Unsupervised nearest neighbor propagation clustering method for airborne radar clutter suppression target detection

Abstract

The application discloses an unsupervised near neighbor propagation clustering airborne radar clutter suppression target detection method, and mainly solves the problem of low target detection probability of the prior art in a complex environment.The implementation scheme is as follows: a radar transmits a signal and obtains received data after matched filtering and pulse stacking; according to the power characteristics of the clutter noise and the signal, data points and weights are designed, a matrix transformation is carried out to obtain a first weighted input data matrix, and then unsupervised weighted near neighbor propagation clustering is carried out on the first weighted input data matrix to obtain a first clustering result; a second weighted input data matrix is re-designed according to the first weighted input data matrix, and then unsupervised weighted near neighbor propagation clustering is carried out on the second weighted input data matrix to obtain a second clustering result; a radar target detection discrimination criterion is designed by using the output results of the two times of clustering, and a target detection result is obtained according to the criterion.The application fully utilizes data characteristics, greatly improves target detection performance, and can be used for airborne radar signal processing and early warning in a clutter background.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, and specifically relates to an airborne radar target detection method that can be used for airborne radar signal processing and early warning in clutter backgrounds. Background Technology

[0002] Airborne radar has attracted much attention due to its high flexibility and long field of view, but its core problem is that strong clutter and its broadened clutter spectrum can overwhelm targets. To address this, the space-time adaptive processing (STAP) technique estimates the accurate clutter and noise covariance matrix, but this also requires sufficient independent and identically distributed (IID) samples. However, due to radar structure and signal environment, actual clutter characteristics are usually non-uniform, significantly reducing the performance of traditional clutter suppression and target detection methods. Especially when the airborne radar antenna array structure is not front-side looking, the range dependence of clutter directly leads to samples not satisfying the IID standard, thus requiring range compensation for the clutter spectrum.

[0003] The goal of range compensation methods is to make the clutter spectra of each range cell consistent after compensation. Existing range compensation methods mainly include the Doppler shift compensation (DW) method and the angle Doppler compensation (ADC) method.

[0004] The DW method is a Doppler frequency shift compensation method proposed for non-frontal side-looking arrays of airborne radars using radar navigation parameters. See IEE proceedings. Radar, sonar and navigation, 2001, 148(5), 390: 253-258. It converts the two-dimensional clutter spectrum into a one-dimensional clutter spectrum, thereby making it consistent with the clutter spectrum of the unit to be detected in the Doppler direction.

[0005] The ADC method uses inertial navigation parameters and linear transformations to improve the broadened clutter spectrum, see Proceedings of IEEE Radar Conference, Long Beach, CA, 2002, 311-317.

[0006] The two distance compensation methods mentioned above mostly rely on multiple parameters known and estimated by the inertial navigation system, and therefore cannot adaptively suppress clutter.

[0007] Besides range compensation methods, in order to overcome the non-independent and identically distributed clutter characteristics of airborne radar non-frontal side-looking arrays, dimensionality reduction and rank reduction methods are often introduced to reduce the system's degrees of freedom (DOF) and select a subset of samples. The main methods include the classic Local Joint Processing (JDL) algorithm and the Auxiliary Channel Processing (ACP) algorithm. The classic JDL algorithm introduces a two-dimensional Fourier transform and selects a rectangular local region for joint processing. This method simultaneously handles beamforming and Doppler frequency (see Electronics Letters, 1996, 32(3): 258). The ACP algorithm reduces the system's DOF ​​from MK to M+K-1 by constructing an auxiliary beam (see Journal of Circuits and Systems, 2004, 9(6): 100-104). However, both of these methods, by utilizing the low-rank characteristics of clutter, reduce the training sample data and system DOF, directly affecting the performance of clutter suppression and target detection.

[0008] For target detection in airborne radar, the classic method is the constant false alarm rate (CFAR) detection method (see IEE Proceedings-Radar, Sonar and Navigation, 2004, 15(3): 135-142). Extended methods include the cell average constant false alarm rate (CA-CFAR), the small constant false alarm rate (SO-CFAR), the large constant false alarm rate (GO-CFAR), and the ordered statistical constant false alarm rate (OS-CFAR). CFAR has been widely used in engineering fields. However, these mean-based CFAR methods suffer from increased false alarm rates, missed detections, or target obstruction in clutter-edge and multi-target environments. Furthermore, the data features extracted and utilized by CFAR-related methods are extremely limited and often quite simplistic.

[0009] Currently, machine learning-based clutter suppression and target detection methods have garnered significant attention in the radar field due to their substantial advantages in higher detection probabilities and better target detection performance. Machine learning methods can directly mine, extract, and learn multidimensional features and even hidden features of data, and are capable of handling high-dimensional and large-scale data.

[0010] To address the target recognition problem in the radar field, researchers have proposed target recognition methods based on coupled neural networks and classification and recognition methods for SAR images based on Faster R-CNN. However, these methods have two shortcomings: first, they are all designed for processing and optimizing synthetic aperture radar (SAR) images, rather than directly processing the radar echo data domain; second, they usually require a large amount of labeled data or label assignment, and cannot be directly applied to clutter suppression and target detection of airborne non-frontal side-looking array radars with non-independent and identically distributed samples.

[0011] To address the clutter suppression problem, researchers have proposed the STAP method based on autoencoder neural networks and the DU-CG-STAP method based on sparse recovery. However, since these methods mainly suppress clutter by improving clutter feature estimation, they cannot directly perform target detection after suppressing clutter. Summary of the Invention

[0012] The purpose of this invention is to address the shortcomings of the prior art by proposing an unsupervised nearest neighbor propagation clustering method for airborne radar clutter suppression and target detection. This method enables direct clustering detection after suppressing clutter in radar echo data without requiring any labeled data or assigned tags, thereby achieving higher target detection probability and better target detection performance.

[0013] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0014] 1) The radar transmits signals and receives the received data x after matched filtering and pulse stacking. cn,l Based on the power characteristics of clutter, noise, and signal, data points are designed. and weight E cns Then, matrix transformations are performed to construct the first weighted input data matrix F. w ;

[0015] 2) Design the first weighted input data matrix F w Reference degree P i And similarity s(i,j), calculate attraction. And the membership degree β(i,j), and perform unsupervised weighted nearest neighbor clustering on its iterative update, when a point j makes When the maximum value is obtained, the cluster center of point i is obtained, and all points i = 1, 2, ..., M are extracted to form different parts with a total of less than M cluster centers, which are output as the first clustering result {I1, I2, ...}, where M is F. w number of rows;

[0016] 3) Based on the output of the first clustering, calculate the number of first cluster centers with a number less than M: Q1 = num{I1,I2,...};

[0017] 4) Based on the power characteristics of clutter plus noise and the structure of the designed weighted input data matrix, redesign the second weighted input data matrix Y. w ;

[0018] 5) Calculate the second weighted input data matrix Y w similarity s γ (i',j'), attractiveness Belonging degree β γ (i',j'), and the and βγ (i',j') Iteratively update the design of the second unsupervised weighted nearest neighbor propagation clustering process, when a point j' makes the second unsupervised weighted nearest neighbor propagation clustering process... When the maximum value is obtained, the cluster center of point i' is obtained, and all points i' = 1, 2, ..., M' are extracted to form different parts with a total of less than M' cluster centers, which are output as the second clustering result. M' is Y w number of rows;

[0019] 6) Calculate the number of second cluster centers whose number is less than M' based on the output of the second clustering:

[0020] 7) Set five discrimination interval values ​​u d1 u d2 u d3 u d4 u d5 And the OS-CFAR detection threshold η0, using the output results {I1,I2,...} of two unsupervised weighted nearest neighbor clustering operations, and Given the cluster center numbers Q1 and Q2, design a radar target detection and discrimination criterion that includes three conditions for the existence of the target:

[0021] Condition 1,

[0022] Condition 2,

[0023] Condition 3, Q1 = 1, Q2 = 2, |x l |>η0, where x l This indicates that the radar has received data;

[0024] 8) Based on the radar target detection criteria designed in step 7), determine the clutter suppression target detection result of the airborne radar:

[0025] The target exists if any of the criteria is met.

[0026] Otherwise, the target does not exist.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] 1. This invention utilizes intelligent clustering of data points for airborne radar clutter suppression and target detection. Weight E cns And a series of input matrix transformations, and twice construct weighted input data F based on matrix transformations. w and Y wBy clustering the two input matrices, clutter can be suppressed, thus making full use of the advantages of machine learning. Compared with existing algorithms such as STAP, ADC, JDL, SO-CFAR, GO-CFAR and OS-CFAR, it has a significantly higher target detection probability.

[0029] 2. This invention employs unsupervised nearest neighbor propagation machine learning clustering to perform two weighted input data F w and Y w Clustering is performed without assigning labels to the input data, and the first and second clustering results are output separately, which fully taps the clustering potential of machine learning and improves the target detection performance compared with existing algorithms such as STAP, ADC, JDL, SO-CFAR, GO-CFAR and OS-CFAR.

[0030] 3. This invention is based on the clustering output results {I1,I2,...} from two clustering operations and... Based on the cluster center numbers Q1 and Q2, a target detection discrimination criterion was designed, which can overcome the range dependence in clutter features, better handle non-independent identically distributed samples of non-frontal side-looking array radar, and obtain accurate target detection results. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0032] Figure 2 This is a comparison graph of the detection probability as a function of signal-to-noise ratio for the present invention and existing detection methods.

[0033] Figure 3 This is a comparison graph showing the change in detection probability as a function of sample number between the present invention and existing detection methods.

[0034] Figure 4 This is a comparison graph showing the change in detection performance of the present invention and existing detection methods with non-frontal side viewing angle;

[0035] Figure 5 This is a curve comparing the contribution of the invention to the detection performance under different testing conditions;

[0036] Figure 6 This is a comparison graph showing the change in computation time of the present invention and existing detection methods as a function of sample number;

[0037] Figure 7 This is a visualization of two clustering operations when the number of cluster centers Q2 = 2, as presented in this invention.

[0038] Figure 8 This is a visual comparison chart of two clustering operations when the number of cluster centers Q2 = 1. Detailed Implementation

[0039] The embodiments and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0040] Reference Figure 1 The implementation steps for this example are as follows:

[0041] Step 1, acquire radar received data x cn,l .

[0042] 1.1) Based on the fact that the element spacing of the uniform linear array in an airborne radar system is half a wavelength d = λ / 2, and given that the number of transmitted pulses is K and the number of receiving elements is N within one coherent processing time, the time steering vector is obtained. and spatial guidance vector

[0043]

[0044]

[0045] in, and These represent the normalized Doppler frequency and the spatial frequency, respectively.

[0046] 1.2) By time-guided vector and spatial guidance vector Obtain the space-time steering vector of the i-th clutter block.

[0047]

[0048] in, It is the Kronecker product;

[0049] 1.3) Based on the spacetime steering vector The radar received data x after pulse stacking is obtained cn,l :

[0050]

[0051] Where l represents the l-th distance ring unit of L training distance unit samples, n l N represents zero-mean noise. c This corresponds to the total number of clutter blocks in the corresponding distance ring, ξ i,l These correspond to complex magnitudes, i = 1...N c .

[0052] Step 2: Design data points based on the power characteristics of clutter noise signals. and weight E cns Then, matrix transformations are performed to construct the first weighted input data matrix F. w .

[0053] 2.1) Design data points

[0054] 2.1.1) From all samples x that may contain the target to be detected cn,l Calculating the covariance matrix

[0055]

[0056] Where L is the total number of training distance unit samples, (·) H Indicates conjugate transpose;

[0057] 2.1.2) Based on the covariance matrix and spacetime steering vector The design includes data points containing clutter, noise, and signal:

[0058]

[0059] in, Represents a data point matrix The (l,m)th element, This represents the normalized Doppler frequency corresponding to the m-th power point in the time domain of interest. The normalized spatial frequency of the l-th power point in the spatial distribution is represented by M, where M is the total number of power points.

[0060] 2.2) Design the weights E of the weighted data cns :

[0061] 2.2.1) Solving the minimum variance problem using the Lagrange multiplier method Obtain the weight vector w of the space-time adaptive processing method STAP opt :

[0062]

[0063] Among them, a st (f s0 ,f d0 ) represents the NK×1 dimensional target spacetime guidance vector, f d0 and f s0 Distinguish between time frequency and spatial frequency; (·) -1 The inverse operation (·) represents the matrix operation. H Represents the conjugate transpose of a matrix;

[0064] For x cn,l The clutter plus noise covariance estimation matrix is ​​obtained using the maximum likelihood estimation method with L-1 training samples, and does not include echo vectors that may contain targets. In the formula ls The unit where the target is located;

[0065] 2.2.2) Based on the weight vector w opt Calculate the weights of the weighted data, which will serve as inputs to the clutter-plus-noise and target clustering processes:

[0066] Calculate the weight E of the l-th weighted data. cns (l):

[0067]

[0068] Where l = 1, 2, ..., L, max{·} and min{·} represent taking the maximum and minimum values, respectively, ω max and ω min These are the normalized expected maximum and minimum values, X. cns Let L be the received echo data, represented as [x cn,1 ,x cn,2 ,...,x cn,l ,...,x cn,L ];

[0069] The weights E of the L weighted data points cns (l) Combine them to obtain the weighted data weight matrix E cns :

[0070]

[0071] 2.3) Design the first cluster weighted input data matrix F w :

[0072] 2.3.1) For data points and weight matrix E cns Perform matrix transformations to obtain the data matrix. The (m,l)th element

[0073]

[0074] in [·] η It is the power of η of [·];

[0075] 2.3.2) Calculate M×L F values ​​sequentially. w (m,l) form a data matrix Then perform matrix transformation on it to obtain the first weighted input data matrix F. w :

[0076]

[0077] Here, |·| represents taking the absolute value of each element of the matrix.

[0078] Step 3, using the constructed first weighted input data matrix F w Similarity s(i,j), attractiveness Belonging degree β(i,j).

[0079] 3.1) Calculate the similarity s(i,j):

[0080] Based on the first weighted input data matrix F w Calculate the similarity s(i,j) between points i and j:

[0081]

[0082] Among them, F w (i,:) represents F w The i-th row, F w (j,:) represents F w The j-th row, {i,j}=1,2,...,M, Representing F w (i,:) and F w The squared distance between (j,:) is given by l = 1, 2, ..., L, where L is the total number of training distance unit samples.

[0083] When i = j, s(i,j) is defined as

[0084] s(i,j)=P i i = j

[0085] Among them, P i =p re • mean{S} is the reference degree, and its value and the maximum number of iterations E can be preset, p re Here, is the reference degree adjustment factor parameter, mean{·} represents the mean of all elements in {·}, and S is a matrix composed of similarity s(i,j) after removing cases i=j, expressed as:

[0086]

[0087] 3.2) Calculate the attraction level And the degree of belonging β(i,j):

[0088] Based on s(i,j), this example further calculates the data F to design the clustering process of clutter, noise, and signal for airborne radar. w attraction And the degree of belonging β(i,j), which represent different categories of data information; for data point i, β(i,j) represents the degree to which data point j is suitable as its cluster center, and β(i,j) represents the degree to which data point i chooses j as its cluster center. They are expressed as follows:

[0089]

[0090]

[0091] Where M is the weighted input data matrix F w The number of rows, the initial value of the membership degree β(i,h) is zero, when i=j, then

[0092]

[0093]

[0094] Step 4, regarding attraction The membership degree β(i,j) is iteratively updated to perform unsupervised weighted nearest neighbor propagation clustering, and the first clustering result {I1,I2,...} is output, thus obtaining the number of first cluster centers Q1.

[0095] 4.1) Let E be the maximum number of iterations, and let E be the number of iterations for the attraction. The membership degree β(i,j) is iteratively updated:

[0096] When i≠j, And β(i,j) are updated as follows:

[0097]

[0098]

[0099] Where λ is the damping coefficient, λ∈[0.5,1). The introduction of the damping coefficient can avoid oscillations and ensure the stability of the iterative process. Let β be the attraction in the (t-1)th iteration. (t) (i,h) represents the degree of belonging in the t-th iteration; For the t-th attraction iteration result, β (t-1) (i,j) represents the result of the (t-1)th iteration of the membership degree;

[0100] When i = j, and β (t) (j,j) are updated as follows:

[0101]

[0102]

[0103] Determine if the current iteration number t satisfies the set maximum iteration number E. If t < E, continue iterating; otherwise, stop iterating.

[0104] 4.2) Output the first clustering result {I1,I2,...}

[0105] According to the updated and β (t) (i,j), the cluster center for point i is taken from such that At a point j when F is at its maximum, for F w Given rows M, extract the distinct parts of all points i = 1, 2, ..., M that form cluster centers less than M, and output the first clustering result {I1, I2, ...} as the result:

[0106]

[0107] 4.3) Based on the output of the first clustering, calculate the number of cluster centers Q1:

[0108] Q1 = num{I1,I2,...}.

[0109] Step 5, construct the second weighted input data matrix Y w .

[0110] Based on the power characteristics of clutter plus noise and the designed weighted input data matrix F w The structure is reconstructed to reconstruct the second weighted input data matrix Y. w , means as follows:

[0111]

[0112] Among them, l se1 and l se2 They are F w Two size parameters with different values, which remove the first clustering input matrix F. w Including F w Multiple points, including (M / 2, M / 2), are reconstructed to form a new data matrix Y. w Y w The dimension is M is the weighted input data matrix F w number of rows.

[0113] Step 6, reconstruct the data matrix Y w Perform unsupervised weighted nearest neighbor clustering and output the second clustering result. The first cluster center number Q2 is obtained.

[0114] 6.1) Using the same process as steps 3 and 4, perform unsupervised weighted nearest neighbor propagation clustering on the reconstructed data matrix Yw, and output the second clustering result: in:

[0115]

[0116] Where, γ i'j' (Y w ,P γ ) and α i'j' (Y w ,P γ ) respectively represent the new input Y w and the new reference matrix P γ The new attractiveness obtained through clustering processing and belongingness β γ (i',j'); γ ij and α ij Represent and β γ The iterative update process of (i',j'); i'=1,2,...,M+1-(l se1 +l se2 ), P γ This is the reference degree matrix for the second clustering, and its elements are... for P i It is the reference value for the first cluster. It is the reference degree adjustment factor for the second clustering, l se1 and l se2 They are F w Two size parameters with different values.

[0117] 6.2) Calculate the number of second cluster centers Q2 based on the output of the second clustering:

[0118]

[0119] Step 7: Using the different clustering results from the two multi-clustering operations, design a target detection discrimination criterion.

[0120] 7.1) Set five discrimination interval values ​​u d1 u d2 u d3 u d4 u d5 The five discrimination interval values ​​are empirical parameters, which are adjusted according to different radar systems and target environments. d1 and u d2 Less than M', u d3 and u d4Less than M, M' is the second weighted input data matrix Y w The number of rows, M is the first weighted input data matrix F w Number of rows, u d1 and u d2 The condition 1 ≤ u is satisfied between them. d2 -u d1 ≤4, u d5 Yes, these are supplementary points indicating the possible locations of the cluster centers;

[0121] 7.2) Set the OS-CFAR detection threshold η0:

[0122] Calculate the power Z of clutter plus noise. cn :

[0123]

[0124] Based on the power Z of clutter plus noise cn Calculate the OS-CFAR detection threshold η0:

[0125] η0=T h ×Z cn

[0126] in, To arrange the R reference unit samples in ascending order, i.e., x1≤x2≤...≤x R The Rth K One sample; α is a scalar factor related to the false alarm probability; T h The threshold factor;

[0127] 7.3) Based on the parameters set in 7.1) and 7.2) and the output results of two unsupervised weighted nearest neighbor clustering operations. Given the cluster center numbers Q1 and Q2, design radar target detection and discrimination criteria:

[0128]

[0129] Step 8: Detect airborne radar clutter suppression targets using discrimination criteria.

[0130] Three conditions for determining whether a target exists in a radar echo:

[0131] Condition 1,

[0132] Condition 2,

[0133] Condition 3, Q1 = 1, Q2 = 2, |x l |>η0, where x l This indicates that the radar has received data;

[0134] If any condition in the discrimination criteria is met, then the target exists.

[0135] Otherwise, the target does not exist.

[0136] The effects of this invention can be further illustrated by the following simulation results:

[0137] I. Simulation Conditions

[0138] Except for some settings specifically indicated in the simulation, the radar system and parameter settings are as follows: airborne radar altitude H = 6000m, number of array elements N = 10, element spacing d = 0.3m. Furthermore, pulse number K = 10, wavelength λ = 0.15m.

[0139] For this invention, the parameter settings are p. re =16.68, l se1 =5,l se2 =6, u d1 =8, u d2 =10, u d3 =11, u d4 =30 and u d5 =42, where p re and This is the reference degree adjustment factor for the first and second clustering in this invention. Furthermore, l se1 =5 and l se2 =6 is used to reconstruct new data input for the second clustering for M=60. Meanwhile, u d1 to u d5 This refers to the corresponding setting of the judgment interval in the designed target detection discrimination criterion. Furthermore, in the following description, SNR represents the signal-to-noise ratio, CNR represents the clutter-to-noise ratio, and the clutter is a complex Gaussian distribution. The simulation results will detail various other parameters required for performance comparison; the simulation results are obtained from 400 Monte Carlo experiments.

[0140] II. Simulation Content

[0141] Simulation 1: Under the above simulation conditions, the non-side viewing angle of the forward-looking array is further set to... The target's azimuth velocity is v c =138m / s, the airborne platform speed is v a =90m / s, the total number of samples collected by the distance unit is L=60, CNR=35dB, and the false alarm probability is PFA=2×10 -3 With the additional condition of this invention, target detection is performed on the radar received echo using the present invention and existing STAP, ADC, JDC, JDL, SO-CFAR, GO-CFAR, and OS-CFAR, and the respective detection probability as a function of signal-to-noise ratio curves are obtained, such as... Figure 2As shown in (a). From Figure 2 (a) It can be seen that the unsupervised clustering method of the present invention successfully suppresses clutter in forward-looking array radar, overcomes the non-independent and identically distributed sample conditions and clutter range dependence, and provides target detection performance that is significantly better than other analysis methods.

[0142] Simulation 2: Under the above simulation conditions, the non-side viewing angle of the forward-looking array is further set to... The target's azimuth velocity is v c =137m / s, the airborne radar platform velocity is v a =91m / s, the total number of samples collected by the distance unit is L=60, CNR=30dB, and the false alarm probability is PFA=2×10 -4 With the additional condition of this invention, target detection is performed on the radar received echo using the present invention and existing STAP, ADC, JDC, JDL, SO-CFAR, GO-CFAR, and OS-CFAR, and the respective detection probability as a function of signal-to-noise ratio curves are obtained, such as... Figure 2 As shown in (b). From Figure 2 (b) It can be seen that when the signal-to-noise ratio is greater than 12dB, all detection methods can successfully achieve accurate detection of moving targets, but when the signal-to-noise ratio is less than about 10dB, the detection probability of the present invention is the highest.

[0143] Simulation 3: Under the above simulation conditions, set a non-frontal side view angle. The target's azimuth velocity is v c =130m / s, the airborne radar speed is v a =80m / s, the total number of samples collected by the distance unit is L=60, CNR=50dB, and the parameters of the unsupervised algorithm are u d1 =35, u d2 =36,u d3 =9,u d4 =28 and u d5 With the additional condition of 1, this invention and existing STAP, ADC, JDC, JDL, SO-CFAR, GO-CFAR, and OS-CFAR methods are used to perform target detection on radar received echoes, obtaining the respective detection probability as a function of signal-to-noise ratio curves, such as... Figure 2 As shown in (c). From Figure 2 (c) It can be seen that when the signal-to-noise ratio is below -1dB, the present invention still maintains a significant advantage over other methods in target detection.

[0144] Simulation 4: Under the above simulation conditions, further set the non-side viewing angle of the forward-looking array to... The target's azimuth velocity is v c =137m / s, the airborne radar platform velocity is v a=91m / s, at SNR=-10dB, the false alarm probability is PFA=2×10 -3 With the additional condition that the sample size ranges from L=48 to L=96, this invention, along with existing STAP, ADC, JDC, JDL, SO-CFAR, GO-CFAR, and OS-CFAR methods, were used to perform target detection on radar received echoes, and the curves showing the change in detection probability as a function of the sample size were obtained. Figure 3 As shown. From Figure 3 As can be seen, the detection probability of this invention gradually stabilizes as the number of samples increases. When the number of samples is less than L = approximately 54, the JDL method has a higher detection probability compared to other methods. When the number of samples is greater than L = approximately 54, this invention has a significant advantage over other methods. This is because the weighted data input proposed in this invention helps the designed detection and discrimination criteria to more accurately determine the number and location of cluster centers.

[0145] Simulation 5: Under the above simulation conditions, further set the non-side viewing angle of the forward-looking array to... The target's azimuth velocity is v c =137m / s, the airborne radar platform velocity is v a =91m / s, the total number of samples collected by the distance unit is L=60, at SNR=-10dB, PFA=5×10 -4 Non-frontal viewpoint is defined as Δθ p From Δθ p =0° changes to Δθ p =15°, the parameters of the unsupervised algorithm are u d1 =8-Δu,u d2 =10-Δu,u d3 =11-Δu,u d4 =30-Δu and u d5 =42-Δu additional condition, where Δu=ceil{(Δθ) p +1) / 5}-1, ceil{·} is an integer not less than {·}. Using this invention and existing STAP, ADC, JDC, JDL, SO-CFAR, GO-CFAR, and OS-CFAR methods, target detection is performed on the radar received echo. The detection performance curves as a function of the non-frontal side viewing angle are obtained, and the results are as follows: Figure 4 As shown. From Figure 4 It can be seen that, with Although the detection performance of this invention may fluctuate due to variations in parameters, it is significantly superior to other methods in terms of target detection probability. Moreover, such fluctuations can be reduced by introducing more precise parameters, further improving the detection effect of this invention.

[0146] Simulation 6: Using this invention, 400 clutter suppression target detections were performed on the radar received echo. The number of targets meeting the radar target detection discrimination criteria for the existence of three targets was counted, among which those meeting Q1=Q2=1 were included. u d3 ≤I1≤u d4 The conditions for the existence of the target are expressed as follows: The condition is met that Q1 = Q2 = 1. u d3 ≤I1≤u d4 The conditions for the existence of the target are expressed as follows: And satisfying the condition that when Q1=1, Q2=2, |x l The condition for the existence of the target |>η0 is expressed as: A higher number of tests indicates a greater contribution to accurate detection of the target, as shown in the results. Figure 5 As shown.

[0147] Depend on Figure 5 It can be seen that, It contributes the most to accurate target detection, especially under high signal-to-noise ratio and signal-to-clutter ratio conditions. This determines the vast majority of accurate tests. It provided approximately one in eight successful tests. It provides a relatively minimum number of detections. To adjust the number of detections, in practice, the number can be adjusted according to different situations and needs. d5 Make flexible adjustments.

[0148] Simulation 7. Under the above simulation conditions, further set the non-side viewing angle of the forward-looking array to... The target's azimuth velocity is v c =138m / s, the airborne platform speed is v a =90m / s, the total number of samples collected by the distance unit is L=60, CNR=35dB, and the false alarm probability is PFA=2×10 -3 With additional conditions, this invention and existing methods such as STAP, ADC, JDC, JDL, SO-CFAR, GO-CFAR, and OS-CFAR were used to detect targets in radar-received echoes. The computation time of each detection method as a function of the number of samples L was obtained, and the results are as follows. Figure 6 As shown. From Figure 6 It can be seen that the JDL method, SO-CFAR method, GO-CFAR method, and OS-CFAR method have relatively short computation times. In contrast, the STAP method, the distance-compensated ADC method, and the present invention require longer computation times. When L = 120, the computation times required by the ADC algorithm and the present invention are similar, but under reasonable computation time conditions, the present invention has superior target detection performance.

[0149] Simulation 8: Under the above simulation conditions, further set the non-side viewing angle of the forward-looking array to... The target's azimuth velocity is v c =138m / s, the airborne platform speed is v a =90m / s, the total number of samples collected by the distance unit is L=60, CNR=35dB, SNR=0dB, and the false alarm probability is PFA=2×10 -3 The additional condition is that the clustering method proposed in this invention is used to cluster the radar received echoes, and the clustering results with Q2=2 cluster centers are visualized, that is, the clustering results are presented intuitively in the form of data points. In order to directly represent different points, the normalized average value of the weighted input corresponding to each range cell is introduced as the main feature, and this feature is used as the vertical axis. In the simulation, the cluster centers are specially marked, and their different center positions in the coordinate system are also indicated. The results are as follows. Figure 7 As shown. Figure 7 The design considers three clustering scenarios: a second clustering with clutter and no target, a second clustering with clutter and a target, and a first clustering with clutter and a target. Clustering visualization is used to compare these three cases. In the case of no target, the second clustering typically splits into two clusters. In this case, the first clustering has a center located at...

[0150] Figure 7 It has been verified that, for the different situations described above, the target detection multiple clustering proposed in this invention can achieve target detection based on the difference in the number of their final cluster centers, and can be used in the designed detection discrimination criterion.

[0151] Simulation 9: Under the above simulation conditions, further set the non-side viewing angle of the forward-looking array to... The target's azimuth velocity is v c =138m / s, the airborne platform speed is v a =90m / s, the total number of samples collected by the distance unit is L=60, CNR=35dB, SNR=0dB, and the false alarm probability is PFA=2×10 -3 As an additional condition, the clustering method proposed in this invention is used to cluster radar received echoes. In a few cases, the second clustering of received echoes containing clutter and without targets will form a single cluster. The clustering results with a cluster center number Q2 = 1 are visualized, such as... Figure 8 As shown. Among them. Figure 8 (a) is the unique cluster center of the second cluster. The clustering results of targetless clutter are shown in the figure. Figure 8 (b) is the unique cluster center of the second cluster. The result of clustering clutter without a target is shown in the figure.

[0152] Figure 8This verifies that the detection and discrimination criteria designed in this invention can detect targets by distinguishing their cluster center locations. Figure 8 Even though the second clustering proposed in this invention produces a cluster center when the target does not exist, its center location is relatively far away, even close to the size boundary of the selected data points. This is because the cluster center is not within the expected location range proposed in the detection and discrimination criteria designed in this invention. Therefore, this invention... Figure 8 The case in the above is determined to be that the target does not exist.

Claims

1. An unsupervised k-means clustering based radar clutter suppression target detection method for airborne radars, characterized in that, The method comprises the following steps: 1) Radar transmits a signal and obtains the received data x after matched filtering and pulse stacking cn,l , according to the power characteristics of the clutter plus noise and the signal, design data points and weights E cns and perform matrix transformation to construct the first weighted input data matrix F w ; 2) design the first weighted input data matrix F w reference degree P i and similarity s(i,j), calculate attraction degree and belonging degree β(i,j), and perform unsupervised weighted near neighbor propagation clustering on their iterative updates, when a point j makes the maximum, obtain the clustering center of point i, extract i=1,2,...,M all points to form different parts of less than M clustering centers respectively, as the first clustering result output {I1,I2,...}, M is the number of rows of F w ​ 3) According to the output result of the first clustering, the number of the first clustering centers less than M is calculated: Q1=num{I1, I2,...}; 4) Redesigning the second weighted input data matrix Y w according to the power characteristics of the clutter plus noise and the designed weighted input data matrix F w ; 5) Calculate the similarity s of the second weighted input data matrix Y w γ (i',j')、attractive degree β of i' to j' γ (i',j') and iteratively update it, and perform the second unsupervised weighted affinity propagation clustering, when a certain point j' makes the maximum, obtain the clustering center of point i', extract all points i' = 1,2,...,M' respectively to form different parts of less than M' clustering centers as the second clustering result output M' is the number of rows of Y w ​​ 6) Calculate the number of second cluster centers whose number is less than M' according to the output result of the second clustering: 7) Set five discriminant interval values u d1 , u d2 , u d3 , u d4 , u d5 and OS-CFAR detection threshold η0, using the output results {I1, I2,...} and the number of clustering centers Q1 and Q2 of twice unsupervised weighted near neighbor propagation clustering, design a radar target detection discriminant criterion containing three conditions of target existence: Condition 1, Q1 = Q2 = 1, u d3 ≤ I1 ≤ u d4 ; Condition 2, Q1 = Q2 = 1, u d3 ≤ I1≤ u d4 ; Condition 3, Q1 = 1, Q2 = 2, |x l | > η0, where x l represents radar reception data; 8) According to the radar target detection standard designed in step 7), the clutter suppression target detection result of the airborne radar is judged: When any condition in the discrimination standard is met, the target exists, Otherwise, the target does not exist.

2. The method according to claim 1, characterized in that in step 1) the radar transmits a signal and obtains received data x after matched filtering and pulse stacking cn,l is represented as follows: wherein l represents the lth range cell of L training range cell samples, n l represents zero mean noise, N c is the total number of clutter patches corresponding to the range cell, ξ i,l is the corresponding complex amplitude, i = 1...N c ; represents the ith clutter patch NKx1 dimensional space-time steering vector, N is the number of uniform linear array receiving elements, K is the number of transmitted pulses, is the Kronecker product, and represent the time steering vector and the spatial steering vector, respectively; and represent the normalized Doppler frequency and spatial frequency, respectively.

3. The method of claim 1, characterized in that the design data points in step 1) are and weights E cns and the matrix transformation is performed to construct a first weighted input data matrix F w as follows: 1a) Design data points which contains clutter plus noise and signal, is represented as follows: wherein for the (l,m)th element of the data point matrix L is the total number of training distance unit samples, and M is the total number of power points; For the covariance matrix estimate, it is obtained from all samples where a detected target can exist; the space-time steering vector, is the Kronecker product, denotes the normalized Doppler frequency corresponding to the mth power point in the time domain of interest, denotes the normalized spatial frequency of the lth power point in the spatial distribution; 1 b) Designing the weights E of the weighted data cns : 1b1 ) Calculate the weight E of the 1th weighted data cns (l): where l = 1, 2,..., L, ω max and ω min are the normalized maximum and minimum expectation values, respectively; x cn,l represents the received data of the lth range cell, X cns is a matrix composed of L received echo data, and is expressed as [x cn,1 , x cn,2 ,..., x cn,l ,..., x cn,L ], w opt is a STAP weight vector; 1b2) the weights E of the l weighted data cns (l) obtaining a weight matrix E cns is: 1 c) Based on the results of steps 1 a) - 1 b), calculate the data matrix for the (m, 1)th element wherein [·] η is the ηth power of [·]; 1d) sequentially calculating M x L F w (m, l), the composition data matrix and matrix transforming it to obtain a first weighted input data matrix F w : Where |·| represents taking the absolute value of each element of the matrix.

4. The method of claim 3, wherein, The STAP weight vector w in step 1b1) opt is obtained by the Lagrange multiplier method and solving the minimum variance problem, which is given by the following equation where a st (f s0 ,f d0 ) is the target space-time steering vector of NKx1 dimension, f d0 and f s0 are the time and spatial frequency, respectively; (·) -1 denotes the inverse operation of a matrix, (·) H denotes the conjugate transpose of a matrix; R X is the clutter-plus-noise covariance estimation matrix, which does not include the echo vector x of the possible target when making the estimation cn,l of x cn,ls is obtained by the maximum likelihood estimation method using L-1 training samples, and is expressed as follows: In the formula, l s The target cell.

5. The method according to claim 1, characterized in that the reference degree P and the similarity degree s(i,j) of the first weighted input data matrix F w i are calculated in step 2) and the attraction degree the belonging degree β(i,j) is achieved as follows:​​​ 2a) design reference degree P i and similarity s(i,j): where F w (i,:) represents the i-th row of F w w (j,:) represents the j-th row of F w , {i,j} = 1,2,...,M, represents the distance square between F w (i,:) and F w (j,:), l = 1,2,...,L, L is the total number of training distance unit samples; when i = j, s(i,j) is defined as,​ s(i,j) = P i i = j where P i = p re · mean{S} is the first cluster reference, p re is the reference adjustment factor parameter, mean{·} represents the mean of all elements in {·}, and S is detailed as follows: 2b) calculating the attraction degree and the belonging degree β(i,j): wherein M is the weighted input data matrix F w the number of rows of F, and the initial value of the degree of belonging β(i, j) is zero, and when i = j, we have directly:

6. The method of claim 1, characterized in that the attraction degree is updated in step 2) The homogeneity β(i,j) is iteratively updated to perform unsupervised weighted affinity propagation clustering, outputting the first clustering result {I1, I2,...}, which is implemented as follows: 2c) when i≠j, update and β(i,j) respectively to: where t < E, λ is a damping factor, is the attraction degree of the (t-1)th iteration, β (t) (i,h) is the membership degree of the tth iteration; is the result of the tth iteration of the attraction degree, β (t-1) (i,j) is the result of the (t-1)th iteration of the membership degree; When i = j, update and β (t) (j, j) are updated as follows, respectively: 2d) according to the updated and β (t) (i,j), the cluster center for point i is taken from the point j that maximizes For the row number M of F w Extract all points i = 1,2,...,M, respectively form different parts of less than M cluster centers as the first clustering result output {I1,I2,...}:

7. The method according to claim 1, characterized in that the second weighted input data matrix Y reconstructed in step 4) w is represented as follows: wherein, l se1 and l se2 They are F w Two size parameters with different values, which remove the first clustering input matrix F. w Including F w Multiple points, including (M / 2, M / 2), are reconstructed to form a new data matrix Y. w Y w The dimension is M is the first weighted input data matrix F w number of rows.

8. The method according to claim 1, characterized in that the second clustering result in step 5) is represented as follows: wherein The second clustering result is output. To extract the values ​​for i' = 1, 2, ..., M+1 - (l se1 +l se2 The different parts of the cluster centers formed by all points, γ i'j' (Y w ,P γ ) and α i'j' (Y w ,P γ ) respectively represent the new input Y w and the new reference matrix P γ The new attractiveness obtained through clustering processing and belongingness β γ (i',j'); γ ij and α ij Represent and β γ The iterative update process of (i',j'); i'=1,2,...,M+1-(l se1 +l se2 ), P γ This is the reference degree matrix for the second clustering, and its elements are... for P i It is the first cluster reference degree. It is the reference degree adjustment factor for the second clustering, l se1 and l se2 They are F w Two size parameters with different values.

9. The method according to claim 1, characterized in that the radar target detection discrimination criterion designed in step 7) and its discrimination interval, the set OS-CFAR detection threshold η0 are specifically represented as follows: wherein Discrimination interval value u d1 , u d2 , u d3 , u d4 , u d5 is an empirical parameter, according to different radar system target environment, u d1 and u d2 are less than M', u d3 and u d4 are less than M, M' is the second weighted input data matrix Y w row number, M is the first weighted input data matrix F w row number, u d1 and u d2 satisfy 1≤u d2 -u d1 ≤4, u d5 is the supplementary point of the cluster center possible location, the OS-CFAR detection gate value η0 is calculated according to the power Z cn of the clutter plus noise: η0 = T h x Z cn wherein, To arrange the R reference unit samples in ascending order, i.e. x1≤ x2≤...≤ xR R , the R K th sample in the sorted sequence; a is a scalar factor related to the false alarm probability; T h is a threshold factor.

Images