A Method for Constructing an Air-to-Air Missile Proxy Model Based on Projection Difference

By generating random perturbation Halton sequences based on projection difference, the randomized state of the target is locked, which solves the problem of incomplete projection coverage of Halton sequences under high-dimensional conditions, and realizes efficient training data acquisition and accurate learning of air-to-air missile proxy models.

CN123069934APending Publication Date: 2026-11-10SICHUAN RES INST OF SHANGHAI JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611591378.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In constructing air-to-air missile proxy models, existing technologies often suffer from strip clustering and local gaps in high-dimensional conditions due to the finite-length Halton sequence prefix, which affects the model's accurate learning of the coupling relationship between configuration variables. Furthermore, during the sequential appending process, a complete set of training inputs may not be obtained due to precision quantization and repetition of historical coordinates.

Method used

A projection difference-based approach is adopted. By determining the physical configuration domain of the surrogate model, an equal-length random perturbation Halton sequence prefix is ​​generated. The projection quality index is calculated to lock the target randomization state. Continuous index scanning and physical domain quantization of the global random perturbation Halton sequence are performed. After deduplication, effective configuration points are obtained, training samples are updated, and a single-objective regression model is constructed.

Benefits of technology

It improves the quality of two-dimensional projection coverage of the surrogate model under high-dimensional conditions, ensures the integrity and reproducibility of training data, and enhances the model's accuracy in learning the coupling relationship of configuration variables and the traceability of training data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069934A_ABST
    Figure CN123069934A_ABST
Patent Text Reader

Abstract

This invention discloses a method for constructing an air-to-air missile surrogate model based on projection dissimilarity, belonging to the field of aerospace numerical modeling technology. This method determines the physical domain of the air-to-air missile configuration, the single-target response, the candidate seed set, and the budget for new samples. For each candidate seed, it generates an equal-length randomized perturbation Halton sequence prefix, calculates the two-dimensional projection dissimilarity without invoking the target response, and locks the target randomization state. Candidate configuration points are generated according to continuous global subscripts, and after physical domain mapping, precision quantization, and deduplication of historical and current samples, valid configuration points are obtained. The finite scalar response of the valid configuration points is obtained, and the single-target surrogate model is updated until the stopping condition is met. This method combines projection coverage pre-screening before response evaluation with sequential appending of fixed configurations, improving the two-dimensional projection coverage of the finite-length prefix without additionally consuming the target response evaluation budget, and enabling the reconstruction of subsequent training inputs based on the same configuration and historical samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of aerospace, specifically relating to a method for constructing an air-to-air missile proxy model based on projection difference. Background Technology

[0002] In the aerospace field, numerical analysis of complex systems such as air-to-air missiles often involves multiple coupled continuous variables, requiring the establishment of surrogate models based on limited input and response data. In the initial stages of research, the number of available samples is limited, and the quality of sample coverage in the parameter space can affect the model's ability to identify variable interactions. Therefore, it is necessary to study sampling and regression modeling methods suitable for high-dimensional, small-sample conditions.

[0003] A single numerical model evaluation may require grid generation, equation solving, and result extraction, and the computational budget limits the number of available samples. When the input dimension is high, the coverage quality of the limited samples can affect the model's predictions at unevaluated locations.

[0004] Commonly used sampling methods include full factorial sampling, Monte Carlo random sampling, Latin hypercube sampling, and Halton sequences. The sample size of full factorial sampling grows rapidly with dimensionality; random sampling may result in point clustering with a finite sample size; Latin hypercube sampling primarily constrains one-dimensional stratification, and in high-dimensional conditions, two-dimensional projective coverage also needs to be considered. The choice of sampling method is also related to sample budgeting and appending methods.

[0005] Classical Halton sequences use different prime cardinality in different dimensions and can be appended with consecutive subscripts. As the dimension increases, some finite-length prefixes may appear as stripes or local gaps in the two-dimensional projection. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a method for constructing an air-to-air missile surrogate model based on projection difference. Existing related technologies, during the surrogate model construction process, under conditions of limited target response evaluation times and high dimensionality of configuration variables, are prone to strip clustering and local gaps in some two-dimensional projections due to the finite-length Halton sequence prefix. Furthermore, during the sequential appending process, due to precision quantification and repetition of historical coordinates, a complete and reproducible number of training inputs may not be obtained, thus affecting the surrogate model's accurate learning of the coupling relationship of configuration variables.

[0007] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for constructing an air-to-air missile proxy model based on projection difference, comprising the following steps: S100. Determine the physical configuration domain and single-objective response of the surrogate model, normalize the configuration variables in the physical configuration domain, determine the candidate seed set, and set the new sampling budget. S200. For each candidate seed in the candidate seed set, generate an equal-length random perturbation Halton sequence prefix according to the newly added sampling budget, calculate the projection quality index based on the unit domain coordinates of the random perturbation Halton sequence prefix, and lock the global random perturbation Halton sequence corresponding to the target randomization state. S300: According to the batch plan, read the global random perturbation Halton sequence by continuous global subscript, map the read unit domain coordinates to the physical configuration domain, and determine the candidate configuration points through capacity check. Filter out the candidate configuration points that are repeated with the unit domain coordinates received in the historical training samples or the current batch, and continue to read the global subscript until the required number of valid configuration points for the current batch are obtained. S400: Perform numerical evaluation on the effective configuration points obtained in the current batch, obtain the finite scalar target response corresponding to each effective configuration point, and combine the effective configuration points and their target responses into new training data and add them to the training samples. S500. Update or train a pre-selected single-objective regression model using the updated training samples, and evaluate whether it meets the preset accuracy condition, stagnation condition or sample budget condition. If so, then obtain and output the target proxy model, the current training sample, the candidate score record, and the reproduction configuration; If not, return to S300 and add valid configuration points while keeping the target randomization state unchanged.

[0008] The beneficial effects of this invention are as follows: This method combines Halton randomization components, coordinate projection pre-screening before response evaluation, and full batch appending under a fixed configuration, which has the following advantages: (1) Before the target response evaluation, the centered L2 difference of all two-dimensional projections is compared only based on the candidate input coordinates and the target randomization state is locked so that the target response evaluation budget can be used for subsequent training data acquisition. The 20-dimensional coordinate verification shows that, within the same number of points and index range, the two-dimensional projection coverage of the selected finite prefix is ​​better than the classic Halton control.

[0009] (2) Lock the sequence type, candidate seed, variable dimension and boundary, number permutation method, cyclic translation state, prime number radix allocation method, starting index and precision setting as the same reproduction configuration; when running again, the candidate coordinates are reconstructed deterministically according to the configuration, and the used points are skipped using the complete historical training coordinates, so as to keep the appended results traceable across runs.

[0010] (3) When multiple cumulative sample sizes are predetermined, the four indicators of the candidate coordinates are normalized at each prefix checkpoint, and then merged according to the fixed indicator weights and prefix stage weights, so that the state selection simultaneously reflects the projection coverage of different additional stages.

[0011] (4) Candidate configuration points are scanned in sequential index order. After physical domain quantization, duplicates are deduplicated with historical coordinates and coordinates already received in this batch. Only valid configuration points that meet the requested number are returned to the current batch. Under fixed precision, the remaining representable grid capacity can be checked before generation. If the scan is still insufficient within the upper limit, the scan is explicitly stopped without returning an incomplete batch.

[0012] (5) When historical samples exist, the normalized historical coordinates and each candidate new prefix can be merged by row, and the cumulative design can be evaluated using the same single endpoint, comprehensive or multi-prefix rules as the new prefix, so that the selection of the target randomization state takes into account both the existing design and the planned new design.

[0013] This method selects and locks the randomization state of the target based on the projection quality index of the unit domain coordinates, and organizes the single-target regression training data through continuous index scanning, physical domain quantization and historical coordinate deduplication. Numerical results verify the improvement of the projection coverage of the tested finite prefix, and engineering reproduction checks verify the deterministic replay and complete batch collection behavior under fixed configuration. Attached Figure Description

[0014] Figure 1 The flowchart of the method for constructing an air-to-air missile proxy model based on projection difference provided by the present invention is shown.

[0015] Figure 2 The 17th and 18th dimension projection comparison diagrams of the conventional Halton and the pre-screened random perturbation Halton (i.e., SORSH) provided for this invention.

[0016] Figure 3 A comparison chart of the 20-dimensional fully two-dimensional projected centralized L2 difference between the conventional Halton and the pre-screened random perturbation Halton (i.e., SORSH) provided by this invention. Detailed Implementation

[0017] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0018] This invention provides a method for constructing an air-to-air missile proxy model based on projection difference, comprising two stages: randomized state pre-screening and fixed-state sequential modeling, connected by continuous global indices of the target randomized states and sequences, referencing... Figure 1 Specifically, it includes the following steps: S100. Determine the physical configuration domain and single-objective response of the surrogate model, normalize the configuration variables in the physical configuration domain, determine the candidate seed set, and set the new sampling budget. S200. For each candidate seed in the candidate seed set, generate an equal-length random perturbation Halton sequence prefix according to the newly added sampling budget, calculate the projection quality index based on the unit domain coordinates of the random perturbation Halton sequence prefix, and lock the global random perturbation Halton sequence corresponding to the target randomization state. S300: According to the batch plan, read the global random perturbation Halton sequence by continuous global subscript, map the read unit domain coordinates to the physical configuration domain, and determine the candidate configuration points through capacity check. Filter out the candidate configuration points that are repeated with the unit domain coordinates received in the historical training samples or the current batch, and continue to read the global subscript until the required number of valid configuration points for the current batch are obtained. S400: Perform numerical evaluation on the effective configuration points obtained in the current batch, obtain the finite scalar target response corresponding to each effective configuration point, and combine the effective configuration points and their target responses into new training data and add them to the training samples. S500. Update or train a pre-selected single-objective regression model using the updated training samples, and evaluate whether it meets the preset accuracy condition, stagnation condition or sample budget condition. If so, then obtain and output the target proxy model, the current training sample, the candidate score record, and the reproduction configuration; If not, return to S300 and add valid configuration points while keeping the target randomization state unchanged.

[0019] In S100 of this embodiment of the invention, it is determined Let there be a continuous configuration variable, denoted as... The physical upper and lower bounds of the i-th configuration variable are respectively and ,and .

[0020] For example, the above configuration variables include the nose slenderness ratio, body slenderness ratio, wing span ratio, wing root chord ratio, wing sweep angle, wing axial position ratio, tail fin span ratio, and tail fin sweep angle of the air-to-air missile. The single-target response can be the drag coefficient, lift-to-drag ratio, pitching moment coefficient, or static stability margin under fixed simulation conditions, or it can be obtained by a predetermined scalarization formula that remains unchanged throughout the test.

[0021] In this embodiment, each method execution builds a corresponding proxy model for only one target response, and the definition of the target response remains consistent during the proxy model construction process.

[0022] When normalizing configuration variables in the configuration parameter space, the configuration variables are mapped to a unit hypercube to obtain normalized coordinates:

[0023] When the physical range of the configuration variables includes an upper bound, the normalized coordinates can be taken as follows: In the normalization domain When determining the candidate seed set, set the historical or initial sample number to... The additional sampling budget (planned total number of new valid inputs) is , No. Approval Request _ valid configuration points, generating a candidate seed set as follows: subscript This represents the number of candidate seeds.

[0024] In S200 of this embodiment of the invention, generating an equal-length random perturbation Halton sequence prefix according to the newly added sampling budget includes: Based on the number of configuration variables Predetermine the order to ascending. A prime number; For each candidate seed in the candidate seed set Determine the dimensional index arrangement Digital substitution and translation amount Construct candidate randomized states ; Based on the candidate randomization states, the unit domain coordinates of each index of the randomly perturbed Halton sequence are determined sequentially through cardinality expansion, digital permutation, perturbation root inversion, cyclic shift, and modulo 1. This generates candidate seeds. The corresponding length is the random perturbation Halton sequence prefix of the newly added sampling budget. , , ; For sequence index, To increase the sampling budget, For configuration variable dimensions, The number of configuration variables; When radix allocation rearrangement is enabled, prime radixes, equal to the number of configuration variables, are obtained in ascending order. A random number generator is initialized with a candidate seed, and the random number generator performs a random permutation without replacement on the dimension index set to obtain the dimension index arrangement. , making the first Each configuration variable is arranged using the index dimension. The Middle The prime base indicated by the term, and each prime base is assigned only once in each dimension; when base assignment rearrangement is not enabled, the th prime base is... The configuration variable uses the first A prime cardinality is used, i.e., an identical dimensional index arrangement. Furthermore, when enabling cardinality allocation rearrangement, the same candidate seed produces the same dimensional index arrangement, and the resulting cardinality allocation is used for number permutation, optional cyclic translation, and candidate sequence coordinate calculation.

[0025] Specifically, the process of generating the above-mentioned randomly perturbed Halton sequence prefix is ​​as follows: Determine the first positive integers in ascending order. prime number Independent of the generation of the candidate seed set; when prime number and dimension rearrangement are enabled, the random number generator initialized by the candidate seed generates the dimension index arrangement. , making the first Each configuration variable is arranged using the index dimension. The Middle The prime base indicated by the term For example, The prime number table for time is , It is the first The prime index corresponding to each configuration variable is represented as: When prime numbers and dimension rearrangement are not enabled, .

[0026] Index of randomly perturbed Halton sequences (corresponding to the subsequent global index) is a non-negative integer, when Time definition ;when When writing the sequence index as a radix... The numbers below, when expanded, are represented as:

[0027] in, corresponding digits, For digit index, It is the highest non-zero digit.

[0028] For the first Dimensional construction of digital permutation This allows the zero digit to remain unchanged while randomly rearranging the remaining digits. Based on this, the root inversion value of the disturbance is determined. for:

[0029] Determine a fixed cyclic translation for each dimension using the same candidate seed. ∈[0,1); Therefore, the unit domain coordinates are:

[0030] In the formula, This represents a Cranley-Patterson modulo 1 translation, used to change the sampling position in low cardinality dimensions. Random numbers are consumed in the order of "optional dimension rearrangement, non-zero number permutation, and optional translation". The resulting randomized state remains unchanged in each sequential batch.

[0031] In S200 of this embodiment of the invention, when locking the global random perturbation Halton sequence corresponding to the pre-target randomization state by calculating the projection quality index, the unit domain sampling coordinates of the random perturbation Halton sequence prefix are used as input, without calling the target response or using the validation set or test set response.

[0032] In one specific embodiment of the present invention, locking the globally random perturbation Halton sequence corresponding to the target randomization state includes: The centered L2 dissimilarity index is calculated for all two-dimensional unit domain coordinate projections of each random perturbation Halton sequence prefix, and the average two-dimensional projection centered L2 dissimilarity index is determined. The candidate randomization states are sorted by average two-dimensional projection-centered L2 dissimilarity index from smallest to largest. The smallest candidate randomization state is determined as the target randomization state, and the corresponding global random perturbation Halton sequence is locked.

[0033] Furthermore, when there are parallel minimum average two-dimensional projection-centered L2 dissimilarity indices, the maximum two-dimensional projection-centered L2 dissimilarity index, the full-dimensional centered L2 dissimilarity index, the maximum absolute Pearson correlation coefficient, and the candidate seed are compared sequentially for each random perturbation Halton sequence prefix. The candidate randomization state that first appears with a unique minimum value is selected as the target randomization state, and the corresponding global random perturbation Halton sequence is locked.

[0034] Specifically, for the unit domain coordinate point set of a randomly perturbed Halton sequence Calculate the square of the centered L2 dissimilarity. In this invention, the unradical value of the centered L2 dissimilarity is denoted as... It is represented as:

[0035]

[0036]

[0037] In the formula, The number of coordinate points contained in the unit domain coordinate point set. Let be the dimension of the set of points being evaluated. For the first The unit domain sample points at the th Coordinates on the dimension For the first The unit domain sample points at the th Coordinates in dimensional space; and All are coordinate point indices. This is the coordinate dimension index. In this embodiment, when calculating the full-dimensional centralized L2 dissimilarity index, When calculating the maximum two-dimensional projection-centered L2 dissimilarity index, .

[0038] When calculating the average two-dimensional projected centered L2 dissimilarity index, the prefix of the random perturbation Halton sequence is denoted as... Take the first Column composition The average two-dimensional projection-centered L2 dissimilarity index was calculated. for:

[0039] according to Sort by size from smallest to largest, and then select the target randomized state. Lock the corresponding globally random perturbation Halton sequence.

[0040] Furthermore, the maximum two-dimensional projection-centered L2 dissimilarity index Full-dimensional centralized L2 difference index Maximum absolute Pearson correlation coefficient They are respectively:

[0041]

[0042]

[0043]

[0044] In the formula, Candidate seeds The first corresponding random perturbation Halton sequence prefix dimensional coordinates and the first Pearson correlation coefficient between dimensional coordinates and The first The unit domain coordinate point at the th unit domain coordinate point peacekeeping Coordinates on the dimension and The first peacekeeping The mean of the 3D coordinates.

[0045] When multiple candidate randomized states When they are equal, compare them sequentially. , , and candidate seeds The first candidate randomized state after sorting according to the lexicographical order is determined as the target randomized state, and the global random perturbation Halton sequence is locked.

[0046] Therefore, the target randomization state is determined. for:

[0047] For example, when multiple candidates When they are equal, compare them first. Compare when they are still equal ; continue comparing when they are equal When all four indicators are equal, select the candidate seed with the smaller value.

[0048] In another specific embodiment of the present invention, locking the globally random perturbation Halton sequence corresponding to the target randomization state includes: For each random perturbation Halton sequence prefix, calculate the projection quality index, including the average two-dimensional projection-centered L2 dissimilarity index, the maximum two-dimensional projection-centered L2 dissimilarity index, the full-dimensional centered L2 dissimilarity index, and the maximum absolute Pearson correlation coefficient; Under the same random perturbation Halton sequence prefix, the range of each projection quality index is normalized; The weight of each projection quality index is obtained by dividing the preset non-negative priority coefficient by the sum of the non-negative priority coefficients. Based on the weights of the projection quality indicators, the normalized projection quality indicators are weighted and summed to obtain the comprehensive projection difference score. The comprehensive projection difference scores are sorted in ascending order, and the candidate randomization state corresponding to the minimum comprehensive projection difference score is identified as the target randomization state.

[0049] Furthermore, when multiple candidates have the same comprehensive projection difference score, the longest prefix is ​​compared sequentially. , , , and candidate seeds .

[0050] Specifically, for each projection quality index The minimum value is taken from the set of prefixes of each randomly perturbed Halton sequence. and maximum value ,like If the range is equal to the given value, then normalization is applied; if the range and the normalization value are equal, then the normalized value of the indicator is set to zero to obtain the normalized value of each indicator. .

[0051] The non-negative priority coefficients of the above four projection quality indicators are... Dividing by the sum of all non-negative priority coefficients yields the weights of the normalized projection quality index. Furthermore, each coefficient is required to be a finite non-negative number and its sum must be greater than zero. For example, the non-negative priority coefficients of the above four projection quality indicators can be set to (0.5, 0.2, 0.2, 0.1) in sequence; since their sum is 1, the normalized indicator weights are the same as those of this set of coefficients.

[0052] Therefore, the overall projection difference score is:

[0053]

[0054] In the formula, Candidate seeds The corresponding comprehensive projection difference score indicates that the smaller the value, the better the comprehensive coordinate coverage quality obtained according to the preset weight. Normalized projection quality index The weight, For projection quality indicators The non-negative priority coefficient, This is the index used when summing the non-negative priority coefficients of the four projection quality indicators.

[0055] In S200 of this embodiment, when multiple corresponding random perturbation Halton sequence prefixes exist for the same candidate randomization state, a projection quality index is calculated based on the unit domain coordinates of the multiple prefixes, and then a score is synthesized according to the index weight and stage weight; specifically, let the predetermined prefix checkpoint be... Furthermore, candidate index normalization was performed at each checkpoint to obtain the multi-prefix comprehensive projection difference score. :

[0056] In the formula, This represents the total number of prefix checkpoints. This indicates that the first part of the sequence prefix is ​​taken cumulatively. Evaluation is performed using unit domain coordinates. Indicates candidate seeds In the The first prefix checkpoint Normalized indicators, For the first Normalized stage weights for each prefix evaluation stage. Normalized projection quality index The weight.

[0057] Each prefix evaluation stage uses the same set of candidate seeds and the same set of projected quality index weights, differing only in the cumulative prefix length. The scores for each stage are then calculated according to... We get the weighted sum. .

[0058] In S200 of this embodiment of the invention, the global random perturbation Halton sequence corresponding to the target randomized state can also be locked through historical joint evaluation; wherein, historical joint evaluation refers to: After merging the unit domain coordinates of the effective configuration points in the historical training samples with the unit domain coordinates of the newly added random perturbation Halton sequence prefix, the joint point set is evaluated using projection quality metrics. Specifically, when historical training samples exist and historical joint evaluation is used, the unit domain coordinates of the effective configuration points in the historical training samples are first normalized according to the same physical boundary to obtain the historical point set. Then With each candidate new prefix The unit domain coordinates are concatenated row by row to obtain the joint point set. Subsequently, the same single-index, comprehensive, or multi-prefix sorting rules as the newly added prefix are applied to each joint point set, and the target randomization state is selected accordingly. Furthermore, historical responses are not included in the coordinate geometric scoring.

[0059] The two evaluation methods provided by this invention for locking the globally random perturbation Halton sequence serve different purposes: the historical joint evaluation determines whether the evaluated point set contains historical coordinates, and the multi-prefix evaluation determines whether to summarize the scores of multiple cumulative prefixes. Therefore, the two methods can be used in combination. When used in combination, at each prefix checkpoint, the same historical point set and the candidate new prefix corresponding to that checkpoint are concatenated row by row. The aforementioned projection quality index is calculated on the resulting joint point set, and a score is synthesized according to the aforementioned index weights and stage weights. It should be noted that both of the above evaluations belong to the candidate randomization state evaluation and selection in step S200, and their output is the selected target randomization state, which is used by step S300 to continue generating candidate configuration points according to the locked globally random perturbation Halton sequence configuration.

[0060] In S300 of this embodiment of the invention, in the first... The approval process involves selecting valid configuration points, including the following steps: Determine the global index of the starting point of the current batch, and read the global index sequentially from the globally random perturbation Halton sequence; For each global index read, calculate the corresponding unit field coordinates according to prime number radix allocation, number permutation, and translation. Map the unit domain coordinates to the physical configuration domain and determine candidate configuration points through a capacity check; Determine whether the unit domain coordinates of the candidate configuration points are duplicates of the unit domain coordinates of historical training samples or those already received in the current batch; If so, the candidate configuration point is duplicated, and it is rejected as a valid configuration point in the current batch. The next global coordinate is then read. If not, the candidate configuration point is not repeated and is accepted as a valid configuration point in the current batch; Repeatedly execute global index reading, physical configuration domain mapping, constraint and duplicate determination until the required number of valid configuration points for the current batch is obtained, and record the actual number of global indices scanned in the current batch. And determine the global index starting point for the next batch of scans as .

[0061] In this embodiment, the unit domain coordinates are mapped to the physical configuration domain, as follows:

[0062] In the formula, Global coordinates The candidate configuration point in the corresponding physical configuration domain is at the th Coordinates on each configuration variable The respective The physical lower and upper boundaries of each configuration variable. For the target randomized state In global coordinates , No. Unit field coordinates in the dimension.

[0063] In S300 of this embodiment of the invention, determining candidate configuration points through boundary constraints includes: Configuration points in the physical configuration domain, execute: Read the physical lower and upper boundaries of the configuration points one dimension at a time, and determine whether any boundary is a non-finite number or whether the physical upper boundary is not greater than the physical lower boundary. If so, stop the determination of valid configuration points in the current batch, record the corresponding dimension, output the boundary verification failure information, and do not proceed to the subsequent coordinate quantization, deduplication and valid configuration point addition steps. After receiving the corrected physical boundary, re-execute the boundary verification and execute the determination of valid configuration points in the current batch. If not, then the unit domain coordinates of the configuration points are quantized, including: Set the quantization precision, determine the quantization interval based on the decimal quantization precision, round up the lower physical boundary of each dimension of the configuration point and round down the upper physical boundary to determine the minimum and maximum allowable grid points located on the original physical boundary; Determine whether there exists a minimum allowed grid point in any dimension that is greater than the maximum allowed grid point; If so, it is determined that there are no representable coordinates under the current quantization precision, the determination of valid configuration points for the current batch is stopped, and the unrepresentable coordinate information containing the corresponding dimension and the current quantization precision is output. The subsequent coordinate quantization, deduplication and valid configuration point addition steps are not entered. After receiving the increased number of decimal places of physical coordinate quantization or the expanded physical boundary, the determination of valid configuration points for the current batch is re-executed from the allowed grid point determination. If not, the physical coordinates of the mapped configuration points are quantized according to the decimal rounding rule and restricted to between the minimum and maximum allowable grid points to complete the coordinate quantization. For configuration points that have completed coordinate quantization, calculate the product of the number of allowed grid points in each dimension as the total number of representable grid points in the physical configuration domain, and compare it with the number of already occupied grid points and the number of effective configuration points required for the current batch to determine whether the capacity condition is met. If so, then this configuration point is taken as a candidate configuration point; If not, stop determining the effective configuration points for the current batch, report insufficient precision grid capacity, and do not return or write a batch of configuration points with insufficient quantity; after receiving the increased number of decimal places for physical coordinate quantization, the expanded physical boundary, or the reduced number of new configuration points in the current batch, recalculate the precision grid capacity and execute the determination of effective configuration points for the current batch.

[0064] Specifically, the physical lower boundary and physical upper boundary of each configuration variable are first converted into numerical values. Then, the physical upper boundary is checked dimension by dimension to see if both boundaries are finite numbers and if the physical upper boundary is greater than the physical lower boundary. If any dimension does not meet the requirements, the corresponding dimension is reported and the generation of this batch is stopped.

[0065] When setting When determining the quantization precision for decimal places, first determine the quantization interval based on the decimal precision. Quantization interval is Based on the quantization interval, for the first The physical lower boundary of each configuration variable is rounded up, and the physical upper boundary is rounded down to obtain the minimum and maximum allowable lattice points within the original physical boundaries. The calculation relationship is as follows:

[0066] If the minimum allowable grid point in any dimension is greater than the maximum allowable grid point, it is determined that there are no representable coordinates at that precision, and generation stops. If allowable grid points exist, the physical coordinates of the mapped configuration points are quantized using decimal rounding rules, and the quantization result is restricted to between the minimum and maximum allowable grid points in the corresponding dimension, specifically:

[0067] After coordinate quantization is completed, before generating the current batch, the product of the allowed number of grid points in each dimension is calculated as the total number of representable grid points in the physical configuration domain. This total is then compared with the sum of the number of already occupied grid points and the number of points requested in the current batch. The calculation and determination relationship is as follows:

[0068] In the formula, To preset the number of decimal places, For quantization interval; and The first The physical lower and upper boundaries of each configuration variable. and These represent the minimum and maximum allowable grid points within the lower boundary of this precision, respectively. For the mapped first The configuration point at the th configuration point Physical coordinates in dimensional space The physical coordinates after quantization and boundary constraints. Indicates reservation Decimal rounding operations for decimal numbers; For configuration variable dimension, This represents the total number of representable lattice points within the physical configuration domain. This refers to the number of unit domain coordinates already received in historical training samples or the current batch, the number of unit domain coordinates reserved for verification, and the number of unit domain coordinates already used for testing. The number of valid configuration points requested in the current batch. Only when... Not less than and If the sum is sufficient, continue generating valid configuration points for the current batch; otherwise, stop generating and report insufficient precision grid capacity.

[0069] Duplicate detection is performed after capacity check: Settings When decimal precision is set, the quantized physical coordinate tuple is directly used as the deduplication key. When no decimal precision is set, the physical coordinates of each dimension are first normalized to the unit interval according to the corresponding physical boundary, and then ten decimal places are retained to form the deduplication key to eliminate the influence of different variable dimensions on collision tolerance. The deduplication key sets of historical training samples or the unit domain coordinates received in the current batch, the reserved verification unit domain coordinates, and the unit domain coordinates used for testing are compared in turn. If a duplicate occurs, the candidate is discarded and the global index is incremented to continue scanning. If no duplicate occurs, the candidate is accepted and its deduplication key is immediately registered until the number of valid configuration points required for the current batch is obtained. If the preset scanning limit is still insufficient, the process terminates and an error is reported, and the training samples remain unchanged.

[0070] In S400 of this embodiment, the valid configuration points of the current batch are written into the dataset to be evaluated, and a single finite scalar response is obtained through a pre-fixed parametric geometry and numerical simulation process, wind tunnel test or engineering data interface; only records with complete input coordinates and corresponding responses are added to the training samples, and records that do not obtain finite responses do not participate in this round of model update.

[0071] In S500 of this embodiment, the selected single-objective surrogate model is updated using the updated training samples; for example, taking a Matérn 5 / 2 Gaussian process as the single-objective surrogate model, its input transformation is determined only by the training data or a pre-given boundary. Let the input of this single-objective surrogate model be... Its kernel function is:

[0072]

[0073] In the formula, For signal variance, The kernel length scale for configuration variables; symbol With respect to the physical lower boundary They are independent of each other. Kernel parameters and noise parameters are learned only using training samples.

[0074] The proxy model described above in this embodiment can be replaced with radial basis Gaussian process, support vector regression, random forest or neural network, etc., depending on the task.

[0075] In S500 of this embodiment, the updating and stopping determination of the surrogate model are evaluated using in-training sample cross-validation or independent validation sets. The evaluation metrics include the coefficient of determination R. 2 The calculation stops when the validation error reaches the threshold, the maximum sample budget is exhausted, several consecutive batches of improvements fall below the threshold, or the engineers terminate the calculation; otherwise, it returns to S300 and reads the next batch of global indices for the same global random perturbation Halton sequence.

[0076] After the stopping condition is met, the surrogate model construction is completed, and the target surrogate model, current training samples, candidate scoring table, and reproduction configuration are output; where: The output target surrogate model record includes the input transformation used, kernel parameters, and noise parameters, and is saved together with the corresponding training data.

[0077] The reproduction configuration records the data range, candidate generation conditions, and evaluation rules used in steps S100 to S300, illustrating the candidate coordinates and the conditions upon which the selected result was based. The meaning of each configuration item and its correspondence with the aforementioned steps are as follows: (1) Configuration variable dimension and physical boundary: Dimension refers to the number of input variables determined in step S100, and physical boundary refers to the physical lower boundary and physical upper boundary of each variable. The two determine the dimension and value range of the input coordinates, which are used for the aforementioned normalization and physical domain mapping in step S300, respectively.

[0078] (2) Seed selection and randomization conditions: The seed selection is the seed corresponding to the target randomization state selected in step S200; the number permutation method refers to the permutation rule used to change the cardinality number mapping; the cyclic translation condition refers to whether to apply the aforementioned fixed translation to the unit domain coordinates; the cardinality allocation rearrangement condition refers to whether to change the allocation between the prime cardinality and the input dimension. The above conditions together correspond to the candidate sequence generation in step S200 and are used in step S300.

[0079] (3) Global index and physical domain quantization decimal places of the first candidate in the sequence: The former refers to the position where the candidate sequence begins to be read, which corresponds to the aforementioned continuous global index; the latter refers to the number of decimal places retained during physical coordinate quantization, which corresponds to the quantization interval, allowed grid points, and repetition determination in step S300. The global index of the sequence and the number of valid configuration points actually received are recorded separately.

[0080] (4) Candidate ranking rules: refers to the candidate comparison criteria predetermined in step S200, including the single index, comprehensive score or multi-prefix score used, and the aforementioned comparison order when scores are the same. This configuration is used to explain how candidate score records form the final ranking and select the target randomization state accordingly.

[0081] (5) Multi-prefix evaluation configuration: The cumulative prefix length refers to the candidate sequence prefix length taken at each checkpoint in step S200; the index weight refers to the weight of each of the four normalized indices within the same checkpoint; the stage weight refers to the weight of each stage when summarizing the scores of different prefix checkpoints. The above records correspond to the aforementioned prefix truncation, in-stage index synthesis, and cross-stage score summarization, respectively.

[0082] (6) Historical Joint Evaluation Configuration: Historical coordinate source refers to the existing coordinate data participating in the joint evaluation in step S200, and normalized boundary refers to the physical upper and lower boundaries used when transforming the data to the unit domain. The two correspond to the formation of the aforementioned historical point set and its merging evaluation with the candidate newly added prefix. When historical joint evaluation and multi-prefix evaluation are used in combination, this item and item (5) configuration are recorded at the same time; the meaning and combination relationship of the two evaluations are consistent with step S200.

[0083] (7) Implementation version: refers to the implementation version identifier used to perform the aforementioned sequence generation, coordinate processing and score calculation. It is saved together with the selected seed and various configurations to determine the implementation conditions corresponding to the current result.

[0084] The candidate score record is a process record generated in step S200 by calculating the four projection quality indicators, summarizing the scores, and ranking the candidates. The randomization state of each candidate is distinguished by the candidate seed. The four projection quality indicators refer to the average two-dimensional projection difference, maximum two-dimensional projection difference, full-dimensional difference, and maximum absolute correlation coefficient calculated on the complete evaluation prefix of each candidate. Their meanings and calculation relationships are consistent with the four indicators in step S200, and they are used to record the coordinate geometric performance of each candidate.

[0085] When using multi-prefix evaluation, the indicators for each prefix checkpoint are the four indicators calculated separately for the corresponding cumulative prefix. The multi-prefix comprehensive score is the summed value obtained according to the normalization, indicator weighting, and stage weighting relationships in step S200. When using historical joint evaluation, the evaluation object of the recorded indicators is the joint point set composed of the historical point set and the candidate newly added prefix; when combining the two evaluation methods, the recording object corresponds to the joint point set of each checkpoint. The above records correspond to the intermediate values ​​and summary results of the aforementioned evaluation steps, respectively.

[0086] The final ranking order is the candidate arrangement obtained in step S200 according to the fixed candidate ranking rules, used to select the corresponding target randomization state; the ranking rules used are consistent with the records in the reproduction configuration. Therefore, the candidate score record stores the evaluation and ranking results, and the reproduction configuration stores the conditions used to generate these results. The two correspond to each other through candidate seeds and ranking rules, jointly explaining the selection basis of step S200 and connecting with the target randomization state used in step S300. The projection quality index in the candidate score record and the prediction error index used for model evaluation in step S500 correspond to different evaluation objects.

[0087] In this embodiment of the invention, eight air-to-air missile configuration variables are used to illustrate the physical boundary mapping, and a normalized analytical function is used to illustrate the data connection between input coordinates, scalar response, batch addition, and model update.

[0088] All eight inputs are dimensionless configuration variables with values ​​in the range [0,1], and their definitions and ranges are as follows:

[0089] To verify the data flow of the method, the eight physical configuration variables were normalized to Then, the single analytical scalar response is defined as follows:

[0090] The above analytical function is only used to verify the coordinate pre-screening, batch addition, and regression modeling process; in air-to-air missile applications, it is replaced with an aerodynamic scalar response obtained by the same numerical evaluation process under fixed conditions.

[0091] For example, a fixed simulation condition is selected and the drag coefficient CD is used as the single objective response. An initial training sample is established using 24 existing and labeled configuration points, and it is planned to add 96 new valid configuration points in 12 batches of 8 points each. The candidate seed set contains 500 candidates, and the sequence starts from a preset global index, with non-zero numerical random permutation and dimension-wise cyclic translation enabled.

[0092] For each candidate seed, 96 new prefixes are generated, which can be merged with 24 initial configuration coordinates to form a joint point set. 28 two-dimensional projection indices for eight configuration variables are calculated, sorted in ascending order according to a pre-fixed average two-dimensional projection index or comprehensive index, and the target randomization state is selected. The drag coefficient evaluation is not invoked during the pre-screening stage. Subsequently, under a fixed configuration, eight unique effective configuration points are obtained in each batch. After all eight points have obtained finite CD responses, the pre-selected single-objective regression model is updated, and R-values ​​are monitored using in-training cross-validation or independent validation data. 2 And RMSE.

[0093] This embodiment outputs a surrogate model from eight-dimensional configuration parameters to a single aerodynamic response CD, and saves the target randomization state, sequence first index, physical domain precision, used training coordinates, candidate score records, and records of each batch of valid configuration points.

[0094] In this embodiment of the invention, a numerical verification example using 20-dimensional coordinate calculation is also provided.

[0095] Spatial filling properties of conventional Halton sequences and pre-screened randomly perturbed Halton sequences were compared in a 20-dimensional unit hypercube. Both types of sequences generated 200 sample points, skipping the first all-zero point. Evaluation metrics included the centered L2 dissimilarity of all 190 two-dimensional projections, the maximum two-dimensional projection dissimilarity, and the maximum inter-dimensional absolute correlation coefficient.

[0096] The random perturbation sequence employs non-zero numerical permutations and dimension-wise cyclic translation. Equal-length sequence prefixes are generated for random seeds 1 to 1000, sorted according to the average dissimilarity of 190 two-dimensional projections, and random seed 67 is selected as the target state. The entire pre-screening process uses sample coordinates as input and does not invoke configurational responses. For ease of illustration, the following text and... Figure 2 , Figure 3 The randomly perturbed Halton sequence selected and locked into the target randomized state according to the above candidate seed pre-screening rules is referred to as SORSH. SORSH corresponds to the target sequence obtained in this embodiment by using non-zero numerical permutation and dimension-wise cyclic translation, and pre-screening by L2 difference centered by all two-dimensional projections; it is not an independent algorithm distinct from the aforementioned pre-screening random perturbed Halton method. Halton in the figure represents the traditional Halton sequence without random perturbation.

[0097] The worst two-dimensional projection of traditional Halton occurs in the 17th and 18th dimensions. Figure 2 Give the sample distribution of this projection. Figure 3 Summarize the differences of all 20-dimensional projections.

[0098] like Figure 2 As shown, traditional Halton samples are concentrated on a few narrow parallel bands, with most areas of the projection plane lacking samples. Pre-screened random perturbation Halton breaks up inter-dimensional correlations, allowing samples to cover a wider range of variable combinations. The centered L2 dissimilarity of the 17th and 18th dimension projections decreased from 0.012235 to 0.000198, a reduction of approximately 98.4%. With reduced band clustering, the sampling of configurational variable combinations becomes more balanced.

[0099] like Figure 3 As shown, after statistical analysis of all 190 two-dimensional projections, the significant dissimilarity of traditional Halton's algorithm is concentrated in high-dimensional variable combinations. The pre-screened random perturbation Halton's algorithm exhibits lower overall dissimilarity and a more balanced distribution. The average centered L2 dissimilarity decreased from 0.001819 to 0.000142, a reduction of approximately 92.2%; the maximum two-dimensional projection dissimilarity decreased by approximately 97.0%; and the maximum inter-dimensional absolute correlation coefficient decreased from 0.6769 to 0.1668. This method reduces the dissimilarity of each variable to the projection, improves the overall coverage of the high-dimensional design space, and supplements the training data for the variable interaction region of the small-sample surrogate model.

[0100] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0101] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for constructing an air-to-air missile proxy model based on projection difference, characterized in that, Includes the following steps: S100. Determine the physical configuration domain and single-objective response of the surrogate model, normalize the configuration variables in the physical configuration domain, determine the candidate seed set, and set the new sampling budget. S200. For each candidate seed in the candidate seed set, generate an equal-length random perturbation Halton sequence prefix according to the newly added sampling budget, calculate the projection quality index based on the unit domain coordinates of the random perturbation Halton sequence prefix, and lock the global random perturbation Halton sequence corresponding to the target randomization state. S300: According to the batch plan, read the global random perturbation Halton sequence by continuous global subscript, map the read unit domain coordinates to the physical configuration domain, and determine the candidate configuration points through capacity check. Filter out the candidate configuration points that are repeated with the unit domain coordinates received in the historical training samples or the current batch, and continue to read the global subscript until the required number of valid configuration points for the current batch are obtained. S400: Perform numerical evaluation on the effective configuration points obtained in the current batch, obtain the finite scalar target response corresponding to each effective configuration point, and combine the effective configuration points and their target responses into new training data and add them to the training samples. S500. Update or train a pre-selected single-objective regression model using the updated training samples, and evaluate whether it meets the preset accuracy condition, stagnation condition or sample budget condition. If so, then obtain and output the target proxy model, the current training sample, the candidate score record, and the reproduction configuration; If not, return to S300 and add valid configuration points while keeping the target randomization state unchanged.

2. The method for constructing an air-to-air missile proxy model based on projection difference degree according to claim 1, characterized in that, In step S200, an equal-length random perturbation Halton sequence prefix is ​​generated according to the newly added sampling budget, including: Based on the number of configuration variables Predetermine the order to ascending. A prime number; For each candidate seed in the candidate seed set Determine the dimensional index arrangement Digital substitution and translation amount Construct candidate randomized states ; Based on the candidate randomization state, the unit domain coordinates of each index of the randomly perturbed Halton sequence are determined by sequentially performing cardinality expansion, digital permutation, perturbation root inversion, cyclic shift, and modulo 1. This generates candidate seeds. The corresponding length is the random perturbation Halton sequence prefix of the newly added sampling budget. , , ; For sequence index, To increase the sampling budget, For configuration variable dimensions, The number of configuration variables; When radix allocation rearrangement is enabled, prime radixes, equal to the number of configuration variables, are obtained in ascending order. A random number generator is initialized with a candidate seed, and the random number generator performs a random permutation without replacement on the dimension index set to obtain the dimension index arrangement. , making the first Each configuration variable is arranged using the index dimension. The Middle The prime base indicated by the term, and each prime base is assigned only once in each dimension; When radix-based rearrangement is not enabled, make the 1st The configuration variable uses the first A prime number base.

3. The method for constructing an air-to-air missile proxy model based on projection difference as described in claim 1, characterized in that, In S200, the globally random perturbation Halton sequence corresponding to the target randomization state includes: The centered L2 dissimilarity index is calculated for all two-dimensional unit domain coordinate projections of each random perturbation Halton sequence prefix, and the average two-dimensional projection centered L2 dissimilarity index is determined. The candidate randomization states are sorted by average two-dimensional projection-centered L2 dissimilarity index from smallest to largest. The smallest candidate randomization state is determined as the target randomization state, and the corresponding global random perturbation Halton sequence is locked.

4. The method for constructing an air-to-air missile proxy model based on projection difference degree according to claim 3, characterized in that, In step S200, when there are parallel minimum average two-dimensional projection-centered L2 dissimilarity indices, the maximum two-dimensional projection-centered L2 dissimilarity index, the full-dimensional centered L2 dissimilarity index, the maximum absolute Pearson correlation coefficient, and the candidate seed are compared sequentially for each random perturbation Halton sequence prefix. The candidate randomization state that first appears with a unique minimum value is selected as the target randomization state, and the corresponding global random perturbation Halton sequence is locked.

5. The method for constructing an air-to-air missile proxy model based on projection difference degree according to claim 1, characterized in that, In S200, the globally random perturbation Halton sequence corresponding to the target randomization state includes: For each random perturbation Halton sequence prefix, calculate the projection quality index, including the average two-dimensional projection-centered L2 dissimilarity index, the maximum two-dimensional projection-centered L2 dissimilarity index, the full-dimensional centered L2 dissimilarity index, and the maximum absolute Pearson correlation coefficient; Under the same random perturbation Halton sequence prefix, the range of each projection quality index is normalized; The weight of each projection quality index is obtained by dividing the preset non-negative priority coefficient by the sum of the non-negative priority coefficients. Based on the weights of the projection quality indicators, the normalized projection quality indicators are weighted and summed to obtain the comprehensive projection difference score. The comprehensive projection difference scores are sorted in ascending order, and the candidate randomization state corresponding to the minimum comprehensive projection difference score is identified as the target randomization state.

6. The method for constructing an air-to-air missile proxy model based on projection difference degree according to claim 1, characterized in that, S300 includes the following steps: Determine the global index of the starting point of the current batch, and read the global index sequentially from the globally random perturbation Halton sequence; For each global index read, calculate the corresponding unit field coordinates according to prime number radix allocation, number permutation, and translation. Map the unit domain coordinates to the physical configuration domain and determine candidate configuration points through a capacity check; Determine whether the unit domain coordinates of the candidate configuration points are duplicates of the unit domain coordinates of historical training samples or those already received in the current batch; If so, the candidate configuration point is duplicated, and it is rejected as a valid configuration point in the current batch. The next global coordinate is then read. If not, the candidate configuration point is not repeated and is accepted as a valid configuration point in the current batch; Repeatedly execute global index reading, physical configuration domain mapping, constraint and repeated judgment until the required number of valid configuration points for the current batch is obtained.

7. The method for constructing an air-to-air missile proxy model based on projection difference as described in claim 6, characterized in that, Mapping unit domain coordinates to the physical configuration domain is represented as: In the formula, Global coordinates The candidate configuration point in the corresponding physical configuration domain is at the th Coordinates on each configuration variable The respective The physical lower and upper boundaries of each configuration variable. For the target randomized state In global coordinates , No. Unit field coordinates in the dimension.

8. The method for constructing an air-to-air missile proxy model based on projection difference degree according to claim 6, characterized in that, In S300, the candidate configuration point is determined through capacity checking, including: For configuration points mapped to the physical configuration domain, perform: Read the physical lower boundary and physical upper boundary of the configuration point dimension by dimension, and determine whether there is any boundary that is not finite or the physical upper boundary is not greater than the physical lower boundary; if so, stop determining the valid configuration points of the current batch. If not, then the unit domain coordinates of the configuration points are quantized, including: Set the quantization precision, determine the quantization interval based on the decimal quantization precision, round up the lower physical boundary of each dimension of the configuration point and round down the upper physical boundary to determine the minimum and maximum allowable grid points located on the original physical boundary; Determine whether there exists a minimum allowed grid point in any dimension that is greater than the maximum allowed grid point; If so, it is determined that there are no representable coordinates under the current quantization precision, and the determination of valid configuration points for the current batch is stopped; If not, the physical coordinates of the mapped configuration points are quantized according to the decimal rounding rule and restricted to between the minimum and maximum allowable grid points to complete the coordinate quantization. For configuration points that have completed coordinate quantization, calculate the product of the number of allowed grid points in each dimension as the total number of representable grid points in the physical configuration domain, and compare it with the number of already occupied grid points and the number of effective configuration points required for the current batch to determine whether the capacity condition is met. If so, then this configuration point is taken as a candidate configuration point; If not, then stop determining the valid configuration points for the current batch.

9. The method for constructing an air-to-air missile proxy model based on projection difference degree according to claim 6, characterized in that, when At that time, the capacity condition is met: In the formula, The number of representable lattice points within the physical configuration domain. This refers to the number of unit domain coordinates already received in historical training samples or the current batch, the number of unit domain coordinates reserved for verification, and the number of unit domain coordinates already used for testing. The number of valid configuration points required for the current batch. and These represent the minimum and maximum allowable grid points within the lower boundary of the quantization precision, respectively. This is the quantization interval.