A true and false identification method based on fast-slow time domain combined features under a false target sample lack condition

Abstract

In the case of low signal-to-noise ratio and lack of false target samples, it is difficult to accurately identify true and false targets, resulting in serious false alarm or missed alarm. In order to solve this problem, the application provides a false target jamming identification method of early warning radar based on fast-slow time domain joint features. Firstly, the radar echo is processed by pulse compression and coherent accumulation; then the peak search algorithm is used to determine the Doppler unit and distance channel of the true and false targets; after the true and false targets are truncated along the fast time dimension, the fast Fourier transform (FFT) is performed to extract the slow time domain joint features; the extracted unmarked true-false target feature mixed data set U and the labeled true target features are used to construct a joint data set, which is input into a support vector machine (SVM) for class prior probability estimation, and the average class prior probability is used to label the data set U to obtain the joint data set P+U; the joint data set is input into a two-channel feature fusion network for training and identification.

Description

Technical Field

[0001] This invention relates to a method for identifying true and false targets based on joint features in the fast and slow time domains when false target samples are scarce, and belongs to the fields of radar electronic countermeasures and radar signal processing. Background Technology

[0002] In modern warfare, a crucial prerequisite for victory is achieving dominance in electronic warfare. With the development of digital radio frequency storage technology, various decoy jamming techniques are constantly emerging. By altering the time delay, Doppler shift, and phase of the real target echo, the enemy can make the decoy echo highly correlated with the real target echo while differing from the real target in terms of position, velocity, and angle information, thereby achieving the purpose of concealing the real target.

[0003] The emergence of decoy target jamming has brought great challenges to the target detection work of our radar. Therefore, timely and accurate identification of the echoes of real and false targets is of great significance to ensuring the radar's detection performance, grasping the battlefield situation, and seizing battlefield opportunities.

[0004] Purpose of the invention

[0005] The purpose of this invention is to propose a true / false target identification method based on joint features in the fast and slow time domains under conditions of scarce false target samples. This method aims to solve the problem of true / false target identification in situations with low signal-to-noise ratio and scarce false target samples, and can provide prior information for radar to adopt corresponding anti-jamming methods. This invention is easy to implement, has promotional value, and can be applied to the field of false target interference identification in early warning radar. Summary of the Invention

[0006] The objective of this invention is achieved as follows: a true / false identification method based on fast-slow time domain joint features is proposed under the condition of missing false target samples, including: fast-slow time domain joint feature extraction, class prior probability estimation, two-channel feature fusion network training, and identification result output.

[0007] The system takes as input a labeled dataset P containing real targets and an unlabeled mixed dataset U containing real and fake targets. It performs feature analysis on the signals of both targets, selecting the joint fast-slow time domain spectrum as a feature. The system performs pulse compression and coherent accumulation processing on the echoes of both targets to obtain range-Doppler images. A peak search algorithm is used to establish the range cells and Doppler channels for both targets. The signals of both targets are windowed and extracted along the fast time dimension. An FFT operation is performed on the truncated signals to extract the joint fast-slow time domain spectrum features of both targets. The labels of all samples in P are uniformly set to 1, and the labels of all samples in U are uniformly set to 0. This data is then input into an SVM for training, and the class prior probability of each sample in U is estimated. Based on the average class prior probability, each sample in U is relabeled, ultimately reconstructing the dataset P+U. The reconstructed dataset P+U is then input into a two-channel feature fusion network composed of a one-dimensional neural network and a long short-term memory network for training, and the prediction results are provided, displaying the category of each peak in the range-Doppler image.

[0008] In one embodiment, the radar transmit signal is a linear frequency modulated signal, and the false target jamming signal is one or more of the three types of deception jamming: range, speed, and angle.

[0009] In one embodiment, pulse compression and coherent accumulation are selected to process the radar echo to obtain a range-Doppler image.

[0010] In one embodiment, a peak search algorithm is used to determine the range cells and Doppler channels of real and false targets in a range-Doppler image.

[0011] In one embodiment, after determining the true and false target range cells and Doppler channels, the signal is truncated by a rectangular window along the range-Doppler image's range dimension (fast time dimension) to extract the true and false target signals.

[0012] In one embodiment, FFT is used to extract the joint fast-slow time-domain spectral features of the real and fake target signals.

[0013] In one embodiment, peak search is used to perform maximum value normalization to address the problem of feature instability caused by power differences, and the features are preprocessed.

[0014] In one embodiment, the labels of all labeled real target fast-slow time spectra are uniformly set to 1 to form dataset P, and the labels of all unlabeled real and fake target time spectra are uniformly set to 0 to form dataset U.

[0015] In one embodiment, the dataset P+U is input into an SVM for training, and the average class prior probability of each sample in U is given. The samples in the dataset U are then re-labeled based on the average class prior probability.

[0016] In one embodiment, the dataset P+U is reconstructed.

[0017] In one embodiment, the dataset P+U is used to train a two-channel feature fusion network.

[0018] In one embodiment, a trained two-channel feature fusion network is used to identify each peak target in the range-Doppler domain, and the prediction results are given, showing the category to which each peak belongs in the range-Doppler image.

[0019] Technical effect

[0020] Compared with existing technologies, this invention employs pulse compression and coherent accumulation processing to extract signal features in both the fast and slow time domains, effectively improving the signal-to-noise ratio (SNR) of the echo signal. Compared to methods that extract signal features only in the fast time domain, this invention can effectively identify real and false targets under lower SNR conditions. This invention can identify individual cases of false target interference such as distance, speed, and angle, as well as multiple composite interferences, providing reliable support for subsequent interference suppression and other processing steps. This invention does not require sample information of false targets, making it more consistent with actual conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart of the real and fake target identification process of the present invention.

[0022] Figure 2 This is an example diagram of radar echo pulse compression according to the present invention.

[0023] Figure 3 This is an example diagram of radar echo range-Doppler for this invention.

[0024] Figure 4 This is the joint spectral feature map of the fast-slow time domain of the true target of this invention.

[0025] Figure 5 This is the joint fast-slow time domain spectral feature map of the pseudo-target of the present invention.

[0026] Figure 6 This is a schematic diagram illustrating the prior probability estimation of a class during a single training iteration of this invention. Detailed Implementation

[0027] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Various embodiments of the present disclosure will be described more fully below. The present disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of the present disclosure to the specific embodiments disclosed herein. Rather, the present disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of the present disclosure.

[0028] The terminology used in the various embodiments of this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the various embodiments of this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this disclosure pertain. Terms (such as those defined in a generally used dictionary) are to be interpreted as having the same meaning as in the context of the relevant technical field and are not to be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this disclosure.

[0029] The overall recognition method of this invention mainly consists of four parts: first, joint feature extraction of real and fake targets in the fast and slow time domains; second, prior probability estimation of the unlabeled dataset U and dataset reconstruction; third, training of the two-channel feature fusion network; and fourth, output of the recognition result. The overall flowchart is as follows: Figure 1 As shown. The specific steps are as follows:

[0030] Step 1: Perform pulse compression and coherent accumulation processing on real and false targets. First, the fast time expression of the radar target echo is as follows: ,but ;

[0031] in, For signal pulse width, The frequency modulation slope of the signal. Let the signal bandwidth be denoted as , and the target echo fast time be expressed as . ,but ;

[0032] in, The symbol for convolution. Given the radar system function in the time domain, the fast-slow time expression for the target echo is: ;

[0033] in, To save time, For slow time, Let the radial distance function between the target and the radar be denoted as . The initial distance to the target. For the target radial velocity, At the speed of light, For carrier frequency; assume the number of coherent cumulative elements is . The pulse repetition interval is We use express The sets of target echoes are respectively

[0034] ;

[0035] In the above formula, it is assumed that the phase difference between two adjacent target echoes is... Then each target echo can be represented by the first target echo, and the results are as follows.

[0036] ;

[0037] The frequency-domain matched filter function for pulse compression is as follows:

[0038] ;

[0039] The above equation satisfies

[0040] Assuming its time-domain representation is ;

[0041] The phase weight matrix for achieving coherent accumulation is as follows:

[0042] ;

[0043] in, Representing different phase factors, , The number of points in the Discrete Fourier Transform. Different Doppler units are represented by matrices. express The results of pulse compression and coherent accumulation processing of continuous target echoes are shown below. Figure 2 , 3 As shown, it can be represented as

[0044] ;

[0045] in, Represents the matrix rank transformation symbol;

[0046] Step 2: Use the peak search algorithm to establish the range cells and Doppler channels of the echoes from the real and false targets, assuming that the real target is found in the fast-slow time domain. The corresponding Doppler unit is Its corresponding The Doppler cell that achieves a large peak value is windowed and truncated along the distance dimension of the Doppler cell containing the real target. The truncated signal is then... ;

[0047] In the formula, the phase compensation factor of the Doppler channel must be equal to or infinitely close to the echo phase difference, i.e. or ,but It can be simplified to

[0048] ;

[0049] Step 3: Extract the joint fast-slow time domain features of the real target echo using FFT;

[0050] Step 4: Repeat steps 2 and 3 to extract the joint features of the fast and slow time domains of the false target echo, and construct a true-false target feature dataset U, setting the label of each sample in it to 0.

[0051] Step 5: Set the label of each sample in the real target feature dataset P collected before the false target appeared to 1, and then apply the SVM to it. The training session, in the... In this training iteration, P is divided into two parts, P1 and P2, proportionally, and the SVM is trained using P1 and U. After each training iteration, the probability matrix is ​​estimated using the SVM. This represents the probability that the sample SVM will estimate each sample in U as 1. Let U be the number of samples in U; ​​and use SVM to estimate the probability. This represents the average probability that the SVM estimates each sample in P2 as 1. Obtain the class prior probability of each sample in U. ; in the After the training session, and will the former The estimated class prior probabilities are accumulated, i.e.

[0052] ;

[0053] Among them, before the first estimation begins As a zero vector, ultimately through After the estimation is completed, we get The cumulative result of the secondary prior probabilities And calculate the average class prior probability. When the average prior probability of a sample is greater than the probability threshold When the condition is met, the sample label is set to 1; otherwise, it is set to 0. This process re-labels each sample in U and completes the reconstruction of the dataset P+U.

[0054] Step 6: Input the reconstructed dataset P+U into the two-channel feature fusion network to automatically extract the joint spectral fluctuation features in the fast and slow time domains, and complete the network training;

[0055] Step 7: Extract the joint fast-slow time domain features of the real and fake targets to be identified, and feed them into the pre-trained two-channel feature fusion network for identification.

Claims

1. A method for identifying true and false targets based on joint features in the fast and slow time domains when false target samples are scarce, characterized in that... Includes the following steps: Step 1: Perform pulse compression and coherent accumulation processing on real and false targets. First, the fast time expression of the radar target echo is as follows: ,but ; in, For signal pulse width, The frequency modulation slope of the signal. Let the signal bandwidth be denoted as , and the target echo fast time be expressed as . ,but ; in, The symbol for convolution. Given the radar system function in the time domain, the fast-slow time expression for the target echo is: s t t t m = h t s t - 2 R t t m / c e - j 4 π f 0 R t m / c ; in, To save time, For slow time, Let be the radial distance function between the target and the radar. The initial distance to the target. For the target radial velocity, At the speed of light, For carrier frequency; assume the number of coherent cumulative elements is . The pulse repetition interval is We use express A set of target echoes, each specific echo can be represented as... : ； In the above formula, it is assumed that the phase difference between two adjacent target echoes is... Then each target echo can be represented by the first target echo, and the results are as follows. ； The frequency-domain matched filter function for pulse compression is as follows: ； The above equation satisfies Assuming its time-domain representation is ; The phase weight matrix for achieving coherent accumulation is as follows: ； in, Representing different phase factors, , The number of points in the Discrete Fourier Transform. To represent different Doppler units, we use matrices. express The result of pulse compression and coherent accumulation of consecutive target echoes can be expressed as follows: ; in, Represents the matrix rank transformation symbol; Step 2: Use the peak search algorithm to establish the range cells and Doppler channels of the echoes from the real and false targets, assuming that the real target is found in the fast-slow time domain. The corresponding Doppler unit is Its corresponding The Doppler cell that achieves a large peak value is windowed and truncated along the distance dimension of the Doppler cell containing the real target. The truncated signal is then... ; In the formula, the phase compensation factor of the Doppler channel must be equal to or infinitely close to the echo phase difference, i.e. or ,but It can be simplified to ; Step 3: Extract the joint fast-slow time domain features of the real target echo using FFT; Step 4: Repeat steps 2 and 3 to extract the joint features of the fast and slow time domains of the false target echo, and construct a true-false target feature dataset U, setting the label of each sample in it to 0. Step 5: Set the label of each sample in the real target feature dataset P collected before the false target appeared to 1, and then apply the SVM to it. The training session, in the... In this training iteration, P is divided into two parts, P1 and P2, proportionally, and the SVM is trained using P1 and U. After each training iteration, the probability matrix is ​​estimated using the SVM. , which represents the probability that SVM estimates each sample in U as 1. Let U be the number of samples in U, and use SVM to estimate the probability. This represents the average probability that the SVM estimates each sample in P2 as 1; Obtain the class prior probability of each sample in U. ; in the After the training session, and will the former The estimated class prior probabilities are accumulated, i.e. ; Among them, before the first estimation begins As a zero vector, ultimately through After the estimation is completed, we get The cumulative result of the secondary prior probabilities And calculate the average class prior probability. When the average prior probability of a sample is greater than the probability threshold When the condition is met, the sample label is set to 1; otherwise, it is set to 0. This process re-labels each sample in U and completes the reconstruction of the dataset P+U. Step 6: Input the reconstructed dataset P+U into the two-channel feature fusion network to automatically extract the joint spectral fluctuation features in the fast and slow time domains, and complete the network training; Step 7: Extract the joint fast-slow time domain features of the real and fake targets to be identified, and feed them into the pre-trained two-channel feature fusion network for identification.

Images