A method and system for clutter edge detection and clutter data classification

Abstract

This invention provides a clutter edge detection and clutter data classification method and system. For sonar systems with multiple array elements, based on the generalized likelihood ratio test and model order selection criterion, it determines the number and location of reverberation edges in auxiliary data and assesses the uniformity of the auxiliary data. The method includes: establishing a binary hypothesis testing problem based on echo data received by the sonar system from the range cell to be detected and the auxiliary cell; estimating the number of reverberation types, the reverberation covariance matrix of the auxiliary data, and the reverberation edge locations using the model order selection criterion; and determining the uniformity of the auxiliary data using the generalized likelihood ratio test. The advantages of this invention are: direct classification of reverberation data by estimating the reverberation edge locations; adaptive estimation of the background reverberation energy of each uniform region; and adaptive estimation of the number and location of reverberation edges in the auxiliary data, making it more practical in engineering applications.

Description

Technical Field

[0001] This invention belongs to the field of sonar, specifically relating to a clutter edge detection and clutter data classification method and system. Background Technology

[0002] For active sonar systems, in addition to inherent underwater noise, reverberation is the main background interference affecting the detection performance of the sonar system. When the sonar platform moves, reverberation exhibits spatiotemporal coupling characteristics, making it difficult to effectively distinguish between the target and the reverberation using only spatial or temporal filtering. Spatiotemporal adaptive detection technology (STAD) uses a spatiotemporal joint processing framework with target detection as its purpose. It can achieve the integration of reverberation suppression and target detection. Compared with the cascaded processing method of filtering first and then detecting and deciding, STAD has a more flexible design, can make full use of received data, and has better detection performance.

[0003] In the STAD method, the accuracy of the reverberation covariance matrix (RCM) estimation directly affects reverberation suppression performance and target detection capability. Traditional RCM estimation methods use data from several range cells near the target range cell (CUT) as auxiliary data to estimate the RCM in the CUT. This method is effective in a uniform reverberation environment, i.e., when the auxiliary data are independent and identically distributed and have the same statistical characteristics as the reverberation in the CUT. However, in real-world environments, the presence of reverberation edges causes uneven reverberation distribution. These reverberation edges divide the sonar observation window into several uniform regions, and the RCMs of different regions are not the same. In this case, directly using the sample covariance matrix of all auxiliary cells as the RCM estimation result will severely degrade the RCM estimation performance.

[0004] To address this problem, scholars have proposed a series of methods, such as reverberation data classification based on the expectation-maximization algorithm and reverberation edge detection algorithms based on the sliding window method. The first type of method introduces a hidden variable representing the reverberation type to determine the category of each distance cell. The expectation-maximization algorithm is used to estimate the unknown parameters. After obtaining the parameter estimates, the posterior probability of the category to which each distance cell belongs is calculated to classify the reverberation data. The second type of method introduces an observation window and moves it within the entire range of distance cells to be detected. The generalized likelihood ratio test (GLRT) is used to determine whether a reverberation edge exists within the window and to estimate the location of the reverberation edge.

[0005] Existing technologies have three main shortcomings:

[0006] 1. Reverberation data classification methods based on the expectation-maximization algorithm only focus on the clustering of reverberation, neglecting the potential problem of locating reverberation edges. In real-world environments, reverberation changes slowly. A more reasonable model is that reverberation edges divide the sonar observation window into several regions, where reverberation within the same region is independent and identically distributed, while reverberation between different regions does not satisfy the uniformity assumption. Therefore, once the location of the reverberation edges is determined, reverberation data clustering can be directly achieved.

[0007] 2. The ability of the reverberation edge detection algorithm based on the sliding window method to extract background reverberation energy information needs to be improved.

[0008] 3. Neither of the two methods can adaptively estimate the number of reverberation edges. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies that only focus on the clustering problem of reverberation, without paying attention to the potential reverberation edge location problem, and cannot adaptively estimate the number of reverberation edges.

[0010] To achieve the above objectives, this invention proposes a clutter edge detection and clutter data classification method. For a sonar system with multiple array elements, based on the generalized likelihood ratio test and model order selection criterion, the method determines the number and location of reverberation edges in auxiliary data, and assesses the uniformity of the auxiliary data. The method includes:

[0011] Step 1: Based on the echo data received by the sonar system from the range unit to be detected and the auxiliary unit, establish a binary hypothesis testing problem;

[0012] Step 2: Using the model order selection criterion, estimate the number of reverberation types, the reverberation covariance matrix of auxiliary data, and the reverberation edge location;

[0013] Step 3: Use the generalized likelihood ratio test to determine whether the auxiliary data are uniform.

[0014] As an improvement to the above method, step 1 specifically includes:

[0015] The binary hypothesis testing problem is as follows:

[0016]

[0017] Where H0 represents the null hypothesis; H 1,m Indicates the alternative hypothesis; z l ∈C N×1 ,l=1,...,L,z l This represents the N-dimensional complex vector formed by sampling the echo data received by the sonar system from each range cell, where N represents the number of sonar elements; L represents the number of auxiliary elements; zl ~CN N (0,M k ) represents z l It is a matrix with a mean of 0 and a covariance matrix of M. k An N-dimensional Gaussian complex vector;

[0018] m represents the number of reverb types; the upper limit for the number of reverb types is set to m. max Then the range of values ​​for m is m = 2, 3, ..., m max ;

[0019] M0∈C N×N M represents the reverberation covariance matrix of the auxiliary data under the uniformity assumption; k ∈C N×N k = 1, 2, ..., m represents the reverberation covariance matrix of the k-th uniform reverberation region under the non-uniformity assumption; L1, L2, ..., L m-1 Indicates the location of the reverberation edge;

[0020] Let Z = [z1, z2, ..., z L ],Ω1={1,...,L1},Ω2={L1+1,..,L2},...,Ω k ={L k-1 +1,..,L k},...Ω m ={L m-1 +1,..,L};

[0021] The probability density function under the H0 hypothesis is:

[0022]

[0023] H 1,m The probability density function under the assumption is:

[0024]

[0025] Where tr[·] denotes the trace of a matrix; det[·] denotes the determinant of a matrix; [·] -1 This represents finding the inverse of a matrix; S0 = Z × Z H , [·] H Σ represents the conjugate transpose operation; m M1, M2, ..., M k ,...,M m Ξ m =[L1,L2,...,L m-1 ].

[0026] As an improvement to the above method, step 2 specifically includes:

[0027] Estimate the number of reverberation types m:

[0028]

[0029] in, This represents an estimated value for the number of reverberation types, m. Representing Σ m Ξ m The estimated value; c represents a constant;

[0030] Estimate the reverberation covariance matrix of the auxiliary data:

[0031]

[0032] in, M represents the reverberation covariance matrix of the auxiliary data. k The estimated value;

[0033] Estimate the location of the reverberation edge:

[0034] First, the reverberation edge positions are initialized, that is...

[0035]

[0036] in, This represents the estimated position of the j-th reverberation edge during initialization;

[0037] The reverberation edge locations are determined using a cyclic search method. The estimation result for the j-th reverberation edge in the t-th iteration is as follows:

[0038]

[0039] in, express The value and The value depends on the value of;

[0040]

[0041]

[0042] The iteration stops when the number of iterations reaches the set upper limit or when the difference between the current estimate and the previous estimate is less than the set convergence threshold, thus obtaining all reverberation edge positions.

[0043] As an improvement to the above method, the constant c takes the following values ​​under the three criteria: the Akaike information criterion, the Bayesian information criterion, and the generalized information criterion:

[0044]

[0045] As an improvement to the above method, step 3 specifically includes:

[0046] To determine whether auxiliary data is uniform, the specific format is as follows:

[0047]

[0048] in, Represents the likelihood ratio function:

[0049]

[0050] η represents the detection threshold; Indicates will The result after substituting the alternative hypothesis; This represents the maximum likelihood estimation result of M0:

[0051]

[0052] The present invention also provides a clutter edge detection and clutter data classification system, implemented based on the above method, the system comprising:

[0053] A binary hypothesis testing module is established to establish a binary hypothesis testing problem based on the echo data received by the sonar system from the range unit to be detected and the auxiliary unit.

[0054] The parameter estimation module is used to estimate the number of reverberation types, the reverberation covariance matrix of auxiliary data, and the reverberation edge locations using the model order selection criterion; and

[0055] The module for determining whether auxiliary data is uniform is used to determine whether auxiliary data is uniform using the generalized likelihood ratio test method.

[0056] Compared with the prior art, the advantages of the present invention are:

[0057] 1. Existing technologies only focus on the clustering problem of reverberation, without addressing the potential problem of locating reverberation edge positions. This invention establishes a more realistic model that directly classifies reverberation data by estimating the reverberation edge positions.

[0058] 2. Existing technologies have limited ability to extract background reverberation energy information. This invention considers data from the entire sonar observation window and can adaptively estimate the background reverberation energy of each uniform region.

[0059] 3. Existing technologies cannot adaptively estimate the number of reverberation edges, limiting their practical value. This invention can simultaneously determine the number and location of reverberation edges in auxiliary data. Through generalized likelihood ratio testing and the MOS algorithm, it achieves adaptive estimation of the number and location of reverberation edges in auxiliary data, making it more practical in engineering applications. Attached Figure Description

[0060] Figure 1 The diagram shows the flowchart of the clutter edge detection and clutter data classification method.

[0061] Figure 2 The figure shows P under three MOS criteria. d A diagram illustrating the changes with CPR;

[0062] Figure 3 The figure shows P under three MOS criteria. cc A diagram illustrating the changes with CPR;

[0063] Figure 4 The figure shows P under three MOS criteria. cc A diagram illustrating how CPR changes. Detailed Implementation

[0064] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0065] This invention discloses a clutter edge detection and clutter data classification method and system. Focusing on the reverberation edge detection problem, based on the generalized likelihood ratio test and the model order selection criterion (MOS), it proposes a method that can simultaneously determine the number and location of reverberation edges in auxiliary data, as well as whether the auxiliary data is uniform. This method does not require any prior information about the number of reverberation edges and can adaptively estimate the number of reverberation edges, locate reverberation edges, and determine whether the environment is uniform.

[0066] First, to determine whether the auxiliary data is uniform, a binary hypothesis testing problem needs to be established, where the alternative hypothesis is a nested hypothesis. Traditional maximum likelihood estimation (ML) methods suffer from severe overfitting in this case. Therefore, this invention employs a two-step cascaded architecture to address the hypothesis testing problem. In the first stage, for the parameter estimation problem under the nested alternative hypothesis, the MOS criterion is used to estimate the number of reverberation edges, while simultaneously providing estimates of the reverberation edge locations and RCM. In the second stage, based on the parameter estimation results under the alternative hypothesis, the GLRT method is used to determine whether reverberation edges exist in the auxiliary data.

[0067] The specific process of the clutter edge detection and clutter data classification method of the present invention is as follows:

[0068] 1. Problem Modeling

[0069] Consider a sonar system with N elements (N≥2). The receiving array is a uniformly spaced linear array with an element spacing of d = λ / 2, where λ is the operating wavelength. This sonar system receives echo data from the cutoff point (CUT) and auxiliary elements, and samples the echo from each range element to form an N-dimensional complex vector z.l ∈C N×1 Let l = 1, ..., L, where L represents the number of auxiliary units. The auxiliary data may be uniform or non-uniform. When the auxiliary data is non-uniform, several reverberation edges exist within the sonar observation window, dividing the auxiliary units into several uniform subsets. To determine whether the auxiliary data is uniform, the following binary hypothesis testing problem is established:

[0070]

[0071] Where z l For the echo data of the l-th range cell, x ~ CN N (μ, M) represents x as an N-dimensional Gaussian complex vector with mean μ and covariance matrix M. m represents the number of reverberation types, assuming an upper limit of m. max Then the range of values ​​for m is m = 2, 3, ..., m max M0∈C N×N RCM of auxiliary data under the uniformity assumption. M k ∈C N×N k = 1, 2, ..., m represents the RCM of the k-th uniform reverberation region under the non-uniformity assumption. L1, L2, ..., L m-1 This represents the reverberation edge position.

[0072] Let Z = [z1, z2, ..., z L ],Ω1={1,...,L1},Ω2={L1+1,..,L2},Ω m ={L m-1 +1,..,L}。Then the probability density function (PDF) under the H0 hypothesis is:

[0073]

[0074] The PDF under the H1 hypothesis is:

[0075]

[0076] In the formula, tr[·] represents the trace of a matrix, det[·] represents the determinant of a matrix, and [·] represents the matrix determinant. -1 Represents the inversion of a matrix, S0 = Z × Z H , [·] H Represents the conjugate transpose operation, Σ m Representing M1, M2, ..., M m Ξ m =[L1,L2,...,L m-1 ].

[0077] 2. Algorithm Derivation

[0078] To solve the binary hypothesis testing problem shown in equation (1), a two-step cascaded method is adopted. In the first step, the parameters under the multivariate nested alternative hypothesis are estimated using the MOS criterion. In the second step, the GLRT method is used to determine whether the auxiliary data are uniform.

[0079] (1) Alternative hypothesis parameter estimation

[0080] 1) First, we need to estimate the parameter m under the H1 assumption. We assume there is an upper limit to the number of uniform regions, denoted as m. max Because the H1 hypotheses are nested, the likelihood function monotonically increases with m. In this case, the estimation result of the ML method is always m. max Therefore, the MOS algorithm is used to estimate m, and its expression is as follows:

[0081]

[0082] In the above formula They are Σ m Ξ m The estimate, c·h(m), is a penalty term added to prevent overfitting, where c is a constant, and its values ​​under the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and Generalized Information Criterion (GIC) are as follows:

[0083]

[0084] h(m) represents the number of unknown parameters, and its values ​​are:

[0085] h(m)=m×(N 2 +N+1)-1 (6)

[0086] 2) Next is Σ m Ξ m The estimation requires addressing the following issues:

[0087]

[0088] Where, Σ m (Ξ m ) represents Σ m The value of Ξ m Related; when the location of the reverberation edge is given, for Σ m The estimate, i.e., M k The maximum likelihood estimate for k = 1, 2, ..., m can be obtained from equation (3), specifically expressed as:

[0089]

[0090] Substituting the result of equation (8) into equation (7) yields:

[0091]

[0092] 3) Finally, the estimation of the reverberation edge position is the optimization problem of solving equation (9). However, the above optimization problem is a high-dimensional search problem. To avoid this high-dimensional search problem, a cyclic search method based on distance cells is used to solve the optimization problem. First, the reverberation edge position is initialized, i.e.

[0093]

[0094] make The estimation result of the j-th reverberation edge in the t-th iteration is:

[0095]

[0096] in,

[0097]

[0098] When the number of iterations reaches the upper limit t max The iteration stops when the difference between the current estimate and the previous estimate is less than the convergence threshold.

[0099] (2) Generalized likelihood ratio test

[0100] Substitute the above parameter estimation results into GLRT to determine whether the auxiliary data is uniform, in the following form:

[0101]

[0102] in Let be the likelihood ratio function, and its expression is:

[0103]

[0104] Taking the logarithm of the above expression, we get:

[0105]

[0106] η is the detection threshold, used to ensure a constant false alarm probability; Indicates will The result after substituting the alternative hypothesis. The maximum likelihood estimate of M0 is expressed as follows:

[0107]

[0108] 3. Performance Analysis

[0109] To verify the effectiveness of the algorithm, performance analysis was performed using simulation data. In the simulation experiment, it was assumed that N = 10, L = 128, and m... max =5. The RCM of the auxiliary data under the H0 assumption and the RCM of the k-th reverberation region under the H1 assumption are generated by the following formula:

[0110]

[0111] Where σ n 2 represents noise power, I N This represents an N×N dimensional identity matrix. Let H0 be the reverberation power of the auxiliary data, expressed as follows: RNR0 is the noise-mixing ratio under the H0 assumption. The reverberation power of the k-th region under assumption H1 is expressed as follows:

[0112]

[0113] Here, RNR1 represents the noise-mixing ratio of the first region, and CPR is used to represent the energy difference between two adjacent regions. v(θ) i )∈C N×1 The guiding vector is expressed as v(θ). i )=[1,exp(jπsinθ i ),...,exp(j(N-1)πsinθ i )] T .

[0114] Θ k Let k = 0, 1, ..., m represent the set of Angles of Arrival (AOA). Different AOAs will result in differences in the RCM structure between uniform regions. Experiments show that, in this case, the algorithm can accurately estimate the number and location of reverberation edges for all CPR values. Therefore, this invention focuses on a more challenging scenario, considering the case where the AOA is the same but the reverberation power differs between uniform regions. In this experiment, it is assumed that... RNR0 = RNR1 = 10 dB. Θ k = {-15°, -5°, 5°, 15°}, k = 0, 1, ..., m. In addition, to ensure a constant false alarm probability (P... fa =10 -3 ), requires running 100 / P fa The Monte Carlo experiment was used to determine the detection threshold η.

[0115] Next, we will use 10. 4 An independent Monte Carlo test was conducted to evaluate the performance of the reverberation edge detection algorithm, specifically including the following performance metrics: reverberation edge detection probability (P...d The probability of correctly estimating the number of reverberation edges (P) cc And the mean squared error (RMSE) of the Hausdorff distance between the estimated edge location and the true edge location. Assume m = 4, Ξ m =[30,60,90], under the GIC criterion, let ρ=2. Figure 2 The figure shows the detection probability P. d The changes with CPR. The graph shows that P... d It increases with increasing CPR, because the larger the CPR, the greater the energy jump between two adjacent uniform reverberation regions. And when CPR ≥ 4 dB, P... d It converges to 1. Figure 3 P is shown cc The changes with CPR. The graph shows that P... cc It increases with increasing CPR. If P... cc If the algorithm converges to 1, then the minimum CPR required for reverberation edge detection under the three MOS criteria (AIC, BIC, and GIC) are 14dB, 26dB, and 18dB, respectively. For the same CPR, the AIC criterion yields a higher probability of accurate estimation of the number of reverberation edges. This is because the AIC criterion introduces a smaller penalty coefficient, effectively preventing underfitting. Figure 4 The figure shows the variation of RMSE between the estimated and true edge locations with CPR. The results indicate that RMSE decreases as CPR increases, and when CPR ≥ 26 dB, the RMSE converges to 0 under different MOS criteria. Similarly, under the AIC criterion, the reverberant edge detection algorithm exhibits the best edge location estimation performance.

[0116] The present invention also provides a clutter edge detection and clutter data classification system, implemented based on the above method, the system comprising:

[0117] A binary hypothesis testing module is established to establish a binary hypothesis testing problem based on the echo data received by the sonar system from the range unit to be detected and the auxiliary unit.

[0118] The parameter estimation module is used to estimate the number of reverberation types, the reverberation covariance matrix of auxiliary data, and the reverberation edge location using the model order selection criterion.

[0119] The module for determining whether auxiliary data is uniform is used to determine whether auxiliary data is uniform using the generalized likelihood ratio test method.

[0120] This invention employs the generalized likelihood ratio test and the MOS method to simultaneously determine the number and location of reverberation edges in auxiliary data, without requiring the number of reverberation edges as prior knowledge.

[0121] This invention proposes a reverberation edge detection method, which enables sonar platforms to effectively classify reverberation data, thereby improving the performance degradation of target detection under non-uniform reverberation backgrounds.

[0122] This invention establishes a more realistic model where the reverberation edge divides the sonar observation window into several regions. Reverberation within the same region is independently and identically distributed, while reverberation between different regions does not satisfy the uniformity assumption. Reverberation data classification is achieved directly by estimating the location of the reverberation edge.

[0123] This invention does not require the number of edges as prior knowledge. It simultaneously determines the number and location of reverberation edges in auxiliary data through generalized likelihood ratio test and MOS algorithm.

[0124] The algorithm proposed in this invention takes into account the data of the entire sonar observation window and can adaptively estimate the background reverberation energy of each uniform region.

[0125] This invention employs a cyclic search method based on distance cells to estimate the reverberation edge location, thereby avoiding the large computational burden caused by solving high-dimensional search problems.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for clutter edge detection and clutter data classification, for a sonar system with multiple array elements, based on generalized likelihood ratio test and model order selection criterion, to determine the number and location of reverberation edges in auxiliary data, and to determine whether the auxiliary data is homogeneous, the method comprising: Step 1: based on the echo data received by the sonar system from the distance unit to be detected and the auxiliary unit, a binary hypothesis testing problem is established; Step 2: using the model order selection criterion, the number of reverberation types, the reverberation covariance matrix of the auxiliary data, and the reverberation edge position are estimated; Step 3: using the generalized likelihood ratio test method to determine whether the auxiliary data is homogeneous; The step 1 specifically comprises: The binary hypothesis testing problem is established as follows: ； wherein, represents a primary hypothesis; represents an alternative hypothesis; , represents the echo data received by the sonar system and sampled for each range cell to form a vector of dimension represents the number of elements of the sonar; represents the number of auxiliary units; represents is a Gaussian complex vector of dimension with mean 0 and covariance matrix . represents the number of reverb types; the upper limit of the number of reverb types is set to , then the value range of is ; represents a reverberation covariance matrix of the auxiliary data under the uniform assumption; represents a reverberation covariance matrix of the first uniform reverberation region under the non-uniform assumption; represents a reverberation edge position; Set , ; The probability density function under the assumption is: ； The probability density function under the assumption is: ； wherein denotes the matrix trace; denotes the matrix determinant; denotes the matrix inverse; ; denotes the conjugate transpose operation; denotes , .

2. The method of clutter edge detection and clutter data classification of claim 1, wherein, The step 2 specifically comprises: On the number of reverb classes Estimation is performed: ； wherein represents the number of reverberation classes represents an estimate of , respectively represent , represents an estimate of ; c represents a constant The reverberation covariance matrix of the auxiliary data is estimated as follows: ； wherein represents an estimate of the reverberation covariance matrix of the auxiliary data represents an estimate of the reverberation covariance matrix of the auxiliary data The reverberation edge position is estimated as follows: First, the reverberation edge position is initialized, that is: ； wherein represents an estimate of the initial position of the reverb edge; and represents an estimate of the initial position of the reverb edge; and The reverberation edge location is solved using a cyclic search method. The first reverberation edge The estimation result of the next iteration is: ； wherein represents the value of the value of ； ； When the number of iterations reaches the set upper limit or the difference between the current estimated value and the previous iteration estimated value is less than the set convergence threshold, the iteration stops, and all reverberation edge positions are obtained.

3. The method of clutter edge detection and clutter data classification of claim 2, wherein, The value of the constant c under the Akaike information criterion, the Bayesian information criterion and the generalized information criterion is respectively: 。 4. The method of clutter edge detection and clutter data classification of claim 3, wherein, The step 3 specifically comprises: Determine whether the auxiliary data is homogeneous, specifically as follows: ； wherein denotes the likelihood ratio function: ； denotes a detection threshold; denotes a decision rule the result after bringing in the alternative hypothesis; denotes the maximum likelihood estimate result: 。 5. A clutter edge detection and clutter data classification system, implemented based on any of the methods of claims 1-4, characterized in that, The system comprises: A binary hypothesis testing module is established for establishing a binary hypothesis testing problem based on the echo data received by the sonar system from the distance unit to be detected and the auxiliary unit; A parameter estimation module is used to estimate the number of reverberation types, the reverberation covariance matrix of the auxiliary data, and the reverberation edge position using the model order selection criterion; and A module for determining whether the auxiliary data is homogeneous is used to determine whether the auxiliary data is homogeneous using the generalized likelihood ratio test method.

Images