Direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target rendezvous targets
By employing non-conformal mesh generation and sparsification, combined with stacked LU decomposition and multi-layer matrix compression algorithms, the electromagnetic characteristics of the target in a missile-target rendezvous are directly solved. This addresses the challenges of mesh generation and low efficiency in iterative solutions, achieving efficient and accurate electromagnetic characteristic calculations.
Patent Information
- Application Number
- CN202111429518.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-29
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2041-11-29
AI Technical Summary
In existing technologies, the fine structure and complex materials of targets during missile-target rendezvous make mesh generation difficult, resulting in ill-conditioned matrices and low solution efficiency. Furthermore, iterative methods require recalculation of each right-hand side term, leading to low analysis efficiency.
The algorithm employs non-conformal mesh generation, IEDG discrete impedance matrix, stacked LU decomposition, and multi-level matrix compression. Through sparsification and compressed storage processing, it directly solves the target electromagnetic properties and uses multi-level sparse matrix equations for fast calculation.
It improves the accuracy and efficiency of electromagnetic characteristic calculation for missile-target rendezvous, reduces calculation time when the trajectory changes, and enables rapid solution of problems with multiple right-hand sides.
Smart Images

Figure CN114239239B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of electromagnetic scattering technology, and in particular to a direct sparse solution method for rapid simulation of electromagnetic characteristics of a missile-target rendezvous target. Background Art
[0002] The missile-target rendezvous is also called the encounter phase between the missile and the target. The target electromagnetic characteristics acquired by the missile fuze during the missile-target rendezvous are one of the key parameters for controlling the timely explosion of the warhead to destroy the target. Therefore, the rapid simulation of the electromagnetic characteristics of the missile-target rendezvous is an important part of missile development.
[0003] In the calculation of actual engineering problems, the traditional integral equation method requires an integrated meshing of the target. Usually, the targets of missile-target intersection contain a large number of fine structures. If the meshing size is too large, it cannot accurately simulate the current distribution of the fine parts, resulting in a loss of accuracy. If the meshing size is too small, the number of meshes will increase rapidly. With the development of aircraft technology, stealth targets contain a large number of non-metallic structures, and the material composition of missile-target intersection targets is becoming more and more complex, resulting in more stringent size requirements for meshing the target model. The non-uniform mesh generated by the integrated meshing can easily lead to an ill-conditioned matrix, which seriously affects the convergence speed and calculation accuracy. This also means that in actual engineering, a lot of time is required to complete the initial geometric modeling and meshing work.
[0004] The moment method commonly used in the industry generates a full matrix after discretization, and the direct storage complexity is O(N 2 ). Direct storage is quite difficult, so some fast algorithms have been proposed one after another. Fast algorithms are mainly divided into two categories: fast multipole algorithm and fast Fourier transform method. Among them, the fast multipole method uses the addition theorem to analytically expand the Green's function, which can divide the long-range interaction matrix into three items: aggregation, transfer and configuration matrix, and reduce the computational complexity of moment-vector multiplication to O(NlogN). The fast Fourier transform method uses the translation invariance of the Green's function to calculate the moment-vector multiplication in the Fourier spectral space, which can reduce the solution complexity to O(NlogN). However, these traditional fast algorithms are based on iterative solutions. When dealing with problems such as missile-target intersection with fine structures and complex material targets, even if preconditions are introduced, there is still the problem of slow convergence or even no convergence.
[0005] The target scattering characteristics acquired by the missile during the missile-target encounter process are one of the key parameters for controlling the timely detonation of the warhead to destroy the target, and are therefore an important input parameter in the missile design phase. However, when studying the target scattering characteristics during the encounter, the angles of the incident and scattered fields change with the missile's trajectory, requiring simulation and calculation of the target's electromagnetic characteristics at multiple angles. In actual encounters, the missile's trajectory has an infinite number of possible possibilities, representing a typical multi-right-hand-term problem. When using iterative methods to solve multi-right-hand-term problems, the calculation must be restarted for each right-hand-term. When studying encounter problems with complex trajectory scenarios, hundreds of thousands of right-hand-terms can be generated. Even if the solution time for a single right-hand-term is kept within one minute, the overall time can reach over two months. Furthermore, when the missile's trajectory changes and requires reanalysis, the change in the right-hand-terms causes the iterative method to recalculate, significantly reducing the efficiency of the solution. Summary of the Invention
[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target rendezvous targets, which solves the following problems:
[0007] 1. When the missile-target rendezvous target contains multi-scale problems such as fine structures and complex materials, integrated segmentation is difficult and it is easy to generate ill-conditioned matrices, resulting in low solution efficiency;
[0008] 2. When using iterative methods to solve the problem of multiple right-hand-side terms generated by projectile-target intersection, the calculation needs to be restarted for each right-hand-side term.
[0009] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a direct sparse solution method for rapid simulation of electromagnetic characteristics of a missile-target rendezvous target, comprising the following steps:
[0010] S1. Establish a simulated target geometric model based on the position and material properties of the projectile-target intersection target;
[0011] S2. Setting simulation parameters, performing non-conformal meshing on the target geometric model, and obtaining a discretized target mesh model;
[0012] S3. Establish the impedance matrix of the IEDG discretization according to the discretized target grid model;
[0013] S4, performing sparse and compressed storage processing on the impedance matrix after discretization of the IEDG to obtain a multi-level sparse matrix equation;
[0014] S5. Use the stacked LU decomposition algorithm to solve the multi-level sparse matrix equation to obtain the target electromagnetic characteristics.
[0015] Furthermore, the simulation parameters in step S2 include: simulation frequency and parameters for solving target electromagnetic characteristics;
[0016] The parameters for solving the target electromagnetic characteristics include: the range and interval of the pitch angle θ of the incident angle in the spherical coordinate system, the range and interval of the azimuth angle φ of the incident angle, the incident wave amplitude, and the polarization mode.
[0017] Furthermore, step S3 includes the following sub-steps:
[0018] S31, dividing the discretized target grid model into multiple independent subdomains;
[0019] S32, defining RWG basis functions on the grids within the subdomains of the target grid model, and defining monopole basis functions on the grids at the junctions of the subdomains of the target grid model;
[0020] S33. Based on the RWG basis function and the monopole basis function, an internal penalty equation is established to constrain the current distribution in the subdomain and obtain the impedance matrix after the IEDG is discretized.
[0021] Furthermore, the impedance matrix after the IEDG is discretized in step S3 is:
[0022] (Z ij +α i Z IP1 +α2Z IP2 )I j =V i
[0023]
[0024]
[0025]
[0026] V i = <f i (r), E inc (r)>
[0027] Among them, Z ij is the subdomain Ω i and subdomainΩ j The coupling matrix, α1 and α2 are weight coefficients, Z IP1 , Z IP2 is the internal penalty matrix, I j is the subdomain Ω j Current coefficient, V i is the subdomain Ω i The excitation vector, r is the observation point, r′ is the source point, k0 is the free space wave number, η0 is the free space wave impedance, f i () is the subdomain Ω i The RWG basis function and monopole basis function, f j() is the subdomain Ω j The RWG basis function and monopole basis function of G(r, r′) are scalar Green’s functions, and ds is the subdomain Ω. j The surface element, <, > is the inner product operation, For two vectors or scalar functions in Ω i The inner product of is the subdomain Ω i The unit tangent vector of is the subdomain Ω j The unit tangent vector, l i is the subdomain Ω i The boundary of l j is the subdomain Ω j The border, is the field point Hamiltonian operator, is the source Hamiltonian operator, N is the number of subdomains, ω is the angular frequency, J j (r′) is the induced current, E inc (r) is the incident field, “·” is the inner product, i and j are the index numbers of the subdomain, and i is the sign of the imaginary part.
[0028] The beneficial effects of the above further solution are as follows: For the problem of the multi-scale problem of missile-target intersection, the integrated decomposition generates a non-uniform grid, which leads to an ill-conditioned matrix. The present invention adopts the discontinuous Galerkin integral equation method to solve it. The target is divided into subdomains according to its geometric characteristics, and each subdomain can be independently non-conformally decomposed. For the problem that the normal continuity of the current at the junction of subdomains is destroyed by means of regionalization and non-conformality, resulting in artificial charge accumulation, the present invention defines a monopole basis function at the junction of the subdomains, uses the discontinuous Galerkin method to discretize the integral equation operator, and establishes an inner penalty equation to constrain the current distribution in the subdomain based on the principle of minimizing the error charge and the electric potential energy generated in space.
[0029] Furthermore, step S4 includes the following sub-steps:
[0030] S41, using stacked matrix algorithm Group the RWG basis functions and the monopole basis functions to obtain a basis function set;
[0031] S42. Establishing a quadtree partitioning structure of the impedance matrix according to the spatial distance between the basis function sets;
[0032] S43. A multi-layer matrix compression algorithm (MLMDA) is used to compress and store the far-field blocks that meet the admissible condition in the quadtree partitioning structure, thereby obtaining a multi-level sparse matrix equation.
[0033] The beneficial effect of the above further solution is: through the stacked matrix algorithm The impedance matrix is sparsely processed. The impedance matrix of the impedance matrix can usually be divided into a near-field matrix and a far-field matrix. The far-field matrix is numerically low-rank and can be compressed and stored with the help of various low-rank compression algorithms. A multi-layer matrix compression algorithm (MLMDA) is used to compress far-field blocks with larger dimensions. Traditional low-rank compression methods such as adaptive cross approximation (ACA) have a "variable rank" characteristic in which the matrix rank increases with the increase of electrical size, resulting in a low compression rate in high-frequency problems. The present invention adopts a "fixed rank" compression algorithm: a multi-layer matrix decomposition algorithm. The storage complexity of the multi-layer matrix decomposition algorithm is O(NlogN), which is independent of the matrix rank.
[0034] Furthermore, the L matrix and the U matrix generated in the process of solving the multi-level sparse matrix equation in step S5 are stored out-of-core.
[0035] The beneficial effect of the above further scheme is: by storing the decomposed L and U matrices outside the core, when it is necessary to analyze a new ballistic problem in the same projectile-target intersection scenario, it is only necessary to update the right-hand term and import the L and U matrix information from the hard disk to quickly solve and analyze the new ballistic problem.
[0036] Furthermore, the step S5 includes the following sub-steps:
[0037] S51, performing multi-level LU decomposition on the multi-level sparse matrix equation to obtain a multi-level LU matrix for compressed storage;
[0038] S52. Calculate target electromagnetic characteristics based on the multi-level LU matrix.
[0039] The beneficial effect of the above-mentioned further solution is that it uses stacked LU decomposition to achieve direct solution, and recursively decomposes the entire multi-level sparse matrix equation after segmentation and compression layer by layer until the finest layer, and solves the lower triangular matrix equation and the upper triangular matrix equation in sequence to quickly obtain the corresponding solution of any right-hand side term. Similar to direct LU decomposition, it is only necessary to replace the original multi-level sparse matrix equation (containing two parts, direct storage block and MLMDA compression block) with the sub-matrices in the L and U positions, without the need to apply for additional storage space. And once the L and U matrices of the matrix are calculated, the corresponding solution of the new right-hand side term can be directly solved by solving the upper and lower triangular matrix equations in sequence when the trajectory changes and needs to be recalculated.
[0040] In summary, the beneficial effects of the present invention are:
[0041] (1) The present invention calculates the electromagnetic characteristics of the target in the case of missile-target intersection by a direct sparse solution method based on multi-layer matrix compression. By storing the L and U matrices in the direct solution process, the computational efficiency is improved when the trajectory changes and needs to be recalculated. Compared with the iterative method, the direct solution method can more efficiently handle multiple right-hand-end problems. Missile-target intersection is a typical multiple right-hand-end problem. The direct sparse solution method can more efficiently realize the calculation and solution of the electromagnetic characteristics of the missile-target intersection target.
[0042] (2) The quality of the mesh generated by the projectile-target intersection target model containing complex multi-scale problems and complex materials is improved through non-conformal meshing. The characteristic spectrum distribution of the discretized impedance matrix is improved through the IEDG method, and the accuracy of the direct solution of extremely ill-conditioned matrices is improved.
[0043] (3) The L matrix and U matrix generated during the solution of the multi-level sparse matrix equation are stored out-of-core. When a new ballistic problem in the same projectile-target intersection scenario needs to be analyzed, it is only necessary to update the right-hand side term and import the L and U matrix information from the hard disk to quickly solve and analyze the new ballistic problem without having to solve it all over again. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 This is a flow chart of a direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target rendezvous targets;
[0045] Figure 2 This is a structural diagram of the quadtree partition structure of the impedance matrix;
[0046] Figure 3 It is the structural diagram of the row matrix block;
[0047] Figure 4 This is the overall process diagram of parallel multi-level LU decomposition;
[0048] Figure 5 This is a comparison chart of the memory overhead of the MLMDA and ACA algorithms;
[0049] Figure 6 This is a comparison chart of the computing time of the MLMDA and ACA algorithms. DETAILED DESCRIPTION
[0050] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.
[0051] like Figure 1 As shown, a direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target rendezvous targets includes the following steps:
[0052] S1. Establish a simulated target geometric model based on the position and material properties of the projectile-target intersection target;
[0053] S2. Setting simulation parameters, performing non-conformal meshing on the target geometric model, and obtaining a discretized target mesh model;
[0054] The simulation parameters in step S2 include: simulation frequency and parameters for solving target electromagnetic characteristics;
[0055] The parameters for solving the target electromagnetic characteristics include: the range and interval of the pitch angle θ of the incident angle in the spherical coordinate system, the range and interval of the azimuth angle φ of the incident angle, the incident wave amplitude, and the polarization mode.
[0056] S3. Establish the impedance matrix of the IEDG discretization according to the discretized target grid model;
[0057] Step S3 includes the following sub-steps:
[0058] S31, dividing the discretized target grid model into multiple independent subdomains;
[0059] In this embodiment, the discretized target grid model Ω is divided into N independent subdomains Ω t ,Ω2,…,Ω N .
[0060] S32, defining RWG basis functions on the grids within the subdomains of the target grid model, and defining monopole basis functions on the grids at the junctions of the subdomains of the target grid model;
[0061] When defining a monopole basis function at the subdomain boundary, the interpolation parameters on both sides of the subdomain boundary are different and cannot cancel each other out, which will introduce artificial charges at the subdomain junction and destroy the normal continuity of the current. Therefore, it is necessary to introduce an internal penalty term (IP) to limit the current distribution in the subdomain and ensure the continuity of the current at the connection between different regions.
[0062] S33. Based on the RWG basis function and the monopole basis function, an internal penalty equation is established to constrain the current distribution in the subdomain and obtain the impedance matrix after the IEDG is discretized.
[0063] In step S33, the steps of establishing the impedance matrix after the IEDG is discretized are:
[0064] S331. Establish an integral equation in the subdomain and use the Galerkin method to test it to obtain the subdomain matrix equation;
[0065] z ij Ij =V i
[0066] S332. Calculate the electric potential generated by the charges accumulated at the boundary in space:
[0067]
[0068] Because the monopole basis function is defined at the boundary of the subdomain, charges will be accumulated in the middle of the boundary. The charges and the boundary surface current are linked by the current continuity equation, and the electric potential generated by the charges accumulated at the boundary in space can be calculated.
[0069] S333. Based on the electric potential generated by the charges accumulated at the boundary in space, two internal penalty equations are obtained:
[0070]
[0071]
[0072] While accumulated charge shouldn't exist at the boundary itself, partitioning introduces charge and creates a potential difference, requiring a boundary condition to enforce this constraint. By testing all basis functions on the subdomain and forcing the inner product of the potential generated by the basis functions and the potential generated by the accumulated charge to be zero, two internal penalty (IP) equations are derived.
[0073] S334. Add the two internal penalty equations to the subdomain matrix equation to obtain the impedance matrix after the IEDG is discretized.
[0074] The impedance matrix after the IEDG discretization in step S3 is:
[0075] (Z ij +α i Z IP1 +α2Z IP2 )I j =V i
[0076]
[0077]
[0078]
[0079] V i = <f i (r), E inc (r)>
[0080] Among them, Z ij is the subdomain Ω i and subdomainΩ jThe coupling matrix, α1 and α2 are weight coefficients, Z IP1 、Z IP2 is the internal penalty matrix, I j is the subdomain Ω j Current coefficient, V i is the subdomain Ω i The excitation vector, r is the observation point, r′ is the source point, k0 is the free space wave number, η0 is the free space wave impedance, f i () is the subdomain Ω i The RWG basis function and monopole basis function, f j () is the subdomain Ω j The RWG basis function and monopole basis function of G(r, r′) are scalar Green’s functions, and ds is the subdomain Ω. j The surface element, <, > is the inner product operation, For two vector or scalar functions in Ω i The inner product of is the subdomain Ω i The unit tangent vector of is the subdomain Ω j The unit tangent vector, l i is the subdomain Ω i The boundary of l j is the subdomain Ω j The border, is the field point Hamiltonian operator, is the source Hamiltonian operator, N is the number of subdomains, ω is the angular frequency, J j (r′) is the induced current, E inc (r) is the incident field, “·” is the inner product, i and j are the index numbers of the subdomain, and i is the sign of the imaginary part.
[0081] In this embodiment, the step of constructing the base function using the MPI parallel method in step S32 includes the following sub-steps:
[0082] A1. Let each processor temporarily obtain all the mesh geometry information after segmentation;
[0083] The geometric information includes the vertex number and vertex coordinates of the patch (the triangular patch mesh after the model is divided).
[0084] A2. Calculate the edge center coordinate tasks of the patches based on the total number of patches, and evenly distribute the edge center coordinate tasks to each process (the edge information of the patch is used to determine the basis function);
[0085] A3. Generate partial box sequences using edge center coordinates, and communicate between processes to obtain the global box sequence.
[0086] A4. Obtain the total number of boxes based on the partial box sequence and the global box sequence.
[0087] A5. Evenly distribute boxes to each process based on the total number of boxes, and classify the side information of the boxes to which each process belongs.
[0088] A6. Based on the edge information of the box to which each process belongs, a basis function is constructed, and then the impedance matrix equation is established.
[0089] The RWG basis function is established based on two triangular patches. Knowing the edge information can establish the corresponding basis function. Through the basis function, the integral equation can be converted into a set of linear algebraic equations, that is, the impedance matrix equation, and then the unknown quantity, that is, the current, can be solved to calculate the target electromagnetic characteristics.
[0090] S4, performing sparse and compressed storage processing on the impedance matrix after discretization of the IEDG to obtain a multi-level sparse matrix equation;
[0091] Step S4 includes the following sub-steps:
[0092] S41, using stacked matrix algorithm Group the RWG basis functions and the monopole basis functions to obtain a basis function set;
[0093] S42. Establishing a quadtree partitioning structure of the impedance matrix according to the spatial distance between the basis function sets;
[0094] The quadtree partitioning structure of the impedance matrix is to divide the impedance matrix into self-block, nearby block and far-field block, such as Figure 2 shown.
[0095] S43. A multi-layer matrix compression algorithm (MLMDA) is used to compress and store the far-field blocks that meet the admissible condition in the quadtree partitioning structure, thereby obtaining a multi-level sparse matrix equation.
[0096] For example, at level 0, the impedance matrix is approximated as the product of two new matrices: U (1) and V (1) , the V matrix is explicitly saved, and the U matrix is sent to the next layer for recompression. After MLMDA compression, all V matrices are unitary matrices. V retains the information of each pair of couplings in the U matrix (i.e., singular values), and U is sent to the next layer for recompression.
[0097] In this embodiment, the MPI method is used to perform steps S42 to S43:
[0098] The impedance matrix of the quadtree partition structure is divided into row matrix blocks according to the number of calculation processes, that is, the number of calculation nodes, such as Figure 3 As shown, in order to avoid conflict with the stacked matrix, the number of processes needs to be set to 2 K , where K is smaller than the stacking matrix The number of layers is L. Thus, The structure will not have a dimension greater than [N / 2 K ] low-rank matrix block, and each dimension is less than or equal to [N / 2 K ] will uniquely belong to a certain row partition block. In the MPI parallel environment, each computing node independently fills and stores the row matrix blocks assigned to it. Since there is no dependency between each row matrix block, there is almost no data communication between nodes during the filling process, which can achieve high MPI parallel efficiency. When each node is equipped with a multi-core processor, the calculation process within the node further executes OpenMP shared memory parallelism. The calculation of each impedance matrix element can generally be used as an OpenMP parallel unit. The calculation process of the impedance matrix elements is independent of each other, so the parallel efficiency of OpenMP will also be higher.
[0099] S5. Solve the multi-level sparse matrix equation using a stacked LU decomposition algorithm to obtain target electromagnetic characteristics. The L matrix and U matrix generated in the process of solving the multi-level sparse matrix equation in step S5 are stored out-of-core.
[0100] Step S5 includes the following sub-steps:
[0101] S51, performing multi-level LU decomposition on the multi-level sparse matrix equation to obtain a multi-level LU matrix for compressed storage;
[0102] S52. Calculate target electromagnetic characteristics based on the multi-level LU matrix.
[0103] The form of the multi-level sparse matrix equation in step S51 is:
[0104] The form of the multi-level sparse matrix equation is: Among them, Z is a multi-level sparse matrix equation, Z 11 、Z 12 、Z 21 、Z 22 Represents the matrices of each part of Z;
[0105] The step S51 includes the following sub-steps:
[0106] S511, Z 11 Perform multi-level LU decomposition to obtain L 11 and U 11 ;
[0107] S512, Solve the lower triangle Equation Z 12 =L 11 U 12 , get U 12 ;
[0108] S513, Solve the upper triangle Equation Z 21 =L 21 U 11 , get L 21 ;
[0109] S514, Update Z 22 :Z′ 22 =Z 22 -L 21 U 12 ;
[0110] S515, Z′ 22 Perform multi-level LU decomposition to obtain L 22 and U 22 ;
[0111] Among them, L 11 , L 21 , L 22 is the matrix Z 11 The decomposed L matrix, U 11 、U 12 、U 22 is the matrix Z 11 The decomposed U matrix; (the L and U matrices here are decomposed matrices and have no actual physical meaning)
[0112] The lower triangular H-matrices equations and upper triangular H-matrices equations of steps S512 and S513 are also solved by sub-layer matrix blocks. Taking equation solving as an example, it can be written in equation form:
[0113] LX=A
[0114] Where X is the value of U in step S512. 12 The sub-layer matrix block, A is the Z in step S512 12 The sub-layer matrix block of . Its block form is:
[0115]
[0116] in,
[0117] 5121. Solve the lower triangle Equation A1 = L 11 X1, get X1;
[0118] 5122. Update A2: A′2=A2-L 21 X1;
[0119] 5123. Solve the lower triangle Equation A′2=L 22 X2, get X2;
[0120] 5124. Update X:
[0121] Upper triangle The solution of the equation is similar and will not be repeated here.
[0122] The step S52 is specifically as follows: after the multi-level LU decomposition in the original step S51, a compressed multi-layer LU matrix is obtained, and any right-hand term V can be quickly obtained by solving the multi-layer LU matrix equation. i The corresponding solution I j , and then calculate the target electromagnetic properties.
[0123] In this embodiment, the MPI method is used to perform multi-level LU decomposition: a threshold based on the size of the matrix block is set. If the matrix block currently being processed is small, only one process is used for calculation / decomposition; when the matrix block being processed is larger than a certain set value, multiple processes are used for parallel calculation / decomposition. If the number of processes is 2 K , the number of unknown quantities is N, we can use N / 2 K Threshold as the matrix block size.
[0124] For a depth of L Take multi-level LU decomposition as an example:
[0125] B1. In the first step of recursion, LU decomposition will go deep into The most refined layer is the Lth layer, and the dimension of the top left corner is approximately (N / 2 L )×(N / 2 L ) performs standard LU decomposition on the full-rank matrix block. Since K≤L, this process only requires one process to perform the operation;
[0126] B2, the recursive process returns to the L-1 layer, and based on the LU decomposition in the previous step, the upper left corner dimension is about (N / 2 L-1 )×(N / 2 L-1 ) performs LU decomposition on the matrix block. If K≤L-1 still holds, this process will continue to be performed by one process;
[0127] B3. Repeat the previous step until the recursion returns to the lth layer, where l≤K. The matrix block processed thereafter will be larger than (N / 2 K )×(N / 2 K ), the operation is shared by multiple processes, that is, the MPI distributed parallel execution begins. For the convenience of expression, use Indicates The first step of the multi-level LU decomposition is itself, its parallelization will be carried out by finer-grained Next is to solve the upper and lower triangular matrix systems, that is, steps S512 and S513 can be written as follows:
[0128]
[0129]
[0130] in For L 21 The sub-layer matrix block, Z 21 The sub-layer matrix block of . is a known matrix block and are the upper and lower block triangular matrices, respectively. and As mentioned above, in a multi-level structure, the solutions of these two equations can be obtained by solving a finer triangular system.
[0131]
[0132] Taking the above formula as an example, by observing the solution process, we can see that the solution column block and are two independent processes, they can be parallelized using MPI. The refinement is Solve row blocks and These are two independent processes that can be parallelized using MPI.
[0133] The overall process of parallel multi-level LU decomposition is as follows Figure 4 As shown:
[0134] C1, before the LU decomposition recursively reaches the K-1th layer, that is, before HLU(K), only a single process P1 is required to participate in the calculation;
[0135] C2, when it reaches the K-1 layer, that is, HLU (K-1), there will be two processes, P1 and P2, to calculate respectively and Update later and pair The LU decomposition of is undertaken by a process P1;
[0136] C3. When recursion reaches the K-2 layer, there will be four processes Pi (i=1~4) participating in the solution of the upper and lower triangular matrix systems. The update process can also be processed in parallel using four processes, that is, each process is responsible for the calculation of a sub-matrix;
[0137] C4. In the coarser layer of computation / decomposition, more processes will participate in parallel. Generally speaking, for the lth layer At most 2K-1 processes can be used in parallel to solve the upper and lower triangular matrix systems, and 4K-1 processes can be used in parallel to solve the upper and lower triangular matrix systems. Updates.
[0138] The MPI method described in the present invention is a method of parallel execution, but the method can be executed serially or in parallel during specific execution. The specific method used to implement the present invention is not specifically limited and does not affect the target electromagnetic characteristics calculation process of the method. The execution method is only a path to implement the method.
[0139] The present invention proposes a direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target rendezvous targets, which has the following advantages:
[0140] (1) Non-conformal meshing is performed on models involving complex multi-scale problems and complex materials, such as missile-target rendezvous targets. Discrete matrix equations are established using the discontinuous Galerkin method, enabling efficient, accurate, and direct solutions to complex multi-scale problems. Table 1 shows a comparison of meshing for a typical complex multi-scale aircraft model. It can be seen that the number of unknowns after the non-conformal meshing proposed by the present invention is only 201,227, while the number after the traditional integrated meshing is 246,079. Furthermore, the matrix compression time, memory consumption, and solution time after non-conformal meshing are all relatively low.
[0141] Table 1 Computational resource consumption of complex multi-scale aircraft
[0142]
[0143] (2) The impedance matrix is sparsely compressed and stored by stacking matrices and MLMDA. Compared with the traditional low-rank compression algorithm ACA, MLMDA can achieve higher global compression rate and computational efficiency.
[0144] like Figures 5-6 The figure shows the comparison of computing resource consumption of the MLMDA and ACA algorithms. It can be seen that MLMDA achieves a higher compression rate, and the compression rate difference becomes more obvious as the circuit size increases.
[0145] (3) By storing the matrix after the stacked LU decomposition out-of-core, it is possible to quickly solve any subsequent right-hand-end terms. Table 2 shows a comparison of resource consumption for solving the electromagnetic characteristics of a typical Cobra cavity target. It can be seen that the calculation method proposed in this paper solves the right-hand-end terms in just 156 seconds, while the FEKO software based on the fast multipole algorithm takes 192,442 seconds to solve 361 right-hand-end terms.
[0146] Table 2 Comparison of computational resource consumption for electromagnetic characteristics of Cobra cavity targets
[0147]
[0148] (4) Parallel processing of the basis functions, the quadtree partitioning structure of the impedance matrix, the multi-level sparse matrix equation, and the LU decomposition improves solution efficiency. Table 3 shows a comparison of the resource consumption of parallel computing for the electromagnetic characteristics of the BGM109 ballistic missile target.
[0149] Table 3 Comparison of parallel computing resource consumption for Cobra cavity target electromagnetic characteristics
[0150]
[0151]
Claims
1. A direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target encounter targets, characterized in that: The following steps are involved: S1. Establish a simulated target geometry model based on the position and material properties of the projectile-target intersection target; S2, setting simulation parameters, performing non-conformal meshing on the target geometric model, and obtaining a discretized target mesh model; S3. According to the discretized target grid model, establish the impedance matrix of the IEDG discretization; S4, performing sparse and compressed storage processing on the impedance matrix after the IEDG is discretized to obtain a multi-level sparse matrix equation; S5, using a stacked LU decomposition algorithm to solve the multi-level sparse matrix equation to obtain the target electromagnetic characteristics; The step S3 comprises the following sub-steps: S31, dividing the discretized target grid model into multiple independent subdomains; S32, defining RWG basis functions in the grids within the subdomains of the target grid model, and defining monopole basis functions in the grids at the junctions of the subdomains of the target grid model; S33, according to the RWG basis function and the monopole basis function, an internal penalty equation is established to constrain the current distribution in the subdomain, and the impedance matrix after the IEDG is discretized is obtained; The impedance matrix after the IEDG is discretized in step S3 is: (Z ij +α1Z IP1 +α2Z IP2 )I j =V i In i = <f i (r), E inc (r)> Among them, Z ij Subdomain Ω i and subdomainΩ j The coupling matrix, α1, α2 are weight coefficients, Z IP1 , Z IP2 is the internal penalty matrix, I j Subdomain Ω j Current coefficient, V i Subdomain Ω i The excitation vector, r is the observation point, r′ is the source point, k0 is the free space wave number, η0 is the free space wave impedance, f i ( ) is the subdomain Ω i The RWG basis function and monopole basis function, f j ( ) is the subdomain Ω j The RWG basis function and monopole basis function of G(r, r′) are scalar Green’s functions, and ds is the subdomain Ω. j The surface element of , <, > is the inner product operation, For two vectors or scalar functions in Ω i The inner product of Subdomain Ω i The unit tangent vector of Subdomain Ω j The unit tangent vector, l i Subdomain Ω i The boundary of j Subdomain Ω j The borders of is the field point Hamiltonian operator, is the source Hamiltonian operator, N is the number of subdomains, ω is the angular frequency, J j (r′) is the induced current, E inc (r) is the incident field, "·" is the inner product, i and j are the index numbers of the subdomain, and i is the sign of the imaginary part.
2. The direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target encounter targets according to claim 1 is characterized in that: The simulation parameters in step S2 include: simulation frequency and parameters for solving target electromagnetic characteristics; The parameters for solving the target electromagnetic characteristics include: the range and interval of the pitch angle θ of the incident angle in the spherical coordinate system, the range and interval of the azimuth angle φ of the incident angle, the incident wave amplitude, and the polarization mode.
3. The direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target encounter targets according to claim 1 is characterized in that: The step S4 comprises the following sub-steps: S41, using stacked matrix algorithm Grouping the RWG basis functions and the monopole basis functions to obtain a basis function set; S42, establishing a quadtree partition structure of the impedance matrix according to the spatial distance between the basis function sets; S43, using a multi-layer matrix compression algorithm MLMDA to compress and store the far-field blocks that meet the admissible conditions in the quadtree partition structure, and obtain a multi-level sparse matrix equation.
4. The direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target encounter targets according to claim 1 is characterized in that: The L matrix and the U matrix generated in the process of solving the multi-level sparse matrix equation in step S5 are stored out-of-core.
5. The direct sparse solution method for rapid simulation of electromagnetic characteristics of missile-target encounter targets according to claim 1 is characterized in that: The step S5 comprises the following sub-steps: S51, performing multi-level LU decomposition on the multi-level sparse matrix equation to obtain a multi-level LU matrix for compressed storage; S52. Calculate target electromagnetic characteristics according to the multi-level LU matrix.
