Unmanned aerial vehicle formation intention recognition method based on fuzzy clustering and random forest

Abstract

The invention discloses an unmanned aerial vehicle formation intention recognition method based on fuzzy clustering and a random forest. The method comprises the following steps: selecting target unmanned aerial vehicle formation intention recognition characteristic quantities in an air combat; target attack tendency prediction based on a fuzzy clustering algorithm; and target unmanned aerial vehicle formation intention identification based on a pigeon inspired optimization random forest algorithm. The attack tendency prediction is introduced, so that the input state quantity of the unmanned aerial vehicle formation intention recognition model is reduced, and the recognition difficulty of the model is reduced; a random forest algorithm is adopted to establish an unmanned aerial vehicle formation intention recognition model, and a pigeon inspired optimization algorithm is introduced to optimize hyper-parameters of the random forest algorithm, so that the accuracy of the intention recognition model is improved. According to the method, through the cooperation of fuzzy clustering dimensionality reduction and pigeon inspired optimization random forest, the input characteristic quantity can be significantly reduced, the accuracy of identifying the tactical intention of the target unmanned aerial vehicle formation can be improved, and effective support is provided for decision making in a complex air combat environment.

Description

Technical Field

[0001] The present invention relates to the problem of tactical intention recognition of UAV formations in air combat, and particularly to a UAV formation intention recognition method based on fuzzy clustering and random forest. Background Art

[0002] In complex air combat environments, understanding the air combat situation and interpreting battlefield information based on air combat information is crucial for air combat decision-making. Furthermore, air combat is a continuous process, and the combat intentions and methods of enemy drones can be reasonably inferred based on acquired air combat information, thereby seizing operational opportunities. Rapidly and accurately identifying enemy aircraft's intentions during the current air combat phase and inferring the enemy drone's future maneuvers will provide our fighters with more time to react, allowing us to effectively allocate air combat resources, thereby seizing the battlefield initiative and achieving overall air combat victory and air superiority.

[0003] The rapid development of modern air combat has complicated the coordinated engagement of fighter jets, placing increasing demands on efficient command systems. Compared to single-aircraft air combat, multi-aircraft coordinated formation operations offer enhanced combat capabilities and have become the predominant form of air combat. Accurately and in real time identifying the tactical intent of a target drone formation based on battlefield situational information, providing reliable decision-making for the enemy, is crucial for achieving victory in air combat. Therefore, identifying the tactical intent of drone formations is of vital importance.

[0004] Drone intent recognition relies on air combat situational information and deeply explores the inherent connections between drone attack and defense decisions and their attack and defense intentions to identify and judge their attack and defense intentions. Currently, commonly used methods for target drone intent recognition include template matching, Bayesian networks, expert systems, and neural networks. Template matching methods leverage the experience of air combat experts and combat commanders to build a model library, then extract air combat features and identify target intent by reasoning about the similarity between these features and the template. Bayesian network methods build a directed acyclic graph (DAG) structure using prior knowledge to solve conditional probabilities between nodes, thereby establishing a tactical intent inference model. Expert system methods build a knowledge base based on expert experience and use rules to describe the correspondence between battlefield situation and air combat intent, thereby constructing an inference model of enemy drone combat intent. Neural network methods use a dataset of air combat target intent features to train a neural network, extract air combat intent recognition rules, and then use these rules to infer enemy combat intent.

[0005] In the scenario of multi-UAV formation coordinated combat, the state quantity of the target UAV formation intention recognition increases, and the difficulty of solving the intention recognition model increases. Therefore, it is necessary to further extract the state quantity of the air combat situation and reduce the number of input feature quantities of the UAV formation intention recognition model, thereby reducing the difficulty of solving the UAV formation intention recognition model and improving the model recognition accuracy. Summary of the Invention

[0006] The present invention aims to address the shortcomings of the aforementioned prior art by proposing a method for identifying UAV formation intentions based on fuzzy clustering and random forests. By predicting the attack propensity of target UAVs and replacing the seven-dimensional air combat situation information features with attack propensity features, the random forest-UAV formation intention recognition model reduces the number of input state variables. Furthermore, a pigeon flock optimization algorithm is introduced to optimize the random forest algorithm hyperparameter K, improving the model's recognition accuracy and providing effective support for decision-making in complex air combat environments.

[0007] The technical solution to achieve the purpose of the present invention is:

[0008] A method for identifying UAV formation intention based on fuzzy clustering and random forest, characterized by comprising the following steps:

[0009] Step S1: Establishing an intention space and intention recognition feature set for the UAV formation. The intention space includes six types of tactical intentions: attack, feint, retreat, reconnaissance, surveillance, and electronic jamming. The intention recognition feature set includes air combat situation information and the target UAV's own state, and establishes a mapping from the target UAV formation state to tactical intentions.

[0010] Step S2: Processing the air combat situation information in the intention recognition feature set based on a fuzzy clustering algorithm to predict the attack tendency of the target UAV, that is, whether the current target is willing to launch an attack on our side;

[0011] Step S3: Based on the pigeon flock optimization random forest algorithm, an intention recognition model is established using the attack tendency and the target's own state characteristics to output the tactical intention of the target UAV formation.

[0012] Furthermore, in step S1:

[0013] The air combat situation information includes the relative distance between the target and our UAV, azimuth, entry angle, target flight speed, our flight speed, target flight altitude, and our flight altitude;

[0014] The target UAV's own state includes each target UAV's motion state, radar state, interference state, interference state, motion trend and maneuvering behavior.

[0015] Furthermore, the target motion state includes the coordinates of the UAV, the target heading angle, the target track angle, and the target flight speed;

[0016] The radar status is a silent penetration feature, and its value is yes or no;

[0017] The jamming state and jammed state reflect whether the target is implementing or being jammed, with the value being yes or no;

[0018] The movement trend includes flying towards our drone, flying away from our drone, flying towards our drone first and then flying away from our drone, and flying away from our drone first and then flying towards our drone. It is determined by the distance between the target drone and our drone over a period of time.

[0019] The maneuvering behavior includes 10 trajectory characteristics: horizontal straight flight, dive, jump, left circle, right circle, half roll, dogfight, upper left combat turn, upper right combat turn and serpentine maneuver, which are judged by the target's movement trajectory.

[0020] Furthermore, the attack tendency in step S2 can be quantified into three levels: strong, medium, and weak. The prediction of the attack tendency depends on the current air combat situation and the target UAV state identification features, including target speed v, self-drone speed v′, target acceleration a, target relative height dz, target azimuth Target entry angle θ, relative distance d between the two parties.

[0021] Furthermore, step S2 includes the following steps:

[0022] Step S2-1: feature quantity normalization processing:

[0023] Assume that the target feature set for attack tendency prediction is X = {x1, x2, ..., x n}, where x i represents the characteristic quantity of the i-th target, x ij is the jth characteristic value of the i-th target, n is the number of attack tendency prediction samples, and each characteristic value in the set is standardized. The formula is as follows:

[0024]

[0025] Where x′ ij For x ij The results after standardization are is the average value of the j-th feature, and the formula is as follows:

[0026]

[0027] Step S2-2, establish fuzzy similarity matrix:

[0028] Establish the fuzzy similarity matrix R0 of the target = (r ij ) n×n ; where r ij represents the target x i With target x j The similarity of the 7 features between ij The calculation formula is as follows:

[0029]

[0030] Where x′ ih is the hth eigenvalue of target i after normalization, x′ jh is the hth eigenvalue of the standardized target j; c is a constant, so that the comprehensive similarity r ij In the range [0,1] and evenly distributed;

[0031] Step S2-3, clustering:

[0032] Before clustering the target, it is necessary to find the transitive closure of R0; use the square method and calculate t(R0) according to the rules of the Zadeh operator (∧,∨): during the square operation, there exists a natural number k that satisfies And 2 k ≤n, then the transitive closure of R0 is the reflexive transitive closure; choose the appropriate confidence level λ∈[0,1], and the classification of attack tendency can be obtained by the truncation matrix; assuming the confidence level λ=λ0, according to the fuzzy rules and transitive closure The corresponding truncation matrix can be obtained, and the fuzzy rule is as follows:

[0033]

[0034] Step S2-4, attack tendency prediction feature vector calculation:

[0035] After clustering, the target set can be divided into t attack tendency combat groups, where each group consists of n l Batch target composition, t = 3; then the characteristic vector s of the lth attack tendency l The calculation formula is:

[0036]

[0037] Among them, lp is the serial number of the pth target in the lth attack tendency group, v lp ,v′ lp ,a lp ,dz lp , θ lp ,dlp is the characteristic quantity of the target;

[0038] Step S2-5: Target attack tendency prediction:

[0039] By comparing the normalized target state x′={x′1,x′2,…,x′7} with each attack tendency feature vector s l Distance D l , select the attack tendency value corresponding to the shortest distance as the target's attack tendency, distance D l The calculation formula is as follows:

[0040]

[0041] Furthermore, the intention recognition feature quantities in step S3 are: attack tendency, maneuvering behavior, radar state, interference state, interfered state, and movement trend; a random forest algorithm is used to establish a UAV formation intention recognition model, and the six state quantities of the target UAVs' attack tendency, maneuvering behavior, radar state, interference state, interfered state, and target movement trend are used as model inputs; the output is the tactical intention encoding of attack, feint, retreat, reconnaissance, surveillance, and electronic interference; after obtaining the model output result, the label value of the model output is converted into the corresponding formation intention.

[0042] Furthermore, the steps to establish a UAV formation intention recognition model based on the random forest algorithm are as follows:

[0043] Step S3-1, randomly extract M samples from the training samples to form a training subset, repeat K times to generate K training subsets;

[0044] Step S3-2, each training subset is trained to generate a decision tree;

[0045] Step S3-3: Repeat step S3-2 to generate a total of K decision trees and establish a random forest formation intention recognition model;

[0046] In step S3-4, the sample to be identified is input into each sub-decision tree of the random forest for identification, and K formation intention identification results are obtained; based on the intention identification results of each decision tree, according to the majority voting principle, the intention with the most votes is taken as the identification result.

[0047] Furthermore, the pigeon flock optimization algorithm is used to optimize the number K of decision trees in the random forest algorithm. The fitness function f of the optimization algorithm is the weighted sum of the training sample recognition accuracy and the test sample recognition accuracy, and the formula is as follows:

[0048]

[0049] Where α is a constant, satisfying 0<α<1, ntrain is the number of training samples, is the number of correctly identified samples in the training sample, n test is the number of test samples, Identify the correct number of samples in the training sample;

[0050] Assume that the number of pigeons in the flock is N and the position of pigeon i is X i =(K i ), the speed is V i =(v i ), at the pigeon's position K i After the update, K i Substitute the value into the random forest algorithm to obtain the recognition accuracy of the random forest model for training samples and test samples, thereby obtaining the fitness f;

[0051] The pigeon flock optimization algorithm divides the iterative process into two stages according to the number of iterations. The ratio of stage 1 to the total number of iterations is NcRate, which is 0.75;

[0052] The number of iterations is in [0,T max NcRate] is stage 1, where T max is the maximum number of iterations; in stage 1, according to the position of pigeon i after the t-1th iteration and speed Update the speed of pigeon i Then at the current location Plus speed Get new location The formula is as follows:

[0053]

[0054] In the formula, R is a constant, rand is a random number uniformly distributed in the range [0,1], is the optimal position of the pigeon group after t-1 iterations, [*] indicates rounding;

[0055] When the number of iterations is greater than T max NcRate, enter phase 2. The number of pigeons in each iteration will be reduced by half, that is, individuals with lower fitness in the pigeon group will be eliminated. Then, the positions of the individual pigeons in the pigeon group will be updated according to the position of the center of the pigeon group. The formula is as follows:

[0056]

[0057] Where N t is the number of pigeons after t iterations, N t-1 is the number of pigeons after t-1 iterations, is the center position of the pigeon flock after t-1 iterations.

[0058] Furthermore, in step S3, the steps of the UAV formation intention recognition algorithm based on the pigeon flock optimization random forest algorithm are as follows:

[0059] The steps of the UAV formation intention recognition algorithm based on the pigeon flock optimization random forest algorithm are as follows:

[0060] Step S3-5, initialize the pigeon population size N, the maximum number of iterations T max , the proportion of stage 1 to the total number of iterations NcRate, the individual position of the pigeon population X i =(K i ) and speed V i =(v i );

[0061] Step S3-6: The position K of each pigeon group i Substitute into the random forest model to obtain the individual fitness f i ;

[0062] Step S3-7: Update the individual position K of the pigeon group according to equations (8) to (12) i and speed v i ;

[0063] Step S3-8: Repeat steps S3-6 and S3-7 to obtain the optimal K value, and substitute the K value into the random forest model to train the UAV formation intention recognition model.

[0064] Compared with the prior art, the present invention adopts the above technical solution and has the following beneficial effects:

[0065] (1) Compared with the existing technology, the method for identifying UAV formation intention based on fuzzy clustering and random forest proposed in this paper introduces attack tendency prediction and replaces seven air combat situation information feature quantities with attack tendency features, thereby reducing the input state quantity of the random forest-UAV formation intention identification model, thereby reducing the number of nodes in the decision tree in the random forest, avoiding the model from falling into overfitting, and improving the model recognition accuracy.

[0066] (2) The random forest algorithm is used to establish the intent recognition model, and the pigeon flock optimization algorithm is introduced to optimize the hyperparameter K of the random forest algorithm, thereby improving the accuracy of the intent recognition model.

[0067] (3) It can accurately identify the tactical intentions of drone formations and provide effective support for decision-making in complex air combat environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1This is the overall flow chart of the UAV formation intention recognition method based on fuzzy clustering and random forest proposed in the present invention;

[0069] Figure 2 A flowchart for target attack tendency prediction based on fuzzy clustering;

[0070] Figure 3 Flowchart of the formation intention recognition model based on random forest. DETAILED DESCRIPTION

[0071] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0072] The method for identifying the intention of a UAV formation comprises the following steps:

[0073] Step S1: Establishing an intention space and intention recognition feature set for the UAV formation. The intention space includes six types of tactical intentions: attack, feint, retreat, reconnaissance, surveillance, and electronic jamming. The intention recognition feature set includes air combat situation information and the target UAV's own state, and establishes a mapping from the target UAV formation state to tactical intentions.

[0074] Step S2: processing the air combat situation information in the intention recognition feature set based on a fuzzy clustering algorithm to predict the attack tendency of the target UAV;

[0075] Step S3: Based on the pigeon flock optimization random forest algorithm, an intention recognition model is established using the attack tendency and the target's own state characteristics to output the tactical intention of the target UAV formation.

[0076] In step S1, the feature quantities for identifying the intention of drone formations in air combat are selected as follows:

[0077] The problem of UAV formation intention recognition requires establishing a mapping from the target UAV formation state in air combat to the UAV formation's tactical intention. In order to accurately identify the target UAV formation's intention, it is necessary to establish an accurate UAV formation intention space and intention recognition state set.

[0078] In different situations, drones will choose different intentions. The air combat situation information includes the relative distance d1, d2 between each target drone and our drone, the target azimuth Target entry angle θ1, θ2, target flight speed v1, v2, our aircraft flight speed v′, target flight altitude h1, h2, our aircraft flight altitude h′.

[0079] The target's status primarily includes its motion state, radar state, jamming state, jammed state, motion trend, and maneuvering behavior. The target's motion state includes the target drone's coordinates {x1, y1, z1} and {x2, y2, z2}, its heading angles ψ1 and ψ2, its track angles γ1 and γ2, and its velocities v1 and v2. The target's radar state indicates a silent penetration feature, with a value of either "yes" or "no." Silent penetration is defined as the target disabling its radar to reduce detection, thereby enabling a surprise attack. The target's jamming state and jammed state are marked as "yes" or "no," reflecting whether the target is implementing or receiving electronic jamming. The target drone's motion trend includes flying toward, flying away from, first toward, then away from, or first away from, then toward. This can be determined by the target's distance from the enemy over a period of time. The target's maneuvering behaviors include 10 types: horizontal straight flight, dive, jump, left circle, right circle, half roll, somersault, upper left combat turn, upper right combat turn, and serpentine maneuver, which can be judged by the target's movement trajectory.

[0080] The above analysis shows that the target drone formation intention recognition requires a large number of features. This results in a large number of nodes in the decision tree of the random forest-drone formation intention recognition model, making the model prone to overfitting. Therefore, the attack tendency feature of each target drone is extracted from the above features to replace the seven air combat situation information features to reduce the input state of the intention recognition model.

[0081] The target attack tendency prediction steps based on the fuzzy clustering algorithm in step S2 are as follows:

[0082] Predicting the target's attack tendency is key to identifying the target's intention. Attack tendency is defined as whether the current target is willing to attack us. Attack tendency characteristics can be quantified as strong, medium, or weak. The prediction of attack tendency depends on the current air combat situation and the state of the target drone. Identification features include target speed v, our own speed v′, target acceleration a, target relative altitude dz, target azimuth The target entry angle θ and the relative distance d between the two sides. During air combat, there is no clear boundary between different attack tendencies, that is, the boundaries are fuzzy. Therefore, to predict attack tendencies, a fuzzy clustering method is used. The basic steps are as follows:

[0083] Step S2-1, feature quantity normalization processing:

[0084] Assume that the target feature set for attack tendency prediction is X = {x1, x2, ..., x n}, where x i represents the characteristic quantity of the i-th target, x ijis the jth feature value of the i-th target, and n is the number of attack tendency prediction samples. The feature values ​​in the set are standardized as follows:

[0085]

[0086] Where x′ ij For x ij The results after standardization are is the average value of the j-th feature, and the formula is as follows:

[0087]

[0088] Step S2-2, establish the fuzzy similarity matrix:

[0089] Establish the fuzzy similarity matrix R0 of the target = (r ij ) n×n . Where r ij represents the target x i With target x j The similarity of the 7 features between ij The calculation formula is as follows:

[0090]

[0091] Where x′ ih is the hth eigenvalue of target i after normalization, x′ jh is the hth eigenvalue of the standardized target j; c is a constant, so that the comprehensive similarity r ij In the range [0,1] and evenly distributed;

[0092] Step S2-3, clustering:

[0093] Before clustering the target, it is necessary to find the transitive closure of R0; use the square method and calculate t(R0) according to the rules of the Zadeh operator (∧,∨): during the square operation, there exists a natural number k that satisfies And 2 k ≤n, then the transitive closure of R0 is the reflexive transitive closure; choose the appropriate confidence level λ∈[0,1], and the classification of attack tendency can be obtained by the truncation matrix; assuming the confidence level λ=λ0, according to the fuzzy rules and transitive closure The corresponding truncation matrix can be obtained, and the fuzzy rule is as follows:

[0094]

[0095] Step S2-4, attack tendency prediction feature vector calculation:

[0096] After clustering, the target set can be divided into t attack tendency combat groups, where each group consists of n l The target is composed of a batch, t = 3. Then the characteristic vector s of the lth attack tendency is l The calculation formula is:

[0097]

[0098] Among them, lp is the serial number of the pth target in the lth attack tendency group, v lp ,v′ lp ,a lp ,dz lp , θ lp ,d lp is the characteristic quantity of the target;.

[0099] Step S2-5, target attack tendency prediction:

[0100] By comparing the target state x′={x′1,x′2,…,x′ m} and each attack tendency feature vector s l Distance D l , select the attack tendency value corresponding to the shortest distance as the target's attack tendency, distance D l The calculation formula is as follows:

[0101]

[0102] The target UAV formation intention identification steps in step S3 are as follows:

[0103] After step S2, the target formation intention identification feature includes each target drone's respective attack tendency, maneuvering behavior, radar status, interference state, interference state, and target motion trend. To identify the drone formation intention, it is necessary to establish a mapping from the above feature to tactical intention. The present invention adopts a pigeon flock optimized random forest algorithm to establish a drone formation intention recognition model, and uses the six state quantities of each target drone's respective attack tendency, maneuvering behavior, radar status, interference state, interference state, and target motion trend as model inputs. The six drone formation tactical intentions of attack, feint, retreat, reconnaissance, surveillance, and electronic interference are encoded as {1, 2, ..., 6} as the output of the intention recognition model. After obtaining the model output result, the label value of the model output is converted into the corresponding formation intention.

[0104] The random forest algorithm selects decision tree models as weak classifiers and then integrates these weak classifiers into a strong classifier. The algorithm first randomly samples the training sample to generate K different sample subsets. It then trains K decision tree models based on these sample subsets. Finally, based on the recognition results of each decision tree model, a majority vote is used to determine the final recognition result. When integrating the recognition results of the decision trees, each decision tree is given the same weight. The K decision trees vote on the sample based on their own recognition results. The classification result with the most votes is the recognition result for that sample.

[0105] The steps of the UAV formation intention recognition model based on the random forest algorithm are as follows:

[0106] Step S3-1, randomly extract M samples from the training samples to form a training subset, repeat K times to generate K training subsets;

[0107] Step S3-2, each training subset is trained to generate a decision tree;

[0108] Step S3-3: Repeat step S3-2 to generate a total of K decision trees and establish a random forest formation intention recognition model;

[0109] In step S3-4, the sample to be identified is fed into each of the random forest's child decision trees for identification, resulting in K formation intent recognition results. Based on the intent recognition results from each decision tree, the intent with the most votes is selected as the recognition result according to the majority voting principle.

[0110] The recognition accuracy of a random forest is affected by the number of decision trees, K. A large K value can easily lead to overfitting, while a small K value can reduce the model's recognition accuracy. Therefore, it is necessary to optimize the random forest hyperparameter, K. The fitness function f of the optimization algorithm is the weighted sum of the recognition accuracy of the training sample and the recognition accuracy of the test sample, as shown in the following formula:

[0111]

[0112] Where α is a constant, satisfying 0<α<1, n train is the number of training samples, is the number of correctly identified samples in the training sample, n test is the number of test samples, To identify the correct number of samples in the training sample, in order to find the optimal K value and thus obtain the best fitness f, the pigeon flock optimization algorithm is used to optimize the K value.

[0113] Pigeon flock optimization algorithm is a swarm intelligence optimization algorithm designed to simulate the homing behavior of pigeons. It has the advantages of simple calculation and strong robustness. Suppose the number of pigeons in the pigeon flock is N, and the position of pigeon i is X. i =(K i), the speed is V i =(v i ), at the pigeon's position K i After the update, K i The value is substituted into the random forest algorithm to obtain the recognition accuracy of the random forest model for training samples and test samples, thereby obtaining the fitness f.

[0114] The pigeon flock optimization algorithm divides the iterative process into two stages according to the number of iterations. The ratio of stage 1 to the total number of iterations is NcRate, which is 0.75.

[0115] The number of iterations is in [0,T max NcRate] is stage 1, where T max is the maximum number of iterations. In stage 1, according to the position of pigeon i after the t-1th iteration and speed Update the speed of pigeon i Then at the current location Plus speed Get new location The formula is as follows:

[0116]

[0117] In the formula, R is a constant, generally taking the value of 0.2, rand is a random number uniformly distributed in the range [0,1], is the optimal position of the pigeon flock after t-1 iterations, and [*] indicates rounding.

[0118] When the number of iterations is greater than T max NcRate, enter phase 2. The number of pigeons in each iteration will be reduced by half, that is, individuals with lower fitness in the pigeon group will be eliminated. Then, the positions of the individual pigeons in the pigeon group will be updated according to the position of the center of the pigeon group. The formula is as follows:

[0119]

[0120] Where N t is the number of pigeons after t iterations, is the center position of the pigeon flock after t-1 iterations, N t-1 is the number of pigeons in the group after t-1 iterations.

[0121] The steps of the UAV formation intention recognition algorithm based on the pigeon flock optimization random forest algorithm are as follows:

[0122] Step S3-5, initialize the pigeon population size N, the maximum number of iterations T max , the proportion of stage 1 to the total number of iterations NcRate, the individual position of the pigeon population X i =(Ki ) and speed V i =(v i );

[0123] Step S3-6, the position K of each pigeon group individual i Substitute into the random forest model to obtain the individual fitness f i ;

[0124] Step S3-7, update the individual position K of the pigeon group according to equations (8) to (12) i and speed v i ;

[0125] In step S3-8, steps S3-6 and S3-7 are repeated to obtain the optimal K value, and the K value is substituted into the random forest model to train the UAV formation intention recognition model.

[0126] 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying UAV formation intention based on fuzzy clustering and random forest, characterized by: The following steps are involved: Step S1: Establishing the intention space and intention recognition feature set of the UAV formation, wherein the intention space includes six types of tactical intentions: attack, feint, retreat, reconnaissance, surveillance, and electronic jamming; The intention recognition feature set includes air combat situation information and the target UAV's own status, establishing a mapping from the target UAV formation status to tactical intention; Step S2: Processing the air combat situation information in the intention recognition feature set based on a fuzzy clustering algorithm to predict the attack tendency of the target UAV, that is, whether the current target is willing to launch an attack on our side; Step S3: Based on the pigeon flock optimization random forest algorithm, an intention recognition model is established using the attack tendency and the target's own state characteristics to output the tactical intention of the target UAV formation.

2. The method for identifying UAV formation intention based on fuzzy clustering and random forest according to claim 1, characterized in that: In the step S1: The air combat situation information includes the relative distance between the target and our UAV, azimuth, entry angle, target flight speed, our flight speed, target flight altitude, and our flight altitude; The target UAV's own state includes each target UAV's motion state, radar state, interference state, interference state, motion trend and maneuvering behavior.

3. The method for identifying UAV formation intention based on fuzzy clustering and random forest according to claim 2, characterized in that: The target motion state includes the coordinates of the UAV, the target heading angle, the target track angle and the target flight speed; The radar status is a silent penetration feature, and its value is yes or no; The jamming state and jammed state reflect whether the target is implementing or being jammed, with the value being yes or no; The movement trend includes flying towards our drone, flying away from our drone, flying towards our drone first and then flying away from our drone, and flying away from our drone first and then flying towards our drone. It is determined by the distance between the target drone and our drone over a period of time. The maneuvering behavior includes 10 trajectory characteristics: horizontal straight flight, dive, jump, left circle, right circle, half roll, dogfight, upper left combat turn, upper right combat turn and serpentine maneuver, which are judged by the target's movement trajectory.

4. The method for identifying UAV formation intention based on fuzzy clustering and random forest according to claim 1, characterized in that: The attack tendency in step S2 can be quantified into three levels: strong, medium, and weak. The prediction of the attack tendency depends on the current air combat situation and the target UAV state identification features, including target speed v, self-drone speed v′, target acceleration a, target relative height dz, target azimuth Target entry angle θ, relative distance d between the two parties.

5. The method for identifying UAV formation intention based on fuzzy clustering and random forest according to claim 1, characterized in that: The step S2 comprises the following steps: Step S2-1: feature quantity normalization processing: Assume that the target feature set for attack tendency prediction is X = {x1, x2, ..., x n }, where x i represents the characteristic quantity of the i-th target, x ij is the jth characteristic value of the i-th target, n is the number of attack tendency prediction samples, and each characteristic value in the set is standardized. The formula is as follows: Where x′ ij For x ij The results after standardization are is the average value of the j-th feature, and the formula is as follows: Step S2-2, establish fuzzy similarity matrix: Establish the fuzzy similarity matrix R0 of the target = (r ij ) n×n ; where r ij represents the target x i With target x j The similarity of the 7 features between ij The calculation formula is as follows: Where x′ ih is the hth eigenvalue of target i after normalization, x′ jh is the hth eigenvalue of the standardized target j; c is a constant, so that the comprehensive similarity r ij In the range [0,1] and evenly distributed; Step S2-3, clustering: Before clustering the target, it is necessary to find the transitive closure of R0; use the square method and calculate t(R0) according to the rules of the Zadeh operator (∧,∨): during the square operation, there exists a natural number k that satisfies And 2 k ≤n, then the transitive closure of R0 is the reflexive transitive closure; choose the appropriate confidence level λ∈[0,1], and the classification of attack tendency can be obtained by the truncation matrix; assuming the confidence level λ=λ0, according to the fuzzy rules and transitive closure The corresponding truncation matrix can be obtained, and the fuzzy rule is as follows: Step S2-4, attack tendency prediction feature vector calculation: After clustering, the target set can be divided into t attack tendency combat groups, where each group consists of n l Batch target composition, t = 3; then the characteristic vector s of the lth attack tendency l The calculation formula is: Among them, lp is the serial number of the pth target in the lth attack tendency group, v lp ,v′ lp ,a lp ,dz lp , θ lp ,d lp is the characteristic quantity of the target; Step S2-5: Target attack tendency prediction: By comparing the normalized target state x′={x1′,x2′,…,x′7′} with each attack tendency feature vector s l Distance D l , select the attack tendency value corresponding to the shortest distance as the target's attack tendency, distance D l The calculation formula is as follows:

6. The method for identifying UAV formation intention based on fuzzy clustering and random forest according to claim 1, characterized in that: In step S3, the intention recognition feature quantities are: attack tendency, maneuvering behavior, radar state, interference state, interference state, and movement trend; a UAV formation intention recognition model is established using a random forest algorithm, and the six state quantities of the target UAVs, namely, attack tendency, maneuvering behavior, radar state, interference state, interference state, and target movement trend, are used as model inputs; and the output is a tactical intention code of attack, feint, retreat, reconnaissance, surveillance, and electronic jamming; After obtaining the model output results, the label values ​​output by the model are converted into corresponding formation intentions.

7. The method for identifying UAV formation intention based on fuzzy clustering and random forest according to claim 1, characterized in that: The steps to establish a UAV formation intention recognition model based on the random forest algorithm are as follows: Step S3-1, randomly extract M samples from the training samples to form a training subset, repeat K times to generate K training subsets; Step S3-2, each training subset is trained to generate a decision tree; Step S3-3: Repeat step S3-2 to generate a total of K decision trees and establish a random forest formation intention recognition model; In step S3-4, the sample to be identified is input into each sub-decision tree of the random forest for identification, and K formation intention identification results are obtained; based on the intention identification results of each decision tree, according to the majority voting principle, the intention with the most votes is taken as the identification result.

8. The method for identifying UAV formation intention based on fuzzy clustering and random forest according to claim 1, characterized in that: The pigeon flock optimization algorithm is used to optimize the number of decision trees K in the random forest algorithm. The fitness function f of the optimization algorithm is the weighted sum of the training sample recognition accuracy and the test sample recognition accuracy, and the formula is as follows: Where α is a constant, satisfying 0<α<1, n train is the number of training samples, is the number of correctly identified samples in the training sample, n test is the number of test samples, Identify the correct number of samples in the training sample; Assume that the number of pigeons in the flock is N and the position of pigeon i is X i =(K i ), the speed is V i =(v i ), at the pigeon's position K i After the update, K i Substitute the value into the random forest algorithm to obtain the recognition accuracy of the random forest model for training samples and test samples, thereby obtaining the fitness f; The pigeon flock optimization algorithm divides the iterative process into two stages according to the number of iterations. The ratio of stage 1 to the total number of iterations is NcRate, which is 0.75; The number of iterations is in [0,T max NcRate] is stage 1, where T max is the maximum number of iterations; in stage 1, according to the position of pigeon i after the t-1th iteration and speed Update the speed of pigeon i Then at the current location Plus speed Get new location The formula is as follows: In the formula, R is a constant, rand is a random number uniformly distributed in the range [0,1], is the optimal position of the pigeon group after t-1 iterations, [*] indicates rounding; When the number of iterations is greater than T max NcRate, enter phase 2. The number of pigeons in each iteration will be reduced by half, that is, individuals with lower fitness in the pigeon group will be eliminated. Then, the positions of the individual pigeons in the pigeon group will be updated according to the position of the center of the pigeon group. The formula is as follows: Where N t is the number of pigeons after t iterations, N t-1 is the number of pigeons after t-1 iterations, is the center position of the pigeon flock after t-1 iterations.

9. The method for identifying UAV formation intention based on fuzzy clustering and random forest according to claim 1, characterized in that: In step S3, the steps of the UAV formation intention recognition algorithm based on the pigeon flock optimization random forest algorithm are as follows: The steps of the UAV formation intention recognition algorithm based on the pigeon flock optimization random forest algorithm are as follows: Step S3-5, initialize the pigeon population size N, the maximum number of iterations T max , the proportion of stage 1 to the total number of iterations NcRate, the individual position of the pigeon population X i =(K i ) and speed V i =(v i ); Step S3-6: The position K of each pigeon group i Substitute into the random forest model to obtain the individual fitness f i ; Step S3-7: Update the individual position K of the pigeon group according to equations (8) to (12) i and speed v i ; Step S3-8: Repeat steps S3-6 and S3-7 to obtain the optimal K value, and substitute the K value into the random forest model to train the UAV formation intention recognition model.

Images