Sea clutter suppression method combining phase processing mechanism and LSTM network model

By combining the phase processing mechanism and the LSTM network model, the phase characteristic information of sea clutter is used to predict and suppress sea clutter, the problem of low sea clutter prediction accuracy in the existing technology is solved, and the radar sea surface target detection performance is significantly improved.

CN117310641BActive Publication Date: 2024-11-26BEIHANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311237415.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2022-09-27
Filing Date
2023-09-25
Publication Date
2024-11-26
Estimated Expiration
2043-09-25

AI Technical Summary

Technical Problem

The existing sea clutter suppression method is difficult to effectively utilize the phase information of sea clutter, resulting in low prediction accuracy of sea clutter and affecting the radar sea surface target detection performance.

Method used

Combining the phase processing mechanism and the LSTM network model, by establishing a phase processing-based LSTM network model, phase feature information is extracted from the chaotic characteristics and nonlinear mapping characteristics of sea clutter, high-precision prediction of sea clutter, and prediction results are achieved. It is eliminated with the original echo to achieve the suppression of sea clutter.

Benefits of technology

It improves the prediction accuracy and suppression effect of sea clutter, enhances the performance of radar on sea surface target detection, and improves the signal-to-miss ratio.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117310641B_ABST
    Figure CN117310641B_ABST
Patent Text Reader

Abstract

The invention discloses a sea clutter suppression method combining a phase processing mechanism with an LSTM network model. An airborne multifunctional maritime surveillance radar system obtains measured radar echo data, performs normalization preprocessing on each distance unit after range pulse compression; selects a distance unit with a high sea clutter intensity and not interfered by a target signal to obtain a one-dimensional time series of sea clutter; estimates the delay time of the sea clutter time series by an autocorrelation function method, and estimates the embedding dimension of the sea clutter time series by a Cao algorithm; constructs a training data set by using the one-dimensional sea clutter time series according to the estimated delay time and embedding dimension, and constructs a verification data set by using the remaining distance unit time series; designs an LSTM network model based on the phase processing mechanism; inputs the training data set into the network, trains the network to obtain a sea clutter prediction model, inputs the verification data set, and obtains sea clutter prediction results in the remaining distance units; and cancels the predicted sea clutter with the original echo to achieve sea clutter suppression.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for suppressing sea clutter, and more particularly, to a method for suppressing sea clutter that combines a phase processing mechanism with an LSTM (Long Short-Term Memory) network model. The method of the present invention can improve the sea surface target detection performance of radar and enhance the target imaging quality. Background Art

[0002] Radar works in the marine environment, and the backscattered signals received from the sea surface are usually called sea clutter. Sea clutter often severely limits the radar's ability to detect sea surface targets. The types of sea surface targets are complex, such as ships, navigation buoys, aircraft missiles above the sea surface, etc. These sea surface targets have different radar scattering characteristics and kinematic characteristics. At the same time, in the presence of complex marine backgrounds such as islands, reefs, and floating ice, sea surface targets will show low observability to varying degrees, making it more difficult for radar to detect sea surface targets. In addition, in medium and high sea conditions with high wind speed and large waves, the scattering of sea clutter is enhanced, and the sea surface target signal is easily submerged in strong sea clutter. There are a large number of spike signals similar to sea surface targets in the sea surface echo, which makes the clutter modeling extremely complicated, seriously affecting the radar's detection and monitoring performance of sea surface targets. By suppressing radar sea clutter, the signal-to-clutter ratio can be improved, and the sea surface target detection performance can be improved. Therefore, radar sea clutter suppression has important research significance.

[0003] In June 2019, the 3rd issue of Volume 8 of the Journal of Radar published an article entitled "Design and Key Technology Research of Airborne Multi-Function Maritime Surveillance Radar System", written by Jiang Qian, Wu Hao, and Wang Yanning. The airborne multi-function maritime surveillance radar consists of two parts: airborne radar equipment and display control / information processing software (hereinafter referred to as Figure 1A ), the airborne radar is installed in the aircraft platform, and as the airborne mission equipment, it performs reconnaissance missions with the support of the data transmission subsystem; the display control / processing software is installed in the mission control station to monitor and control the status of the airborne radar, receive the reconnaissance data sent in real time through the data transmission analysis chart, form a preliminary situation display, and send the processed reconnaissance information to the back end for intelligence processing.

[0004] At present, the methods for suppressing sea clutter mainly include cyclic cancellation method, wavelet transform method, subspace decomposition method and neural network model method. The cyclic cancellation method is based on the measured sea clutter, and different sinusoidal signals are combined to approximate the measured sea clutter signal. The parameters of these sinusoidal signals are estimated according to the measured sea clutter. The signal composed of these sinusoidal signals is used as the sea clutter background signal, and the original echo is canceled to suppress the sea clutter. The wavelet transform method is to suppress the sea clutter by taking corresponding threshold processing according to the different characteristics of the sea clutter signal and the target signal in the wavelet domain. This method does not rely on the statistical characteristics of the clutter and has good adaptability to complex sea clutter. The subspace decomposition method separates the sea clutter in the radar echo according to the aggregation characteristics of the sea clutter in the subspace to suppress the sea clutter. The neural network model method is proposed based on the chaotic prediction characteristics of sea clutter. The inherent chaotic dynamic change law of sea clutter can be learned, so it can be used to predict and suppress sea clutter. In recent years, with the improvement of computer performance, deep learning technology has developed rapidly in all walks of life. The high fault tolerance of neural networks and the ability to approximate complex nonlinear relationships with arbitrary precision can be used to predict sea clutter under different sea conditions.

[0005] At present, the research on predicting sea clutter time signal based on neural network model mainly focuses on the prediction of sea clutter amplitude, but the predicted sea clutter amplitude is difficult to use for sea clutter suppression, because the sea clutter time signal is a complex signal, and the result after cancellation with the sea clutter amplitude prediction result is a real signal, which is difficult to perform subsequent imaging correction, etc. Some also separate the real and imaginary parts of the sea clutter time series, input them into the neural network model respectively to obtain the corresponding output prediction results, and recombine the prediction results of the real and imaginary parts to obtain the output complex prediction results, which ignores the phase information between the real and imaginary parts. Summary of the invention

[0006] The present invention proposes a sea clutter suppression method combining a phase processing mechanism with an LSTM network model, which utilizes the chaotic characteristics of sea clutter and the nonlinear mapping characteristics of the LSTM network model; by establishing an LSTM network model based on a phase processing mechanism, a certain chaotic generation law is grasped from the seemingly random sea clutter time series to achieve sea clutter prediction; then the phase characteristics of the sea clutter are used to improve the sea clutter prediction accuracy, thereby overcoming the deficiency that the phase characteristic information is ignored in the existing sea clutter prediction method; finally, the sea clutter prediction result is compared with the original echo (radar echo data SAR in ) cancellation to achieve sea clutter suppression. The present invention adopts the LSTM network model to effectively retain phase information, which helps to improve the sea clutter prediction results and further enhance the sea clutter suppression effect.

[0007] like Figure 1The sea clutter suppression method of the present invention combining the phase processing mechanism and the LSTM network model is embedded in Figure 1A The data terminal unit in the airborne multi-function maritime surveillance radar system is used to receive radar echo data transmitted by the airborne data terminal, denoted as SAR in In the present invention, a portion of the radar echo data SAR in Used to construct a training data set, denoted as MSAR 训练 , the other part of the radar echo data SAR in Used to construct a validation data set, denoted as MSAR 验证 .

[0008] like Figure 1 As shown, a sea clutter suppression method combining a phase processing mechanism with an LSTM network model of the present invention specifically includes two parts: constructing an LSTM sea clutter prediction model and completing real-time processing of sea clutter based on the LSTM sea clutter prediction model.

[0009] In the first part, the process of building the LSTM sea clutter prediction model includes the following steps:

[0010] Construction step 1: performing range pulse compression processing on the measured radar echo data;

[0011] In the data terminal unit, the measured radar echo data SAR is first stored in , and then the radar echo data SAR in The two-dimensional data formed after the range pulse compression is recorded as radar echo-range pulse compression data D, which is expressed in matrix form as follows:

[0012]

[0013] N a Indicates the number of pulse points in azimuth.

[0014] N f Indicates the number of sampling points in the distance direction.

[0015] a 1,1 Indicates the first distance sampling point at the first azimuth moment collected.

[0016] a 1,2 Indicates the second distance sampling point at the first azimuth time collected.

[0017] Indicates the Nth position at the first time point collected f distance sampling points.

[0018] a 2,1Indicates the first distance sampling point at the second azimuth moment collected.

[0019] a 2,2 Indicates the second distance sampling point collected at the second azimuth time.

[0020] Indicates the Nth position at the second time point collected f distance sampling points.

[0021] Indicates the Nth a The first distance sampling point at the azimuth time.

[0022] Indicates the Nth a The second distance sampling point at the azimuth time.

[0023] Indicates the Nth a The Nth position at the time f distance sampling points.

[0024] In the present invention, the distance-directed compressed echo is normalized by distance unit. k The processing method is:

[0025]

[0026] k represents the identification number of the distance sampling point, satisfying k=1,2,…,N f .

[0027] a k represents the one-dimensional sea clutter pulse sequence in the k-th range unit

[0028] a 1,k It represents the kth distance sampling point at the first azimuth moment collected.

[0029] a 2,k It represents the kth distance sampling point collected at the second azimuth time.

[0030] Indicates the Nth a The kth range sampling point at the azimuth time.

[0031] abs(a k ) indicates a k All sampling points in the modulus value.

[0032] max(abs(a k )) means ak The maximum value of all sampling points after taking the modulus value.

[0033] In the present invention, the radar echo-range pulse compression data D is normalized so that the radar echo data SAR in The amplitudes of all range samples in are concentrated between -1 and 1.

[0034] In the second construction step, a distance unit with a large sea clutter intensity and not interfered by the target signal is selected as the selected one-dimensional sea clutter pulse sequence;

[0035] In the present invention, a distance unit with a large sea clutter intensity and not interfered by the target signal is selected from the radar echo-range pulse compression data D. If the selected distance unit is the sth distance unit, it is recorded as the selected one-dimensional-sea clutter pulse sequence SS s ; The SS s The sequence form is expressed as:

[0036]

[0037] a 1,s Indicates the sth distance sampling point collected at the first azimuth time.

[0038] a 2,s Indicates the sth distance sampling point collected at the second azimuth time.

[0039] Indicates the Nth a The sth distance sampling point at the sth azimuth time.

[0040] In the third construction step, the delay time of the sea clutter time series is estimated using the autocorrelation function method;

[0041] In the present invention, the autocorrelation function R(τ) used to estimate the delay time of the sea clutter time series is defined as:

[0042]

[0043] τ represents the delay time.

[0044] n represents the identification number of the azimuth sampling point, satisfying n=1,2,...,N a .

[0045] a n,s Indicates the sth distance sampling point collected at the nth azimuth time.

[0046] a n+τ,s It represents the s-th distance sampling point collected at the n+τ-th azimuth time.

[0047] Indicates a n,s The complex conjugate of .

[0048] In the present invention, the delay time τ is the value of n corresponding to when the autocorrelation function R(τ) drops to 1-1e of the initial value.

[0049] In the present invention, formula (4) is used to calculate the selected one-dimensional sea clutter pulse sequence SS s Calculate the delay time and get the estimated delay time Time_SS s .

[0050] In step 4, the Cao algorithm is used to estimate the embedding dimension of the sea clutter time series;

[0051] In the present invention, the Cao algorithm for estimating the embedding dimension of the sea clutter time series is expressed as follows:

[0052] 1) Define the time delay vector:

[0053] A i,s,m ={a i,s ,a i+τ,s ,…,a i+(m-1)τ,s} (5)

[0054] τ represents the delay time.

[0055] m represents the embedding dimension that needs to be estimated.

[0056] A i,s,m represents the i-th time delay vector constructed according to the embedding dimension m.

[0057] a i,s It represents the sth distance sampling point collected at the ith azimuth time.

[0058] a i+τ,s It represents the sth distance sampling point collected at the i+τth azimuth time.

[0059] a i+(m-1)τ,s It represents the sth distance sampling point collected at the i+(m-1)τ,sth azimuth time.

[0060] 2) Determine the function x(i,m) of the embedding dimension m:

[0061]

[0062] A i,s,m represents the i-th time delay vector constructed according to the embedding dimension m.

[0063] A i,s,m+1represents the i-th time delay vector constructed according to the next embedding dimension m+1.

[0064] A o(i,m),s,m represents the o(i,m)th time delay vector constructed according to the embedding dimension m.

[0065] A o(i,m),s,m+1 represents the o(i,m)th time delay vector constructed according to the next embedding dimension m+1.

[0066] Where · is a Euclidian distance calculation formula, which is defined as:

[0067]

[0068] A p,s,m represents the pth time delay vector constructed according to the embedding dimension m. According to formula (5), A p,s,m ={a p,s ,a p+τ,s ,…,a p+(m-1)τ,s}, a p,s represents the sth distance sampling point at the pth azimuth time collected, a p+τ,s represents the sth distance sampling point at the p+τth azimuth time collected, a p+(m-1)τ,s It represents the s-th range sampling point collected at the p+(m-1)τ-th azimuth time.

[0069] A q,s,m represents the qth time delay vector constructed according to the embedding dimension m. According to formula (5), A q,s,m ={a q,s ,a q+τ,s ,…,a q+(m-1)τ,s}, a q,s represents the sth distance sampling point at the qth azimuth time collected, a q+τ,s represents the sth distance sampling point at the q+τth azimuth time collected, a q+(m-1)τ,s It represents the s-th range sampling point collected at the q+(m-1)τ-th azimuth time.

[0070] a p+wτ,s Denotes the time delay vector A p,s,m The w-th element of , where w is an integer and its value range satisfies 0≤w≤m-1.

[0071] a q+wτ,s A represents the time delay vector q,s,m The w-th element of , where w is an integer and its value range satisfies 0≤w≤m-1.

[0072] 3) Define the mean of x(i,m) as:

[0073]

[0074] The value of E(m) depends on the time delay τ and the embedding dimension m.

[0075] 4) Define the function used to study the change of E(m) from m to m+1:

[0076]

[0077] In the present invention, when m increases to a certain value m0, m continues to increase, and the function value maintains E σ (m0) hardly changes, and at this time m0+1 is the estimated optimal embedding dimension.

[0078] In step five, a training data set is constructed using the selected one-dimensional-sea clutter pulse sequence according to the estimated delay time and the optimal embedding dimension;

[0079] In the present invention, the training data set MSAR 训练 The construction method is as follows: (A) using the estimated delay time Time_SS s and the optimal embedding dimension m0+1 will be selected as the one-dimensional sea clutter pulse sequence SS s Split into short sequences of shorter length (B) Short sequence In the input LSTM network model, the length of the sea clutter time series of the input LSTM network satisfies L ≥ mτ. If the length of the sea clutter time series of a single distance unit is N a , at most N a -mτ short sequences of length mτ. During the construction process, each short sequence The next sampling point value is used as its corresponding label data, so the maximum number of training sets that can be constructed by a single distance unit is N. a -mτ.

[0080] In the present invention, the LSTM network model refers to the content of "Deep Learning" written by Ian Goodfellow (USA) and published by People's Posts and Telecommunications Press, pages 248-250. LSTM (convolutional long short-term memory network) adds long-term memory units on the basis of gated RNN, controls the flow of timing information through forgetting gates, input gates and output gates, updates the weights of the network through back propagation, and learns long-term dependencies. The learning step length of the RNN neural network using the tanh activation function and the Relu activation function is too short and unstable, while the LSTM neural network can stably learn multi-step timing relationships.

[0081] In construction step six, a pair of real LSTM networks are used to implement the phase processing mechanism and obtain the LSTM sea clutter prediction model;

[0082] In the present invention, a pair of real LSTM networks is used to implement the phase processing mechanism, and a sea clutter prediction model of the LSTM network combined with the phase processing mechanism is obtained:

[0083]

[0084] f represents the LSTM network model, f1 and f2 represent a pair of real LSTM networks, Z represents a complex vector, A and B are the real and imaginary parts of the complex vector Z, which are input into the two real network models respectively, and then the output result of the LSTM network model combined with the phase processing mechanism is calculated in a way similar to the complex multiplication operation rule:

[0085]

[0086] The second part, based on the LSTM sea clutter prediction model, completes the real-time processing of sea clutter, including the following steps:

[0087] Step A, performing range pulse compression processing on the measured radar echo data;

[0088] In the present invention, the airborne multifunctional maritime surveillance radar system acquires the measured radar echo data SAR in , the radar echo data SAR in The two-dimensional data formed after the range pulse compression is recorded as radar echo-range pulse compression data

[0089] In the present invention, the radar echo-range pulse compression data D is normalized so that the radar echo data SAR in The amplitudes of all range samples in are concentrated between -1 and 1.

[0090] Step B, normalizing the distance units one by one;

[0091] In the present invention, the radar echo-range pulse compression data D is processed according to the construction steps 3 and 4 to obtain the MSAR data set for constructing the verification data set. 验证 .

[0092] In the present invention, the training data set MSAR 训练 With the validation dataset MSAR 验证 The difference is that: the validation dataset MSAR 验证 The radar echo data used does not contain the selected one-dimensional sea clutter pulse sequence SS in step 2. s .

[0093] Step C, running the LSTM sea clutter prediction model;

[0094] In this invention, the validation dataset MSAR 验证 Input into the LSTM sea clutter prediction model, After processing, the predicted value of sea clutter at the next moment is obtained Should The sequence representation is:

[0095]

[0096] F is the LSTM sea clutter prediction model, which is a nonlinear mapping function.

[0097] v represents the distance sampling point identification number, and its value range is 1, 2, ..., N f (v≠s).

[0098] {a j,v ,a j+1,v ,…,a j+(mτ-1),v} represents the amount of data in the validation dataset that is input into the model each time.

[0099] a j,v It represents the vth distance sampling point at the jth azimuth time collected.

[0100] a j+1,v It represents the vth distance sampling point at the j+1th azimuth time collected.

[0101] a j+(mτ-1),v It represents the vth range sampling point collected at the j+(mτ-1)th azimuth time.

[0102] In the present invention, the value This is the output of the LSTM sea clutter prediction model.

[0103] Step D, canceling the sea clutter prediction value with the original range compression echo to suppress the sea clutter;

[0104] In the present invention, the sea clutter prediction value Compress echo with original distance Cancellation is performed. Since the prediction result of sea clutter is also a complex value, the complex difference operation is performed during the cancellation process. If there is a target signal in the original range compression echo, then after the sea clutter signal is cancelled and suppressed, the signal retained in the result is the target signal.

[0105] The LSTM sea clutter prediction model constructed by the present invention can be embedded in a data terminal unit of an airborne multifunctional maritime surveillance radar system alone, or can be embedded in a processing device that needs to perform radar echo data simulation, such as a computer, a digital processing device, etc.

[0106] Compared with the prior art, the advantages of the sea clutter suppression method combining the phase processing mechanism with the LSTM network model in the present invention are:

[0107] (1) The phase characteristic information of the sea clutter pulse sequence is used to provide the LSTM network model with more effective information to describe the sea clutter, which is beneficial to the training of the LSTM sea clutter prediction model and the application of the LSTM sea clutter prediction model to improve the sea clutter prediction accuracy.

[0108] (2) Most of the existing sea clutter prediction methods either directly predict the sea clutter amplitude and discard the phase information, or first separate the sea clutter pulse sequence into real and imaginary parts and then input them into the network model for training to obtain the sea clutter prediction model for the real part and the imaginary part. The output prediction results need to be recombined into complex data. Compared with the LSTM sea clutter prediction model proposed in the present invention, the input and output of this model are both complex data, the prediction process of sea clutter is more convenient, and the phase feature information and amplitude feature information of sea clutter are utilized.

[0109] (3) The sea clutter prediction network model obtained by complex data training has higher sea clutter prediction accuracy than the real and virtual separation method, which can improve the sea clutter suppression effect in the LSTM sea clutter prediction model.

[0110] (4) Since the phase characteristic information of the sea clutter time pulse sequence is significantly different from the phase characteristic of the target signal, when the target signal and the sea clutter signal exist in the echo data in the same range, the prediction error of the sea clutter prediction model will increase. By designing a reasonable error threshold, the target can be detected from the increased prediction error. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] Figure 1 It is a principle block diagram of the sea clutter suppression method combining the phase processing mechanism with the LSTM network model.

[0112] Figure 1A This is a schematic diagram of the composition of the airborne multi-function maritime surveillance radar system.

[0113] Figure 2 is the autocorrelation function method to estimate the time delay (echo 1).

[0114] Figure 3 The autocorrelation function method estimates the time delay (echo 2).

[0115] Figure 4The time delay is estimated by the autocorrelation function method (echo 3).

[0116] Figure 5 is the Cao algorithm estimating the embedding dimension (echo 1).

[0117] Figure 6 It is the Cao algorithm that estimates the embedding dimension (echo 2).

[0118] Figure 7 It is Cao's algorithm that estimates the embedding dimension (echo 3).

[0119] Figure 8 It is an LSTM network model based on the phase processing mechanism.

[0120] Fig. 9 is the result of sea clutter suppression of echo 1. Sub-figures (a), (b), (c), and (d) represent range compression echo 1, imaging result of range compression echo 1, sea clutter suppression result, and imaging result after sea clutter suppression, respectively. Fig. 9A This is a cross-sectional comparison diagram before and after sea clutter suppression (cross-sectional diagram at the 4000th pulse of each distance unit).

[0121] Fig.10 is the result of sea clutter suppression of echo 2. Sub-figures (a), (b), (c), and (d) represent range compression echo 2, imaging result of range compression echo 2, sea clutter suppression result, and imaging result after sea clutter suppression, respectively. Fig. 10A This is a cross-sectional comparison diagram before and after sea clutter suppression (cross-sectional diagram at the 43rd pulse of each distance unit).

[0122] Fig.11 is the result of sea clutter suppression of echo 3. Sub-figures (a), (b), (c), and (d) represent range compression echo 3, imaging result of range compression echo 3, sea clutter suppression result, and imaging result after sea clutter suppression, respectively. Fig.11A This is a cross-sectional comparison diagram before and after sea clutter suppression (cross-sectional diagram at the 69869th pulse of each distance unit). DETAILED DESCRIPTION

[0123] This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments. The specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings:

[0124] This embodiment uses three groups of distance compressed echo data. The file numbers of echo 1 and echo 2 are 20210106155330_01_staring and 20210106172511_02_staring, and the file number of echo 3 is 1931107_135603_starea. These three groups of data are actual sea measurement data. Echo 1 does not contain a target signal, echo 2 has a target signal in the 940-953, 1482-1492, and 2140-2810 distance units, and echo 3 has a target signal in the 9th distance unit, and the adjacent 8-11 distance units are also affected by the target signal.

[0125] The relevant parameters of these three sets of radar echo data are shown in the following table.

[0126] Echo 1 Echo 2 Echo 3 Antenna polarization HH HH VV Azimuth angle (°) 42.18 8.01 170.26 Sea conditions 3-4 3-4 4 Radar frequency (GHz) 9.3-9.5 9.3-9.5 9.39 Pulse repetition frequency (Hz) 3000 3000 1000 Azimuth pulse number 6940 4390 131072 Distance Units 4346 4346 14

[0127] The specific operation process of applying the sea clutter suppression method combining the phase processing mechanism and the LSTM network model proposed by the present invention is as follows:

[0128] 1. Perform normalization preprocessing on the range compressed echo signal collected in the data terminal unit for each range unit.

[0129] 2. Select a range echo data sequence with stronger sea clutter. Here, the sea clutter time pulse sequence of the 1st range unit is selected for echo 1, the sea clutter time pulse sequence of the 933rd range unit is selected for echo 2, and the sea clutter time pulse sequence of the 1st range unit is selected for echo 3.

[0130] 3. Calculate the corresponding autocorrelation function for the selected sea clutter time pulse signal. When the autocorrelation function value drops to 1-1e of the initial value, the corresponding value of n is the delay time. For the sea clutter time pulse signal selected for echo 1, the value of n is 5; so the time delay τ is 5; for the sea clutter time pulse signal selected for echo 2, the value of n is 3, and for the sea clutter time pulse signal selected for echo 3, the value of n is 5. The corresponding normalized autocorrelation function diagram is as follows: Figure 2 , Figure 3 , Figure 4 shown.

[0131] 4. The Cao algorithm is used to estimate the embedding dimension m of the selected sea clutter time series. For echo 1, the sea clutter time pulse signal tends to be stable after m is greater than 9, so the minimum embedding dimension value is 10; for echo 2, the sea clutter time pulse signal tends to be stable after m is greater than 13, so the minimum embedding dimension value is 14; for echo 2, the sea clutter time pulse signal tends to be stable after m is greater than 11, so the minimum embedding dimension value is 12. The corresponding curves are as follows: Figure 5 , Figure 6 , Figure 7 shown.

[0132] 5. The estimated time delay and embedding dimension are used to construct the training data set. For echo 1, since the length of the selected sea clutter time series is 6940, 6890 short sequences of length 50 are used to construct the training data set. Similarly, the remaining distance units that are not selected are also used to construct the corresponding verification data set. For echo 2, since the length of the selected sea clutter time series is 4390, 4348 short sequences of length 42 are used to construct the training data set. The remaining distance units that are not selected are also used to construct the corresponding verification data set. For echo 3, since the length of the selected sea clutter time series is 131072, 131012 short sequences of length 60 are used to construct the training data set. The remaining distance units that are not selected are also used to construct the corresponding verification data set.

[0133] 6. Use a pair of real LSTM networks to implement the phase processing mechanism and obtain an LSTM network model combined with the phase processing mechanism. The implementation of this network model is as follows: Figure 8 shown.

[0134] 7. Use the constructed training data set to train the LSTM network model. The learning rate of the network is set to 0.0002. The size of a batch of data input to the network is 64. The loss function is the mean square error loss function, and the number of training rounds is set to 25. The input and output of the LSTM prediction network model for a single training, the input of the network is the constructed short sequence, the output of the network is the predicted value of the short sequence at the next moment, the mean square error is calculated between the network output prediction value and the label data corresponding to the input, and the network is optimized in the direction of reducing the mean square error until the network training is completed. At this time, the network has learned the changing law of sea clutter and has the ability to predict sea clutter.

[0135] 8. Use the validation data set to validate the trained LSTM sea clutter prediction network model, input the validation data, and output the predicted sea clutter. The sea clutter prediction result is cancelled with the original echo. Fig. 9 (a) is the range compression echo 1, Fig. 9 (b) is the imaging result of range compression echo 1. Fig. 9 (c) shows the result of sea clutter suppression. Fig. 9 (d) The imaging result after sea clutter suppression. Fig. 9A This is a cross-sectional comparison diagram before and after sea clutter suppression; Fig.10 (a) is the range compression echo 2, Fig.10 (b) is the imaging result of range compression echo 2. Fig.10 (c) shows the result of sea clutter suppression. Fig.10 (d) The imaging result after sea clutter suppression. Fig. 10AThis is a cross-sectional comparison diagram before and after sea clutter suppression; Fig.11 (a) is the range compression echo 3, Fig.11 (b) is the imaging result of range compression echo 3. Fig.11 (c) shows the result of sea clutter suppression. Fig.11 (d) The imaging result after sea clutter suppression. Fig.11A The cross-sectional comparison diagram before and after sea clutter suppression. In addition, through calculation, after sea clutter suppression, the sea clutter energy of echo 1 is reduced by 12.53dB, the signal-to-clutter ratio of echo 2 is increased by 5.34dB, and the signal-to-clutter ratio of echo 3 is increased by 12.35dB. The experimental results show that the sea clutter suppression method combining the phase processing mechanism with the LSTM network model has certain practicality.

Claims

1. A sea clutter suppression method combining phase processing mechanism and LSTM network model to obtain measured radar echo data through an airborne multi-function maritime surveillance radar system ; Before performing sea clutter suppression, it is necessary to construct an LSTM sea clutter prediction model in the data terminal unit; It is characterized by: Building an LSTM sea clutter prediction model includes the following steps: Construction step 1: performing range pulse compression processing on the measured radar echo data; In the data terminal unit, the measured radar echo data is first stored , then After performing range pulse compression, radar echo is formed - range pulse compression data , expressed in matrix form as: ; Indicates the number of pulse points in azimuth; Indicates the number of sampling points in the distance direction; Indicates the first distance sampling point at the first azimuth moment collected; Indicates the second distance sampling point at the first azimuth time collected; Indicates the first position at the time of the first acquisition. distance sampling points; Indicates the first distance sampling point at the second azimuth time collected; Indicates the second distance sampling point at the second azimuth time collected; Indicates the second position at the time of the collection. distance sampling points; Indicates the collected The first distance sampling point at the azimuth time; Indicates the collected The second distance sampling point at the azimuth time; Indicates the collected The first position at the time distance sampling points; The range-to-range compression echo is normalized per range unit. The processing method is: ; Indicates the identification number of the distance sampling point, satisfying ; Indicates One-dimensional sea clutter pulse sequence within a range unit ; Indicates the first position at the time of the first acquisition. distance sampling points; Indicates the second position at the time of the collection. distance sampling points; Indicates the collected The first position at the time distance sampling points; Express Take the modulus value of all sampling points in; express The maximum value of all sampling points after taking the modulus value; In the second construction step, a distance unit with a large sea clutter intensity and not interfered by the target signal is selected as the selected one-dimensional sea clutter pulse sequence; From the stated Select a distance unit with greater sea clutter intensity and not interfered by target signals. distance units, denoted as the selected one-dimensional sea clutter pulse sequence ; The sequence form is expressed as: ; Indicates the first position at the time of the first acquisition. distance sampling points; Indicates the second position at the time of the collection. distance sampling points; Indicates the collected The first position at the time distance sampling points; In the third construction step, the delay time of the sea clutter time series is estimated using the autocorrelation function method; Autocorrelation function for estimating the delay time of sea clutter time series Defined as: ; Indicates the delay time; Indicates the identification number of the azimuth sampling point, satisfying ; Indicates the collected The first position at the time distance sampling points; Indicates the collected The first position at the time distance sampling points; express The complex conjugate of In step 4, the Cao algorithm is used to estimate the embedding dimension of the sea clutter time series; The Cao algorithm for estimating the embedding dimension of sea clutter time series is expressed as follows: 1) Define the time delay vector: ; Indicates the embedding dimension that needs to be estimated; Indicates Embedding Dimension The constructed time delay vector; Indicates the collected The first position at the time distance sampling points; Indicates the collected The first position at the time distance sampling points; Indicates the collected The first position at the time distance sampling points; 2) Determine the embedding dimension Function : ; Indicates According to the next embedding dimension The constructed time delay vector; Indicates Embedding Dimension The constructed time delay vector; Indicates According to the next embedding dimension The constructed time delay vector; In the formula is a Euclidian distance calculation formula, which is defined as: ; Indicates Embedding Dimension The time delay vector constructed; according to formula (5), we have , Indicates the collected The first position at the time distance sampling points, Indicates the collected The first position at the time distance sampling points, Indicates the collected The first position at the time distance sampling points; Indicates Embedding Dimension The time delay vector constructed; according to formula (5), we have , Indicates the collected The first position at the time distance sampling points, Indicates the collected The first position at the time distance sampling points, Indicates the collected The first position at the time distance sampling points; represents the time delay vector No. elements, is an integer and its value range satisfies ; represents the time delay vector No. elements, is an integer and its value range satisfies ; 3) Definition The mean of is: ; The value depends on the time delay and embedding dimension ; 4) Definition for research from Change to Function: ; In step five, a training data set is constructed using the selected one-dimensional-sea clutter pulse sequence according to the estimated delay time and the optimal embedding dimension; Training Dataset The construction method is as follows: (A) using the estimated delay time and the optimal embedding dimension The one-dimensional sea clutter pulse sequence will be selected Split into short sequences of shorter length ; (B) short sequence In the input LSTM network model, the length of the sea clutter time series of the input LSTM network satisfies , if the length of the sea clutter time series of a single range unit is , at most The length is During the construction process, each short sequence The next sampling point value is used as its corresponding label data, so the maximum number of training sets that can be constructed by a single distance unit is ; In construction step six, a pair of real LSTM networks are used to implement the phase processing mechanism and obtain the LSTM sea clutter prediction model; A pair of real LSTM networks are used to implement the phase processing mechanism, and the sea clutter prediction model of the LSTM network combined with the phase processing mechanism is obtained: ; Represents the LSTM network model, and represents an LSTM network of a pair of real numbers, represents a complex vector, and is a complex vector The real and imaginary parts are input into the two real network models respectively, and then the output result of the LSTM network model combined with the phase processing mechanism is calculated by the complex multiplication operation rule: 。 2. The sea clutter suppression method combining the phase processing mechanism and the LSTM network model according to claim 1 is characterized in that: The process of completing real-time sea clutter processing based on the LSTM sea clutter prediction model includes the following steps: Step A, performing range pulse compression processing on the measured radar echo data; Radar echo data Perform range pulse compression to obtain radar echo-range pulse compression data ; Step B, normalizing the distance units one by one; According to construction steps 3 and 4, the radar echo-range pulse compression data D is processed to obtain the data set used to construct the verification data. ; Training Dataset With the validation dataset The difference is that: the validation dataset The radar echo data used does not contain the selected one-dimensional sea clutter pulse sequence in step 2. ; Step C, running the LSTM sea clutter prediction model; The validation dataset Input into the LSTM sea clutter prediction model, After processing, the predicted value of sea clutter at the next moment is obtained ; Said The sequence representation is: ; It is the LSTM sea clutter prediction model; Indicates the distance sampling point identification number, its value range is ; Indicates the amount of data in the validation dataset that is input into the model each time; Indicates the collected The first position at the time distance sampling points; Indicates the collected The first position at the time distance sampling points; Indicates the collected The first position at the time distance sampling points; Step D, canceling the sea clutter prediction value with the original range compression echo to suppress the sea clutter; Sea clutter prediction value and Cancellation is performed. Since the prediction result of sea clutter is also a complex value, the complex difference operation is performed during the cancellation process. If there is a target signal in the original range compression echo, then after the sea clutter signal is cancelled and suppressed, the signal retained in the result is the target signal.

3. The sea clutter suppression method combining the phase processing mechanism and the LSTM network model according to claim 1 or 2, characterized in that: After normalizing the radar echo-range pulse compression data D, the radar echo data The amplitudes of all range samples in are concentrated between -1 and 1.