A high-frequency ground wave radar clutter and target identification method

Abstract

The application discloses a high-frequency ground wave radar clutter and target identification method, and belongs to the technical field of high-frequency ground wave radar clutter and target identification.The application solves the problems that the prior art cannot quickly and accurately distinguish target and clutter components and the problem of missing alarm in target identification.The application trains a deep learning model based on time-frequency analysis results and echo components corresponding to the time-frequency analysis results;for radar echo data to be detected, time-frequency analysis results corresponding to the radar echo data to be detected are input into the trained deep learning model, echo components contained in the radar echo data to be detected are output through the deep learning model, and identification of the clutter and the target is realized.The method can be applied to identification of high-frequency ground wave radar clutter and targets.

Description

Technical Field

[0001] This invention belongs to the field of high-frequency ground wave radar clutter and target recognition technology, specifically relating to a high-frequency ground wave radar clutter and target recognition method based on time-frequency analysis and deep learning. Background Technology

[0002] High-frequency ground-wave radar operates at frequencies between 3 and 30 MHz, with a maximum detection range of 200 to 500 kilometers. This demonstrates that its detection capability is sufficient to cover the exclusive economic zone (EEZ), enabling beyond-line-of-sight detection of surface ships and low-flying aircraft. With its advantages of long detection range, wide coverage, and all-weather operation, high-frequency ground-wave radar has become an effective facility for long-range detection and real-time early warning within the EEZ. However, one of the most challenging problems facing high-frequency ground-wave radar is the impact of clutter on target identification. First-order sea clutter and E-layer ionospheric clutter have energy far exceeding that of the target. Both types of clutter extend in both the Doppler and range dimensions, and this extended clutter can overwhelm moving targets. Consequently, a large number of targets will fail to be identified, leading to missed detections.

[0003] Time-frequency analysis is a primary method for analyzing clutter echo components, applied in various radar systems, such as X-band solid-state fully coherent radar and high-frequency ground wave radar, in both military and civilian applications. Short-time Fourier Transform (STFT) is widely used. However, due to the Heisenberg uncertainty principle, STFT cannot simultaneously achieve optimal time-domain and frequency-domain resolution. Furthermore, the Wigner-Ville distribution is also widely used in signal processing. However, when analyzing multi-component signals, the Wigner-Ville distribution introduces cross-interference terms. While wavelet transform can fully reveal signal details, its performance depends heavily on the chosen wavelet basis functions. Because ionospheric clutter and sea clutter change in real time, the above methods cannot effectively analyze the components of ionospheric clutter, sea clutter, and target echoes. The constantly changing conditions on the sea surface and in the ionosphere also affect the time-frequency analysis results of clutter components. Previous target recognition and detection methods based on time-frequency analysis required human judgment and were subjective.

[0004] In summary, existing technologies suffer from limitations in quickly and accurately distinguishing targets from clutter components, and they also exhibit issues with missed detections in target identification. Therefore, identifying dispersed targets and accurately identifying each type of clutter under the interference of sea clutter and ionospheric clutter in high-frequency ground wave radar remains a significant challenge, greatly impacting the performance of existing systems. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of existing technologies being unable to quickly and accurately distinguish between targets and clutter components, and the existence of missed detections in target identification. It proposes a high-frequency ground wave radar clutter and target identification method, which can be used to identify various E-layer ionospheric clutter, sea clutter and targets of different shapes under high-frequency ground wave radar, thereby increasing the probability of target detection.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is: a method for identifying clutter and targets in high-frequency ground wave radar, the method specifically including the following steps:

[0007] Step 1: Obtain a dataset containing M batches of radar echo data, and preprocess each batch of radar echo data in the obtained dataset to obtain the range-Doppler image corresponding to each batch of radar echo data.

[0008] Step 2: For any batch of radar echo data in Step 1, normalize the echo energy of each range cell in the range-Doppler chart corresponding to the batch of radar echo data to obtain normalized echo data; then perform time-frequency analysis on each range cell in the normalized echo data to obtain the time-frequency analysis results for each range cell.

[0009] Similarly, the above processing is performed on the range-Doppler plots corresponding to each batch of radar echo data in step one;

[0010] Step 3: Use the time-frequency analysis results as the input of the training set, and use the echo components contained in the distance cells corresponding to the time-frequency analysis results as the output of the training set. Use the input and output of the training set to train the YOLO-v5 network to obtain the trained YOLO-v5 network.

[0011] Step 4: Preprocess the radar echo data to be detected;

[0012] Step 5: Normalize the preprocessing results from Step 4 to obtain normalized echo data;

[0013] Step 6: Perform time-frequency analysis on each distance cell in the normalized echo data from Step 5 to obtain the time-frequency analysis results;

[0014] Step 7: Input the time-frequency analysis results obtained in Step 6 into the trained YOLO-v5 network. The trained YOLO-v5 network outputs echo components, thus completing the identification of different types of clutter and different targets.

[0015] Furthermore, each batch of radar echo data in the acquired dataset is preprocessed. The specific preprocessing process is as follows:

[0016] For any batch of radar echo data in the acquired dataset, the radar echo data is sequentially processed by filtering, amplification, demodulation, pulse compression and beamforming.

[0017] Similarly, each batch of radar echo data is preprocessed separately.

[0018] Furthermore, the normalization method used in step two is global threshold normalization.

[0019] Furthermore, the specific process of global threshold normalization is as follows:

[0020] Find the maximum echo energy value of all range cells from the range-Doppler map, define the maximum echo energy value as Max, and regularize the echo energy value of each range cell according to the maximum echo energy value, mapping the echo energy value of each range cell to the range [0,1].

[0021] Furthermore, the time-frequency analysis results for each distance cell include the following:

[0022] 1) Time-frequency analysis results when the range cell contains only extended ionospheric clutter;

[0023] 2) Time-frequency analysis results when the range cell contains only uniformly distributed specular ionospheric clutter;

[0024] 3) Time-frequency analysis results when the range cell contains only non-uniformly distributed specular ionospheric clutter;

[0025] 4) Time-frequency analysis results when the range cell contains only a single target;

[0026] 5) Time-frequency analysis results when multiple targets are fused within a range cell;

[0027] 6) Time-frequency analysis results when the range cell contains targets near sea clutter;

[0028] 7) Time-frequency analysis results when the range cell contains targets near extended ionospheric clutter;

[0029] 8) Time-frequency analysis results when the range cell contains targets with uniformly distributed mirror delamination clutter;

[0030] 9) Time-frequency analysis results when the range cell contains targets with nearby non-uniformly distributed specular delamination clutter;

[0031] 10) Time-frequency analysis results when the range cell contains extended ionospheric clutter and sea clutter;

[0032] 11) Time-frequency analysis results when the range cell contains uniformly distributed specular ionospheric clutter and sea clutter;

[0033] 12) Time-frequency analysis results when the distance cell contains non-uniformly distributed specular ionospheric clutter and sea clutter.

[0034] Furthermore, when a target near sea clutter is contained in a certain range cell of the normalized echo data, the specific process of performing time-frequency analysis on that range cell is as follows:

[0035] The radar echo data on this range cell is represented as:

[0036] χ(t)=c(t)+n(t)+s(t) (1)

[0037] Where χ(t) represents radar echo data, c(t) represents sea clutter signal, s(t) represents target signal, and n(t) represents noise signal;

[0038] At this distance cell, the received target signal s(t) is represented by the superposition of the target components at time t:

[0039]

[0040] s m (t)=α m exp(j2πf m t+ψ m (3)

[0041] Where M is the number of targets, s m (t) is the echo component of the m-th target, α m Let f be the amplitude of the m-th target signal, j be the imaginary unit, and f be the amplitude of the m-th target signal. m ψ is the frequency of the m-th target. m It is the random phase of the m-th target, and t is time;

[0042] At this range cell, sea clutter is represented as:

[0043]

[0044] Among them, A i+ A is the amplitude modulation coefficient of the first-order peak affected by wave height. i- Let be the i-th amplitude modulation factor affected by wave height for the negative first-order peak, e be the base of the natural logarithm, and v be the amplitude modulation factor. t (i) represents the radial current velocity of the surface source in the i-th direction, f B For the Bragg frequency, The initial phase is represented by N, the number of sea scattering patches is represented by λ, and the carrier wavelength of the radar is represented by λ.

[0045] The received signal model is simplified to:

[0046] f(t)=c(t)+s(t) (5)

[0047] Where f(t) is the simplified model of the received signal;

[0048]

[0049] In the formula, ρ is the frequency corresponding to the Fourier transform. The Fourier transform result of the received signal f(t);

[0050] The Hamming window method (6) is used for processing:

[0051]

[0052] Among them, Γ f (η,t) represents the short-time Fourier transform of the received signal f, g(·) is the Hamming window, τ represents the center point of the window function g(t) after shifting along the time t axis, and η represents the frequency corresponding to the short-time Fourier transform. It is the field of real numbers;

[0053] The radar echo signal is represented at a fixed time t0 as:

[0054]

[0055] Substituting equation (8) into equation (7), the radar echo signal is expressed as equation (9):

[0056]

[0057] in, For the estimated representation of g(·), define

[0058] According to Γ f (η,t) is used to obtain the local instantaneous frequency of the radar echo signal f(t).

[0059]

[0060] in, On behalf of the Pragmatic Department, Represents Γ f Partial derivatives of (η,t);

[0061] according to The mapping relationship, FSST will Γ f The frequency variables (η,t) are redistributed to obtain the final time-frequency plane representation T. f (ω,t):

[0062]

[0063] Where δ is the Dirac function and ω represents the frequency after the synchronous compression transform based on Fourier.

[0064] Furthermore, when a target near extended ionospheric clutter is contained in a certain range cell of the normalized echo data, the specific process of performing time-frequency analysis on that range cell is as follows:

[0065] The radar echo data on this range cell is represented as:

[0066] χ(t)=ξ(t)+n(t)+s(t) (12)

[0067] Where χ(t) represents radar echo data, ξ(t) represents extended ionospheric clutter signal, s(t) represents target signal, and n(t) represents noise signal;

[0068] The extended ionospheric clutter signal ξ(t) is:

[0069]

[0070] Where ε(t) is the time-varying clutter amplitude, The Doppler frequency of the time-varying ionosphere;

[0071] Single target signal s m (t) is represented as:

[0072] s m (t)=α m exp(j2πf m t+ψ m (14)

[0073] Where, α m Let ψ be the amplitude of the m-th target signal. m Let f be the random phase of the m-th target signal. m Let m be the frequency of the m-th target;

[0074] At time t, when the range cell receives M target echo signals, the target signal s(t) is expressed as:

[0075]

[0076] Among them, s m (t) is the echo component of the m-th target, and n(t) is Gaussian white noise;

[0077] The received signal model is simplified to:

[0078] f(t)=s(t)+ξ(t) (16)

[0079] Where f(t) is the simplified model of the received signal;

[0080]

[0081] In the formula, ρ is the frequency corresponding to the Fourier transform, and t is time. The Fourier transform result of the received signal f(t);

[0082] The Hamming window method (17) is used for processing:

[0083]

[0084] Among them, Γ f (η,t) represents the short-time Fourier transform of the received signal f, g(·) is the Hamming window, τ represents the center point of the window function g(t) after shifting along the time t axis, and η represents the frequency corresponding to the short-time Fourier transform. It is the field of real numbers;

[0085] The radar echo signal is represented at a fixed time t0 as:

[0086]

[0087] Where ε(t0) is the time-varying clutter amplitude, and ε'(t0) is the derivative of the time-varying clutter amplitude function. It is the Doppler frequency of the time-varying ionosphere. It is the derivative of the Doppler frequency function of the time-varying ionosphere;

[0088] Substituting equation (19) into equation (18), the radar echo signal is expressed as equation (20):

[0089]

[0090] in, For the estimated representation of g, define and η = ε'(t);

[0091] According to Γ f (η,t) is used to obtain the local instantaneous frequency of the radar echo signal f(t).

[0092]

[0093] in, On behalf of the Pragmatic Department, Represents Γ f Partial derivatives of (η,t);

[0094] according to The mapping relationship, FSST will Γ fThe frequency variables (η,t) are redistributed to obtain the final time-frequency plane representation T. f (ω,t):

[0095]

[0096] Where δ is the Dirac function and ω represents the frequency after the synchronous compression transform based on Fourier.

[0097] Furthermore, the absolute value of the amplitude of the m-th target signal |α m |for:

[0098]

[0099] Among them, SCR t It is the signal-to-noise ratio at time t.

[0100] The beneficial effects of this invention are:

[0101] This invention trains a deep learning model based on time-frequency analysis results and the corresponding echo components. For radar echo data to be detected, the time-frequency analysis results corresponding to the radar echo data are input into the trained deep learning model, which then outputs the echo components contained in the radar echo data. Based on time-frequency analysis and deep learning techniques, this invention can identify targets approaching clutter, weak targets, and fused targets. It accurately and quickly identifies ionospheric clutter and sea clutter, which is beneficial for target detection and track tracking, effectively reducing missed detections. Furthermore, the method of this invention is simple and convenient to implement, and highly adaptable. Attached Figure Description

[0102] Figure 1 This is a flowchart of a high-frequency ground wave radar clutter and target identification method according to the present invention;

[0103] Figure 2a A schematic diagram showing the time-frequency representation of the echo signal when it includes sea clutter and extended ionospheric clutter;

[0104] Figure 2b A schematic diagram showing the time-frequency representation of the echo signal when only extended ionospheric clutter exists.

[0105] Figure 2c A schematic diagram showing the time-frequency representation of the echo signal when the target and sea clutter are included;

[0106] Figure 3 This is a schematic diagram of the recognition results obtained using the method of the present invention. Detailed Implementation

[0107] Specific Implementation Method 1: Combination Figure 1This embodiment describes a method for identifying clutter and targets using high-frequency ground wave radar. The method specifically includes the following steps:

[0108] Step 1: Obtain a dataset containing M batches of radar echo data, and preprocess each batch of radar echo data in the obtained dataset to obtain the range-Doppler spectrum (RD spectrum) corresponding to each batch of radar echo data.

[0109] Step 2: For any batch of radar echo data in Step 1, normalize the echo energy of each range cell in the range-Doppler chart corresponding to the batch of radar echo data to obtain normalized echo data; then perform time-frequency analysis on each range cell in the normalized echo data to obtain the time-frequency analysis results for each range cell.

[0110] Similarly, the above processing is performed on the range-Doppler plots corresponding to each batch of radar echo data in step one;

[0111] Step 3: Use the time-frequency analysis results as the input of the training set, and use the echo components contained in the distance cells corresponding to the time-frequency analysis results as the output of the training set. Use the input and output of the training set to train the YOLO-v5 network to obtain the trained YOLO-v5 network.

[0112] We used a YOLO network based on transfer learning to learn features and complete the detection of different echo components.

[0113] (1) When training data is limited, it is insufficient to distinguish between targets and clutter. Therefore, this invention employs a data augmentation method. Two different schemes are proposed to address the differences in data distribution in heterogeneous scenarios. The first scheme aims to increase the amount of data on targets adjacent to clutter by adding simulated targets near real clutter. That is, given the known range-Doppler coordinates, randomly generated simulated targets are superimposed on the clutter edges. The second scheme aims to increase the amount of data on multiple fused targets by simulating two adjacent targets at specified range-Doppler coordinates.

[0114] (2) Considering the target's motion and rapidly changing clutter scene, we adopted a YOLO network combined with transfer learning to learn the time-frequency spectrum of clutter with multiple components. We first determined several batches of data to be detected, and collected a large amount of previous data starting from these batches. We performed time-frequency analysis on the large amount of previous data. We labeled the time-frequency spectrum of different components under each range unit, and the output magnitude of the spectrum of different echo components was different. Thus, we created a training sample of 2000 range units and set and adjusted the YOLO-v5 network parameters. The 2000 range unit training samples covered approximately 40 minutes of echo data. The test set consisted of 500 range unit training samples, and the training set and test set were independent.

[0115] Step 4: Preprocess the radar echo data to be detected; the preprocessing method is the same as in Step 1.

[0116] Step 5: Normalize the preprocessing results from Step 4 to obtain normalized echo data;

[0117] Step 6: Perform time-frequency analysis on each distance cell in the normalized echo data from Step 5 to obtain the time-frequency analysis results;

[0118] Step 7: Input the time-frequency analysis results obtained in Step 6 into the trained YOLO-v5 network. The trained YOLO-v5 network outputs echo components. Based on the differences in the time-frequency distribution characteristics of different echo components, the identification of different clutter and different targets is completed.

[0119] like Figure 3 The diagram shown is a schematic representation of the recognition results obtained using the method of the present invention.

[0120] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that it preprocesses each batch of radar echo data in the acquired dataset. The specific preprocessing process is as follows:

[0121] For any batch of radar echo data in the acquired dataset, the radar echo data is sequentially processed by filtering, amplification, demodulation, pulse compression and beamforming.

[0122] Similarly, each batch of radar echo data is preprocessed separately.

[0123] The other steps and parameters are the same as in Specific Implementation Method 1.

[0124] Specific Implementation Method 3: This implementation method differs from Specific Implementation Method 1 or 2 in that the normalization method used in step 2 is global threshold normalization.

[0125] Other steps and parameters are the same as in specific implementation method one or two.

[0126] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the specific process of global threshold normalization is as follows:

[0127] Find the maximum echo energy value of all range cells from the range-Doppler map, define the maximum echo energy value as Max, and regularize the echo energy value of each range cell according to the maximum echo energy value, mapping the echo energy value of each range cell to the range [0,1].

[0128] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0129] Specific Implementation Method Five: Combining Figure 2a , Figure 2b and Figure 2c This embodiment is described below. The difference between this embodiment and one of the specific embodiments one to four is that the time-frequency analysis results on each distance unit include the following:

[0130] 1) Time-frequency analysis results when the range cell contains only extended ionospheric clutter;

[0131] 2) Time-frequency analysis results when the range cell contains only uniformly distributed specular ionospheric clutter;

[0132] 3) Time-frequency analysis results when the range cell contains only non-uniformly distributed specular ionospheric clutter;

[0133] 4) Time-frequency analysis results when the range cell contains only a single target;

[0134] 5) Time-frequency analysis results when multiple targets (two or more targets) are fused in a range cell;

[0135] 6) Time-frequency analysis results when the range cell contains targets near sea clutter;

[0136] 7) Time-frequency analysis results when the range cell contains targets near extended ionospheric clutter;

[0137] 8) Time-frequency analysis results when the range cell contains targets with uniformly distributed mirror delamination clutter;

[0138] 9) Time-frequency analysis results when the range cell contains targets with nearby non-uniformly distributed specular delamination clutter;

[0139] 10) Time-frequency analysis results when the range cell contains extended ionospheric clutter and sea clutter;

[0140] 11) Time-frequency analysis results when the range cell contains uniformly distributed specular ionospheric clutter and sea clutter;

[0141] 12) Time-frequency analysis results when the distance cell contains non-uniformly distributed specular ionospheric clutter and sea clutter.

[0142] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0143] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that, when a target near sea clutter is present in a certain range cell of the normalized echo data, the specific process for performing time-frequency analysis on that range cell is as follows:

[0144] The radar echo data of the range cell to be identified mainly includes the target, sea clutter, and background noise. The radar echo data of this range cell is represented as follows:

[0145] χ(t)=c(t)+n(t)+s(t) (1)

[0146] Where χ(t) represents radar echo data, c(t) represents sea clutter signal, s(t) represents target signal, and n(t) represents noise signal;

[0147] At this distance cell, the received target signal s(t) is represented by the superposition of the target components at time t:

[0148]

[0149] s m (t)=α m exp(j2πf m t+ψ m (3)

[0150] Where M is the number of targets, s m (t) is the echo component of the m-th target, α m Let f be the amplitude of the m-th target signal, j be the imaginary unit, and f be the amplitude of the m-th target signal. m ψ is the frequency of the m-th target. m It is the random phase of the m-th target, and t is time;

[0151] At this range cell, sea clutter is represented as:

[0152]

[0153] Among them, A i+ A is the amplitude modulation coefficient of the first-order peak affected by wave height. i- Let be the i-th amplitude modulation factor affected by wave height for the negative first-order peak, e be the base of the natural logarithm, and v be the amplitude modulation factor. t (i) represents the radial current velocity of the surface source in the i-th direction, f B For the Bragg frequency, The initial phase is represented by N, the number of sea scattering patches is represented by λ, and the carrier wavelength of the radar is represented by λ.

[0154] Assuming the current range cell to be identified mainly consists of sea clutter and the target component, interference, noise, and ionospheric clutter will be temporarily ignored in the following study. Therefore, the received signal model is simplified to:

[0155] f(t)=c(t)+s(t) (5)

[0156] Where f(t) is the simplified model of the received signal;

[0157] Use functions The Fourier transform of the received signal f is normalized as follows:

[0158]

[0159] In the formula, ρ is the frequency corresponding to the Fourier transform. The Fourier transform result of the received signal f(t);

[0160] Gabor proposed the Short-Time Fourier Transform (SFT) in 1946. Essentially, it multiplies the signal by a time-finite sliding window *g* before performing the SFT, assuming that the non-stationary signal is stationary within a short time interval of the window. By shifting the window function along the time axis, a set of local "spectrums" of the signal is obtained through segment-by-segment analysis. This experiment uses the Hanning window.

[0161] The Hamming window method (6) is used for processing:

[0162]

[0163] Among them, Γ f (η,t) represents the short-time Fourier transform of the received signal f, known as the spectrum of the received signal f. g(·) is the Hamming window, τ represents the center point of the window function g(t) after shifting along the time t axis, and η represents the frequency corresponding to the short-time Fourier transform. It is the field of real numbers;

[0164] In this radar echo, the target component changes slowly in instantaneous amplitude and frequency, while the sea clutter component changes to some extent in instantaneous amplitude and frequency. The radar echo signal can be represented at a fixed time t0 (ignoring noise effects) as follows:

[0165]

[0166] Substituting equation (8) into equation (7), the radar echo signal is expressed as equation (9):

[0167]

[0168] in, Let Γ be the estimated representation of g(·). f (η,t) represents the STFT (Short Time Fourier Transform) of the radar echo signal f(t). From formula (9), it can be concluded that the peak is concentrated near two ridges, and the definition is...

[0169] Furthermore, SST, as a decomposition method, can separate and demodulate different components of a signal. According to Γ... f (η,t) is used to obtain the local instantaneous frequency of the radar echo signal f(t).

[0170]

[0171] in, On behalf of the Pragmatic Department, Represents Γ f Partial derivatives of (η,t);

[0172] according to The mapping relationship, FSST (Fourier-Based Synchrosqueezing Transform) will Γ f The frequency variables (η,t) are redistributed to obtain the final time-frequency plane representation T. f (ω,t):

[0173]

[0174] Where δ is the Dirac function, and ω represents the frequency after the synchronous compression transform based on Fourier.

[0175] FSST sharpens the time-frequency plane with respect to... and Information on the components surrounding the spine.

[0176] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0177] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that, when a certain range cell in the normalized echo data contains a target near extended ionospheric clutter, the specific process for performing time-frequency analysis on that range cell is as follows:

[0178] The radar echo data of the range cell to be identified mainly includes the target, ionospheric clutter, and background noise. The radar echo data of this range cell is represented as follows:

[0179] χ(t)=ξ(t)+n(t)+s(t) (12)

[0180] Where χ(t) represents radar echo data, ξ(t) represents extended ionospheric clutter signal, s(t) represents target signal, and n(t) represents noise signal;

[0181] For a specific range cell subjected to strong ionospheric clutter interference, the extended ionospheric clutter signal ξ(t) is:

[0182]

[0183] Where ε(t) is the time-varying clutter amplitude, The Doppler frequency of the time-varying ionosphere;

[0184] Single target signal s m (t) is represented as:

[0185] s m (t)=α m exp(j2πf m t+ψ m (14)

[0186] Where, α m Let ψ be the amplitude of the m-th target signal. m Let f be the random phase of the m-th target signal. m Let m be the frequency of the m-th target;

[0187] At time t, when the range cell receives M target echo signals, the target signal s(t) is expressed as:

[0188]

[0189] Among them, s m (t) is the echo component of the m-th target, and n(t) is Gaussian white noise;

[0190] Our focus is on the issue of the ionosphere covering the target. Therefore, interference, noise, and other types of clutter are temporarily ignored. The received signal model is simplified to:

[0191] f(t)=s(t)+ξ(t) (16)

[0192] Where f(t) is the simplified model of the received signal;

[0193] Use functions The Fourier transform of the received echo component f is normalized as follows:

[0194]

[0195] In the formula, ρ is the frequency corresponding to the Fourier transform, and t is time. The Fourier transform result of the received signal f(t);

[0196] Gabor proposed the Short-Time Fourier Transform (SFT) in 1946. Essentially, it multiplies the signal by a time-finite sliding window *g* before performing the SFT, assuming that the non-stationary signal is stationary within a short time interval of the window. By shifting the window function along the time axis, a set of local "spectrums" of the signal is obtained through segment-by-segment analysis. This experiment uses the Hanning window.

[0197] The Hamming window method (17) is used for processing:

[0198]

[0199] Among them, Γ f (η,t) represents the short-time Fourier transform of the received signal f, known as the spectrum of the received signal f. g(·) is the Hamming window, τ represents the center point of the window function g(t) after shifting along the time t axis, and η represents the frequency corresponding to the short-time Fourier transform. It is the field of real numbers;

[0200] We now study the STFT conversion of a multi-component radar echo signal. In this radar echo, the target component changes slowly in instantaneous amplitude and frequency, while the ionospheric clutter component changes rapidly in instantaneous amplitude and frequency. The radar echo signal is represented at a fixed time t0 as:

[0201]

[0202] Where ε(t0) is the time-varying clutter amplitude, and ε'(t0) is the derivative of the time-varying clutter amplitude function. It is the Doppler frequency of the time-varying ionosphere. It is the derivative of the Doppler frequency function of the time-varying ionosphere;

[0203] Substituting equation (19) into equation (18), the radar echo signal is expressed as equation (20):

[0204]

[0205] in, Let Γ be an estimate of g. f (η,t) represents the STFT (Short Time Fourier Transform) of the radar echo signal f(t). From formula (20), it can be concluded that the peak is concentrated near two ridges. and η = ε'(t);

[0206] Furthermore, SST, as a decomposition method, can separate and demodulate different components of a signal. According to Γ...f (η,t) is used to obtain the local instantaneous frequency of the radar echo signal f(t).

[0207]

[0208] in, On behalf of the Pragmatic Department, Represents Γ f Partial derivatives of (η,t);

[0209] according to The mapping relationship, FSST will Γ f The frequency variables (η,t) are redistributed to obtain the final time-frequency plane representation T. f (ω,t):

[0210]

[0211] Where δ is the Dirac function, and ω represents the frequency after the synchronous compression transform based on Fourier.

[0212] FSST sharpens the time-frequency plane with respect to... Information on the components surrounding the ε'(t) ridge.

[0213] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0214] This invention takes "targets containing near-sea clutter in the range cell" and "targets containing near-ionospheric clutter in the range cell" as examples for time-frequency analysis. The time-frequency analysis method is similar for cases where the range cell contains only E-layer extended ionospheric clutter, only uniformly distributed E-layer specular ionospheric clutter, only non-uniformly distributed E-layer specular ionospheric clutter, only a single target, and multiple targets (greater than or equal to 2 targets) fused together.

[0215] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that the absolute value of the amplitude of the m-th target signal |α m |for:

[0216]

[0217] Among them, SCR t It is the signal-to-noise ratio at time t.

[0218] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0219] Experimental Section

[0220] The training results at epochs=160 are shown in Table 1:

[0221] Table 1

[0222] category precision recall overall 0.898 0.898 Sea clutter 0.953 0.885 Ionospheric clutter 0.856 0.879 Single target 0.77 0.727 Two integration goals 1 1 Targets integrated into ionospheric clutter 0.909 1

[0223] In summary, this invention proposes a device and method for identifying clutter and targets in high-frequency ground wave radar based on time-frequency analysis and deep learning. This invention can identify different clutter components and various forms of targets in the echo under non-stationary background of high-frequency ground wave radar, and can also find targets integrated in clutter-contaminated areas. It is beneficial for target detection and track tracking, and has the characteristics of simple and convenient implementation and adaptability.

[0224] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A high-frequency ground wave radar clutter and target discrimination method, characterized by, The method specifically comprises the following steps: Step one, obtaining a data set containing M batches of radar echo data, and respectively pre-processing each batch of radar echo data in the obtained data set to obtain a distance-Doppler graph corresponding to each batch of radar echo data; Step two, for any one batch of radar echo data in step one, normalizing the echo energy of each distance unit in the distance-Doppler graph corresponding to the batch of radar echo data to obtain normalized echo data; and then performing time-frequency analysis on each distance unit in the normalized echo data to obtain time-frequency analysis results on each distance unit; The time-frequency analysis results on each distance unit include the following: 1) time-frequency analysis results when the distance unit only contains extended ionospheric clutter; 2) time-frequency analysis results when the distance unit only contains uniformly distributed specular ionospheric clutter; 3) time-frequency analysis results when the distance unit only contains non-uniformly distributed specular ionospheric clutter; 4) time-frequency analysis results when the distance unit only contains a single target; 5) time-frequency analysis results when the distance unit contains multiple targets; 6) time-frequency analysis results when the distance unit contains a target close to sea clutter; 7) time-frequency analysis results when the distance unit contains a target close to extended ionospheric clutter; 8) time-frequency analysis results when the distance unit contains a target close to uniformly distributed specular ionospheric clutter; 9) time-frequency analysis results when the distance unit contains a target close to non-uniformly distributed specular ionospheric clutter; 10) time-frequency analysis results when the distance unit contains extended ionospheric clutter and sea clutter; 11) time-frequency analysis results when the distance unit contains uniformly distributed specular ionospheric clutter and sea clutter; 12) time-frequency analysis results when the distance unit contains non-uniformly distributed specular ionospheric clutter and sea clutter; Similarly, the distance-Doppler graph corresponding to each batch of radar echo data in step one is processed in the above manner; Step three, taking the time-frequency analysis results as the input of the training set, taking the echo components contained in the distance unit corresponding to the time-frequency analysis results as the output of the training set, training the Yolo-v5 network using the input and output of the training set, and obtaining the trained Yolo-v5 network; Step four, pre-processing the radar echo data to be detected; Step five, normalizing the pre-processing result in step four to obtain normalized echo data; Step six, performing time-frequency analysis on each distance unit in the normalized echo data in step five to obtain time-frequency analysis results; Step seven, inputting the time-frequency analysis results obtained in step six into the trained Yolo-v5 network, outputting echo components through the trained Yolo-v5 network, and completing the identification of different kinds of clutter and different targets.

2. The method according to claim 1, wherein The specific process of pre-processing each batch of radar echo data in the obtained data set is as follows: For any one batch of radar echo data in the obtained data set, sequentially performing filtering, amplification, demodulation, pulse compression, and beam forming processing on the batch of radar echo data; Similarly, each batch of radar echo data is pre-processed respectively.

3. The method according to claim 2, wherein The normalization method used in the second step is global threshold normalization.

4. The method of claim 3, wherein the method further comprises: The specific process of the global threshold normalization is as follows: Find the maximum echo energy value of all distance units from the range-Doppler map, define the found maximum echo energy value as Max, and normalize the echo energy value of each distance unit according to the maximum echo energy value, and map the echo energy value of each distance unit to [0, 1].

5. The HF ground wave radar clutter and target identification method of claim 4, wherein, When the target close to sea clutter is contained in a certain distance unit in the normalized echo data, the specific process of time-frequency analysis on the distance unit is as follows: The radar echo data on the distance unit is expressed as: χ(t)=c(t)+n(t)+s(t) (1) Wherein, χ(t) is the radar echo data, c(t) represents the sea clutter signal, s(t) represents the target signal, and n(t) represents the noise signal. On the distance unit, the received target signal s(t) is expressed by the superposition of target components at time t: s m (t) = a m exp(j2πf m t + ψ m ) (3) where M is the number of targets, s m (t) is the echo component of the mth target, a m is the amplitude of the mth target signal, j is the imaginary unit, f m is the frequency of the mth target, ψ m is the random phase of the mth target, t is time; On the distance unit, the sea clutter is expressed as: where A i+ is the i th amplitude modulation coefficient of the positive first-order peak affected by wave height, A i- is the i th amplitude modulation coefficient of the negative first-order peak affected by wave height, e is the base of natural logarithm, v t (i) is the radial current velocity of the i th direction surface source, f B is the Bragg frequency, indicates the initial phase, N indicates the number of sea scattering patches, and λ is the carrier wavelength of the radar. The received signal model is simplified as: f(t)=c(t)+s(t) (5) Wherein, f(t) is the simplified received signal model. where p is the frequency corresponding to the Fourier transform, is the Fourier transform result of the received signal f(t); The Hamming window is used to process formula (6): where Γ f (η, t) is a short-time Fourier transform result representation of the received signal f, g(·) is a Hamming window, τ represents the center point position of the window function g(t) after moving on the time t axis, η represents the frequency corresponding to the short-time Fourier transform, is a real number field; The radar echo signal is expressed as: The radar echo signal is expressed as formula (9) by substituting formula (8) into formula (7): wherein, is an estimate of g(·), defined as According to Γ f (η, t) obtains the local instantaneous frequency of the radar echo signal f(t) wherein, representing the practicality department, representing Γ f partial derivative of (η, t); According to the mapping relationship of , FSST reassigns the frequency variable of f (η, t) to obtain the final time-frequency plane representation f (ω, t): Wherein, δ is Dirac function, and ω represents the frequency after Fourier-based synchronous compression transformation.

6. The HF ground wave radar clutter and target identification method of claim 4, wherein, When the target close to extended ionospheric clutter is contained in a certain distance unit in the normalized echo data, the specific process of time-frequency analysis on the distance unit is as follows: The radar echo data on the distance unit is expressed as: χ(t)=ξ(t)+n(t)+s(t) (12) Wherein, χ(t) represents the radar echo data, ξ(t) represents the extended ionospheric clutter signal, s(t) represents the target signal, and n(t) is Gaussian white noise. The extended ionospheric clutter signal ξ(t) is: where ε(t) is the time-varying clutter amplitude, is the Doppler frequency of the time-varying ionosphere; Single target signal s m (t) is represented as: s m (t) = a m exp(j2πf m t + ψ m ) (14) wherein α m is the amplitude of the mth target signal, ψ m is the random phase of the mth target signal, f m is the frequency of the mth target. When M target echo signals are received by the distance unit at time t, the target signal s(t) is expressed as: where s m (t) is the echo component of the mth target, and n(t) is Gaussian white noise. The received signal model is simplified as: f(t)=s(t)+ξ(t) (16) Wherein, f(t) is the simplified received signal model. where p is the frequency corresponding to the Fourier transform, t is time, is the Fourier transform result of the received signal f(t); The Hamming window is used to process formula (17): where Γ f (η, t) is a short-time Fourier transform result representation of the received signal f, g(·) is a Hamming window, τ represents the center point position of the window function g(t) after moving on the time t axis, η represents the frequency corresponding to the short-time Fourier transform, is the real number field; The radar echo signal is expressed as: where ε(t0) is the time-varying clutter amplitude, ε'(t0) is the derivative of the time-varying clutter amplitude function, is the Doppler frequency of the time-varying ionosphere, is the derivative of the Doppler frequency function of the time-varying ionosphere; The radar echo signal is expressed as formula (20) by substituting formula (19) into formula (18): wherein is an estimate of g, defined as and η = ε'(t); According to Γ f (η,t) obtains the local instantaneous frequency of the radar echo signal f(t) wherein, representing the realization department, representing Γ f partial derivative of (η, t); According to the mapping relationship , FSST reassigns the frequency variable of T f (η,t) to obtain the final time-frequency plane representation T f (ω,t): Wherein, δ is Dirac function, and ω represents the frequency after Fourier-based synchronous compression transformation.

7. The HF ground wave radar clutter and target discrimination method according to claim 5 or 6, characterized in that, An absolute value of the amplitude of the mth target signal |α m | is: wherein SCR t is the signal-to-clutter ratio at time t.

Images