Dsp direction finding method and system based on power iteration and multi-core parallel processing
Patent Information
- Application Number
- CN202611042756.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-14
- Publication Date
- 2026-10-09
AI Technical Summary
[0007]本发明目的是为了解决现有DSP方向估计方法中完整特征分解计算量大、信号子空间提取耗时长、多核共享存储效率低以及谱峰搜索阶段访存与运算负担较重的问题，本发明提供一种基于幂迭代与多核并行处理的DSP测向方法及系统
[0058]降低子空间提取阶段的计算复杂度与运算量：本发明针对现有DSP测向方法中特征分解计算量大的问题进行结构性重构，采用幂迭代方法直接提取测向所需的主特征子空间，而不对协方差矩阵的全部特征值和全部特征向量进行完整求解，避免对全部特征值与全部特征向量进行完整求解，从而显著减少信号子空间提取阶段的冗余矩阵运算量，适于DSP平台上的实时测向处理。
Smart Images
Figure CN122883087A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of array signal processing, embedded real-time computing and intelligent sensing technology. Background Technology
[0002] In applications such as intelligent sensing, array signal processing, wireless information processing, and complex environment information analysis, direction estimation systems need to complete the estimation of incoming signal direction parameters within a short time to provide timely and reliable direction information for subsequent signal analysis, target perception, state recognition, and spatial information extraction. For two-dimensional MUSIC direction estimation of vector arrays, although commonly used methods can theoretically achieve high angular resolution, their engineering implementation is often limited by the following factors.
[0003] First, traditional DSP direction finding typically involves constructing an equivalent real symmetric matrix from the complex covariance matrix, then using Householder and Givens transforms to obtain all eigenvalues and eigenvectors. Afterward, eigenvalue sorting, noise subspace reconstruction, and spectral function calculation are required, resulting in a large computational load. While this process is algorithmically complete, it necessitates performing numerous intermediate calculations in real-time systems that do not directly contribute to the actual direction finding results. This is especially problematic when only a small number of principal eigenvectors are needed; solving for the entire eigenstructure still leads to significant computational redundancy.
[0004] Secondly, most existing DSP implementations focus their optimizations on the peak search stage, while retaining the covariance matrix decomposition stage as a complete eigenvalue decomposition process. This results in the main time overhead of the system still being concentrated in the signal subspace extraction stage. The signal subspace extraction is time-consuming and makes it difficult to fully utilize the parallel processing advantages of multi-core DSPs for complex matrix operations and direction search tasks. The memory access and computation burden in the peak search stage is also heavy.
[0005] Furthermore, traditional spectral function calculations often require explicitly constructing the noise subspace matrix or noise projection matrix, followed by complex matrix multiplication, determinant inversion, or square root operations. For embedded DSP platforms, these operations not only incur significant instruction overhead but also introduce additional memory relocation and numerical computation burdens.
[0006] Therefore, a DSP direction finding method is proposed that can avoid complete feature decomposition, directly extract the principal feature subspace required for direction estimation, be suitable for parallel implementation in multi-core DSPs, and enable fast spectral peak determination, thus having clear engineering application value. Summary of the Invention
[0007] The purpose of this invention is to solve the problems of large computational load of complete feature decomposition, long time consumption of signal subspace extraction, low efficiency of multi-core shared storage, and heavy memory access and operation burden in the spectrum peak search stage in existing DSP direction estimation methods. This invention provides a DSP direction finding method and system based on power iteration and multi-core parallel processing.
[0008] A DSP direction finding method based on power iteration and multi-core parallel processing is implemented using multiple parallel DSP cores, where one DSP core serves as the master DSP core and the remaining DSP cores are slave DSP cores. The method includes:
[0009] The preset azimuth search interval is divided into multiple non-overlapping azimuth sub-intervals that all cover the preset azimuth search interval, and each DSP core uses the same preset pitch search interval in the corresponding azimuth sub-interval.
[0010] For each DSP core, the azimuth sub-interval and the preset pitch angle search interval are discretized and paired with equal intervals of corresponding step size to generate all angle grid points of the azimuth sub-interval corresponding to the DSP core. Each angle grid point consists of an azimuth angle and a pitch angle.
[0011] The main DSP core constructs an iterative enhancement matrix based on the complex sampled signals acquired by all array elements. It then uses a power-law iterative method with deterministic initial vectors and column-by-column orthogonal projection elimination to extract the principal eigenvectors of the iterative enhancement matrix, thus constructing the signal subspace matrix. The principal feature vectors corresponding to each column, and Stored in the shared memory area and broadcast to all DSP cores;
[0012] Each DSP core is based on a signal subspace matrix. The spectral peak search is performed on the steering matrix and self-inner product matrix of each angle grid point corresponding to the constructed DSP core. The local spectral value of each angle grid point corresponding to the DSP core is calculated, and the maximum value of the local spectral value is taken as the local peak value of the corresponding azimuth sub-interval of the DSP core.
[0013] The main DSP core compares all local peak values and selects the azimuth and elevation angles corresponding to the maximum local peak value as the final azimuth and elevation angles of the target signal.
[0014] Preferably, the main DSP core constructs the iterative enhancement matrix based on the complex sampled signals acquired by all array elements in the array in the following manner:
[0015] Construct the Hermitian covariance matrix based on the complex sampled signals acquired by all array elements. ;
[0016] Using Hermitian covariance matrix Constructing an iterative enhancement matrix , For conjugate.
[0017] Preferably, a power-law iterative method using deterministic initial vectors and column-by-column orthogonal projection elimination is employed to extract the principal eigenvectors of the iterative enhancement matrix, thereby constructing the signal subspace matrix. The first one to be obtained in the middle The process of creating column principal eigenvectors includes:
[0018] S11. Initialize the current column index and deterministic initialization vector:
[0019] Let the principal eigenvector to be obtained be the signal subspace matrix. The The column is randomly initialized to generate the current column to be calculated. The deterministic original initial vector corresponding to the column principal eigenvector ; , For the signal subspace matrix The total number of columns, The total number of array elements, , For the signal subspace matrix Middle Column principal eigenvectors ;
[0020] S12. Perform column-by-column orthogonal projection culling on the deterministic original initial vector:
[0021] when At that time, let the deterministic initial vector after projection removal be... ;
[0022] when At that time, In the already obtained 1 to 1 Column principal eigenvectors to Projective culling is performed on Zhang Cheng's subspace to obtain a deterministic initial vector after projection culling. ; Indicates that it has been written into the signal subspace matrix The The column principal eigenvector is also used as the obtained first eigenvector. Column principal eigenvectors Indicates conjugate;
[0023] S13. Perform validity assessment and normalization on the initial vector after projection removal:
[0024] like ,make ,and Execute step S12;
[0025] For the first The standard basis vectors corresponding to the column, and the innermost one of them. Each component The value is 1, and all other components are 0. for Inner One portion, The minimum norm threshold is preset;
[0026] like Then for Normalization is performed to obtain the normalized deterministic initial vector. , Also as the first The initial normalized deterministic initial vector is obtained after the nth iteration, and step S14 is executed;
[0027] S14, No. In the next iteration, perform a first-order iterative matrix multiplication:
[0028] according to Calculate the intermediate vector Execute step S15; where, For the iterative enhancement matrix, The initial value is 0;
[0029] S15, regarding the intermediate vector Perform column-by-column orthogonal projection culling:
[0030] when At that time, let the intermediate vector after projection removal be... Proceed to step S16;
[0031] when At that time, let the intermediate vector after shadow removal Proceed to step S16;
[0032] S16. The intermediate vector after projection removal Perform normalization update:
[0033] right Perform normalization, and use the normalized result as the first... Normalized deterministic initial vector after iteration number n Proceed to step S17;
[0034] S17. Calculate the consistency parameter between two adjacent iteration vectors:
[0035] according to and Calculate the consistency parameter Proceed to step S18;
[0036] S18. Perform convergence judgment:
[0037] like Then determine the current number to be obtained. When the column principal eigenvectors converge, Write signal subspace matrix The Column principal eigenvectors In the middle, making ; The preset convergence threshold is used;
[0038] like Then let Then return to step S14.
[0039] Preferably, the implementation method for constructing the guiding matrix of each angle grid point corresponding to each DSP core is as follows:
[0040] ;
[0041] ;
[0042] ;
[0043] in, This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core. This represents any pitch angle to be searched within the preset pitch angle search range. Represents the angle grid points corresponding to the DSP core. The corresponding guidance matrix, Indicates the total number of array elements. Indicates the first Each element corresponds to an angle grid point. The guiding phase response, , Represents the imaginary unit. Indicates the wavelength of the received signal. Indicates the first The position vectors of each array element relative to the array reference point They represent the first The horizontal and vertical coordinates of each array element in the horizontal plane Represents angle grid points The corresponding unit direction vector, This is a transpose.
[0044] Preferably, the self-inner product matrix of each angle grid point corresponding to each DSP core is constructed. ;
[0045] in, The angle grid point corresponding to the current DSP core The guiding matrix, This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core. This represents any pitch angle to be searched within the preset pitch angle search range. For conjugate.
[0046] Preferably, the method for calculating the local spectral values of each angle grid point corresponding to each DSP core is as follows:
[0047] Based on the steering matrix, inner product matrix, and signal subspace matrix of the angle grid points corresponding to the current DSP core Construct the grid points at the current angle. Complex decision matrix ; This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core. This represents any pitch angle to be searched within the preset pitch angle search range;
[0048] Based on the complex decision matrix real part matrix and the imaginary part of the matrix Calculate the current angle grid point Local spectral values .
[0049] Preferably, ;
[0050] in, and These represent the angle grid points corresponding to the current DSP core. The guiding matrix and the self-inner product matrix Indicates conjugate.
[0051] Preferably, .
[0052] Preferably, the method for generating all angle grid points for the azimuth sub-interval and preset pitch angle search interval corresponding to each DSP core by performing equally spaced discretization sampling and pairing with corresponding step sizes is as follows:
[0053] According to the preset azimuth step size, the azimuth sub-intervals responsible for the current DSP core are discretized at equal intervals to obtain a one-dimensional azimuth sequence;
[0054] According to the preset pitch step size, the preset pitch angle search interval corresponding to the current DSP core is discretized at equal intervals to obtain a one-dimensional pitch sequence;
[0055] Each azimuth angle in the one-dimensional azimuth sequence is paired with each pitch angle in the one-dimensional pitch sequence to generate a two-dimensional angle grid point set corresponding to the current DSP core.
[0056] A DSP direction finding system based on power iteration and multi-core parallel processing includes a storage device, a processor, and a computer program stored in the storage device and executable on the processor. The processor executes the computer program to implement the DSP direction finding method based on power iteration and multi-core parallel processing as described above.
[0057] The beneficial effects of this invention are:
[0058] Reducing the computational complexity and workload of the subspace extraction stage: This invention addresses the problem of high computational cost in eigenvalue decomposition in existing DSP direction finding methods by structurally reconstructing the main feature subspace required for direction finding. It adopts a power iteration method to directly extract the main feature subspace required for direction finding, without fully solving for all eigenvalues and eigenvectors of the covariance matrix. This avoids the need for a complete solution for all eigenvalues and eigenvectors, thereby significantly reducing the redundant matrix computation in the signal subspace extraction stage, making it suitable for real-time direction finding processing on DSP platforms.
[0059] Improving the convergence speed of power iteration and reducing the influence of random initial values: This invention constructs an enhanced iteration matrix based on the Hermitian covariance matrix, which further amplifies the advantage of the principal eigenvalues over the non-principal eigenvalues, thereby accelerating the convergence speed of the principal eigenvectors during the power iteration process. At the same time, deterministic initial vectors are used instead of random initial vectors, and combined with column-by-column orthogonal projection elimination and normalized convergence criteria, the possibility of the iteration results being affected by random initial value disturbances is reduced, thereby improving the stability, reproducibility and engineering consistency of the principal feature subspace extraction process.
[0060] This invention constructs an enhanced iteration matrix, which further amplifies the advantage of the principal eigenvalue, improves the convergence speed of the power iteration, and enhances the feasibility of the algorithm in real-time engineering.
[0061] Improving the collaborative efficiency of multi-core DSP parallel direction finding: This invention utilizes a shared memory area and a master-slave core event notification mechanism to enable the master DSP core to broadcast the completed signal subspace matrix to all slave DSP cores. This shared memory area and master-slave core event notification mechanism improve data consistency and task collaboration efficiency during multi-core DSP parallel direction finding. Each DSP core independently performs local spectral peak search within its assigned azimuth sub-interval and only writes back its own local optimal results, thereby reducing redundant calculations and invalid data exchange between cores and improving data consistency and task collaboration efficiency during multi-core DSP parallel direction finding.
[0062] Reducing the computational burden of local spectral values: This invention improves the local spectral value calculation process in the peak search process. First, a complex decision matrix for the current angle grid point is constructed using the steering matrix, the self-inner product matrix, and the signal subspace matrix. Then, the real and imaginary parts are obtained from this complex decision matrix, and the negative of the sum of their squares is used as the local spectral value for peak comparison. This method transforms the denominator-type spectral function, which originally required minimum value judgment, into a local spectral value form that can directly perform maximum value comparison. This reduces division, square root, and additional reciprocal operations, lowering the computational burden of peaks on the DSP platform and improving real-time search efficiency.
[0063] This invention employs a pre-storage method of guiding moments and self-inner product matrices at grid points at various angles, reducing the repetitive trigonometric function calculations and matrix construction processes during the spectral peak search stage, thereby lowering memory access and computational pressure. Attached Figure Description
[0064] Figure 1 This is a schematic diagram illustrating the principle of the DSP direction finding method based on power iteration and multi-core parallel processing described in this invention.
[0065] Figure 2 This is a flowchart of multi-DSP core orientation partition search. Detailed Implementation
[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0067] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0068] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.
[0069] Specific Implementation Method 1: Combination Figure 1 and Figure 2 This embodiment describes a DSP direction finding method based on power iteration and multi-core parallel processing. It is implemented using multiple parallel DSP cores, where one DSP core serves as the master core and the remaining cores are slave cores. The method includes:
[0070] S1. Divide the preset azimuth search interval into multiple non-overlapping azimuth sub-intervals that all cover the preset azimuth search interval, and each DSP core uses the same preset pitch search interval in its corresponding azimuth sub-interval.
[0071] For each DSP core, the azimuth sub-interval and the preset pitch angle search interval are discretized and paired with equal intervals of corresponding step size to generate all angle grid points of the azimuth sub-interval corresponding to the DSP core. Each angle grid point consists of an azimuth angle and a pitch angle.
[0072] S2. The main DSP core constructs an iterative enhancement matrix based on the complex sampled signals acquired by all array elements. It then uses a power-law iterative method with deterministic initial vectors and column-by-column orthogonal projection elimination to extract the principal eigenvectors of the iterative enhancement matrix, constructing a signal subspace matrix. The principal feature vectors corresponding to each column, and Stored in the shared memory area and broadcast to all DSP cores;
[0073] S3, each DSP core is based on the signal subspace matrix. The spectral peak search is performed on the steering matrix and self-inner product matrix of each angle grid point corresponding to the constructed DSP core. The local spectral value of each angle grid point corresponding to the DSP core is calculated, and the maximum value of the local spectral value is taken as the local peak value of the corresponding azimuth sub-interval of the DSP core.
[0074] S4. The main DSP core compares all local peak values and selects the azimuth and elevation angles corresponding to the maximum local peak value as the final azimuth and elevation angles of the target signal.
[0075] This invention avoids complete feature decomposition and reduces computational load. Specifically, it uses a power iteration method to extract the main feature vector, replacing the traditional complete feature decomposition operation, which significantly reduces the computational complexity and amount of computation required for matrix decomposition.
[0076] By combining deterministic initial vectors with column-by-column orthogonal projection, the construction of the signal subspace is accelerated. The iteration is initiated using deterministic initial vectors, and the projections in the obtained directions are eliminated column by column when calculating the principal eigenvectors of each column, which ensures convergence stability and reduces the number of iterations, thereby shortening the extraction time of the signal subspace.
[0077] The main core constructs the signal subspace once and shares it via broadcast, improving the storage efficiency of multi-core systems. Only the main DSP core completes the construction of the signal subspace and stores it in the shared area. Then, it notifies all slave cores via broadcast, avoiding repeated calculations or frequent access to shared storage by each slave core. This reduces storage access conflicts and overhead during multi-core collaboration.
[0078] Distributed local spectral peak search reduces memory access and computational burden. The azimuth search interval is divided into non-overlapping sub-intervals. Each DSP core only calculates local spectral values within its own sub-interval and only needs to return the local peak value to the main core. This reduces the search range and computational load of a single core, while also reducing the access pressure on shared memory and the overall computational burden during the spectral peak search stage.
[0079] Furthermore, the main DSP core constructs the iterative enhancement matrix based on the complex sampled signals acquired by all array elements in the array in the following manner:
[0080] Construct the Hermitian covariance matrix based on the complex sampled signals acquired by all array elements. ;
[0081] Using Hermitian covariance matrix Constructing an iterative enhancement matrix , For conjugate.
[0082] In this preferred approach, a Hermitian covariance matrix is first constructed from the array's complex sampled signals, and then this covariance matrix and its conjugate are used to construct an iterative enhancement matrix. This implementation provides a structurally regular and mathematically sound input matrix for subsequent power iterations, enabling the principal eigenvector extraction process to proceed stably and consistently. It avoids unnecessary computational biases that might be introduced by directly processing the original sampled data and ensures that the signal subspace matrices used by each DSP core in subsequent spectral peak searches have a unified numerical basis, which is beneficial for the reliable execution of the entire parallel direction finding process across multiple DSP cores.
[0083] Furthermore, a power-law iterative method using deterministic initial vectors and column-by-column orthogonal projection elimination is employed to extract the principal eigenvectors of the iterative enhancement matrix, thereby constructing the signal subspace matrix. The first one to be obtained in the middle The process of creating column principal eigenvectors includes:
[0084] S11. Initialize the current column index and deterministic initialization vector:
[0085] Let the principal eigenvector to be obtained be the signal subspace matrix. The The column is randomly initialized to generate the current column to be calculated. The deterministic original initial vector corresponding to the column principal eigenvector ; , For the signal subspace matrix The total number of columns, The total number of array elements, , For the signal subspace matrix Middle Column principal eigenvectors ;
[0086] S12. Perform column-by-column orthogonal projection culling on the deterministic original initial vector:
[0087] when At that time, let the deterministic initial vector after projection removal be... ;
[0088] when At that time, In the already obtained 1 to 1 Column principal eigenvectors to Projective culling is performed on Zhang Cheng's subspace to obtain a deterministic initial vector after projection culling. , Indicates that it has been written into the signal subspace matrix The The column principal eigenvector is also used as the obtained first eigenvector. Column principal eigenvectors Indicates conjugate; Represents the deterministic original initial vector In the The complex projection coefficients along the direction of the principal eigenvector have been obtained;
[0089] S13. Perform validity assessment and normalization on the initial vector after projection removal:
[0090] like ,make ,and Execute step S12;
[0091] For the first The standard basis vectors corresponding to the column, and the innermost one of them. Each component The value is 1, and all other components are 0. for Inner One portion, The minimum norm threshold is preset;
[0092] By using the standard basis vector backoff mechanism described above, we can avoid the deterministic original initial vector from degenerating into a zero or near-zero vector after removing the obtained principal eigenvector components, thereby ensuring that the subsequent power iteration process has an effective non-zero initial direction.
[0093] like Then for Normalization is performed to obtain the normalized deterministic initial vector. , Also as the first The initial normalized deterministic initial vector is obtained after the nth iteration, and step S14 is executed;
[0094] S14, No. In the next iteration, perform a first-order iterative matrix multiplication:
[0095] according to Calculate the intermediate vector Execute step S15; where, For the iterative enhancement matrix, The initial value is 0;
[0096] S15, regarding the intermediate vector Perform column-by-column orthogonal projection culling:
[0097] when At that time, let the intermediate vector after projection removal be... Proceed to step S16;
[0098] when At that time, let the intermediate vector after shadow removal Proceed to step S16;
[0099] In this step S15, the components of the intermediate vector in the direction of the obtained principal eigenvector are removed in order to maintain the orthogonality between the current principal eigenvector to be determined and the obtained principal eigenvector.
[0100] S16. The intermediate vector after projection removal Perform normalization update:
[0101] right Perform normalization, and use the normalized result as the first... Normalized deterministic initial vector after iteration number n Proceed to step S17;
[0102] S17. Calculate the consistency parameter between two adjacent iteration vectors:
[0103] according to and Calculate the consistency parameter Proceed to step S18;
[0104] S18. Perform convergence judgment:
[0105] like Then determine the current number to be obtained. When the column principal eigenvectors converge, Write signal subspace matrix The Column principal eigenvectors In the middle, making ; The preset convergence threshold is used;
[0106] like Then let Then return to step S14.
[0107] This preferred embodiment provides the specific steps for extracting the principal eigenvectors of each column of the signal subspace matrix using a power-law iterative method employing a deterministic initial vector and column-by-column orthogonal projection elimination. By setting a deterministic initial vector (i.e., a non-random initial vector), inconsistencies in results across different iterations due to random initialization are avoided, as well as the problem of iteration direction deviation or even non-convergence caused by a random initialization approaching a non-principal eigenvector. The deterministic starting point makes the iteration process mathematically deterministic, eliminating the need for multiple attempts or additional verification steps to determine the reliability of the results, as is required with random initialization. In each iteration and before and after, column-by-column orthogonal projection elimination is performed on the current vector, orthogonalizing it with the already calculated principal eigenvectors of each column. This ensures that each newly extracted column represents a previously uncovered directional component, preventing duplication or excessive correlation between different columns. Simultaneously, normalization is performed in each iteration, and a consistency parameter between adjacent iteration vectors is calculated. Whether this parameter is less than a preset threshold is used as a convergence criterion, thereby accurately controlling the iteration termination time, avoiding wasting computational resources on redundant loops and preventing loss of accuracy due to premature exit. The entire process also includes judging the validity of the initial vector after projection removal. When the vector norm is too small, it is replaced with standard basis vectors and reprojected to avoid subsequent calculation failures due to numerical degradation. Finally, through this orderly set of iteration, projection, normalization and convergence judgment mechanisms, the principal eigenvectors of each column of the signal subspace matrix can be stably and efficiently obtained one by one, providing reliable signal subspace data for subsequent spectral peak search.
[0108] Furthermore, the self-inner product matrix of each angle grid point corresponding to each constructed DSP core is... ;
[0109] in, The angle grid point corresponding to the current DSP core The guiding matrix, This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core. This represents any pitch angle to be searched within the preset pitch angle search range. For conjugate.
[0110] In this preferred embodiment, the direction finding matrix is obtained by multiplying the guidance matrix of the current angle grid point with its conjugate transpose. This self-inner product matrix is a square matrix whose elements reflect the inner product relationships between the columns of the guidance matrix. After pre-calculating and storing this matrix, it can be directly reused when calculating the local spectral values of each angle grid point in the subsequent calculation. This avoids repeatedly performing the multiplication operation of the guidance matrix with its conjugate transpose at each grid point, thereby reducing the amount of redundant calculation, reducing the computational burden in the spectral peak search stage, and improving the overall direction finding efficiency.
[0111] Furthermore, the implementation method for constructing the guiding matrix of each angle grid point corresponding to each DSP core is as follows:
[0112] ;
[0113] ;
[0114] ;
[0115] in, This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core, and the azimuth angle to be searched is taken as the azimuth angle of the incoming wave signal. This represents any pitch angle to be searched within the preset pitch angle search range, and the pitch angle to be searched is taken as the pitch angle of the incoming signal. Represents the angle grid points corresponding to the DSP core. The corresponding guidance matrix, Indicates the total number of array elements. Indicates the first Each element corresponds to an angle grid point. The guiding phase response, , Represents the imaginary unit. Indicates the wavelength of the received signal. Indicates the first The position vectors of each array element relative to the array reference point They represent the first The horizontal and vertical coordinates of each array element in the horizontal plane Represents angle grid points The corresponding unit direction vector, This is a transpose.
[0116] In this preferred embodiment, a formula for calculating the steering matrix is provided. This formula calculates the propagation phase difference of each element relative to the array reference point in the candidate direction based on the element position vector and the unit direction vector of the incoming wave, thereby forming the steering matrix. Each row of the matrix corresponds to the phase response of an element in that direction. This construction method clarifies the physical meaning and numerical generation path of the steering matrix, enabling different DSP cores to independently generate the required steering matrix according to unified and determined calculation rules within their respective azimuth sub-intervals, without relying on data or intermediate results from other cores. This facilitates independent operation and load balancing of each core in parallel computing and provides standardized input data for subsequent calculation of local spectral values.
[0117] Furthermore, the method for calculating the local spectral values of each angle grid point corresponding to each DSP core is as follows:
[0118] Based on the steering matrix, inner product matrix, and signal subspace matrix of the angle grid points corresponding to the current DSP core Construct the grid points at the current angle. Complex decision matrix ;
[0119] This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core. This represents any pitch angle to be searched within the preset pitch angle search range. and These represent the angle grid points corresponding to the current DSP core. The guiding matrix and the self-inner product matrix Indicates conjugate;
[0120] Based on the complex decision matrix real part matrix and the imaginary part of the matrix Calculate the current angle grid point Local spectral values .
[0121] Furthermore, for each DSP core's corresponding azimuth sub-interval and preset pitch angle search interval, equal-interval discretization sampling and pairing with corresponding step sizes are performed to generate all angle grid points for the azimuth sub-interval corresponding to that DSP core. The implementation method is as follows:
[0122] According to the preset azimuth step size, the azimuth sub-intervals responsible for the current DSP core are discretized at equal intervals to obtain a one-dimensional azimuth sequence;
[0123] According to the preset pitch step size, the preset pitch angle search interval corresponding to the current DSP core is discretized at equal intervals to obtain a one-dimensional pitch sequence;
[0124] Each azimuth angle in the one-dimensional azimuth sequence is paired with each pitch angle in the one-dimensional pitch sequence to generate a two-dimensional angle grid point set corresponding to the current DSP core.
[0125] The above method decomposes the two-dimensional joint search into two independent one-dimensional sampling steps. The logic is clear and the implementation is simple. Each DSP core can independently generate grid points within its own range according to the same step size rule without interacting or coordinating with other cores. This is beneficial for the independent operation and task division of each core under the parallel architecture, and also provides a regular and definite angle sampling basis for subsequent spectral peak search.
[0126] Specific Implementation Method 2: A DSP direction finding system based on power iteration and multi-core parallel processing, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor. The processor executes the computer program to implement the DSP direction finding method based on power iteration and multi-core parallel processing as described in Specific Implementation Method 1.
[0127] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.
Claims
1. A DSP direction finding method based on power iteration and multi-core parallel processing, characterized in that, This method is implemented using multiple parallel DSP cores, where one DSP core serves as the master DSP core and the remaining DSP cores are slave DSP cores. The method includes: The preset azimuth search interval is divided into multiple non-overlapping azimuth sub-intervals that all cover the preset azimuth search interval, and each DSP core uses the same preset pitch search interval in the corresponding azimuth sub-interval. For each DSP core, the azimuth sub-interval and the preset pitch angle search interval are discretized and paired with equal intervals of corresponding step size to generate all angle grid points of the azimuth sub-interval corresponding to the DSP core. Each angle grid point consists of an azimuth angle and a pitch angle. The main DSP core constructs an iterative enhancement matrix based on the complex sampled signals acquired by all array elements. It then uses a power-law iterative method with deterministic initial vectors and column-by-column orthogonal projection elimination to extract the principal eigenvectors of the iterative enhancement matrix, thus constructing the signal subspace matrix. The principal feature vectors corresponding to each column, and Stored in the shared memory area and broadcast to all DSP cores; Each DSP core is based on a signal subspace matrix. The spectral peak search is performed on the steering matrix and self-inner product matrix of each angle grid point corresponding to the constructed DSP core. The local spectral value of each angle grid point corresponding to the DSP core is calculated, and the maximum value of the local spectral value is taken as the local peak value of the corresponding azimuth sub-interval corresponding to the DSP core. The main DSP core compares all local peak values and selects the azimuth and elevation angles corresponding to the maximum local peak value as the final azimuth and elevation angles of the target signal.
2. The DSP direction finding method based on power iteration and multi-core parallel processing according to claim 1, characterized in that, The main DSP core constructs an iterative enhancement matrix based on the complex sampled signals acquired by all array elements as follows: Construct the Hermitian covariance matrix based on the complex sampled signals acquired by all array elements. ; Using Hermitian covariance matrix Constructing an iterative enhancement matrix , For conjugate.
3. The DSP direction finding method based on power iteration and multi-core parallel processing according to claim 1, characterized in that, The iterative enhancement matrix is extracted using a power-law iterative method that employs deterministic initial vectors and column-by-column orthogonal projection elimination to construct the signal subspace matrix. The first one to be obtained in the middle The process of creating column principal eigenvectors includes: S11. Initialize the current column index and deterministic initialization vector: Let the principal eigenvector to be obtained be the signal subspace matrix. The The column is randomly initialized to generate the current column to be calculated. The deterministic original initial vector corresponding to the column principal eigenvector ; , For the signal subspace matrix The total number of columns, The total number of array elements. , For the signal subspace matrix The Middle Column principal eigenvectors ; S12. Perform column-by-column orthogonal projection culling on the deterministic original initial vector: when At that time, let the deterministic initial vector after projection removal be... ; when At that time, In the already obtained 1 to 1 Column principal eigenvectors to Projective culling is performed on Zhang Cheng's subspace to obtain a deterministic initial vector after projection culling. ; Indicates that it has been written into the signal subspace matrix The The column principal eigenvector is also used as the obtained first eigenvector. Column principal eigenvectors Indicates conjugate; S13. Perform validity assessment and normalization on the initial vector after projection removal: like ,make ,and Execute step S12; For the first The standard basis vectors corresponding to the column, and the innermost one of them. Each component The value is 1, and the values of all other components are 0. for Inner One portion, The minimum norm threshold is preset; like Then for Normalization is performed to obtain the normalized deterministic initial vector. , Also as the first The initial normalized deterministic initial vector is obtained after the nth iteration, and step S14 is executed; S14, No. In the next iteration, perform a first-order iterative matrix multiplication: according to Calculate the intermediate vector Execute step S15; where, For the iterative enhancement matrix, The initial value is 0; S15, regarding the intermediate vector Perform column-by-column orthogonal projection culling: when At that time, let the intermediate vector after projection removal be... Proceed to step S16; when At that time, let the intermediate vector after shadow removal Proceed to step S16; S16. The intermediate vector after projection removal Perform normalization update: right Perform normalization, and use the normalized result as the first... Normalized deterministic initial vector after iteration number n Proceed to step S17; S17. Calculate the consistency parameter between two adjacent iteration vectors: according to and Calculate the consistency parameter Proceed to step S18; S18. Perform convergence judgment: like Then determine the current number to be obtained. When the column principal eigenvectors converge, Write signal subspace matrix The Column principal eigenvectors In the middle, making ; The preset convergence threshold is used; like Then let Then return to step S14.
4. The DSP direction finding method based on power iteration and multi-core parallel processing according to claim 1, characterized in that, The implementation method for constructing the guidance matrix for each angle grid point corresponding to each DSP core is as follows: ； ； ； in, This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core. This represents any pitch angle to be searched within the preset pitch angle search range. Represents the angle grid points corresponding to the DSP core. The corresponding guidance matrix, Indicates the total number of array elements. Indicates the first Each element corresponds to an angle grid point. The guiding phase response, , Represents the imaginary unit. Indicates the wavelength of the received signal. Indicates the first The position vectors of each array element relative to the array reference point and They represent the first The horizontal and vertical coordinates of each array element in the horizontal plane Represents angle grid points The corresponding unit direction vector, This is a transpose.
5. The DSP direction finding method based on power iteration and multi-core parallel processing according to claim 1, characterized in that, The self-inner product matrix of each angle grid point corresponding to each constructed DSP core ; in, The angle grid point corresponding to the current DSP core The guiding matrix, This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core. This represents any pitch angle to be searched within the preset pitch angle search range. For conjugate.
6. The DSP direction finding method based on power iteration and multi-core parallel processing according to claim 1, characterized in that, The method for calculating the local spectral values of each angle grid point corresponding to each DSP core is as follows: Based on the steering matrix, inner product matrix, and signal subspace matrix of the angle grid points corresponding to the current DSP core Construct the grid points at the current angle. Complex decision matrix ; This represents any azimuth angle to be searched within the azimuth sub-interval corresponding to the current DSP core. This represents any pitch angle to be searched within the preset pitch angle search range; Based on the complex decision matrix real part matrix and the imaginary part of the matrix Calculate the current angle grid point Local spectral values .
7. The DSP direction finding method based on power iteration and multi-core parallel processing according to claim 6, characterized in that, ； in, and These represent the angle grid points corresponding to the current DSP core. The guiding matrix and the self-inner product matrix Indicates conjugate.
8. The DSP direction finding method based on power iteration and multi-core parallel processing according to claim 6, characterized in that, 。 9. The DSP direction finding method based on power iteration and multi-core parallel processing according to claim 1, characterized in that, The implementation method for generating all angle grid points for the azimuth sub-interval and preset pitch angle search interval responsible for each DSP core is as follows: (This is achieved by performing equally spaced discretization sampling and pairing with corresponding step sizes for each DSP core's responsible azimuth sub-interval.) According to the preset azimuth step size, the azimuth sub-intervals responsible for the current DSP core are discretized at equal intervals to obtain a one-dimensional azimuth sequence; According to the preset pitch step size, the preset pitch angle search interval corresponding to the current DSP core is discretized at equal intervals to obtain a one-dimensional pitch sequence; Each azimuth angle in the one-dimensional azimuth sequence is paired with each pitch angle in the one-dimensional pitch sequence to generate a two-dimensional angle grid point set corresponding to the current DSP core.
10. A DSP direction finding system based on power iteration and multi-core parallel processing, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that, The processor executes a computer program to implement the DSP direction finding method based on power iteration and multi-core parallel processing as described in any one of claims 1 to 9.