Weapon target assignment method based on neural network

Abstract

The application provides a weapon target allocation method based on a neural network, and mainly solves the problem that the prior art cannot be applied to the continuous combat scene of unmanned aerial vehicle group attacking ground targets which has extremely high time efficiency requirement.The scheme is as follows: threat evaluation is carried out on the discovered targets, the targets are sorted according to the threat degree, and the targets are allocated with weapons one by one; the damage probability estimation is used to obtain the attack suitability of each weapon on the current target to be allocated, the commander adjusts part of the suitability according to the actual combat experience, and a training data set and a test data set are formed; the full connection network model is trained through the training set, the best full connection network model is selected through the test data set, and the selected best full connection network model is used to allocate the weapon target for each target one by one.The application not only can utilize the historical combat information and the combat experience of the commander, but also can reduce the allocation time, and can be applied to the continuous combat scene of unmanned aerial vehicle group attacking ground targets which has extremely high time efficiency requirement.

Description

Technical Field

[0001] This invention belongs to the field of computer simulation technology, and specifically relates to a weapon target allocation method that can be used for unmanned aerial vehicle (UAV) swarms to strike ground targets. Background Technology

[0002] The key to the weapon target allocation problem lies in solving the problem of allocating weapons with different lethal capabilities to targets with different threats in order to achieve the maximum combat effectiveness. It belongs to a class of optimization scheduling problems with multiple resource constraints.

[0003] In large-scale combat scenarios, the complexity of the weapon target allocation problem increases dramatically with the number of weapons and targets. The inherent complexity of determining the optimal solution includes limited weapon resources, time windows for responding to threats, load balancing among weapon systems, and complex interdependencies among various tasks and resources. The weapon target allocation problem has been proven to be NP-complete, and traditional methods cannot efficiently solve large-scale allocation problems in polynomial time. Currently used genetic algorithms struggle to adapt to fast-paced, dynamic, and continuous combat scenarios, and their timeliness often fails to meet practical requirements.

[0004] Furthermore, there is a current practical need to utilize historical combat information, that is, to gain experience from historical combat information to better guide future operations. Traditional target allocation methods cannot utilize historical combat information. For example, weapon target allocation methods based on genetic algorithms obtain allocation results by optimizing a fixed objective function each time, and cannot utilize allocation information generated from previous operations.

[0005] Target allocation for UAV swarms can be viewed as an assignment problem. Existing methods for solving this problem mainly include matrix methods, heuristic genetic simulated annealing algorithms, and the Hungarian algorithm. These methods treat the weapon target allocation problem as static, that is, they optimize and solve it within a fixed time window. Since the positions of both UAVs and targets change during the optimization process, the longer the solution time, the more significant the position changes, and the greater the reduction in allocation efficiency. Therefore, these highly complex methods are difficult to apply to continuous combat scenarios with extremely high time-sensitivity requirements, such as UAV swarms attacking ground targets. In addition, these methods are based on deterministic rules and are difficult to dynamically adjust as the operation progresses or the commander's strategy changes. That is, they cannot utilize historical combat information and the commander's combat experience, and cannot meet the requirements of intelligent decision-making in actual combat. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art by proposing a weapon target allocation method based on neural networks, which utilizes historical combat information and commanders' combat experience to improve the timeliness and responsiveness of weapon target allocation, thereby meeting the requirements for intelligent decision-making in actual combat.

[0007] To achieve the above objectives, the technical solution of the present invention includes the following:

[0008] (1) Based on the acquired information on enemy target types and status, the analytic hierarchy process (AHP) is used to assess the threat level of enemy targets. j ∈[0,1] represents the threat level of target j;

[0009] (2) Based on the threat level of the enemy target, determine the strike priority of the currently discovered target and sort them from largest to smallest to obtain the strike order table. The greater the threat level, the higher the strike priority.

[0010] (3) Select targets to engage in sequence according to the attack order table:

[0011] 3a) For the selected target j, based on our weapon information and environmental information, and taking into account the enemy target information, estimate the probability of damage to the target by the weapons, and calculate the probability P of damage to the target by each weapon. ij ;

[0012] 3b) Obtain the damage effect P of weapon i on the currently selected target based on the damage probability. ij ×T j Cost-effectiveness Calculate the suitability F for each weapon to engage the target. i :

[0013]

[0014] Where α∈[0,1] is the damage effect weight, v j For the value of target j, c i The cost of weapon i;

[0015] 3c) Select the available weapon with the highest suitability as the allocation result for striking the target. Finally, the commander judges whether the allocation result needs to be adjusted based on combat experience and actual situation. If adjustment is required, the commander manually adjusts the corresponding strike suitability.

[0016] (4) Repeat the core steps of weapon target allocation (1)-(3) to obtain two types of training data, A and B. A is the rule-based combat data, and B is the data adjusted by the commander according to combat experience and actual situation. The A and B data are combined into a training set and a test set in a ratio of 9:1. The training set has 90,000 data A and 10,000 data B, and the test set has 900 data A and 100 data B.

[0017] (5) Assume each UAV carries two types of weapons, and obtain the spatial coordinates [x] of the i-th UAV. wi ,y wi ,z wi ] and the spatial coordinates of the currently selected target [x j ,y j ,z j The input vector X of the network model with dimension 3*M+7 is obtained under M drones:

[0018] X = [x w1 ,y w1 ,z w1 ,...,x wi ,y wi ,z wi ,...,x wM ,y wM ,z wM ,N,E,x j ,y j ,z j Belong j Type j ]

[0019] Where N is the natural environment impact factor, E is the electromagnetic environment impact factor, and Belong j For the target's company, Type j The type of target;

[0020] (6) Construct a neural network model consisting of multiple fully connected layers. The hidden layers have k layers, and each layer has L = [l1, l2, ..., l...]. k The output vector is: The dimension is 2*M, representing the suitability of each weapon for striking the currently selected target;

[0021] (7) Using the training set, train the network model using the stochastic gradient descent algorithm. The fitting error of the trained network is verified using a test set, and the optimal k and L are selected to obtain the trained network model.

[0022] (8) As the combat progresses and the data B continues to expand, repeat steps (4)-(7) at regular intervals to retrain the network and obtain a new network model.

[0023] (9) For the feature vector X of the currently selected target j, use the trained model Obtain the output vector Using this neural network model The suitability F in steps 3a) and 3b) is replaced i The calculation process involves repeating weapon target allocation steps (1) to (3). Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall implementation of the present invention;

[0025] Figure 2 This is a flowchart of the threat assessment sub-process in this invention;

[0026] Figure 3 For threat assessment, a multi-layered set of factors is used.

[0027] Figure 4 A sub-flowchart for target allocation using a neural network-based target allocation method;

[0028] Figure 5 This is a schematic diagram of the simulation environment;

[0029] Figure 6 The accuracy on the training and test sets varies with the number of training algebras. Detailed Implementation

[0030] This invention addresses a scenario of dynamic, continuous strikes by a swarm of drones against ground targets, in which all ground targets do not appear simultaneously, and assumes that:

[0031] The enemy targets are ground forces, moving on the ground in company units. There are 12 types of enemy targets, and the 6 companies are equipped with several targets of different types according to actual needs. The enemy targets move along a certain route in company units.

[0032] After detecting enemy movements, M drones will take off and patrol over the target area, awaiting an opportunity to strike. Each drone can carry four types of missiles, with a certain number of two types per drone. Our drones will patrol from the given starting point to the destination, randomly and intermittently detecting targets based on the intelligence system, and allocating weapon targets during the patrol.

[0033] During its patrol, the drone selects and strikes currently detected ground targets to maximize strike effectiveness. How to allocate targets in this scenario is the problem this invention aims to solve.

[0034] The embodiments and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0035] Reference Figure 1 The implementation steps for this example are as follows:

[0036] Step 1: Target threat assessment.

[0037] Reference Figure 2 The specific steps for implementing target threat assessment are as follows:

[0038] 1.1) Establish a set of threat assessment factors, such as Figure 3 :

[0039] Based on the scenario of drone swarm offensive operations, the common targets of enemy ground forces include tanks, armored vehicles, troop carriers, and reconnaissance vehicles. The target capability, target intent, and distance are selected as the three primary factors for target threat assessment.

[0040] The impact of target capabilities on the threat level is subdivided into three secondary factors: maneuverability, jamming capability, and air defense capability. The distance factor is further divided into two secondary factors: the distance between the target and the UAV and the distance between the target and the key location.

[0041] 1.2) Quantify and standardize the indicators of each factor:

[0042] 1.2.1) The target’s mobility, jamming capability and air defense capability are set as qualitative indicators, denoted as A1, A2 and A3 respectively, and the three capabilities are divided into four levels: strong, medium, weak and none. The quantitative values ​​corresponding to each level are 0.9, 0.6, 0.3 and 0 respectively.

[0043] 1.2.2) Target intention I is divided into three categories: attack, reconnaissance and support, with corresponding quantification values ​​of 0.9, 0.7 and 0.4, respectively;

[0044] 1.2.3) Distance factors are set as quantitative indicators, which are divided into the distance d1 from the UAV and the distance d2 from the key location. The calculation method is as follows: Among them, d' k Let d1 be the distance between the target and the kth drone or key location, n be the number of drones or key locations, i = 1 represents the distance d1 from the drone, and i = 2 represents the distance d2 from the key location.

[0045] 1.2.4) Normalize the distances d1 and d2 to values ​​D1 and D2 between [0-1]. The specific calculation method is as follows:

[0046]

[0047] μ = 1.0 is the maximum value of the distance index. When the distance is less than 10km, the threat is very high, and when the distance is greater than 100km, the threat is very low. The minimum value of μ is 0.1. When the distance is greater than 10km and less than 100km, the threat level decreases as the distance increases, and the index μ decreases linearly with the distance.

[0048] 1.3) The Analytic Hierarchy Process (AHP) is used to assign weights to each factor based on their perceived importance:

[0049] 1.3.1) Determine the judgment matrix:

[0050] Based on the judgment matrix A = (a ij ) n×n To utilize the properties of symmetric matrices, using the threat assessment factor set in step 1.1), three judgment matrices A, B1, and B2 are set to assess the importance of indicators. A is the judgment matrix for the three primary factors of target capability, target intent, and distance; B1 is the judgment matrix for the three secondary factors of target capability; and B2 is the judgment matrix for the two secondary factors of distance. A and B1 are 3-order square matrices, and B2 is a 2-order square matrix.

[0051] 1.3.2) The definition table of the judgment matrix scale is shown in Table 1:

[0052] Table 1. Definition of Matrix Scale

[0053] Compared to factor i, factor j Scale value Equally important 1 Slightly important 3 Stronger and more important 5 Strongly important 7 Extremely important 9 The median value between two adjacent judgments 2，4，6，8

[0054] 1.3.3) Experts determine the scale values ​​of the three judgment matrices A, B1, and B2 based on the judgment matrix scaling definition table and the actual situation. That is, experts compare the importance of each pair of factors based on experience and give the scale values ​​of each judgment matrix.

[0055] For example, the scale value of the judgment matrix A for the target capability is given as follows:

[0056] If experts consider target capability and target intent to be more important than target capability, and their importance falls between equal and slightly more important, then according to Table 1, the value A in the first row and second column of the matrix is... 12 =2;

[0057] If experts consider distance to be more important than target capability, and the degree of importance falls between slightly important and strongly important, then according to Table 1, the value A in the first column of the third row of this matrix is... 31 =4;

[0058] If experts consider distance to be more important than target intent, and the degree of importance falls between strong and strong, then according to Table 1, the value A in the third row, second column of the matrix is... 32 =6;

[0059] Finally, the judgment matrix is ​​obtained.

[0060] Following the same method as judgment matrix A, the experts determined judgment matrices B1 and B2 as follows:

[0061]

[0062]

[0063] 1.3.4) Perform a consistency check on the three judgment matrices A, B1, and B2:

[0064] Perform consistency checks on the instances of judgment matrices A, B1, and B2 given in step 1.3.3), that is, for each judgment matrix, perform consistency checks based on the consistency index. Calculate the consistency test ratio based on the average random consistency index (RI) standard value in Table 2. Where λ max To test the largest eigenvalue of a matrix, n is the dimension of the matrix to be tested;

[0065] If CR < 0.10, the judgment matrix is ​​considered to have passed the consistency test.

[0066] Otherwise, modify the judgment matrix until the consistency check is passed.

[0067] Table 2. Standard values ​​of the average random consistency index (RI)

[0068] Matrix order n 1 2 3 4 5 6 7 8 9 10 RI 0 0 0.58 0.90 1.12 1.24 1.32 1.41 1.45 1.49

[0069] This example determines the largest eigenvalue λ of matrix A. max =3.013, consistency test ratio CR = 0.011 < 0.10, the consistency test is passed; determine the largest eigenvalue λ of matrix B1. max =3.004, consistency check ratio CR = 0.004 < 0.10, the consistency check is passed. B2 dimension has 2, so consistency is guaranteed;

[0070] 1.3.5) The arithmetic mean method is used to approximate the calculation of the largest eigenvalue λ of each judgment matrix. max The corresponding eigenvector W:

[0071]

[0072] This example calculates the eigenvectors of the judgment matrices A, B1, and B2 given in step 1.3.3), and obtains the following results:

[0073] Determine the largest eigenvalue λ of matrix A max =3.013, corresponding to the eigenvector W A = [0.204, 0.105, 0.691];

[0074] Determine the largest eigenvalue λ of matrix B1 max =3.004, corresponding eigenvector

[0075] Determine the largest eigenvalue λ of matrix B2 max =2.000, the corresponding eigenvector is

[0076] 1.3.6) The threat level T of target j is obtained by weighting each factor. j :

[0077]

[0078] Where A1, A2, and A3 are the quantified values ​​of the target's maneuverability, jamming capability, and air defense capability, respectively; I is the quantified value of the target's intent; D1 and D2 are the quantified values ​​of the distance to the UAV and the distance to the key location, respectively; and W... A [1] indicates that the first value of the eigenvector corresponding to the largest eigenvalue of the target capability judgment matrix A is 0.204.

[0079] Step 2: Based on the threat level assessment of enemy targets, determine the strike priority of currently detected targets and sort them from highest to lowest to obtain the strike order table Attack_List. The higher the threat level, the higher the strike priority.

[0080] Step 3: Select targets in sequence according to the Attack_List and engage them.

[0081] 3.1) Damage probability estimation:

[0082] For the selected target j, based on our weapon information, environmental information, and combined with enemy target information, the probability of damage to the target by each weapon is estimated, and the probability of damage P of each weapon to the target is calculated. ij The probability of damage is influenced by multiple factors, and is represented by the product of the quantified values ​​of each factor:

[0083] P ij =N×E×W ij ×H j ×M ij

[0084] Among them, P ij N represents the probability of weapon i damaging enemy target j, E represents the probability of weapon i damaging enemy target j, and W represents the probability of weapon i damaging enemy target j. ij H represents the factor influencing the ability of weapon i to strike enemy target j. j M represents the factor influencing the defensive capabilities of enemy target j. ij The factor representing the compatibility between weapon i and enemy target j is quantified as follows:

[0085] 3.1.1) Without considering the effects of the natural environment and electromagnetic countermeasures, N = 1, E = 1;

[0086] 3.1.2) Based on the weapon's basic hit accuracy, warhead weight, and strike range, calculate the capability impact factor W of our weapon. ij :

[0087] W ij =P base ×P w ×P d

[0088] Among them, P base The base hit probability of weapon i is the probability of the missile hitting under the conditions of good environment, no electromagnetic interference, appropriate strike distance, and no missile system malfunction.

[0089] P w The quantification method for the influence of weapon power on accuracy is as follows: w represents the weight of the missile warhead (kg). The larger w is, the greater P becomes. w It's close to 1;

[0090] P d The influence factor of distance on strike accuracy is given below, with the following values:

[0091]

[0092] r is the weapon's range, d is the distance between the weapon and the target. Within the range range [0, r], the effect of distance on accuracy is negligible. The distance influence factor P d =1; Within the range and 1.5 times the range [r, 1.5r], the influence of distance on accuracy gradually increases. When the distance is greater than 1.5r, the distance influence factor P... d =0, the weapon cannot hit;

[0093] 3.1.3) Calculate the enemy target's defensive capability influence factor H based on the target's armor level and air defense capability. j :

[0094]

[0095] in and These are the quantified values ​​of the armor level and anti-air capability of target j, respectively, and are integers between [1, 10]. and The smaller the sum, the better H j The closer it is to 1; and The larger the sum, the greater H j The closer the value is to 0.95, the stronger the target's armor and air defense capabilities, and the lower the probability of damage.

[0096] 3.1.4) Calculate the influence factor of weapon-target matching degree on the probability of destruction:

[0097] The fact that the same weapon has different probabilities of damaging different targets, and different weapons have different probabilities of damaging the same target, indicates that weapon operations are targeted. For example, armor-piercing projectiles have a higher probability of damaging heavily armored targets like tanks than scattered infantry. Therefore, armor-piercing projectiles have a correspondingly higher probability of damaging tank targets. This characteristic is defined as the weapon-target matching degree.

[0098] Based on matching degree Calculate the influence factor M of weapon target matching degree on the probability of destruction. ij :

[0099]

[0100] in, The matching degree value between weapon i and target j is an integer between [1, 10], which is derived from historical combat experience and expert knowledge.

[0101] 3.2) Determine the objective function:

[0102] In the scenario of drone offensive operations, the basic principle of weapon target allocation is to maximize combat effectiveness and minimize combat costs while satisfying all constraints.

[0103] 3.2.1) Calculate the damage effect

[0104] To maximize combat effectiveness, it is necessary to strike high-threat targets with weapons that have a high probability of destruction, i.e., to maximize the damage effect c, as calculated below:

[0105] c = P ij ×T j ,

[0106] Among them, T j P represents the threat level of target j. ijLet be the probability of weapon i damaging target j;

[0107] 3.2.2) Define the cost-effectiveness ratio:

[0108] To maximize the effectiveness of target allocation, it's crucial to consider not only how to destroy the target to the greatest extent possible, but also the combat effectiveness generated per unit of weaponry and ammunition. For example, large missiles and small missiles may differ in cost by a factor of ten. When both strike the same small target, the large missile might have a slightly higher probability of destruction. In this case, if only the goal of better target destruction is considered, without regard to cost, the large missile would be chosen. Furthermore, if the probability of a missile destroying a target is very low, prioritizing destruction without considering cost would result in significant waste and negatively impact subsequent combat effectiveness.

[0109] To avoid waste and enable more rational use of weapons, the cost-effectiveness ratio e is defined as follows:

[0110]

[0111] Where v j The destruction value of target j is the cost of the target equipment; c j The cost of launching weapon i, i.e., the cost of manufacturing the weapon, is P. ij Let be the probability of weapon i damaging target j;

[0112] 3.2.3) Construct the objective function based on the damage effect and cost-effectiveness ratio:

[0113] For an assignment task involving m weapons and n targets during combat, an assignment scheme X is an m×n matrix. The value X in the i-th row and j-th column of this matrix is... ij for:

[0114]

[0115] In target allocation scenarios, failure to destroy enemy targets will result in friendly forces being attacked and suffering losses. Therefore, multiple weapons are typically allocated to strike targets to destroy them as much as possible. However, target allocation tasks in UAV-based ground strike scenarios differ from general target allocation. The "urgency" of the target strike is not high, and even if the target is not destroyed, there will be no immediate losses. Therefore, constraints need to be set for each allocation task, namely, each target can be allocated a maximum of one weapon, and each weapon can be fired a maximum of one time.

[0116] Taking into account the degree of damage and the cost-effectiveness ratio, under the set constraints, the objective function F is obtained by weighted summation:

[0117]

[0118]

[0119] Where α and β are the weights of the damage effect and the cost-effectiveness ratio, constrained by... This means that a weapon can be used at most once in a single task assignment, which is a constraint. This means that in a single task assignment, a target can be assigned a maximum of one weapon.

[0120] 3.2.3) Simplify the objective function F to a single-objective allocation objective function F i :

[0121] To reduce the complexity of solving the above objective function, the multi-target allocation is simplified to a single-target allocation. Specifically, based on the attack order list Attack_List, individual targets are selected sequentially for attack. The damage probability of each weapon on that target is calculated, and then the most suitable weapon is selected by comprehensively considering the damage effect and cost-effectiveness. The simplified objective function is as follows:

[0122]

[0123] Where α∈[0,1] is the damage effect weight, v j For the value of target j, c i For the cost of weapon i, F i The suitability of weapon i for striking the currently selected target;

[0124] 3.3) Select the most suitable available weapon as the allocation result for striking the target, and let the commander judge whether the allocation result needs to be adjusted based on combat experience and actual situation:

[0125] If adjustments are needed, the commander can manually adjust the corresponding strike suitability F. i For F i ′;

[0126] If no adjustment is needed, use the calculated strike suitability F. i Distribute the strikes accordingly.

[0127] Step 4: Construct the training dataset and the test dataset.

[0128] The core steps 1-3 of the repeated weapon target allocation process will yield two types of data, A and B, where A is the rule-based operational data F. i B represents data F adjusted by the commander based on combat experience and actual conditions, building upon A. i ';

[0129] 4.1) Dataset format:

[0130] In data A, each data sample is in the form of (X, Y). A In data B, each data sample is in the form of (X, Y). B ), where X represents the characteristics of a target and all weapons, Y A Y B The suitability vector for each weapon to engage the target is as follows:

[0131] Assuming each drone carries two types of weapons, obtain the spatial coordinates [x] of the i-th drone. wi ,y wi ,z wi ] and the spatial coordinates of the currently selected target [x j ,y j ,z j ], thus obtaining the feature vector X with dimension 3*M+7 for M drones:

[0132] X = [x w1 ,y w1 ,z w1 ,...,x wi ,y wi ,z wi ,...,x wM ,y wM ,z wM ,N,E,x j ,y j ,z j Belong j Type j ]

[0133] Where N is the natural environment impact factor, E is the electromagnetic environment impact factor, and Belong j For the target's company, Type j The type of target;

[0134] For M drones, each drone has two types of weapons, Y A Y B The dimension is 2*M, representing the suitability of each weapon for striking the currently selected target;

[0135] Y A =[F1,F2,···,F i ,···,F n ],

[0136] Y B =[F1,F2,···,F i ',···,F n ],

[0137] Among them, Y of data AA This represents the suitability vector for each weapon to strike the currently selected target;

[0138] B data Y B This represents the strike suitability vector for the currently selected target after the commander adjusts the strike suitability of the i-th weapon;

[0139] 4.2) Constructing the training and testing datasets:

[0140] In the feasible region of each feature of feature vector X, different feature vectors X are randomly generated according to a uniform distribution, and then the strike suitability F of each weapon is calculated. i Composition Y A This, in turn, constitutes data A;

[0141] In data A, the commander needs to select the data to be adjusted, and set F... i Adjust to F i 'Composition Y B This constitutes data B;

[0142] The training set and test set are constructed using data A and B in a 9:1 ratio. The training set contains 90,000 data points of data A and 10,000 data points of data B, for a total of 100,000 data points.

[0143] The test set contains 900 data points A and 100 data points B, for a total of 1000 data points. If the number of data points B adjusted by the commander is less than 10,000 in the training set or less than 100 in the test set, then the data points B in the training set are constructed by repeated copying so that the number of data points B in the test set reaches 10,000 and the number of data points B in the test set reaches 100.

[0144] Step 5: Construct a neural network model consisting of multiple fully connected layers.

[0145] Construct a neural network model consisting of multiple fully connected layers. The hidden layers have k layers, and each layer has L = [l1, l2, ..., l...]. k The activation function uses the ReLU function, and the output vector is: The dimension is 2*M, representing the suitability of each weapon for striking the currently selected target; network model. The structure is as follows:

[0146] Input layer → 1st hidden layer → 2nd hidden layer → … → kth hidden layer → Output layer.

[0147] Step 6: Train and select the best network model

[0148] 6.1) Train the network model using the stochastic gradient descent algorithm on the training set.

[0149] 6.1.1) Add the latest data B to the training and test sets, set the initial learning rate α = 0.1, and the mini_batch size = 200;

[0150] 6.1.2) For the input feature vector X = [x1, x2, ..., x...] i ,···,x n Each dimension x i Normalization is performed to obtain the normalized eigenvector: X'=[x′1,x'2,···,x′ i ,···,x' n ],in:

[0151]

[0152] in, Let be the minimum value of the i-th dimension feature. The maximum value of the i-th dimension feature;

[0153] 6.1.3) Define the multi-output mean squared error loss function as a neural network. The loss function L(θ):

[0154]

[0155] Where θ is the neural network model The parameter, Y is F N It is a network model The predicted value of Y is Y A Y B , is the true value of the sample, and n is the sample size of a mini-batch;

[0156] 6.1.4) Set the decay coefficient k = 0.01, λ = epoch = 100, and apply it to each θ of the learning rate α and model parameters θ. j To update, that is:

[0157]

[0158]

[0159] 6.1.5) Repeat step 6.1.4) until λ = 100, then stop the training process to obtain the model with determined k and L.

[0160] 6.2) Select the optimal k and L to obtain the trained model.

[0161] To make it more intuitive, define a network. The accuracy rate is defined as the ratio of the most suitable weapon output to the actual weapon allocation. Where n is the network The number of weapons with the highest suitability output is the same as the number of weapons actually assigned, where N is the total number of samples in the test set;

[0162] First, different neural network models were obtained by setting different k and L. Each model is then trained using the process in step 6.1.

[0163] Then, using the test set, calculate the error or accuracy of each model, and select the model corresponding to k and L with the smallest error or accuracy as the trained model.

[0164] Step 7: Neural Network Model Update.

[0165] As the operation progressed, the commander adjusted the network output intermittently. Get Y B This allows data B to be continuously expanded while data A remains unchanged. Every certain period of time, through steps 4-6, the expanded training set and test set are used to train and select the network model.

[0166] Network models updated as the battle progresses. Including historical combat information and commanders' combat experience, it not only greatly speeds up target allocation but also makes target allocation more flexible and intelligent.

[0167] Step 8: In a scenario where a swarm of drones attacks ground targets, use the trained network model. Target allocation.

[0168] Considering the fast forward propagation speed of neural networks, a neural network model is used. Given a target, this example outputs the suitability of each weapon for engaging that target. It rapidly allocates weapons to targets detected in each time window in sequence to meet the stringent time requirements for target allocation in dynamic UAV combat scenarios. The specific implementation is as follows:

[0169] For the feature vector X of the currently selected target j, the trained model... Obtain the output vector Using this neural network model Replace the fitness value F in steps 3.1) and 3.2). i Calculated Y AThe process is repeated, and then the core steps 1-3 of weapon target allocation are repeated.

[0170] Reference Figure 4 The specific implementation of this step is as follows:

[0171] 8.1) Initialize the combat scenario:

[0172] A single operation can be divided into multiple time windows, with each time window serving as a task allocation window, which can be assigned to several targets.

[0173] Set the number, initial position, movement route, and movement speed of each type of enemy target; set the number of friendly drones, the type and quantity of weapons carried by each drone, initial position, movement route, and movement speed, and initialize the time window T=1;

[0174] 8.2) Conduct a threat assessment on the targets that have been discovered but not yet destroyed, and sort them in descending order of threat level to obtain the Attack_List;

[0175] 8.3) Determine the number of targets n to be attacked in the current time window. a The condition 0≤n is satisfied. a ≤min(N u N attack_List ), that is, the number of targets n attacked in each time window. a The number of available weapons cannot exceed N. u And does not exceed the number of targets N in the current attack order list Attack_List. attack_List .

[0176] 8.4) Select the top n from Attack_List a Each target was assigned for attack;

[0177] 8.5) Update the current battlefield status:

[0178] A: Based on the target allocation results and the damage probability, determine whether the attacked target has been destroyed, and update the enemy target status;

[0179] B: Update the remaining ammunition quantity for each drone;

[0180] C: Based on the probability of damage from enemy air defense weapons and the distance, determine whether each of our drones has been destroyed and update the status of our drones.

[0181] D: Update the positions of all undisturbed targets based on their movement route and speed;

[0182] E: Update the drone's position based on its flight path and speed.

[0183] 8.6) Determination of Operation Termination Conditions:

[0184] The commander can set termination conditions according to the specific situation: running out of ammunition, all drones being destroyed, no targets being detected for 5 consecutive time windows, ammunition usage reaching 80% of the total carrying capacity, or a specific key target being destroyed.

[0185] 8.7) Let T = T + 1, enter the next time window, and determine whether the combat termination condition has been triggered:

[0186] If the conditions for terminating the operation are triggered, then the operation will cease.

[0187] If the conditions for terminating the operation are not triggered, return to step 8.2.

[0188] The effectiveness of this invention can be further illustrated by the following simulation experiments.

[0189] I. Simulation Conditions

[0190] The simulation experiment uses a combat scenario where an unmanned aerial vehicle (UAV) platform carrying various types of weapons strikes enemy ground targets as the main simulation scenario, and sets up a confrontation scenario between friendly and enemy sides. The scenario description is as follows:

[0191] Suppose that our arsenal contains several different types of weapons, and the missile types and parameters are shown in Table 3, with a total of four missile types.

[0192] Table 3. Summary of our missiles and parameters

[0193] Weapon (missile) model Basic accuracy Power (warhead weight / kg) Range / km Cost per 10,000 yuan <![CDATA[w1]]> 0.90 10 10 40 <![CDATA[w2]]> 0.85 5 8 15 <![CDATA[w3]]> 1.00 20 18 100 <![CDATA[w4]]> 0.95 20 40 250

[0194] The combat platform is a certain type of UAV, capable of carrying various types of air-to-ground missiles, with each UAV limited to carrying two types of missiles. Two configuration schemes are set as shown in Table 4.

[0195] Table 4. Missile Configuration Scheme for Unmanned Aerial Vehicles

[0196] Drone Configuration Scheme <![CDATA[w1]]> <![CDATA[w2]]> <![CDATA[w3]]> <![CDATA[w4 <!-- 11 -->]]> Configuration 1 0 13 0 3 Configuration 2 10 0 6 0

[0197] Configuration 1: The UAV carries 13 W2 missiles and 3 W4 missiles.

[0198] Configuration 2 involves the drone carrying 10 W1 missiles and 6 W3 missiles.

[0199] Assume the enemy targets consist of several ground armored companies, equipped with various armored vehicles, tanks, air defense weapons, etc. The target types and parameters are shown in Table 5, with a total of 12 target types.

[0200] Table 5. Summary of Target Types and Parameters

[0201]

[0202] The matching degree between the 4 types of weapons and 12 types of targets is shown in Table 6.

[0203] Table 6. Weapon-Target Matching Degree M ij

[0204]

[0205] The enemy's basic combat unit is the company. Different companies have different combat functions and are equipped with various types of equipment. The company configuration table is shown in Table 7.

[0206] Table 7. Configuration of Different Combat Companies

[0207]

[0208] The simulation environment sets up a total of 6 enemy combat companies. Companies 2 and 3, as well as companies 5 and 6, have the same configuration, indicating that the two companies perform the same combat functions and belong to the same type of company.

[0209] Simulation scenario of drone swarm attacking ground targets Figure 5 The enemy units are combat companies, and our units are drones carrying different types and quantities of missiles. The positions of each unit on both sides are described in a coordinate system. The simulation environment is established in three-dimensional coordinates, and the battlefield space is set to 40km*40km*4km. Six enemy combat companies are distributed on the ground. Figure 5 The blue dots represent six clusters, with different shapes within each cluster representing different target types, totaling 12 types. Enemy companies move along specific routes and speeds. Red dots represent friendly drones, with five drones in each of two configurations, for a total of ten. They are distributed in the air at a range of 2km to 4km, flying along specific routes. Green squares represent key friendly defensive positions that require protection to a certain extent. The lines connecting drones and targets indicate the allocation results within that time window. A ground enemy target changing from blue to red indicates that the target has been destroyed.

[0210] Our reconnaissance system conducts ground reconnaissance, periodically transmitting target information to drones. The drones then allocate targets based on the received enemy target information and engage them. Since the reconnaissance system cannot detect all enemy targets simultaneously, the simulation environment uses a random selection of a certain number of targets to simulate this scenario.

[0211] The operational constraints in the simulation environment are set as follows:

[0212] Firstly, due to the influence of unstable factors such as weather and electromagnetic fields, reconnaissance and detection systems cannot detect all ground targets at the same time; within a specific time window, they can only detect a limited number of targets.

[0213] Secondly, each type of missile on each drone can only be launched once per time window.

[0214] Third, if the ammunition quantity corresponding to the most suitable weapon is 0, or the drone carrying the weapon is destroyed, or the weapon is already occupied in the current time window, then the next most suitable weapon will be used.

[0215] II. Simulation Experiment Content

[0216] Simulation 1: Determine the network structure.

[0217] Construct a training set of 100,000 samples and a test set of 1,000 samples. Train the network on a network with 1 hidden layer and 50 nodes, and simulate the change in accuracy on the training and test sets as the number of training iterations increases. The results are as follows: Figure 6 As shown. From Figure 6 It can be seen that such a neural network model can fit such a dataset with an accuracy K1 greater than 90%, indicating that the target assignment algorithm based on neural networks is feasible.

[0218] To determine the optimal number of hidden layers and nodes per layer for the fully connected network model, different numbers of hidden layers k and hidden nodes L were set and trained separately for 1000 generations each. The results are shown in Table 8.

[0219] Table 8. Accuracy of different network structures

[0220]

[0221] As shown in Table 8, the accuracy reached 1.0 on the training set of the network with 1 hidden layer and 500 nodes. This is due to overfitting caused by the excessive number of network nodes. The network with 1 hidden layer and 200 nodes achieved the highest accuracy on the test set, therefore this network structure can be selected. Weapon target allocation.

[0222] Simulation 2: Target allocation experiments were conducted in two different combat scenarios, static and dynamic, using the target allocation methods of this invention and existing genetic algorithm-based methods, respectively.

[0223] The genetic algorithm is configured with the following parameters: population size N = 100, maximum number of iterations T. max =100. The neural network model is a fully connected network with one hidden layer and 200 nodes.

[0224] The static scenario is set as follows: all targets are detected at once, and 20 missiles from 10 drones carry out their strike missions within the same time window. During this time window, the positions of all targets and weapon platforms remain unchanged, i.e., static.

[0225] The dynamic scenario is set up as follows: all targets are detected at intervals, and 20 missiles from 10 drones carry out strike missions in multiple time windows. In each time window, 5 targets are detected, and the positions of all targets and weapon platforms are changing. This dynamic scenario is closer to a real battlefield environment.

[0226] Experiment 1: Comparing the allocation performance of the two methods within a time window in a static scenario, the performance evaluation index uses the objective function F. The experiment was repeated 10 times under the objective function α = [0.4, 0.5, 0.6, 0.7, 0.8] to obtain the running time T(s) and objective function value F of the allocation results of the two methods. The experimental results are shown in Tables 9 and 10.

[0227] Table 9. Allocation time T(s) under different weighting parameters α

[0228]

[0229] Table 10. Objective function values ​​F under different weight parameters α in static scenarios

[0230]

[0231] As shown in Table 9 of the experimental results, the time efficiency of the allocation method proposed in this invention is three orders of magnitude higher than that of the weapon target allocation method based on genetic algorithm. It has high timeliness and can be applied to rapid and continuous combat scenarios such as UAVs striking ground targets.

[0232] As shown in Table 10 of the experimental results, the allocation efficiency is slightly worse than that of the weapon target allocation method based on the genetic algorithm because the objective function F of this invention is simplified. However, this efficiency gap is acceptable in the combat scenario of UAV swarms attacking ground targets with extremely high timeliness requirements, and the smaller reduction in efficiency brings a great improvement in time efficiency.

[0233] Experiment 2: In a dynamic scenario, since 5 targets are detected in each time window, this experiment gathers 20 targets every four time windows. The target allocation method based on genetic algorithm is used to allocate targets once every four windows. This invention is used to allocate targets in each time window, so each type of missile can be used multiple times, making the use of missiles more flexible.

[0234] To compare the allocation effectiveness of the two methods under four time windows, the objective function F was used as the performance evaluation index. Ten repeated experiments were conducted under the objective function α = [0.4, 0.5, 0.6, 0.7, 0.8] to obtain the objective function values ​​F for the allocation results of the two methods. The experimental results are shown in Table 11.

[0235] Table 11. Objective function values ​​F under different weight parameters α in dynamic scenarios

[0236]

[0237] As shown in Table 11 of the experimental results, under different weight parameters α, the evaluation index F of this invention is higher than that of the target allocation method based on genetic algorithms. This indicates that the allocation performance of this invention in dynamic scenarios is superior to that of existing genetic algorithms. The high timeliness of this invention can lead to improved allocation performance, while existing algorithms with poor timeliness cannot be applied to dynamically changing battlefield environments.

Claims

1. A neural network-based weapon target assignment method, characterized in that, Comprise the following steps: (1) According to the acquired enemy target type and state information, the threat degree of the enemy target is evaluated by using the analytic hierarchy process, T j ∈[0,1] represents the threat degree of target j; (2) According to the threat degree estimation result of the enemy target, determine the strike priority of the current found target, and sort from large to small to get the strike order table, the greater the threat degree, the higher the strike priority; (3) According to the strike order table, select the target for strike in turn: 3a) For the selected target j, according to our weapon information and environment information, and integrated with the enemy target information, the damage probability of weapon attack target is estimated, and the damage probability P of each weapon to the target is calculated ij ; 3b) The damage effect P of the attack of weapon i on the currently selected target according to the damage probability ij x T j and the cost-effectiveness ratio The fitness F of the attack of each weapon on the target is calculated i : where, a ∈ [0, 1] is the damage effect weight, v j is the value of target j, c i is the cost of weapon i; 3c) Select the available weapon with the largest suitability as the allocation result for striking the target, and finally the commander judges whether the allocation result needs to be adjusted according to the combat experience and actual situation, if it needs to be adjusted, the commander adjusts the corresponding strike suitability artificially; (4) Repeat the weapon target allocation core steps (1)-(3) to obtain two kinds of training data A and B, that is, A is the rule-based combat data, and B is the data adjusted by the commander according to the combat experience and actual situation, and A and B data are combined into training set and test set according to the proportion of 9:1, wherein there are 90000 data A and 10000 data B in the training set, and there are 900 data A and 100 data B in the test set; (5) Set each UAV carries two types of weapons, get the spatial coordinates of the ith UAV [x wi ,y wi ,z wi ] and the spatial coordinates of the current selected target [x j ,y j ,z j ], get the network model input vector X with dimension 3*M+7 under M UAVs: X = [x w1 ,y w1 ,z w1 ,...,x wi ,y wi ,z wi ,...,x wM ,y wM ,z wM ,N,E,x j ,y j ,z j ,Belong j ,Type j ] Wherein, N is the natural environment influence factor, E is the electromagnetic environment influence factor, Belong j The target belongs to the unit, Type j The type of the target; (6) Construct a neural network model composed of multiple layers of full connection wherein the number of layers of the hidden layer is k, the number of nodes of each layer is L = [l1, l2, ···, lk], and the output vector is: k The dimension is 2*M, which represents the degree of fitness of each weapon to the currently selected target.​ (7) Using the training set, the network model is trained by the stochastic gradient descent algorithm The fitting error of the trained network is verified using the test set, the best k and L are selected, and the trained network model is obtained (8) As the data B of the ongoing campaign is continuously expanded, steps (4) - (7) are repeated at intervals to retrain the network and obtain a new network model (9) The feature vector X of the current selected target j is passed through the trained model The output vector is obtained The neural network model is used The fitness F in steps 3a) and 3b) is replaced by i The process is repeated, weapon target assignment steps (1) - (3).

2. The method of claim 1, wherein, The threat degree of the enemy target is evaluated by using the analytic hierarchy process in the step (1), and the following is realized: (1a) Establish the evaluation factor set: In the scene of unmanned aerial vehicle group attack combat, according to the common targets of enemy ground forces, such as tanks, armored vehicles, troop carriers and reconnaissance vehicles, the target capability, target intention and distance are selected as the first level factors of target threat evaluation; The influence of target capability on threat degree is divided into three second level factors: mobility, interference capability and air defense capability; The distance factor is divided into two second level factors: the distance of the target from the unmanned aerial vehicle and the distance of the target from the stronghold; (1b) Quantize and normalize each factor index: The target's mobility, interference capability and air defense capability are set as qualitative indexes, denoted as A1, A2 and A3 respectively, and the three capabilities are divided into four levels: strong, medium, weak and none, and the quantized values of each level are 0.9, 0.6, 0.3 and 0 respectively; The target intention I is divided into three categories: attack, reconnaissance and support, and the corresponding quantized values of each intention are 0.9, 0.7 and 0.4 respectively; The distance factor is set as a quantitative index, which is divided into a distance d1 from the unmanned aerial vehicle and a distance d2 from the target, and the calculation manner is Wherein, d' k is the distance of the target and the kth unmanned aerial vehicle or the target, and n is the number of unmanned aerial vehicles or targets. The distances d1 and d2 are normalized to values D1 and D2 between 0 and 1, and the specific calculation method is: (1c) The weights of each factor are given according to the importance of each factor by using AHP method; (1c1) The judgment matrix A, B1 and B2 about the importance of the indexes are obtained through expert experience, wherein A is the judgment matrix of the three first level factors: target capability, target intention and distance, B1 is the judgment matrix of the three second level factors of target capability, and B2 is the judgment matrix of the two second level factors of distance, A and B1 are 3-order square matrices, and B2 is a 2-order square matrix; (1c2) According to the judgment matrix scale definition table, the experts compare the importance of each two factors according to experience, and give the scale values of each judgment matrix; (1c3) Consistency check is performed on the three judgment matrices A, B1 and B2: According to the consistency index CI and the average random consistency index RI standard value, the consistency check ratio is calculated According to the consistency index CI and the average random consistency index RI standard value, the consistency check ratio is calculated where λ max is the maximum eigenvalue of the matrix to be checked, and n is the dimension of the matrix to be checked, If CR<0.10, it is considered that the judgment matrix passes the consistency check; Otherwise, modify the judgment matrix until it passes the consistency check. (1c4) Approximate calculation of the maximum eigenvalue λ of each judgment matrix using the arithmetic mean method max The corresponding eigenvector: The weight vectors W of the three judgment matrices are obtained A , W B1 , W B2 ; (1c5) Weighted to obtain the threat degree T of the target j j :

3. The method of claim 1, wherein, The damage probability of the weapon attack target in the step (3) is estimated by quantifying each factor influencing the damage probability and then using the product of the quantified values of the factors to represent the damage probability: P ij = N x E x W ij x H j x M ij wherein P ij represents the damage probability of weapon i to target j, N represents the natural environment influence factor, E represents the electromagnetic countermeasure influence factor, W ij represents the weapon i capability influence factor of our side, H j represents the target j defense capability influence factor of the enemy, M ij represents the weapon i and target j matching degree influence factor.

4. The method of claim 1, wherein, The neural network model in the step (6) The dimension of the input feature vector X of the neural network model is 3*M+7, the number of layers k of the hidden layer, the number of nodes L of each layer is L=[l1, l2, ···, l k ], and the output vector F N The dimension of the input feature vector X of the neural network model is 3*M+7, the number of layers k of the hidden layer, the number of nodes L of each layer is L=[l1, l2, ···, l k ], and the output vector F N The dimension of the input feature vector X of the neural network model is 3*M+7, the number of layers k of the hidden layer, the number of nodes L of each layer is Input layer→1st hidden layer→2nd hidden layer→…→kth hidden layer→output layer.

5. The method of claim 1, wherein, The step (7) utilizes the training set to train the network model by a stochastic gradient descent method This is implemented as follows: (7a) add the latest data B to the training set and the test set, set the initial learning rate α=0.1, and the number of small batches of samples mini_batch=200; (7b) Normalizing each dimension x n of the input feature vector X = [x1, x2, ···, xn]T i : The normalized feature vector X' = [x1', x'2, ···, x'N] is obtained. n ] (7c) Defining the multi-output mean squared error loss function as the loss function L(0) of the neural network ​ wherein, θ is a parameter of the neural network model Y is a network prediction value, Y is a sample true value, and n is a number of samples in a small batch. (7d) Model parameter update: Set the decay coefficient k = 0.01, λ = epoch = 100, and update each θ of the learning rate a and the model parameter θ j is updated, i.e.: 。 6. The method of claim 1, wherein, In the step (7), the fitting error of the trained network is verified by using the test set, and the best k and L are selected, and the following is realized: Different neural network models are obtained by setting different k and L and train each model thereof; Recycle the test set to calculate the error of each model, select the model corresponding to the minimum error of k and L as the trained model

Images