Rapid target redistribution method for patrol ammunition cluster combat

Abstract

The invention provides a rapid target redistribution method for patrolling ammunition cluster combat, and relates to the technical field of patrolling ammunition cluster cooperative combat. The invention relates to an agent-assisted greedy algorithm. In order to ensure the prediction performance of the agent model, a learning algorithm and a special algorithm for feature extraction and model management are introduced; meanwhile, in order to improve the optimization process efficiency, a special greedy algorithm is provided. The method has a plurality of remarkable advantages. Firstly, the prediction precision of the agent model is superior to that of other algorithms. And secondly, the adopted greedy algorithm has obvious advantages in the aspect of optimization efficiency, and the algorithm can efficiently solve the task allocation problem. The advantages show that the algorithm has high practical value under the background of large-scale tasks, and has reliability and feasibility in practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of cruise missile cluster coordinated combat, and in particular to a cruise missile cluster combat rapid target reallocation method. Background Art

[0002] In large-scale air combat, there are many types of targets and a large number of them, coupled with a powerful defense system, which makes the types and number of threat sources continue to increase. These factors make it difficult for the current single-soldier cruise missile combat mode to adapt to the battlefield environment. In this context, the concept of cruise missile cluster combat has been proposed. It is particularly important to use the collaborative ability of multiple cruise missiles to perform complex tasks and assign appropriate attack targets to each cruise missile (target allocation) to maximize the effectiveness of cluster combat. However, most of the existing target allocation technologies are aimed at static battlefield environments, and there is no special design for dynamic battlefield environments, resulting in poor practicality. In a complex battlefield environment, whether it is the attack target and threat source, or our cruise missiles, there is uncertainty, and the optimal allocation plan may change continuously. An efficient target reallocation technology is very necessary.

[0003] The essence of target allocation is a combinatorial optimization problem. As the number of targets and threats increases, the search space grows exponentially, making it more difficult to find the optimal allocation solution. Existing methods for target allocation problems include exact algorithms, meta-heuristic algorithms, and intelligent algorithms. Exact algorithms can produce optimal solutions in most cases, but are inefficient in large-scale and real-time tasks. Heuristic algorithms can produce satisfactory solutions, but often require manual parameter adjustment and have limited applicability to real-time tasks. In recent years, intelligent algorithms such as deep reinforcement learning have shown great prospects, but their training process is very time-consuming and has obvious defects in robustness. In addition, existing algorithms often focus only on optimizing the search process, while ignoring the important impact of the "expensive target" problem on computational efficiency. Expensive targets are a typical type of problem in the field of optimization, specifically referring to the situation where the value of the objective function is difficult to obtain. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for rapid target reallocation in cruise missile cluster combat in view of the deficiencies of the above-mentioned prior art, which is used for target reallocation in cruise missile cluster combat in a complex battlefield environment, and solves the problems of low efficiency and failure to consider expensive targets in large-scale target allocation problems.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A method for rapid target reallocation in cruise missile cluster combat comprises the following steps:

[0007] Step 1: Generate a training dataset to simulate various battlefield scenarios by changing the number, location, and radius of threat sources as input data;

[0008] Step 2: Feature extraction, that is, extracting information about the target, cruise missile, and threat source. In the feature extraction stage, the A* algorithm is used as the path optimization algorithm, which includes:

[0009] Step 2.1: Take the straight line from the starting point to the end point of the cruise missile as the baseline of the trajectory, define the threat source intersecting the baseline as the dominant threat source, i.e. DTS, and define the remaining threat sources as non-dominant threat sources, i.e. NDTS;

[0010] Determine whether the baseline intersects with the threat source;

[0011] Step 2.2: If the baseline does not intersect any threat area, the A* algorithm will find the baseline source that does not intersect any threat area as the actual trajectory;

[0012] Step 2.3: If the baseline intersects the threat area, a curved path is used to bypass the threat area, and the trajectory length is greater than the baseline length;

[0013] Step 2.3.1: If the baseline intersects a single DTS, the threat radius and the distance from the center of the DTS to the baseline are used as input features, and the threat coefficient index is defined as:

[0014]

[0015] Among them, λ represents the threat coefficient index, r and d represent the threat radius and the distance from the DTS center to the baseline respectively, and λ∈[0,1] is the threat coefficient, which represents the impact on the trajectory increment;

[0016] Step 2.3.2: If the baseline intersects multiple DTSs, use the average value as the input feature, as follows:

[0017]

[0018]

[0019] in, is the set of all DTSs, N DT is the number of DTS; R DT and λ DT Represent the average threat radius and average threat coefficient respectively; the input features include the number of DTS N in addition to the above average values DT ;

[0020] Step 2.4: There are three types of intersections between the threat sources, namely, the intersection between different DTSs, the intersection between different NDTSs, and the intersection between DTS and NDTS. The first two types of intersections are called homogeneous intersections, and the third type of intersection is called heterogeneous intersections.

[0021] In uniform DTS intersection, the DTS that intersects the baseline is defined as the core DTS; the DTS whose center is on the same side of the baseline as the center of the core DTS is called the positive DTS; the DTS whose center is on the other side of the baseline relative to the center of the core DTS is called the negative DTS;

[0022] When the positive DTS intersects with the core DTS, the impact of the positive DTS is adjusted by reducing the threat radius of the positive DTS, as shown in the following formula:

[0023]

[0024] Among them, r j ′ is the adjusted threat radius, r j is the original threat radius, r i is the threat radius of the core DTS; α j is the attenuation coefficient, which means that the closer to the core DTS, the greater the attenuation of the threat radius;

[0025] For negative DTS to reduce the trajectory increment, not only the threat radius needs to be modified, but also the threat coefficient λ needs to be adjusted. The adjusted threat coefficient is defined as:

[0026]

[0027] Among them, λ′∈[1,2] is the adjusted threat coefficient, which represents the impact on the trajectory increment, r and d represent the threat radius and the distance from the DTS center to the baseline, respectively;

[0028] For the scenario where multiple DTSs intersect, a sequential modification algorithm is proposed based on steps 2.3 and 2.4;

[0029] Step 2.5: In the heterogeneous intersection scenario, when the threat source closest to the starting point is the DTS, the search direction of the A* algorithm depends on the position of the DTS center. Based on the heuristic search mechanism of A*, the opposite side of the DTS center is the optimal direction, and the features of the NDTS are extracted.

[0030] Step 2.6: After all NDTS are identified, relevant features are extracted from the information of all NDTS;

[0031] Step 3: Establish an integrated agent model, which includes the following steps:

[0032] Step 3.1: Use linear regression, ridge regression, support vector regression, CART regression tree and BP neural network to build a heterogeneous ensemble model, and learn the parameters of each regression algorithm through cross-validation;

[0033] Step 3.2: Propose the concept of dynamic integration and construct a dynamic fusion algorithm;

[0034] Step 3.2.1: Given the test data, determine the ability region of the test data from the training set using the k-nearest neighbor method or clustering method;

[0035] Step 3.2.2: Calculate the performance of all base learners on the capability region determined in step 3.2.1 to describe the capability of the base learners on the test data;

[0036] Step 3.2.3: Select the most competitive base learner for prediction;

[0037] Step 4: Propose a voting scheme as a model management strategy;

[0038] The performance of each base learner on the training set is defined as the capability baseline. Given a test data, the capability of each base learner is compared with its capability baseline. If the capability is lower than the capability baseline, it indicates that the base learner believes that the test data is uncertain. After obtaining the results of all base learners, a majority vote is used to obtain the final selection result. If the voting result confirms that a specific sample shows promise, the algorithm A* is used to obtain its true trajectory. The obtained samples are merged into the training set for subsequent updating of the base learners.

[0039] Step 5: Under the premise of the assistance of the proxy model, a fast target allocation algorithm based on the greedy search idea is proposed;

[0040] Two scenarios are defined: the number of cruise missiles is greater than the number of targets and the number of cruise missiles is less than the number of targets. A fast target allocation algorithm based on the greedy search idea allocates the most suitable cruise missile to each target, and in the scenario where the number of cruise missiles is greater than the number of targets, checks are made to see if there are more suitable targets to place the remaining cruise missiles.

[0041] Furthermore, in step 2.4, for the scenario where multiple DTSs intersect, the proposed sequence modification algorithm is specifically as follows:

[0042] Sort all DTSs in ascending order according to their distance from the starting point;

[0043] Assume that one DTS is the core DTS, and determine whether the remaining DTSs are positive or negative by checking whether the line segment formed by the center of each DTS and the center of the core DTS intersects the baseline;

[0044] If it is a positive DTS, then only the threat radius needs to be adjusted;

[0045] If it is a negative DTS, the threat radius and threat coefficient need to be adjusted.

[0046] Furthermore, in step 2.5, in the heterogeneous crossover scenario, the specific steps for feature extraction are as follows:

[0047] Step 2.5.1: Find the threat source closest to the starting point and define it as TS*;

[0048] Step 2.5.2: If TS* is NDTS, determine that the other side of the baseline is the valid NDTS, defined as NDTS+; calculate the threat radius and threat coefficient of all NDTS+, and modify the threat radius of all NDTS+ using the sequential modification algorithm;

[0049] Step 2.5.3: If TS* is DTS, identify the valid NDTS as NDTS+ according to the NDTS center, DTS center, and baseline; calculate the threat radius and threat coefficient of all NDTS+, and modify the threat radius of all NDTS+ using the sequential modification algorithm.

[0050] Furthermore, the specific method of step 5 is as follows:

[0051] Step 5.1: For the updated scenario information, use the feature extraction in step 2 to obtain the input features acceptable to the surrogate model;

[0052] Step 5.2: Combine the trajectory length matrix output by the surrogate model with the target value to obtain the cost matrix C;

[0053] Step 5.3: Apply the greedy algorithm to the cost matrix C to obtain the final allocation result, represented by the decision matrix D;

[0054] Step 5.4: Assume that the current situation information includes M targets and N loitering munitions. The target allocation optimization problem is divided into two different forms: N>M and N<M; it is required that all available loitering munitions must perform attack tasks, and each target can be allocated at most one loitering munition. The constraint conditions are as follows:

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] Among them, x ijis a decision variable, which takes 1 when the j-th target is assigned to the i-th loitering munition, and 0 otherwise; when the i-th loitering munition attacks the j-th target, L ij , T ij , G ij represent the ballistic length, threat level, and attack gain respectively; assuming that all loitering munitions can complete the attack, G ij can be represented by the value of the j-th target;

[0061] Define the trajectory cost as:

[0062] C ij = αL ij -(1-α)G ij

[0063] where α is a balance parameter that controls the balance between the trajectory length and the target value of the trajectory cost;

[0064] Step 5.5: When the number of loitering munitions is greater than the number of targets, i.e., N>M, the greedy algorithm is represented by the following steps:

[0065] Step 5.5.1: Initialize the decision matrix D by setting all elements to zero;

[0066] Step 5.5.2: Find the minimum value and its index in the cost matrix C, and set all elements in the row and column where the minimum value is located to infinity;

[0067] Step 5.5.3: Set the element in the decision matrix D corresponding to the above index to 1;

[0068] Step 5.5.4: Repeat steps 5.5.2 and 5.5.3 M times, and let the resulting decision matrix be D', with each target assigned the most suitable loitering munition;

[0069] Step 5.5.5: Find the row subscript of the element with a value of 1 in the matrix D', and remove the row corresponding to this row subscript from the original cost matrix C to obtain the matrix C';

[0070] Step 5.5.6: Find the minimum value and its index in each row of the matrix C'; if the obtained minimum value is less than 0, set the corresponding element in the matrix D' to 1;

[0071] Step 5.6: When the number of loitering munitions is less than the number of targets, i.e., N<M, the greedy algorithm is represented by the following steps:

[0072] Step 5.6.1: Initialize the decision matrix D by setting all elements to zero;

[0073] Step 5.6.2: Find the minimum value and its index in the cost matrix C, and set all elements in the row and column where the minimum value is located to infinity;

[0074] Step 5.6.3: Set the element corresponding to the above index in the decision matrix D to 1;

[0075] Step 5.6.4: Repeat step 5.6.2 and step 5.6.3 N times.

[0076] The beneficial effects of adopting the above technical solution are: the rapid target reallocation method for cruise missile cluster combat provided by the present invention is used for the reallocation of cruise missile cluster combat targets in complex battlefield environments, and provides an agent-assisted greedy algorithm based on cruise missile cluster target allocation to solve the problems of low efficiency and failure to consider expensive targets in large-scale target allocation problems. In order to ensure the predictive performance of the proxy model, special algorithms for learning algorithms, feature extraction and model management are introduced. At the same time, in order to improve the efficiency of the optimization process, a special greedy algorithm is proposed, which has several significant advantages. First, the prediction accuracy of its proxy model is better than other algorithms. Secondly, the adopted greedy algorithm has obvious advantages in optimization efficiency, and the algorithm can efficiently solve the task allocation problem. These advantages show that the algorithm has high practical value in the context of large-scale tasks and is reliable and feasible in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 The overall framework of the proxy model-assisted greedy algorithm provided by the embodiment of the present invention;

[0078] Figure 2 A feature extraction flow chart provided for an embodiment of the present invention;

[0079] Figure 3 The trajectory diagram generated by the A* algorithm in the scenario where the baseline provided by the embodiment of the present invention intersects with a single threat source; wherein, Figure (3a) is the trajectory diagram generated by the A* algorithm in the scenario where the DTS center is on the right side of the baseline, and Figure (3b) is the trajectory diagram generated by the A* algorithm in the scenario where the DTS center is on the left side of the baseline;

[0080] Figure 4 The trajectory diagram generated by the A* algorithm in the scene where different DTSs intersect with each other provided by the embodiment of the present invention; wherein, Figure (4a) shows the intersection of two independent DTSs, and Figures (4b) and (4c) show the intersection of two DTSs at different locations;

[0081] Figure 5 Trajectory diagrams of two heterogeneous intersection scenarios provided by an embodiment of the present invention; wherein, FIG. (5a) is a trajectory diagram generated by the A* algorithm when the threat source closest to the starting point is NDTS, and FIG. (5b) is a trajectory diagram generated by the A* algorithm when the threat source closest to the starting point is DTS;

[0082] Figure 6A schematic diagram of dynamic integration selection provided by an embodiment of the present invention;

[0083] Figure 7 A flowchart of a voting-based model management strategy provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0084] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0085] like Figure 1 As shown, the method of this embodiment is described as follows.

[0086] Step 1: Generate training data set: Various battlefield scenarios are simulated by changing the number, location and radius of threat sources. All of this information constitutes the input data. In each simulated scenario, the A* algorithm is used to generate trajectories and calculate the corresponding length as its output.

[0087] Step 2: Feature extraction: Extract information about targets, cruise missiles, and threat sources. Get informative features from the original scenario information as input to the machine learning algorithm. The feature extraction flowchart is shown below: Figure 2 shown.

[0088] Step 2.1: Take the straight line from the starting point to the end point of the cruise missile as the baseline of the trajectory, define the threat source that intersects with the baseline as the dominant threat source (DTS), and define the remaining threat sources as non-dominant threat sources (NDTS). Determine whether the baseline intersects with the threat area.

[0089] Step 2.2: If the baseline does not intersect any threat region, the A* algorithm will find the baseline source that does not intersect any threat region as the actual trajectory.

[0090] Step 2.3: If the baseline intersects the threat area, a curved path is used to bypass the threat area, and the trajectory length is greater than the baseline length.

[0091] Step 2.3.1: Consider the scenario where the baseline intersects with a single threat source. For each DTS, the main factor affecting the trajectory increment is its threat radius. A larger threat radius may lead to a larger trajectory increment. The intersection position between the baseline and the DTS also affects the size of the increment.

[0092] Under a certain threat radius, the closer the baseline is to the DTS center, the larger the corresponding trajectory increment. If the baseline intersects a single DTS, the threat radius and the distance from the DTS center to the baseline are used as input features, and the threat coefficient index is defined as:

[0093]

[0094] Among them, r and d represent the threat radius and the distance from the DTS center to the baseline, respectively, and λ∈[0,1] is the threat coefficient, which represents the impact on the trajectory increment. The trajectory diagram generated by the A* algorithm in the scenario where the baseline intersects with a single threat source is as follows: Figure 3 shown.

[0095] Step 2.3.2: In a complex environment with multiple DTS, assume that all threat sources are independent of each other and do not have any intersection. In different scenarios, the number of DTS will vary. Integrate the information of all DTS to obtain a comprehensive feature. If the baseline intersects with multiple DTS, use the average value as the input feature, as follows:

[0096]

[0097]

[0098] in, is the set of all DTSs, N DT is the number of DTS; R DT and λ DT Represent the average threat radius and average threat coefficient respectively; the input features include the number of DTS N in addition to the above average values DT .

[0099] Step 2.4: There are three types of intersections between threat sources, namely, the intersection between different NDTS, the intersection between different DTS, and the intersection between DTS and NDTS. The first two types of intersections are called homogeneous intersections, and the third type of intersection is called heterogeneous intersections. The trajectory diagrams generated by the A* algorithm in the latter two scenarios are shown in Figure 2. Figure 4 shown.

[0100] Step 2.4.1: In the case of uniform NDTS intersection, since all threat sources do not intersect the baseline, their effect on the trajectory length is almost negligible.

[0101] Step 2.4.2: In the uniform DTS intersection, the DTS that intersects with the baseline determines the initial direction of the trajectory when facing the entire threat area. This special DTS is called the core DTS. When the center of the DTS is on the other side of the baseline relative to the center of the core DTS, the influence of the uniform intersection on the increment is enhanced, and this type of DTS is called negative DTS; when the center of the DTS is on the same side of the baseline as the center of the core DTS, the influence of the uniform intersection on the increment is weakened, and this type of DTS is called positive DTS. The trajectory diagram generated by the A* algorithm in this scenario is as follows Figure 4 As shown in Figure (4a), two independent DTSs are shown. In Figure (4b) and Figure (4c), the two DTSs intersect, but at different locations.

[0102] Step 2.4.3: When the positive DTS intersects with the core DTS, the impact on the trajectory increment is weakened. The impact of the positive DTS is adjusted by reducing the threat radius of the positive DTS, as shown in the following formula:

[0103]

[0104] Among them, r j ′ is the adjusted threat radius, r j is the original threat radius, r i is the threat radius of the core DTS; α j is the attenuation coefficient, indicating that the closer to the core DTS, the greater the attenuation of the threat radius. For negative DTS to reduce the trajectory increment, not only the threat radius needs to be modified, but also the threat coefficient λ needs to be adjusted.

[0105]

[0106] The original coefficient range is [0,1], and the higher the value, the stronger the impact. After adjustment, the range becomes [1,2], but the principle remains the same: the larger the value, the greater the impact.

[0107] Step 2.4.4: For the scenario of designing multiple DTS homogeneous crossover, a sequential modification algorithm is proposed:

[0108] Sort all DTSs in ascending order according to their distance from the starting point;

[0109] Assume that one DTS is the core DTS, and determine whether the remaining DTSs are positive or negative by checking whether the line segment formed by the center of each DTS and the center of the core DTS intersects the baseline;

[0110] If it is a positive DTS, then only the threat radius needs to be adjusted;

[0111] If it is a negative DTS, the threat radius and threat coefficient need to be adjusted.

[0112] Step 2.5: In heterogeneous crossover, the intersection of NDTS and DTS will also affect the trajectory increment. Figure 5 Trajectory diagram of the scene with two heterogeneous intersections.

[0113] Step 2.5.1: When the threat source closest to the starting point is NDTS, the algorithm A* will not search for a path on that side. Therefore, all NDTS on this side will not have any impact on the track. This type of NDTS is called invalid NDTS. The trajectory diagram generated by the A* algorithm is shown in Figure (5a).

[0114] Step 2.5.2: When the threat source closest to the starting point is a DTS, the search direction of A* depends on the position of the center of the DTS. Based on the heuristic search mechanism of A*, the opposite side of the DTS center is the optimal direction. Therefore, all NDTS on this side are considered to be valid NDTS. The trajectory diagram generated by the A* algorithm is shown in Figure (5b).

[0115] Step 2.5.3: In heterogeneous cross-scenario, perform feature extraction on all NDTS:

[0116] Step 2.5.1: Find the threat source closest to the starting point, defined as TS*;

[0117] Step 2.5.2: If TS* is NDTS, determine that the other side of the baseline is a valid NDTS, defined as NDTS+; calculate the threat radius and threat coefficient of all NDTS+, and modify the threat radius of all NDTS+ using the sequential modification algorithm;

[0118] Step 2.5.3: If TS* is DTS, then identify the valid NDTS as NDTS+ based on the NDTS center, DTS center and baseline; calculate the threat radius and threat coefficient of all NDTS+, and modify the threat radius of all NDTS+ using the sequential modification algorithm.

[0119] Step 2.6: After all NDTS are identified, extract relevant features from their information. Similar to the features related to DTS, consider the average threat radius, average threat factor, and number of NDTS. If the NDTS does not intersect the baseline, use the distance from the center to the baseline to calculate the threat factor.

[0120] Step 3: Select the most appropriate learning algorithm as the agent model and establish an integrated agent model.

[0121] Step 3.1: Use linear regression, ridge regression, support vector regression, CART regression tree and BP neural network to build a heterogeneous ensemble model, and learn the parameters of each regression algorithm through cross-validation. By combining different learners, diversity is enhanced.

[0122] Step 3.2: Propose the concept of dynamic integration and construct a dynamic fusion algorithm. The dynamic integration selection diagram is as follows: Figure 6 shown.

[0123] Step 3.2.1: Given the test data, determine the capability region of the test data from the training set using the k-nearest neighbor method or clustering method. The capability region is considered as a validation set to evaluate the base learner relative to the capability region test data. In the field of dynamic ensembles, two commonly used methods for determining capability regions are the k-nearest neighbor method and the clustering method.

[0124] Step 3.2.2: Calculate the performance of all base learners on the ability region determined in Step 3.2.1 to describe the ability of the base learners for this test data.

[0125] Step 3.2.3: Select the most competitive base learner for prediction.

[0126] Step 4: Propose a voting scheme as a model management strategy, and its process is as Figure 7 shown.

[0127] Step 4.1: Assume that the base learners have been obtained using the training set. Obtain the performance of each base learner on the training set, which is defined as the ability baseline.

[0128] Step 4.2: Given a test data, compare the ability of each base learner with its ability baseline. If the ability is lower than its baseline, it indicates that the base learner considers this test data to have a high degree of uncertainty.

[0129] Step 4.3: After obtaining the results of all base learners, use majority voting to obtain the final selection result.

[0130] Step 4.4: If the voting result confirms that a specific sample shows promise, use algorithm a* to obtain its true trajectory.

[0131] Step 4.5: Incorporate the obtained samples into the training set for subsequent update of the base learners.

[0132] Step 5: Under the premise of assisted by the surrogate model, propose a fast target assignment algorithm based on the greedy search idea.

[0133] Step 5.1: For the updated scenario information, use the feature extraction in the previous Step 2 to obtain the input features acceptable to the surrogate model.

[0134] Step 5.2: The surrogate model will output a trajectory length matrix. Combine this matrix with the target value to obtain the cost matrix C.

[0135] Step 5.3: Apply the greedy algorithm to the cost matrix C to obtain the final assignment result, represented by the decision matrix D.

[0136] Step 5.4: Assume that the current situation information includes M targets and N loitering munitions. The target assignment optimization problem can be divided into two different forms: N>M and N<M. It is required that all available loitering munitions must perform attack tasks, and each target can be assigned at most one loitering munition. The constraint conditions are as follows:

[0137]

[0138]

[0139]

[0140]

[0141]

[0142] Among them, x ij is a decision variable. When the jth target is assigned to the i-th cruise missile, the decision variable takes 1, otherwise it takes 0. When the i-th cruise missile attacks the jth target, L ij 、T ij , G ij Represents trajectory length, threat level and attack gain respectively; Assuming that all cruise missiles can complete the attack, G ij It can be expressed by the value of the jth target;.

[0143] Define the trajectory cost as:

[0144] C ij =αL ij -(1-α)G ij

[0145] Among them, α is a balance parameter between the control trajectory length and trajectory cost target value.

[0146] Step 5.5: When the number of cruise missiles is greater than the number of targets (N>M), the greedy algorithm is represented by the following steps:

[0147] Step 5.5.1: Initialize the decision matrix D by setting all elements to zero;

[0148] Step 5.5.2: Find the minimum value and its index in the cost matrix C, and set all elements in the row and column where the minimum value is located to infinity to prevent the target from occupying too many cruise missiles, thereby causing other targets to be ignored;

[0149] Step 5.5.3: Set the element corresponding to the above index in the decision matrix D to 1;

[0150] Step 5.5.4: Repeat steps 5.5.2 and 5.5.3 M times, and let the resulting decision matrix be D'. Each target is equipped with the most appropriate cruise missile;

[0151] Step 5.5.5: Find the row subscript of the element with a value of 1 in the matrix D', remove the row corresponding to the row subscript from the original cost matrix C, and obtain the matrix C';

[0152] Step 5.5.6: Find the minimum value and index of each row in matrix C'; if the minimum value is less than 0, set the corresponding element in matrix D' to 1;

[0153] The purpose of the first four steps is to match the most suitable loitering munition for each target, and the purpose of the last two steps is to check if there are more suitable targets to place the remaining loitering munitions to achieve higher combat efficiency.

[0154] Step 5.6: When the number of loitering munitions is less than the number of targets (N < M), the greedy algorithm is represented by the following steps:

[0155] Step 5.6.1: Initialize the decision matrix D by setting all elements to zero;

[0156] Step 5.6.2: Find the minimum value and its index in the cost matrix C, and set all elements in the row and column where the minimum value is located to infinity;

[0157] Step 5.6.3: Set the element in the decision matrix D corresponding to the above index to 1;

[0158] Step 5.6.4: Repeat steps 5.6.2 and 5.6.3 N times.

[0159] Although this operation is used in the greedy algorithms of both step 5.5 and step 5.6, steps 5.5.5 and 5.5.6 in step 5.5 ensure that high-value targets are assigned to multiple loitering munitions, thus achieving higher combat efficiency.

[0160] In this embodiment, Figure 3 、 Figure 4 、 Figure 5 The figures respectively show the trajectory generated by the A* algorithm in the scenario where the baseline intersects with a single threat source, the trajectory generated by the A* algorithm in the scenario where different DTSs intersect, and the trajectory diagram of the scenario of two heterogeneous intersections. It can be seen from the figures that the method of the present invention has better prediction accuracy of the proxy model than other algorithms, and has obvious advantages in optimization efficiency, and can efficiently solve the task assignment problem.

[0161] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present invention.

Claims

1. A method for rapid target reallocation in cruise missile swarm combat, characterized in that: The following steps are involved: Step 1: Generate a training dataset to simulate various battlefield scenarios by changing the number, location, and radius of threat sources as input data; Step 2: Feature extraction, that is, extracting information about the target, cruise missile, and threat source. In the feature extraction stage, the A* algorithm is used as the path optimization algorithm; Step 3: Establish an integrated agent model, which includes the following steps: Step 3.1: Use linear regression, ridge regression, support vector regression, CART regression tree and BP neural network to build a heterogeneous ensemble model, and learn the parameters of each regression algorithm through cross-validation; Step 3.2: Propose the concept of dynamic integration and construct a dynamic fusion algorithm; Step 3.2.1: Given the test data, determine the ability region of the test data from the training set using the k-nearest neighbor method or clustering method; Step 3.2.2: Calculate the performance of all base learners on the capability region determined in step 3.2.1 to describe the capability of the base learners on the test data; Step 3.2.3: Select the most competitive base learner for prediction; Step 4: Propose a voting scheme as a model management strategy; The performance of each base learner on the training set is defined as the capability baseline. Given a test data, the capability of each base learner is compared with its capability baseline. If the capability is lower than the capability baseline, it indicates that the base learner believes that the test data is uncertain. After obtaining the results of all base learners, a majority vote is used to obtain the final selection result. If the voting result confirms that a specific sample shows promise, the algorithm A* will be used to obtain its true trajectory. The obtained samples are merged into the training set for subsequent updating of the base learners. Step 5: Under the premise of the assistance of the proxy model, a fast target allocation algorithm based on the greedy search idea is proposed; Two scenarios are defined: the number of cruise missiles is greater than the number of targets and the number of cruise missiles is less than the number of targets. A fast target allocation algorithm based on the greedy search idea allocates the most suitable cruise missile to each target, and in the scenario where the number of cruise missiles is greater than the number of targets, checks are made to see if there are more suitable targets to place the remaining cruise missiles.

2. The method for rapid target reallocation in cruise missile swarm combat according to claim 1, characterized in that: The step 2 specifically includes: Step 2.1: Take the straight line from the starting point to the end point of the cruise missile as the baseline of the trajectory, define the threat source intersecting the baseline as the dominant threat source, i.e. DTS, and define the remaining threat sources as non-dominant threat sources, i.e. NDTS; Determine whether the baseline intersects with the threat source; Step 2.2: If the baseline does not intersect with any threat source, the A* algorithm will find the baseline source that does not intersect with any threat source as the actual trajectory; Step 2.3: If the baseline intersects the threat source, a curved path is used to bypass the threat area, and the trajectory length is greater than the baseline length; Step 2.3.1: If the baseline intersects a single DTS, the threat radius and the distance from the center of the DTS to the baseline are used as input features, and the threat coefficient index is defined as: Among them, λ represents the threat coefficient index, λ∈[0,1], which represents the impact on the trajectory increment; r and d represent the threat radius and the distance from the DTS center to the baseline, respectively; Step 2.3.2: If the baseline intersects multiple DTSs, use the average value as the input feature, as follows: in, is the set of all DTSs, N DT is the number of DTS; R DT and λ DT Represent the average threat radius and average threat coefficient respectively; the input features include the number of DTS N in addition to the above average values DT ; Step 2.4: There are three types of intersections between the threat sources, namely, the intersection between different DTSs, the intersection between different NDTSs, and the intersection between DTS and NDTS. The first two types of intersections are called homogeneous intersections, and the third type of intersection is called heterogeneous intersections. In uniform DTS intersection, the DTS that intersects the baseline is defined as the core DTS; the DTS whose center is on the same side of the baseline as the center of the core DTS is called the positive DTS; the DTS whose center is on the other side of the baseline relative to the center of the core DTS is called the negative DTS; When the positive DTS intersects with the core DTS, the impact of the positive DTS is adjusted by reducing the threat radius of the positive DTS, as shown in the following formula: Among them, r j ′ is the adjusted threat radius, r j is the original threat radius, r i is the threat radius of the core DTS; α j is the attenuation coefficient, which means that the closer to the core DTS, the greater the attenuation of the threat radius; For negative DTS to reduce the trajectory increment, not only the threat radius needs to be modified, but also the threat coefficient λ needs to be adjusted; Among them, λ′ represents the adjusted threat coefficient index, λ′∈[1,2]; r and d represent the threat radius and the distance from the DTS center to the baseline, respectively; For the scenario where multiple DTSs intersect, a sequential modification algorithm is proposed based on steps 2.3 and 2.4; Step 2.5: In the heterogeneous intersection scenario, when the threat source closest to the starting point is the DTS, the search direction of the A* algorithm depends on the position of the DTS center. Based on the heuristic search mechanism of A*, the opposite side of the DTS center is the optimal direction, and the features of the NDTS are extracted. Step 2.6: After all NDTS are identified, relevant features are extracted from the information of all NDTS.

3. The method for rapid target reallocation in cruise missile swarm combat according to claim 2, characterized in that: In step 2.4, for the scenario where multiple DTSs intersect, the proposed sequence modification algorithm is specifically as follows: Sort all DTSs in ascending order according to their distance from the starting point; Assume that one DTS is the core DTS, and determine whether the remaining DTSs are positive or negative by checking whether the line segment formed by the center of each DTS and the center of the core DTS intersects the baseline; If it is a positive DTS, then only the threat radius needs to be adjusted; If it is a negative DTS, the threat radius and threat coefficient need to be adjusted.

4. The method for rapid target reallocation in cruise missile swarm combat according to claim 3, characterized in that: In step 2.5, in a heterogeneous cross-scenario, the specific steps for feature extraction are: Step 2.5.1: Find the threat source closest to the starting point, defined as TS*; Step 2.5.2: If TS* is NDTS, determine that the other side of the baseline is a valid NDTS, defined as NDTS+; calculate the threat radius and threat coefficient of all NDTS+, and modify the threat radius of all NDTS+ using the sequential modification algorithm; Step 2.5.3: If TS* is DTS, then identify the valid NDTS as NDTS+ based on the NDTS center, DTS center and baseline; calculate the threat radius and threat coefficient of all NDTS+, and modify the threat radius of all NDTS+ using the sequential modification algorithm.

5. The method for rapid target reallocation in cruise missile swarm combat according to claim 4, characterized in that: The specific method of step 5 is: Step 5.1: For the updated context information, use the feature extraction in step 2 to obtain the input features acceptable to the proxy model; Step 5.2: Combine the trajectory length matrix output by the proxy model with the target value to obtain the cost matrix C; Step 5.3: Apply the greedy algorithm to the cost matrix C to obtain the final allocation result, denoted by the decision matrix D; Step 5.4: Assume that the current situation information includes M targets and N loitering munitions. The target allocation optimization problem is divided into two different forms: N > M and N < M; it is required that all available loitering munitions must perform attack tasks, and each target can be allocated at most one loitering munition. The constraints are as follows: Among them, x ij is a decision variable. When the jth target is assigned to the i-th cruise missile, the decision variable takes 1, otherwise it takes 0. When the i-th cruise missile attacks the jth target, L ij , T ij , G ij Represents trajectory length, threat level and attack gain respectively; Assuming that all cruise missiles can complete the attack, G ij It can be expressed by the value of the jth target; Define the trajectory cost as: C ij =αL ij -(1-α)G ij where α is a balance parameter that controls the balance between the trajectory length and the trajectory cost target value; Step 5.5: When the number of loitering munitions is greater than the number of targets, i.e., N > M, the greedy algorithm is represented by the following steps: Step 5.5.1: Initialize the decision matrix D by setting all elements to zero; Step 5.5.2: Find the minimum value and its index in the cost matrix C, and set all elements in the row and column where the minimum value is located to infinity; Step 5.5.3: Set the element corresponding to the above index in the decision matrix D to 1; Step 5.5.4: Repeat steps 5.5.2 and 5.5.3 M times, and let the obtained decision matrix be D', and each target is equipped with the most suitable loitering munition; Step 5.5.5: Find the row subscript of the element with a value of 1 in the matrix D', and remove the row corresponding to the row subscript from the original cost matrix C to obtain the matrix C'; Step 5.5.6: Find the minimum value and its index in each row of the matrix C'; if the obtained minimum value is less than 0, set the corresponding element in the matrix D' to 1; Step 5.6: When the number of loitering munitions is less than the number of targets, i.e., N < M, the greedy algorithm is represented by the following steps: Step 5.6.1: Initialize the decision matrix D by setting all elements to zero; Step 5.6.2: Find the minimum value and its index in the cost matrix C, and set all elements in the row and column where the minimum value is located to infinity; Step 5.6.3: Set the element corresponding to the above index in the decision matrix D to 1; Step 5.6.4: Repeat steps 5.6.2 and 5.6.3 N times.

Images