An improved robust adaptive beamforming algorithm for data dimensionality reduction
Through the improved Lagrangian factor solution method, combined with conjugate gradient and Newton iteration method, the data dimension reduction and stable adaptive beamforming algorithm are optimized, which solves the problem of broadband detection performance decline caused by invalid loading factors, and improves the stability of the algorithm and computational efficiency.
Patent Information
- Application Number
- CN202210732222.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-24
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-06-24
AI Technical Summary
The existing data dimensionality reduction and robust adaptive beamforming algorithms are prone to obtain invalid loading factors during the calculation process, resulting in a degradation of broadband detection performance, especially when the Heiser matrix is not positive and initial point selection is not selected, resulting in the direction of the direction vector being not. The downward direction causes the loading volume to not converge.
The improved Lagrangian factor solution method is adopted, combined with the conjugate gradient algorithm and Newton iteration method, and optimize the Newton iteration method through dichotomy and error judgment mechanism to ensure the calculation efficiency and stability, avoid invalid loading factors, and solve the optimal one Lagrangian factor to obtain the estimated weighted vector.
The stability and computing efficiency of the iterative process are improved, the broadband detection performance is improved, the negative impact of invalid loading factors on detection performance is avoided, and the robustness of the robust adaptive beamforming algorithm is significantly improved.
Smart Images

Figure CN115114574B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of passive sonar signal processing, relates to research on sonar array adaptive broadband warning, and more specifically to a robust adaptive beamforming method for data dimensionality reduction. Background Art
[0002] In view of the high signal-to-noise ratio and real-time requirements of the new generation of low-frequency passive sonar systems, it is difficult for the minimum variance distortionless response (MVDR) beamformer under ideal conditions to achieve the optimal spatial filtering performance. The MVDR beamformer is very sensitive to errors. The limited number of snapshots, the covariance matrix model, and the steering vector error will cause a sharp decline in signal detection capabilities. In actual engineering applications, the small sample size and model errors are inevitable, which will cause the expected signal to cancel each other, deteriorating the interference suppression capability of the adaptive beamformer. In order to achieve robust and reliable engineering implementation of adaptive beamformers, several methods to improve the robustness of MVDR beamformers have emerged, among which the diagonal loading method has been the most widely used. However, the biggest problem with the diagonal loading method is that it is difficult to determine the size of the diagonal loading based on the steering vector error information. Using the same fixed amount for diagonal loading of the covariance matrix of all frequency points can easily cause some frequency components with good detection performance to be overloaded, affecting the broadband detection performance. The weighted vector norm constraint method, the worst performance best method and the robust adaptive beamforming algorithm can accurately calculate the size of the diagonal loading according to the steering vector error ellipse, so that the model error problem can be solved to a certain extent.
[0003] With the demand for long-range target detection, passive sonar systems are developing towards large-scale, large-array sonar systems. This makes the real-time performance of the sonar system worse. In particular, the robust adaptive beamforming algorithm (RCB) based on elliptical uncertainty sets uses the Newton iteration method to solve the optimal weight vector, which requires huge computation. In view of the contradiction between filtering technology and matrix operation computation in matrix spatial domain filtering, the robust adaptive beamforming algorithm based on Krylov data dimensionality reduction is a low-dimensional filter matrix design method.
[0004] The diagonal loading of the data dimension reduction robust adaptive beamforming algorithm (RDRCB) can determine the appropriate loading according to different steering vector errors. The required diagonal loading factor is not a simple closed-form solution. The Newton method can quickly solve the approximate weighted amount. The Newton method is an iterative algorithm with the advantage of fast convergence speed, which alleviates the computing resource burden of the sonar signal processing system. However, the Newton method is not a descent algorithm. When the Hessian matrix is not positive definite and the initial point is not selected properly, the direction of the direction vector is often not the descent direction, resulting in the loading amount not being a converged value. In order to avoid obtaining an invalid divergent solution, it is necessary to provide a fast and robust Lagrangian factor solution method. Summary of the invention
[0005] One of the purposes of the invention is to provide a fast and robust Lagrangian factor solution method to solve the problem that the RDRDB method proposed in the background technology easily obtains a divergent value of the loading, thereby affecting the broadband detection information.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] An improved data dimension reduction robust adaptive beamforming algorithm, the algorithm comprising the following steps:
[0008] Step 1: Calculate the covariance matrix of each sub-band sample
[0009] Step 2: Use the conjugate gradient algorithm to calculate the covariance matrix Solve the dimension reduction matrix D;
[0010] Step 3: Use the dimension reduction matrix to calculate the diagonal matrix Λ CG ;
[0011] Step 4: Solve the estimated expected signal steering vector after dimensionality reduction and constraint matrix M;
[0012] Step 5, perform eigenvalue decomposition on the constraint matrix M;
[0013] Step 6, solving the Lagrangian factor;
[0014] Step 7: Obtain an estimated weight vector based on the optimal Lagrangian factor solution.
[0015] Preferably, the covariance matrix Where K is the number of snapshots, x k is the array snapshot data received by the kth array element, and H represents transposition.
[0016] Preferably, the dimension reduction matrix D=[d1, ..., d N ] includes the following steps:
[0017] Step 2.1, initialize the iteration counter j = 1, initialize the residual vector in, To estimate the desired signal steering vector;
[0018] Step 2.2, calculate the learning rate α for the jth iteration j and step length β j , learning rate Step Length Among them, d j is the residual of the jth iteration, rj is the vector of the jth iteration, H is the transpose, is the covariance matrix of each sub-band sample;
[0019] Step 2.3, based on the learning rate α of the jth iteration j and step length β j Calculate the vector r of the j+1th iteration j+1 and residual amount d j+1 ,vector residual amount d j+1 =-r j+1 +β j d j ;
[0020] Step 2.4, add 1 to the iteration counter j, and return to step 2.2 until j=N-1, where N is the dimensionality reduction order.
[0021] Preferably, the diagonal matrix is calculated by using the dimension reduction matrix Among them, N is the dimensionality reduction order, H is the transpose, is the covariance matrix, and the diag function is used to extract the diagonal elements of the matrix and create a diagonal matrix.
[0022] Preferably, the estimated expected signal steering vector after dimensionality reduction is and the constraint matrix M are expressed as
[0023]
[0024]
[0025] in, To estimate the desired signal steering vector, E = ε -1 I N , ε is the error between the preset steering vectors, I N represents the identity matrix, Λ CG is a diagonal matrix, D represents the dimension reduction matrix, and H represents the transpose.
[0026] Preferably, the eigenvalue decomposition of the constraint matrix M after dimension reduction is performed as follows: U=[u1,...,u n , ..., u N ], U represents the eigenvector matrix of the constraint matrix M after dimensionality reduction, u is the eigenvector, n is the matrix index, λ n for u n The corresponding eigenvalues.
[0027] Preferably, solving the Lagrangian factor comprises the following steps:
[0028] Step 6.1, set the initial value of the Lagrangian factor μ1 = eps, set the convergence error err, and set the iteration counter i = 1;
[0029] Step 6.2: Newton iteration updates the Lagrange factor. The process is: make c n is the nth element of c;
[0030] Step 6.3, judge If yes, repeat step 6.2 until the solution is Optimal Lagrangian factor μ opt =u i+1 , otherwise go to step 6.4;
[0031] Step 6.4, set the Lagrangian factor μ i =(μ max +μ min ) / 2, and set the upper limit of the Lagrange factor and the lower limit μ min =0;
[0032] Step 6.5, judge If so, we get the optimal Lagrangian factor μ opt =μ i , otherwise go to step 6.6;
[0033] Step 6.6, judge If so, then the upper limit μ max Update to μ i , otherwise the lower limit μ min Update to μ i ;
[0034] Step 6.7, the iteration counter is incremented by 1 and the Lagrangian factor μ is updated i =(μ max +μ min ) / 2 and return to step 6.5.
[0035] Preferably, in step 7, the optimal Lagrangian factor is substituted into the formula Get the estimated dimension reduction guidance vector
[0036] Substitute the estimated dimension reduction guidance vector into the formula Get the estimated weight vector W RDRCB .
[0037] Compared with the prior art, the present invention has the following beneficial effects:
[0038] The present invention aims at the problem that a data dimension reduction robust adaptive beamformer easily obtains an invalid loading factor by using the Newton iteration method. It is proposed to add a dichotomy method and an error detection mechanism to optimize the Newton iteration method, which can ensure the calculation efficiency while improving the stability of the iteration process and solve the problem that the invalid loading factor affects the broadband detection performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 The present invention is a flowchart of an improved robust adaptive beamforming method for data dimensionality reduction.
[0040] Figure 2 The figure is a flow chart of the Lagrangian factor solving process of the present invention.
[0041] Figure 3 Non-convergence curve for solving the Lagrangian factor for the Newton iteration method.
[0042] Figure 4 Convergence curve of the improved Newton iteration method for solving the Lagrange factor.
[0043] Figure 5 Output signal-to-noise ratio gain curves (zero error) for various algorithms.
[0044] Figure 6 Output signal-to-noise ratio gain curves of various algorithms (with errors).
[0045] Figure 7 This is a diagram of the azimuth process of MVDR beamformer sea trial data processing.
[0046] Figure 8 Azimuth history diagram for improving RDRCB beamformer sea trial data processing.
[0047] Fig. 9 It is the azimuth spectrum at the 10th second of the sea trial data.
[0048] Fig.10 It is the azimuth spectrum at the 30th second of the sea trial data. DETAILED DESCRIPTION
[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.
[0050] An improved data dimension reduction robust adaptive beamforming algorithm includes the following 7 steps.
[0051] Step 1: Calculate the covariance matrix of each sub-band sample
[0052] Step 2: Use the conjugate gradient algorithm to calculate the covariance matrix Solve the dimension reduction matrix D = [d1, ..., d N ];
[0053] Step 3: Use the dimension reduction matrix to calculate the diagonal matrix Λ CG ;
[0054] Step 4: Solve the estimated expected signal steering vector after dimensionality reduction and constraint matrix M;
[0055] Step 5, perform eigenvalue decomposition on the constraint matrix M;
[0056] Step 6, solving the Lagrangian factor;
[0057] Step 7: Obtain an estimated dimension reduction steering vector based on the optimal Lagrangian factor solution.
[0058] In step 1 of the present invention, the covariance matrix Calculated by formula (1):
[0059]
[0060] In formula (1), K is the number of snapshots, x k is the array snapshot data received by the kth array element, and H represents transposition.
[0061] In the present invention, a covariance matrix is calculated corresponding to each sub-band sample, and a covariance matrix is calculated for each frequency point.
[0062] Step 2 of the present invention specifically includes the following 4 sub-steps.
[0063] Step 2.1, initialize the iteration counter j = 1, initialize the residual vector in, To estimate the desired signal steering vector.
[0064] Step 2.2, calculate the learning rate α for the jth iteration j and step length β j , calculated by formula (2) and formula (3) respectively:
[0065]
[0066]
[0067] Where, d j is the residual of the jth iteration, r jis the vector of the jth iteration, H is the transpose, is the covariance matrix.
[0068] Step 2.3, based on the learning rate α of the jth iteration j and step length β j Calculate the vector r of the j+1th iteration j+1 and residual amount d j+1 , calculated by equation (4) and equation (5) respectively.
[0069]
[0070] d j+1 =-r j+1 +β j d j (5)
[0071] Step 2.4, add 1 to the iteration counter j, and return to step 2.2 until j=N-1, where N is the dimensionality reduction order.
[0072] In step 3, the diagonal matrix is calculated using the reduced dimension matrix Among them, N is the dimensionality reduction order, H is the transpose, is the covariance matrix, and the diag function is used to extract the diagonal elements of the matrix and create a diagonal matrix.
[0073] In step 4, the estimated expected signal steering vector after dimensionality reduction is and constraint matrix M are expressed as equation (6) and equation (7) respectively:
[0074]
[0075]
[0076] In the formula, To estimate the desired signal steering vector, E = ε -1 I N , ε is the error between the preset steering vectors, I N represents the identity matrix, Λ CG is a diagonal matrix, D represents the dimension reduction matrix, and H represents the transpose.
[0077] Step 5: Perform eigenvalue decomposition on the constraint matrix M after dimensionality reduction according to formula (8):
[0078]
[0079] In the formula, U=[u1,...,u n , ..., u N ], U represents the eigenvector matrix of the constraint matrix M after dimensionality reduction, u is the eigenvector, n is the matrix index, λn for u n The corresponding eigenvalues.
[0080] Figure 2 The flowchart of the Lagrangian factor solving process of the present invention is shown, and the Lagrangian factor divergence is avoided by using the fusion error correction mechanism and the dichotomy method.
[0081] Combination Figure 2 The Lagrangian factor solution of step 6 of the present invention specifically includes the following 6 sub-steps.
[0082] Step 6.1, set the initial value of the Lagrangian factor μ1=eps, the convergence error err, and set the iteration counter j=1.
[0083] Step 6.2: Newton iteration updates the Lagrangian factor. The process is shown in equation (9):
[0084]
[0085] In the formula, c n is the nth element of c.
[0086] Step 6.3, judge If yes, repeat step 6.2 until the solution is Optimal Lagrangian factor μ opt =u i+1 , otherwise proceed to step 6.4.
[0087] Step 6.4, set the Lagrangian factor μ i =(μ max +μ min ) / 2, and set the upper limit of the Lagrange factor and the lower limit μ min =0.
[0088] Step 6.5, judge If so, we get the optimal Lagrangian factor μ opt =μ i , otherwise proceed to step 6.6.
[0089] Step 6.6, judge If so, then the upper limit μ max Update to μ i , otherwise the lower limit μ min Update to μ i .
[0090] Step 6.7, the iteration counter is incremented by 1 and the Lagrangian factor μ is updated i =(μ max +μ min) / 2 and return to step 6.5.
[0091] Step 7 specifically includes the following two sub-steps.
[0092] Step 7.1, substitute the optimal Lagrangian factor into equation (10) to obtain the estimated dimension reduction guidance vector
[0093]
[0094] In the formula, Λ=diag(λ1, λ2,..., λ N ).
[0095] Step 7.2: Substitute the estimated dimension reduction guidance vector into equation (11) to obtain the estimated weighted vector W RDRCB .
[0096]
[0097] Figure 3 The non-convergence curve of Newton iteration method for solving Lagrange factor, Figure 4 This is the convergence curve of the improved Newton iteration method for solving the Lagrange factor. It can be found that the Newton method is not a descent algorithm. When the Hessian matrix is not positive and the initial point is not selected properly, the direction of movement of the direction vector is not the descent direction, and the obtained load is not a converged value. The improved Newton iteration method can correct the wrong direction in time to avoid obtaining an invalid value.
[0098] Figure 5 and Figure 6 The output signal-to-noise ratio gain curves of various algorithms are shown in Figure 2. The signal-to-noise ratio gain curves are tested using MVDR, RCB, and improved RDRCB algorithm beamforming algorithms. The expected input signal-to-noise ratio varies in the range of -20dB to 40dB, and the two plane wave interference signals are 30dB and 35dB respectively. The expected signal, interference signal, and noise are Gaussian random processes that are uncorrelated with each other. Figure 5 is the signal-to-noise ratio gain curve when the error between the real steering vector and the imaginary steering vector is 0, Figure 6 The signal-to-noise ratio gain curve when the error between the real steering vector and the imaginary steering vector is 0.5° can be seen that the improved RDRCB algorithm improves the robustness of the robust adaptive beamforming algorithm.
[0099] Figure 7 and Figure 8 These are the azimuth process diagrams of the sea trial data processing of MVDR and improved RDRCB beamformer. Fig. 9 and Fig.10The azimuth spectra at the 10th and 100th seconds of the sea trial data are shown in Table 1. From the continuity of the target detection trajectory and the snapshot results, it can be seen that the detection performance of the present invention is better than that of the MVDR algorithm.
[0100] The invention is described above in conjunction with the best example, but the present invention is not limited to the examples disclosed above, but should cover various modifications and equivalent combinations based on the essence of the present invention.
Claims
1. An improved data dimension reduction robust adaptive beamforming algorithm, characterized in that: The algorithm comprises the following steps: Step 1: Calculate the covariance matrix of each sub-band sample Step 2: Use the conjugate gradient algorithm to calculate the covariance matrix Solve the dimension reduction matrix D = [d1, ..., d N ], including the following steps: Step 2.1, initialize the iteration counter j = 1, initialize the residual vector in, To estimate the desired signal steering vector; Step 2.2, calculate the learning rate α for the jth iteration j and step length β j , learning rate Step Length Among them, d j is the residual of the jth iteration, r j is the vector of the jth iteration, H is the transpose, is the covariance matrix of each sub-band sample; Step 2.3, based on the learning rate α of the jth iteration j and step length β j Calculate the vector r of the j+1th iteration j+1 and residual amount d j+1 ,vector residual amount d j+1 =-r j+1 +β j d j ; Step 2.4, add 1 to the iteration counter j, and return to step 2.2 until j = N-1, where N is the dimensionality reduction order; Step 3: Use the reduced dimension matrix to calculate the diagonal matrix Among them, N is the dimensionality reduction order, H is the transpose, is the covariance matrix, and the diag function is used to extract the diagonal elements of the matrix and create a diagonal matrix; Step 4: Solve the estimated expected signal steering vector after dimensionality reduction and constraint matrix M, respectively, are expressed as in, To estimate the desired signal steering vector, E = ε -1 I N , ε is the error between the preset steering vectors, I N represents the identity matrix, Λ CG is a diagonal matrix, D represents the dimension reduction matrix, and H represents the transpose; Step 5: Perform eigenvalue decomposition on the constraint matrix M: U=[u1,...,u n , ..., u N ], U represents the eigenvector matrix of the constraint matrix M after dimensionality reduction, u is the eigenvector, n is the matrix index, λ n for u n The corresponding eigenvalues; Step 6, solving the Lagrangian factor, includes the following steps: Step 6.1, set the initial value of the Lagrangian factor μ1 = eps, set the convergence error err, and set the iteration counter i = 1; Step 6.2: Newton iteration updates the Lagrange factor. The process is: make c n is the nth element of c; Step 6.3, judge If yes, repeat step 6.2 until the solution is Then the optimal Lagrangian factor μ opt =u i+1 , otherwise go to step 6.4; Step 6.4, set the Lagrangian factor μ i =(μ max +μ min ) / 2, and set the upper limit of the Lagrange factor and the lower limit μ min =0; Step 6.5, judge If so, we get the optimal Lagrangian factor μ opt =μ i , otherwise go to step 6.6; Step 6.6, judge If so, then the upper limit μ max Update to μ i , otherwise the lower limit μ min Update to μ i ; Step 6.7, the iteration counter is incremented by 1 and the Lagrangian factor μ is updated i =(μ max +μ min ) / 2 and return to step 6.5; Step 7: Substitute the optimal Lagrangian factor into the formula Get the estimated dimension reduction guidance vector Substitute the estimated dimension reduction guidance vector into the formula Get the estimated weight vector W RDRCB .
2. The improved data dimension reduction robust adaptive beamforming algorithm according to claim 1, characterized in that: The covariance matrix Where K is the number of snapshots, x k is the array snapshot data received by the kth array element, and H represents transposition.
