UAV swarm intention inference algorithm and system based on dynamic Bayesian network

Abstract

The present invention discloses an algorithm and system for inferring the intention of a drone cluster based on a dynamic Bayesian network. The method includes S1, establishing an intention feature comparison table according to the combat intention and behavior characteristics of the drone cluster; S2, taking the combat intention in the intention feature comparison table as the root node and the behavior characteristics as the child node, and making the root node first-order state transfer to establish a dynamic Bayesian network inference model; S3, inputting the observation data into the dynamic Bayesian network inference model to infer the combat intention. The present invention infers the cluster intention characteristics through the observation data of the behavior characteristics of the drone cluster. There are many ways to obtain the observation data, and the reliability of inference based on the data is strong.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle technology, and more specifically, to an unmanned aerial vehicle cluster intention inference algorithm and system based on a dynamic Bayesian network. Background Art

[0002] Identification of battlefield target combat intent has always been the focus of commanders at all levels and is the basic basis for deciding the next combat action. Rapidly and accurately analyzing the linkage between battlefield situation elements and grasping the enemy's combat intent can make the confrontation between the two sides develop according to the trajectory of our ideal battle situation, thereby seizing the combat initiative and achieving a quick victory. Under the conditions of informatization, battlefield information has increased dramatically, the enemy-us confrontation is complex, and the air threat is becoming increasingly severe. It is difficult to quickly and accurately identify the intention of enemy targets from multi-source battlefield data simply relying on expert experience. How to use advanced data fusion technology to assist commanders in identifying the enemy's combat intent has become an urgent problem to be solved in the current battlefield situation assessment. The current methods used to solve the analysis of enemy target combat intent in complex battlefield environments mainly include template matching, expert systems, Bayesian networks, neural networks, etc.

[0003] With the continuous promotion of the application of drones, single drones and traditional multi-drone systems have exposed the shortcomings of poor flexibility and low efficiency in task completion when performing tasks, and the autonomous intelligence of drones is constantly improving. Therefore, the use of drone clusters to perform related tasks will definitely become the main trend of future drone development. Many wars in recent years have also demonstrated that the traditional air defense system is almost ineffective against large-scale drone cluster operations, and research on intention recognition for drone clusters is imminent. Existing research on battlefield intention inference problems mainly analyzes the enemy's offensive intention based on the overall battlefield situation. The battlefield environment information required for inference is relatively large, and research on intention inference for drone clusters, a new type of combat style, is still relatively lacking. Xue Xirui et al. established a cluster motion model based on Markov bridging distribution, and then proposed a Bayesian intention inference method based on reachable domain optimization. This method focuses more on the prediction of cluster attack targets and has high requirements for cluster motion model assumptions.

[0004] The composition of drone swarms is complex and diverse, with strong mobility. They can achieve the purpose of saturation attack in a short period of time with high intensity. If they do not rely on continuous observation data and inference, it is very easy to miss key details and delay the opportunity. At the same time, the characteristics of drone swarms, such as low, slow and small, bring great difficulties to the continuous and accurate observation of radar. Existing methods such as template matching and Bayesian networks only infer the battlefield status at a single moment, and cannot infer the moment when they are interfered and observation is restricted. Drone swarm combat can take advantage of the cluster and cooperate to complete tasks that a single drone cannot complete, or complete multiple tasks continuously. Distributed killing will significantly increase the difficulty of air defense and greatly increase the difficulty and complexity of air defense decision-making. To this end, it is necessary to continuously observe and dynamically infer drone swarms, and existing static methods are difficult to achieve this goal.

[0005] Different types of drone swarms often have different combat missions and different levels of threat to us. For different types of threats, targeted countermeasures are designed to help improve the cost-effectiveness of the confrontation. At the same time, the combat intention of drone swarms is not static. Due to the self-organization of drone swarms, the same swarm can adjust the formation and change the flight mode to achieve multiple functions and perform different tasks. Drone swarms can also make adaptive adjustments according to changes in the battlefield environment, or change behavioral characteristics and combat intentions according to ground instructions, and perform multiple combat missions at different stages of an operation. Affected by environmental climate and other conditions, it may not be possible to continuously detect the behavior of drone swarms, and there will be observation omissions or observation errors; and the combat operations of both sides have certain concealment, deception and confrontation. In order to hide their combat intentions, battlefield targets usually release some deceptive false information. Therefore, relying on battlefield situation information at a single moment to infer the target's combat intention is obviously not scientific enough.

[0006] Therefore, in order to target incoming unknown drone swarms, infer the swarm's combat intentions based on the swarm behavior characteristics observed by radar and other detection equipment, and thus make appropriate responses such as interception or anti-interference to the incoming swarm, it is necessary to develop a drone swarm intention inference algorithm and system based on dynamic Bayesian networks. Summary of the invention

[0007] The purpose of the present invention is to provide a drone cluster intention inference algorithm and system based on a dynamic Bayesian network to overcome the defects of the prior art.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] The UAV swarm intention inference algorithm based on dynamic Bayesian network includes:

[0010] S1. Establish an intention feature comparison table based on the combat intention and behavior characteristics of the drone cluster;

[0011] S2. Taking the combat intention in the intention feature comparison table as the root node and the behavior feature as the child node, the first-order state transition of the root node is used to establish a dynamic Bayesian network inference model;

[0012] S3. Input the observation data into the dynamic Bayesian network inference model to infer the combat intent.

[0013] Furthermore, the establishment of an intention feature comparison table according to the combat intention and behavior characteristics of the drone cluster includes:

[0014] Obtain the combat intentions and M behavioral characteristics of N types of drone swarms;

[0015] Discretizing the value of the behavior feature;

[0016] The values ​​of the behavioral characteristics of the drone cluster under different combat intentions are matched to form an intention feature comparison table.

[0017] Furthermore, the intention feature comparison table is represented by a matrix, and for the behavior feature C r The intention feature matching matrix is

[0018]

[0019] Furthermore, the step S2 is specifically as follows: in a single time slice, the combat intention in the intention feature comparison table is used as the root node, the behavior feature is used as the child node, each node is connected to the root node, and a first-order state transfer exists between the root nodes of adjacent time slices to establish the intention feature comparison table.

[0020] Furthermore, the step S2 also includes: setting parameters of the dynamic Bayesian network inference model, including the prior probability of the root node in a single time slice, the conditional probability of the child node in a single time slice, and the state transition probability of the root node;

[0021] The prior probability of the root node in the single time slice is obtained by statistically analyzing the occurrence probability of each combat intention in the combat case database, and is expressed as:

[0022]

[0023] Where n i Intent I in the case database i The frequency of occurrence, n is the total number of cases in the database;

[0024] The conditional probability of the child node in the single time slice is:

[0025]

[0026] In the formula, Represents feature C r The value of the intention I i The number of matches, n F =p r -n T ;

[0027] The state transition probability of the root node is:

[0028]

[0029] Where a is the probability that the drone cluster continues to maintain its previous combat intention, i, j are the states of node I, and n is the state dimension of I.

[0030] Furthermore, the step S3 includes: performing discretization preprocessing on the observed data and inputting the preprocessed data into the dynamic Bayesian network inference model for inference.

[0031] The present invention also provides a system according to the above-mentioned drone cluster intention inference algorithm based on dynamic Bayesian network, comprising:

[0032] A construction module is used to establish an intention feature comparison table based on the combat intention and behavior characteristics of the drone cluster;

[0033] A model module is used to establish a dynamic Bayesian network inference model by taking the combat intention in the intention feature comparison table as the root node and the behavior feature as the child node, and making the first-order state transition of the root node;

[0034] The inference module is used to input the observation data into the dynamic Bayesian network inference model to infer the combat intent.

[0035] Compared with the prior art, the advantages of the present invention are: a drone cluster intention inference algorithm and system based on a dynamic Bayesian network provided by the present invention infers cluster intention characteristics by observing data of drone cluster behavior characteristics. There are many ways to obtain observation data, and the reliability of inference based on data is strong. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1It is a flow chart of the drone cluster intention inference algorithm based on dynamic Bayesian network of the present invention.

[0038] Figure 2 It is a dynamic Bayesian network inference model diagram of the present invention.

[0039] Figure 3 The present invention utilizes GeNIe software to establish an inference network.

[0040] Figure 4 It is a schematic diagram of the inference results of the present invention.

[0041] Figure 5 The present invention utilizes GeNIe software to establish a dynamic Bayesian network inference model.

[0042] Figure 6 It is a schematic diagram of the inference results of the dynamic Bayesian network model of the present invention.

[0043] Figure 7 This is a comparison chart of the inference results of the DBN model and the BN model.

[0044] Figure 8 It is a framework diagram of the drone cluster intention inference system based on dynamic Bayesian network of the present invention. DETAILED DESCRIPTION

[0045] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0046] Embodiment 1

[0047] See also Figure 1 As shown, this embodiment discloses a drone cluster intention inference algorithm based on a dynamic Bayesian network, including:

[0048] Step S1: Establish an intention feature comparison table based on the combat intention and behavior characteristics of the drone cluster.

[0049] Specifically, step S1 includes:

[0050] Obtain the combat intentions and M behavioral characteristics of N types of drone swarms. With the development of technology, the tasks performed by drone swarms are constantly expanding, and air defense detection equipment is also constantly upgrading, so the mainstream combat intentions and swarm behavioral characteristics should be summarized according to the specific situation of the current era.

[0051] Discretize the values ​​of the behavioral features. The behavioral features of drones are mostly continuous variables, and it is more complicated for Bayesian networks to process continuous variables. The behavioral feature values ​​of the drone cluster should be divided first and discretized into qualitative variables. The number of qualitative values ​​p can be determined according to the size of the data set. The more values, the higher the inference accuracy, but too many values ​​will greatly increase the complexity of network inference.

[0052] The values ​​of the behavioral characteristics of the drone cluster under different combat intentions are matched to form an intention characteristic comparison table. Through statistical analysis of the battle case database and combined with expert experience, the values ​​of the behavioral characteristics of the drone cluster under different combat missions are matched. The behavioral characteristic values ​​corresponding to each combat intention can be a single value, or a double value or multiple values.

[0053] Moreover, to facilitate subsequent processing, the intention feature comparison table is represented by a matrix. r The intention feature matching matrix is

[0054]

[0055] Step S2: Take the combat intention in the intention feature comparison table as the root node, the behavior feature as the child node, and make the first-order state transition of the root node to establish a dynamic Bayesian network inference model.

[0056] Specifically, in a single time slice, the combat intention in the intention feature comparison table is taken as the root node I, and the value is the N combat intentions summarized in step S1; the behavior feature is taken as the child node C r ,r=1,2,…,M, child node C r The value of is the discretized p in step S1 r Qualitative values ​​are taken, each node is connected to the root node, and the root node has a first-order state transition between adjacent time slices to establish an intention feature comparison table. The specific network structure is as follows Figure 2 shown.

[0057] Specifically, step S2 also includes: setting parameters of the dynamic Bayesian network inference model, including the prior probability of the root node in a single time slice, the conditional probability of the child node in a single time slice, and the state transition probability of the root node;

[0058] The prior probability of the root node in the single time slice is obtained by statistically analyzing the occurrence probability of each combat intention in the combat case database, and is expressed as:

[0059]

[0060] Where n i Intent I in the case database iThe frequency of occurrence, n is the total number of cases in the database;

[0061] Conditional probability of child nodes in a single time slice: The conditional probability of child nodes is set according to the intent feature comparison table, and a leakage probability of 0.05 is set for unconnected values ​​to ensure that all nodes are open during the inference process. i Next, feature C r The probability of each value of is:

[0062]

[0063] In the formula, Represents feature C r The value of the intention I i The number of matches, n F =p r -n T ;

[0064] The state transition probability of the root node is:

[0065]

[0066] Where a is the probability that the drone cluster continues to maintain its previous combat intention, i, j are the states of node I, and n is the state dimension of I.

[0067] Step S3: input the observation data into the dynamic Bayesian network inference model to infer the combat intent.

[0068] Specifically, step S3 includes:

[0069] The observed data is discretized and preprocessed. The input data from the observed data to the model must first be preprocessed according to the discretization scale in step S1 to make it conform to the value set of the node in the network model. For the observed drone cluster, if the drones show convergence characteristics between clusters, they are considered to be the same cluster, otherwise they are considered to be different clusters. For each drone in the same cluster, if the values ​​of the characteristic parameters are not much different and are in the same qualitative interval, then the value is taken as the qualitative feature of the cluster. If the difference is large and is in multiple qualitative intervals, the characteristic node is ignored.

[0070] The preprocessed data is input into the dynamic Bayesian network inference model for inference. The inference process can be regarded as a given set of T sequential time feature observation variables. For the hidden variable sequence of combat intention The conditional probability distribution of The inference process can be completed through forward propagation or backward propagation algorithms.

[0071] In this embodiment, the network model establishment in step S2 and the inference process in step S3 can be implemented using GeNIe software. GeNIe software is a development environment for building graphical decision theory models. It can realize structural modeling, inference and learning of graph models such as Bayesian networks, and use it for probabilistic reasoning and decision-making under uncertainty. The network model established in the software and the schematic diagram of the inference result are shown in the attached figure. Figure 3 and attached Figure 4 .

[0072] See also Figure 8 As shown, this embodiment also provides a system according to the above-mentioned UAV cluster intention inference algorithm based on the dynamic Bayesian network, including: a construction module 1, used to establish an intention feature comparison table according to the combat intention and behavior characteristics of the UAV cluster; a model module 2, used to use the combat intention in the intention feature comparison table as the root node, the behavior characteristics as the child nodes, and make the root node first-order state transition to establish a dynamic Bayesian network inference model; an inference module 3, used to input the observation data into the dynamic Bayesian network inference model to infer the combat intention.

[0073] The present invention starts from the observation of the behavioral characteristics of drone clusters, and based on the matching of cluster behavioral characteristics and combat intentions, infers the combat intentions of the incoming drone clusters. The required observation information all depends on the observation of the drone clusters, which can be relied on the air defense detection system, is easy to obtain, and has little dependence on the battlefield situation.

[0074] When defining the conditional probability of the network, the present invention sets a certain leakage probability to ensure that all nodes are open during the inference process.

[0075] The present invention establishes a dynamic Bayesian network inference model, and connects the previous and next moments through the state transfer of the root node. It can process observations with continuous time series, and can effectively handle the situation where there are missing and erroneous observations of the behavioral characteristics of the drone cluster, and the model has high robustness. In addition, the root node is defined to have a first-order Markov property. On the basis of the static Bayesian network, the dynamic Bayesian network combines time information and introduces the state transfer probability of the successive time slices, taking into account the impact of the previous time slice on its subsequent events. By introducing the concept of time series, the Markov property of the node in the dynamic Bayesian network inference effectively compensates for the impact of a single observation on the inference result, thereby improving the robustness of the system. At the same time, the temporal continuity of the dynamic Bayesian network can effectively compensate for the problems of missing and erroneous observations of the behavioral characteristics of the drone cluster, greatly improving the robustness of the model.

[0076] The present invention sets a certain leakage probability for non-matching features to ensure that all nodes are opened during the inference process, thereby enhancing the fault tolerance and universality of the model.

[0077] Embodiment 2

[0078] Establishing a Bayesian network inference model can also basically achieve the purpose of using the behavioral characteristics of drone clusters to infer their combat intentions, but the Bayesian network has no temporal continuity, poor dynamic inference ability, and the robustness of the model is worse than that of the dynamic model. Figure 5 The inferred effect is shown in the attached figure. Figure 6 , under the same input, the comparison of the inference results of the dynamic Bayesian network inference model and the Bayesian network inference model is shown in the attached figure. Figure 7 The specific implementation plan is:

[0079] In the second embodiment, a Bayesian network inference model is established, and the network establishment method is the same as the network structure in a single time slice of the dynamic Bayesian network.

[0080] The network parameter definition is the same as the probability definition in the prior probability of the root node in a single time slice and the conditional probability of the child node in a single time slice in the first embodiment, and there is no state transfer.

[0081] The inference process in step S3 of Example 1 uses a variational inference method, taking the true intention of the drone cluster as the latent variable, constructing the likelihood function of the observed variable, and rewriting the likelihood function as the sum of the variational lower bound and the KL divergence based on variational inference. The distribution of the latent variable is iteratively solved so that the posterior probability of the observation takes a maximum value.

[0082] The remaining steps are the same as those in the dynamic Bayesian network inference model technical solution.

[0083] Although the embodiments of the present invention are described in conjunction with the accompanying drawings, the patent owner may make various variations or modifications within the scope of the appended claims. As long as they do not exceed the protection scope described in the claims of the present invention, they should be within the protection scope of the present invention.

Claims

1. The UAV cluster intention inference algorithm based on dynamic Bayesian network is characterized by: include: S1. Establish an intention feature comparison table based on the combat intention and behavior characteristics of the drone cluster; S2. Taking the combat intention in the intention feature comparison table as the root node and the behavior feature as the child node, the first-order state transition of the root node is used to establish a dynamic Bayesian network inference model; S3, inputting the observation data into the dynamic Bayesian network inference model to infer the combat intent; The intention feature comparison table is represented by a matrix, and the behavior feature C r The intention feature matching matrix is The step S2 also includes: setting parameters of the dynamic Bayesian network inference model, including the prior probability of the root node in a single time slice, the conditional probability of the child node in a single time slice, and the state transition probability of the root node; The prior probability of the root node in the single time slice is obtained by statistically analyzing the occurrence probability of each combat intention in the combat case database, and is expressed as: Where n i Intent I in the case database i The frequency of occurrence, n is the total number of cases in the database; The conditional probability of the child node in the single time slice is: In the formula, Represents feature C r The value of the intention I i The number of matches, n F =p r -n T ; The state transition probability of the root node is: Where a is the probability that the drone cluster continues to maintain its previous combat intention, i, j are the states of node I, and N is the state dimension of I.

2. The UAV cluster intention inference algorithm based on dynamic Bayesian network according to claim 1 is characterized in that: The intention feature comparison table established according to the combat intention and behavior characteristics of the drone cluster includes: Obtain the combat intentions and M behavioral characteristics of N types of drone swarms; Discretizing the value of the behavior feature; The values ​​of the behavioral characteristics of the drone cluster under different combat intentions are matched to form an intention feature comparison table.

3. The UAV cluster intention inference algorithm based on dynamic Bayesian network according to claim 1 is characterized in that: The step S2 is specifically as follows: in a single time slice, the combat intention in the intention feature comparison table is used as the root node, the behavior feature is used as the child node, each node is connected to the root node, and a first-order state transfer exists between the root nodes of adjacent time slices to establish the intention feature comparison table.

4. The UAV cluster intention inference algorithm based on dynamic Bayesian network according to claim 1 is characterized in that: The step S3 includes: performing discretization preprocessing on the observed data and inputting the preprocessed data into the dynamic Bayesian network inference model for inference.

5. A system for the drone cluster intention inference algorithm based on a dynamic Bayesian network according to any one of claims 1 to 4, characterized in that: include: A construction module is used to establish an intention feature comparison table based on the combat intention and behavior characteristics of the drone cluster; A model module is used to establish a dynamic Bayesian network inference model by taking the combat intention in the intention feature comparison table as the root node and the behavior feature as the child node, and making the first-order state transition of the root node; The inference module is used to input the observation data into the dynamic Bayesian network inference model to infer the combat intent.

Images