sample set needs to be established for each carrier frequency for training, and the number of training samples is significantly increased. Therefore, it is extremely difficult to construct a complete training sample library, and the number of samples that can be collected in the actual application environment is much smaller than the number of required training samples.

[0004] Therefore, a generative adversarial network deep learning direction finding method under a small sample condition is proposed. SUMMARY

[0005] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application proposes a generative adversarial network deep learning direction finding method under a small sample condition, which solves the problem that it is extremely difficult to construct a complete training sample library in the prior art, and the number of samples that can be collected in the actual application environment is much smaller than the number of required training samples.

[0006] To achieve the above-mentioned purpose, according to the first aspect of the present application, a generative adversarial network deep learning direction finding method under a small sample condition is proposed, which comprises the following steps:

[0007] Step 1: constructing a generative adversarial network using a nonlinear mapping function; wherein the generative adversarial network includes a generator G and a discriminator D;

[0008] Step two: the generator G and the discriminator D reach Nash equilibrium through mutual confrontation and mutual promotion, and generate generated samples, including the following steps:

[0009] Inputting the noise sequence into the generator to generate generated samples subject to the real data distribution;

[0010] Inputting the generated samples and the real samples into the discriminator;

[0011] Fixing the generator G, training the discriminator D by minimizing the cross-entropy, so that it can effectively distinguish the real samples and the generated samples, and the loss function of the discriminator D is:

[0012]

[0013] In the formula, In the formula, x is the real sample, and D(x) is the probability of the real sample x input by the discriminator D;

[0014] In the formula, z represents the noise sequence input into the generator G, G(z) represents the generated sample, and D(G(z)) represents the probability of the generated sample being judged as a real sample by the discriminator D;

[0015] The generator G then learns the formal data distribution P data to deceive the discriminator D, and the generator G wants to increase the probability of the generated sample being judged as a real sample, so that the output of the discriminator D when the generated sample is input is as close to 1 as possible, and the loss function of the generator G is:

[0016]

[0017] Step three: preprocessing the radiation source received by each array element to obtain sampling data; wherein the preprocessing includes power amplification, frequency conversion and data sampling processing;

[0018] Step four: taking the relative amplitude and covariance characteristics of each array element as the input data of the CNN;

[0019] Step five: the convolution layer realizes nonlinear feature mapping of the data and feature extraction of the data by convolution operation with the input data through different convolution kernels;

[0020] Step six: the pooling layer divides the input feature mapping into multiple non-overlapping regions;

[0021] Step seven: based on the generative adversarial network, the training data set under the condition of small samples is expanded.

[0022] Preferably, the discriminator D is a binary classification model.

[0023] Preferably, the discriminator D outputs the probability that the input sample is a real sample;

[0024] When the input is a real sample, the output is as close to 1 as possible;

[0025] When the input is a generated sample, the output is as close to 0 as possible.

[0026] Preferably, the convolutional layer adopts a local connection mode.

[0027] Preferably, the pooling layer compresses data through downsampling operation.

[0028] Compared with the prior art, the present application has the following beneficial effects:

[0029] The present application uses a nonlinear mapping function to construct a generative adversarial network; the generator G and the discriminator D reach Nash equilibrium through mutual confrontation and mutual promotion, generating a generated sample; the small sample condition for the actual application environment is realized, and the generative adversarial network is used to generate a generated sample approximating the real sample distribution;

[0030] The radiation sources received by each array element are preprocessed to obtain sampling data; the relative amplitude and covariance characteristics of each array element are used as the input data of the CNN; the convolutional layer performs convolution operation on the input data through different convolution kernels to realize nonlinear feature mapping of the data, and then realizes feature extraction of the data; the pooling layer divides the input feature mapping into multiple non-overlapping regions; based on the generative adversarial network, the training data set under the small sample condition is expanded; the training sample data set is effectively expanded, and then the convolutional neural network direction finding model is used to realize stable high-precision direction finding based on learning and training. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a schematic diagram of the generative adversarial network;

[0032] Figure 2 It is a flowchart of the present application. DETAILED DESCRIPTION

[0033] The technical solutions of the present application will be described below in conjunction with the embodiments, obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0034] As shown in the drawings, the small sample condition generative adversarial network deep learning direction finding method comprises the following steps: Figures 1-2

[0035] Step one: construct a generative adversarial network​

[0036] The principle of constructing the generative adversarial network using the nonlinear mapping function is shown in Figure 1 The generative adversarial network is composed of a generator G and a discriminator D, wherein the generator G generates a generated sample G(z) according to an input noise sequence z, and the discriminator D is a binary classification model, which judges whether the input sample is a real sample x or a generated sample G(z).

[0037] Step two: generating and

[0038] First, the noise sequence z is input into the generator G to generate a generated sample G(z) conforming to the real data distribution, and then the generated sample G(z) and the real sample x are input into the discriminator D, and the discriminator D outputs the probability that the input sample is a real sample. When the real sample x is input, the output is as close to 1 as possible, and when the generated sample G(z) is input, the output is as close to 0 as possible.

[0039] The discriminator D is trained by minimizing the cross-entropy to make it effectively distinguish the real sample x and the generated sample G(z), and the loss function of the discriminator D is:

[0040]

[0041] In the formula, In the formula, x is a real sample, and D(x) is the probability that the discriminator D judges the input real sample x;

[0042] In the formula, z represents the noise sequence input into the generator G, G(z) represents the generated sample, and D(G(z)) represents the probability that the generated sample is judged as a real sample by the discriminator D.

[0043] The generator G then learns the formal data distribution P data to deceive the discriminator D, and the generator G increases the probability that the generated sample G(z) is judged as a real sample, so that when the generated sample G(z) is input, the output of the discriminator D is as close to 1 as possible. The loss function of the generator G is:

[0044]

[0045] The generator G and the discriminator D are mutually opposed and promote each other, and in the continuous confrontation, both reach a Nash equilibrium, and finally can generate a generated sample G(z) that is difficult to distinguish from the real sample.

[0046] Step three: data preprocessing

[0047] The data preprocessing includes power amplification, frequency conversion and data sampling processing.

[0048] Suppose the number of array antenna elements of the receiving array is L, the number of incident signals is M, and the array output is represented by a vector x(t) = [x1(t), x2(t), …, xL(t)]. L (t)] T represents:

[0049]

[0050] wherein A = [a1, a1, …, aM] is an array response matrix corresponding to the M incident signals, M is the array response vector corresponding to the incident angle of the i-th signal. i

[0051] The received radiation sources of each array element are subjected to power amplification, frequency down-conversion and data sampling processing to obtain sampling data.

[0052] Step four: generating CNN input data

[0053] Generally, each signal is independent of each other, and the covariance matrix of the array output is:

[0054] R = E[x(t)x H (t)] = AE[s(t)s H (t)]A H +E[v(t)v H (t)] = AR s A H + R v

[0055] wherein R s and R v represent the signal covariance matrix and the noise covariance matrix, respectively.

[0056] The relative amplitude of each array element is related to the incident direction of the signal, and the covariance matrix retains all the angle information of the target, so the relative amplitude of each array element and the covariance characteristics are used as the input data of the CNN.

[0057] Step five: convolution processing

[0058] The convolution layer mainly analyzes each small block of data more deeply to abstract higher-level features. The convolution layer adopts a local connection mode, performs convolution operation on the input data through different convolution kernels to realize nonlinear feature mapping of the data, and then realizes feature extraction.

[0059] Step six: pooling processing

[0060] ​​The pooling layer realizes data compression through a downsampling operation, and is used to reduce the feature dimension after convolution and reduce the number of neurons required by the network. The pooling process is based on the translational invariance, and divides the input feature map into multiple non-overlapping regions. After the pooling process, the automatic extraction of the target angle feature vector is realized.

[0061] Step seven: small sample CNN direction finding

[0062] Based on the generative adversarial network, learning and training under the condition of small samples are realized. When the real data is input, the CNN direction finding model is used, and after preprocessing, convolution processing and pooling processing, the distributed generated features are coupled to different spatial angles in the full connection layer, so that stable high-precision direction finding is realized.

[0063] The above formulas are all dimensionless values, and the formulas are obtained by software simulation of a large amount of data to obtain a formula closest to the actual situation. The preset parameters and the preset threshold in the formula are set by the person skilled in the art according to the actual situation or obtained by a large amount of data simulation.

[0064] In order to verify the performance of the method, a 9-element uniform linear array is sampled, and under the condition of different angle sample numbers (the number of samples at a single angle), the patent "arbitrary array direction of arrival estimation method based on deep learning" (hereinafter referred to as method 1) and the method of the present application are applied respectively to perform direction finding statistical test, and the statistical results of the direction finding accuracy are shown in Table 1.

[0065] Table 1 Direction finding accuracy statistical results

[0066]

[0067] The experimental results show that when the number of training samples is large, the direction finding accuracy of the two methods is very high, but under the condition of small samples, the direction finding accuracy of the method is about 3% higher than that of method 1, and the direction finding stability is better.

[0068] The above embodiments are only used to illustrate the technical method of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A deep learning direction-finding method for generative adversarial networks under small sample conditions, characterized in that, Includes the following steps: Step 1: Construct a generative adversarial network using a nonlinear mapping function; wherein the generative adversarial network includes a generator G and a discriminator D; Step 2: The generator G and discriminator D reach Nash equilibrium through mutual competition and promotion, generating samples. This includes the following steps: The noise sequence is input into the generator to produce generated samples that follow the distribution of the real data. The generated samples and real samples are input into the discriminator; With a fixed generator G, a discriminator D is trained using the method of minimizing cross-entropy to effectively distinguish between real and generated samples. The loss function of discriminator D is: In the formula, In this context, x is the real sample, and D(x) is the probability that the discriminator D judges the input sample x to be real. In this context, z represents the noise sequence in the input generator G, G(z) represents the generated sample, and D(G(z)) represents the probability that the generated sample is judged as a real sample after passing through the discriminator D. The generator G learns the distribution of real data P. data To deceive the discriminator D, with the discriminator D fixed, the generator G needs to increase the probability that the generated sample is judged as a real sample, so that when the input sample is generated, the output of the discriminator D is as close to 1 as possible. The loss function of the generator G is: Step 3: Preprocess the radiation sources received by each array element to obtain sampled data; wherein, the preprocessing includes power amplification, down-conversion and data sampling processing; Step 4: Use the relative amplitude and covariance features of each array element as input data for the CNN; Step 5: The convolutional layer performs convolution operations with the input data using different convolutional kernels to achieve non-linear feature mapping of the data, thereby realizing feature extraction from the data; Step 6: The pooling layer divides the input feature map into multiple non-overlapping regions; Step 7: Expand the training dataset under small sample conditions based on generative adversarial networks.

2. The generative adversarial network deep learning direction finding method under small sample conditions according to claim 1, characterized in that, The discriminator D is a binary classification model.

3. The generative adversarial network deep learning direction finding method under small sample conditions according to claim 1, characterized in that, The discriminator D outputs the probability that the input sample is a real sample; When inputting real samples, the output should be as close to 1 as possible; When generating samples from input, the output should be as close to 0 as possible.

4. The deep learning direction finding method for generative adversarial networks under small sample conditions according to claim 1, characterized in that, The convolutional layer uses local connectivity.

5. The deep learning direction finding method for generative adversarial networks under small sample conditions according to claim 1, characterized in that, The pooling layer compresses the data through downsampling operations.

Images