A Method for Estimating Composite Jamming Parameters of Air Defense Warning Radar Based on YOLOv5

Abstract

The present invention belongs to the technical field of radar anti-jamming, and specifically relates to a method for estimating composite interference parameters of an air defense warning radar based on YOLOv5. First, the radar received signal containing composite interference is input, and the time-frequency feature maps of single-type interference and interactive composite interference are extracted; the time-frequency joint distribution map of single-type interference is data-labeled, anchor boxes are calibrated, and a small amount of time-frequency joint distribution maps of interactive composite interference are added to improve the diversity of the data set. Then, the YOLOv5 convolutional neural network is used for training to extract the features of the interference signals in the time-frequency joint distribution map; the features automatically extracted by the convolutional neural network are used to identify the interference signals, the judgment probability is given, the types of single-type interference and interactive composite interference are displayed, and the time-frequency information is preliminarily located using the anchor boxes; finally, the interference carrier frequency, relative time delay, and Doppler frequency shift are accurately estimated by comprehensively using the chi-square statistical test, local search and regression, and frequency-domain slice interpolation methods.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar anti-jamming, and specifically relates to a method for estimating composite interference parameters of an air defense warning radar based on YOLOv5. Background Art

[0002] The electronic active jamming affecting radar operation can be mainly divided into two categories: suppression jamming and deception jamming. Especially with the development of digital radio frequency storage technology (DRFM), deception jamming is often used in combination with suppression jamming, and even may be an interactive combination of multiple deception jammings and suppression jammings, which brings great challenges to interference suppression. And interference environment perception is the premise and foundation for the radar to implement effective anti-jamming measures. By analyzing the interference signal in multiple domains to master the interference situation information and accurately estimating the time-frequency parameters, it can provide reliable technical support for the radar to take anti-jamming measures, which is of great significance for the early warning radar to obtain a good working environment. Summary of the Invention

[0003] The purpose of the present invention is to solve the problem that it is difficult to obtain the interference delay and Doppler frequency information under the radar composite deception jamming environment, and to provide a method for estimating composite interference parameters of an air defense warning radar based on YOLOv5. The present invention can effectively extract the parameter information of the interference signal of the radar under a complex background, can simultaneously extract the time-frequency information of multiple interferences, and is easy to implement in engineering.

[0004] A method for estimating composite interference parameters of an air defense warning radar based on YOLOv5 includes the following steps:

[0005] Step 1: Receive the target echo signal; use the short-time Fourier transform to extract the features of the received signal;

[0006] The mathematical expression of the target echo signal is:

[0007]

[0008] where A is the signal amplitude; f j is the carrier frequency; k is the frequency modulation rate; T P is the pulse width; k is the frequency modulation rate; τ is the target delay, R is the relative distance between the target and the radar, C is the speed of light; the frequency transformation relationship of the linear frequency modulation signal is:

[0009] f = f j + kt

[0010] First, frame and window the signal, then perform discrete Fourier transform on the windowed signal respectively, and finally sum up the transformed results to generate a time-frequency diagram and the energy spectral density corresponding to each time-frequency point;

[0011]

[0012] Among them, N is the number of sampling points of the window function; x(m) is the windowed signal, x(m) = x(n)ω * (n - m); ω * (n - m) is the window function;

[0013] Step 2: Send the time-frequency feature maps of single-type interference and the time-frequency feature map data of composite interference into the YOLOv5 convolutional neural network for training;

[0014] Step 3: Input the time-frequency map data of the composite interference into the trained network, and output the recognition and localization results, including the composite interference category and the anchor box position information; calculate the mean average precision mAP;

[0015]

[0016] Among them, P is the accuracy rate, and R is the recall rate;

[0017] Step 4: Convert the output anchor box position information to obtain the time-frequency parameter ranges of each composite interference in the time-frequency map;

[0018] The frequency conversion relationship follows the formula:

[0019]

[0020] The time delay conversion relationship follows the formula:

[0021]

[0022] Among them, y max and y min are the anchor box positions corresponding to the maximum and minimum values of the frequency domain axis coordinates respectively; y pos is the anchor box position corresponding to the interference carrier frequency; f is the entire coordinate frequency range; x max and x min are the anchor box positions corresponding to the maximum and minimum values of the time domain axis coordinates respectively; x pos is the anchor box position corresponding to the interference time delay; t is the time length;

[0023] Step 5: In the local range obtained in Step 4, combine the energy spectral density time-frequency matrix after STFT for local search, find the probability density distribution of the noise energy spectral density, and obtain the corresponding energy spectral density according to the significance level α as the discrimination threshold E d ; Use the threshold detection to obtain the time delay τ, the carrier frequency f i and the sum of the rough Doppler frequency shift and the carrier frequency;

[0024] Step 6: Using local interpolation method, at the quadratic time-frequency localization point, take the frequency-domain slice at the frequency of this time-frequency point, fit the energy spectral density at each frequency point position of the slice with a quadratic polynomial, and then perform N i point linear interpolation;

[0025] N i Satisfy:

[0026]

[0027] where k is a proportionality coefficient, k > 1; f s is the sampling frequency; f c is the sum of the estimated Doppler frequency and the carrier frequency before interpolation;

[0028] Search according to the threshold within the interpolated frequency-domain slice to locate the sum of the fine Doppler frequency and the carrier frequency; the discrimination criterion is:

[0029] E(t, fi) > E α

[0030] where α is the significance level; E α is the test threshold, which is determined according to the significance level;

[0031] Step 7: Subtract the carrier frequency estimated in Step 2 from the sum of the carrier frequency and the Doppler frequency to obtain the Doppler frequency shift.

[0032] The beneficial effects of the present invention are as follows:

[0033] The present invention first combines the image processing method for preliminary parameter estimation, so it can estimate the time-frequency parameters of the composite interference, and obtain the time-frequency parameters of each interference participating in the composite; it has good generalization ability in the case of dynamic combination of multiple interferences, and can obtain good estimation results of interference signal parameters under low signal-to-noise ratio, which can provide reliable support for subsequent processing links such as interference suppression. The present invention can solve the problem that it is difficult to obtain the interference delay and Doppler frequency information in the radar composite deception interference environment, can effectively extract the parameter information of the interference signal of the radar in a complex background, can simultaneously extract the time-frequency information of multiple interferences, and is easy to implement in engineering. Brief Description of the Drawings

[0034] Figure 1 is the overall flowchart of the present invention.

[0035] Figure 2 is the flowchart of time-frequency feature extraction.

[0036] Figure 3 is an example diagram of a time-frequency search matrix.

[0037] Figure 4 is the time-frequency diagram of a chirp signal.

[0038] Figure 5 It is the time-frequency diagram of noise amplitude modulation interference.

[0039] Figure 6 It is the time-frequency diagram of intermittent sampling and forwarding interference.

[0040] Figure 7 It is the time-frequency diagram of dense false target interference.

[0041] Figure 8 It is the time-frequency diagram of comb spectrum interference.

[0042] Figure 9 It is the time-frequency diagram of composite interference. Detailed implementation manners

[0043] The present invention will be further described below with reference to the accompanying drawings.

[0044] The present invention proposes a method for estimating deception interference parameters of an air defense warning radar based on YOLOv5, including: extracting time-frequency characteristics of composite interference, preliminarily estimating composite interference parameters, and accurately estimating composite interference parameters.

[0045] The input of the system is the radar received signal containing composite interference. The radar signal is analyzed for characteristics, and the time-frequency joint distribution diagram is selected as the two-dimensional feature; the time-frequency characteristic diagrams of single-type interference and interactive composite interference are extracted as the input data of the convolutional neural network; the time-frequency joint distribution diagram of single-type interference is data-labeled, the anchor boxes are calibrated, and a small amount of time-frequency joint distribution diagrams of interactive composite interference are added to improve the diversity of the data set. The YOLOv5 convolutional neural network is used for training to extract the characteristics of the interference signal in the time-frequency joint distribution diagram, such as shape, texture, etc.; the characteristics automatically extracted by the convolutional neural network are used to identify the interference signal, the judgment probability is given, the types of single-type interference and interactive composite interference are displayed, and the time-frequency information is preliminarily located using the anchor boxes; finally, methods such as chi-square statistical test, local search and regression, and interpolation of frequency-domain slices are comprehensively used to accurately estimate the interference carrier frequency, relative time delay, and Doppler frequency shift.

[0046] The specific implementation steps are as follows:

[0047] Step 1: Use the short-time Fourier transform to extract the characteristics of the received signal as Figure 2 shown. First, after frame division and windowing, the Hamming window is selected as the window function, the window length N = 128 sampling points, and the sampling frequency f s corresponds to twice the bandwidth B; then the discrete Fourier transform is performed on the windowed signal respectively:

[0048]

[0049] where x(m) is the windowed signal, and its expression is:

[0050] x(m) = x(n)ω * (n - m)

[0051] where ω * (n - m) is the window function. Finally, the results after transformation are summed up to generate the time-frequency diagram and the energy spectral density corresponding to each time-frequency point.

[0052] Step 2: Feed the time-frequency feature map data of single-type interference and the time-frequency feature map data of composite interference into the YOLOv5 convolutional neural network for training; fuse the Focus structure and the CSP structure in the Backbone module to achieve the extraction of basic general features of the time-frequency image; then add the FPN + PAN module through the Neck network to further extract deep features.

[0053] Step 3: Input the time-frequency diagram data of composite interference into the trained network, and use the loss function: CIoU Loss , and DIOU selected by the prediction box nms to estimate the detection accuracy, output the recognition and positioning results, give the composite interference category and the anchor box position information, and calculate the mean average precision mAP. Its expression is as follows:

[0054]

[0055] where P and R are the precision rate and recall rate respectively.

[0056] Step 4: Convert the anchor box information output by the composite interference recognition module to obtain the range of each composite interference time-frequency parameter in the time-frequency diagram. The frequency conversion relationship follows the formula:

[0057]

[0058] The time delay conversion relationship follows the formula:

[0059]

[0060] where u max and u min are the anchor box positions corresponding to the maximum and minimum values of the frequency domain axis coordinates respectively, y pos is the anchor box position corresponding to the interference carrier frequency, f is the entire coordinate frequency range; x max and x min are the anchor box positions corresponding to the maximum and minimum values of the time domain axis coordinates respectively, x pos is the anchor box position corresponding to the interference time delay, and t is the time length.

[0061] Step 5: In the transformed local range, perform local search in combination with the energy spectral density time-frequency matrix after STFT, such asFigure 3 As shown in the figure; find the probability density distribution of the noise energy spectral density, and obtain the corresponding energy spectral density according to the significance level α as the discrimination threshold E d , and use the threshold detection to obtain the time delay τ and the carrier frequency f i as well as the sum of the rough Doppler frequency shift and the carrier frequency.

[0062] Step 6: Adopt the local interpolation method. At the above-mentioned quadratic time-frequency positioning points, take the frequency-domain slice at the frequency where the time-frequency point is located, fit the energy spectral density at each frequency point position of the slice with a quadratic polynomial, and then perform N i point linear interpolation. N i satisfies:

[0063]

[0064] where k (k>1) is a proportionality coefficient. The larger k is, the shorter the step size is; N is the number of sampling points of the window function. If it is a Hamming window, then N = 128; f s is the sampling frequency; f c is the sum of the Doppler frequency and the carrier frequency estimated before interpolation.

[0065] Search according to the threshold in the interpolated frequency-domain slice to locate the sum of the fine Doppler frequency and the carrier frequency. The discrimination criterion is:

[0066] E(t, fi)>E α

[0067] where α is the significance level and E α is the test threshold, which is determined according to the significance level.

[0068] Step 7: Subtract the carrier frequency estimated in Step 2 from the sum of the carrier frequency and the Doppler frequency to obtain the Doppler frequency shift.

[0069] In a specific embodiment of Step 1, the received signal selected by the present invention uses a linear frequency modulation signal as the target echo, and the composite interference signal is randomly composed of four types of interference, with a total of 11 composite methods. The four types of interference are divided into noise amplitude modulation interference, dense false target interference, comb-shaped spectrum interference, and intermittent sampling and forwarding interference according to different action mechanisms.

[0070] The target echo signal is as Figure 4 shown, and its mathematical expression is:

[0071]

[0072] where is the signal amplitude A, f j is the carrier frequency, k is the frequency modulation rate, T P is the pulse width, k is the frequency modulation rate, τ is the target time delay, and its expression is:

[0073]

[0074] Wherein, R is the relative distance between the target and the radar, and C is the speed of light. The frequency transformation relationship of the linear frequency modulation signal is:

[0075] f = f j + kt

[0076] That is, its frequency changes linearly with time, and in the time-frequency diagram, the frequency shows a linear increase from the carrier frequency f j starting to f, and the duration is limited by the pulse width, which is expressed as T P .

[0077] The noise amplitude modulation interference signal is as Figure 5 shown. It is a common suppression-type active interference. In the entire time domain, it can stably cover the target echo signal completely, so it is very difficult to obtain the specific information of the target signal. Its mathematical expression is:

[0078]

[0079] Wherein, the constant U0 is the carrier voltage, which is the DC bias set to meet the needs of the transmitter; U n (t) is the modulation noise, whose mean value is 0 and variance is σ n 2 ; f j is the carrier frequency of the noise interference; is the phase, and the change range is (0, 2π).

[0080] When the sampling frequency satisfies f s = 2f j , in its time-frequency joint distribution diagram, it is shown as a spectral line, and the distance between the two spectral lines depends on |f s - 2f j |, and the larger |f s - 2f j | is, the farther the distance is. Since it is a single-carrier frequency signal, its carrier frequency information can be clearly seen in the time-frequency diagram.

[0081] The intermittent sampling and forwarding interference is as Figure 6 shown. It is an interference that combines deception and suppression. Compared with traditional interference, it has a shorter sampling duration, generally in the mode of instantaneous sampling and instantaneous transmission, and has good interference timeliness. Its mathematical expression is:

[0082]

[0083] Wherein, s(t) is the linear frequency modulation signal, rect represents the gate function, that is, the envelope of the intermittent sampling and forwarding interference, f is the pulse width of the intermittent sampling, Ts is the sampling period, and δ(t - nT s ) represents the impulse function. n determines the number of times of intermittent sampling and forwarding, which is expressed as the number of interfering signals of sampling and forwarding in the time-frequency diagram. This interference is also sampled and forwarded with the chirp signal as the prototype, and its time-frequency characteristics are highly similar to those of the chirp signal.

[0084] Dense false target interference is as Figure 7 shown. It needs to preset a series of time intervals and modulation rules to forward multiple deception signals. For the radar target echo, it has both suppression and deception, and it is very difficult to identify the dense false target signal from a single time domain or frequency domain. If the dense false target interference is not suppressed at the signal level and allowed to enter the data layer, it will seriously affect the subsequent target tracking. The modeling of dense false target interference is as follows:

[0085]

[0086] In the formula, N is the number of dense false targets, the Doppler frequency shift is f d , the forwarding time delay is t i , k is the frequency modulation rate, T P is the pulse width, is the initial phase, and its variation range is (0, 2π). Since the dense false target interference is also modulated by the chirp signal, it is highly consistent with the chirp signal in the time-frequency diagram. Its frequency changes with time. When the Doppler frequency or the forwarding time delay is large, multiple separated false target signals can be seen, while when the Doppler frequency or the forwarding time delay is small, the energy spectrum characteristics will overlap.

[0087] Comb-like spectrum interference is as Figure 8 shown. It is mainly modulated by the product of the comb-like spectrum signal and the chirp signal. The generated comb-like spectrum interference signal has both deception and suppression. The mathematical expression of the comb-like spectrum interference is as follows:

[0088]

[0089] In the formula, A ii is the amplitude at the i-th frequency point; f i corresponds to the frequency point where each comb tooth appears; k is the frequency modulation rate. Since the comb-like spectrum interference is obtained by modulating the comb-like signal:

[0090]

[0091] Its frequency changes with time, and clear comb teeth can be seen on the time-frequency characteristic diagram.

[0092] Compound interference is as Figure 9 shown, and its mathematical expression is as follows:

[0093]

[0094] Wherein, N(t) is additive white Gaussian noise, whose amplitude follows a Gaussian distribution and power spectral density follows a uniform distribution, and n is a specific type of interference (n ≤ 4).

[0095] Compared with the prior art, the present invention first combines the image processing method to perform preliminary parameter estimation, so it can estimate the time-frequency parameters of the composite interference and obtain the time-frequency parameters of each interference participating in the composite; it has good generalization ability under the condition of dynamic combination of multiple interferences, and can achieve good estimation effect of interference signal parameters under low signal-to-noise ratio, which can provide reliable support for subsequent processing links such as interference suppression.

[0096] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for estimating the composite interference parameters of an air defense warning radar based on YOLOv5, characterized in that, It includes the following steps: Step 1: Receive the target echo signal; extract features from the received signal using short-time Fourier transform; The mathematical expression of the target echo signal is: Among them, A is the signal amplitude; f j is the carrier frequency; k is the frequency modulation rate; T P is the pulse width; k is the frequency modulation rate; τ is the target time delay, R is the relative distance between the target and the radar, C is the speed of light; the frequency transformation relationship of the linear frequency modulation signal is: f = f j + kt First, perform frame division and windowing, then perform discrete Fourier transform on the windowed signal respectively, and finally sum up the transformed results to generate a time-frequency diagram and the energy spectral density corresponding to each time-frequency point; Among them, N is the number of sampling points of the window function; x(m) is the windowed signal, x(m) = x(n)ω * (n - m); ω * (n - m) is the window function; Step 2: Feed the time-frequency feature map data of single-type interference and the time-frequency feature map data of composite interference into the YOLOv5 convolutional neural network for training; Step 3: Input the time-frequency diagram data of the composite interference into the trained network, and output the recognition and localization results, including the composite interference category and the anchor box position information; calculate the mean average precision mAP; mAP = ∫0 1 P(R) dR where P is the accuracy and R is the recall rate; Step 4: Convert the output anchor box position information to obtain the range of each composite interference time-frequency parameter in the time-frequency diagram; The frequency conversion relationship follows the formula: The time delay conversion relationship follows the formula: Among them, y max , y min are the anchor box positions corresponding to the maximum and minimum values of the frequency domain axis coordinates respectively; y pos is the anchor box position corresponding to the interfering carrier frequency; f is the entire coordinate frequency range; x max , x min are the anchor box positions corresponding to the maximum and minimum values of the time domain axis coordinates respectively; x pos is the anchor box position corresponding to the interfering time delay; t is the time length; Step 5: Within the local range obtained in Step 4, perform local search in combination with the time-frequency matrix of the energy spectral density after STFT, find the probability density distribution of the noise energy spectral density, and obtain the corresponding energy spectral density according to the significance level α as the discrimination threshold E d ; Use the threshold detection to obtain the time delay τ and the carrier frequency f i and the sum of the rough Doppler frequency shift and the carrier frequency; Step 6: Using local interpolation method, at the quadratic time-frequency localization point, take the frequency-domain slice at the frequency where the time-frequency point is located, fit the energy spectral density at each frequency point position of the slice with a quadratic polynomial, and then perform N i point linear interpolation; N i Satisfy: where k is a proportionality coefficient, k > 1; f s is the sampling frequency; f c is the sum of the estimated Doppler frequency before interpolation and the carrier frequency; Search within the interpolated frequency-domain slice according to the threshold to locate the sum of the fine Doppler frequency and the carrier frequency; the discrimination criterion is: E(t,fi)>E α where α is the significance level; E α is the test threshold, which is determined according to the significance level; Step 7: Subtract the carrier frequency estimated in Step 2 from the sum of the carrier frequency and the Doppler frequency to obtain the Doppler frequency shift.

Images