CN118071161B - Method and system for evaluating threat of air cluster target under small sample condition - Google Patents

Abstract

The invention discloses an air cluster target threat assessment method and system under a small sample condition, wherein the method comprises the following steps: combining expert experience and a discrete firefly algorithm, establishing a dynamic Bayesian network based on the air cluster target characteristics, and determining the structure of a threat assessment model; the parameters of the threat assessment model are learned by adopting a small sample parameter learning method based on data expansion, so that the construction of the threat assessment model is completed; and discretizing the enemy target information, inputting the discretized enemy target information into an established dynamic Bayesian network to obtain threat level probability distribution, and obtaining a specific threat value through a good-bad solution distance method. The threat assessment method provided by the invention can be used for carrying out threat assessment on the air cluster target in real time, and still has higher assessment accuracy under the condition of lack of sample information.

Description

Aerial cluster target threat assessment method and system under small sample conditions

Technical Field

The present invention relates to the field of situation awareness of aerial cluster targets, and in particular to a threat assessment method and system for aerial cluster targets under small sample conditions.

Background technique

With the rapid development of swarm intelligence technology and communication technology, cluster targets have begun to be widely used in various aerial combat scenarios such as reconnaissance and penetration. Compared with single aerial targets, aerial cluster targets have advantages such as strong anti-destruction and distributed functions, and can more efficiently coordinate and complete various combat tasks such as fire strikes. Therefore, future aerial combat will be combat between clusters. In this context, accurate situation assessment of enemy cluster targets is the premise for us to make reasonable combat decisions. Threat assessment, as an important part of situation assessment, is the basis for firepower allocation.

However, there are many difficulties in the threat assessment process. First, the powerful agile coordination capability of enemy air cluster targets makes combat operations more flexible and increases the variability of battlefield situations. Specifically, the speed, position and other information of enemy targets change rapidly, which places high demands on the real-time and accuracy of cognitive models. Secondly, the model structure of most current threat assessment methods is built entirely based on expert experience. However, this approach completely ignores the information contained in battlefield sample data and is too subjective. Finally, the construction of threat assessment models requires a large amount of sample data as support. However, enemy information on the real battlefield is difficult to obtain. Therefore, the data used for model construction is usually less in the early stages of combat, which places high demands on the ability to build models with small samples.

Summary of the invention

In view of the above-mentioned difficulties faced by aerial cluster target threat assessment, the present invention proposes a method and system for aerial cluster target threat assessment under small sample conditions, which overcomes the difficulties faced by the problem, such as high real-time requirements, overly subjective models, and small number of samples, through methods such as dynamic Bayesian networks, structural learning based on discrete firefly algorithm, and small sample parameter learning algorithm based on data expansion.

The technical solution to achieve the purpose of the present invention is: a method for assessing the threat of aerial cluster targets under small sample conditions, comprising the following steps:

Step 1: Obtain basic information of enemy air cluster targets, including target position, target speed, target type, target quantity, and cluster formation;

Step 2: Select the features of the aerial cluster targets and learn the dynamic Bayesian network model structure of threat assessment through the discrete firefly algorithm;

Step 3, based on the model structure in step 2, determine the model parameters using a small sample parameter learning algorithm based on data expansion;

Step 4: Input the enemy target information in step 1 into the threat assessment model for inference to obtain the threat level probability distribution, and then process it through the superior and inferior solution distance method to obtain the final threat assessment value;

Step 5: Output the threat value and ranking of all air cluster targets in the current time slice.

An air cluster target threat assessment system under small sample conditions, comprising:

The first module is used to obtain basic information about enemy air cluster targets, including target position, target speed, target type, target quantity, and cluster formation;

The second module selects the features of aerial cluster targets based on expert experience and learns the dynamic Bayesian network model structure of threat assessment through the discrete firefly algorithm;

The third module, based on the model structure in the second module, uses a small sample parameter learning algorithm based on data expansion to determine the model parameters;

The fourth module is used to input the enemy target information in the first module into the threat assessment model for inference, obtain the threat level probability distribution, and then obtain the final threat assessment value through the superior and inferior solution distance method;

The fifth module is used to output the threat value and ranking of all air cluster targets in the current time slice.

An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the threat assessment method for aerial cluster targets under small sample conditions is implemented.

A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for assessing the threat of aerial cluster targets under small sample conditions.

The present invention adopts the above technical solution, and can achieve the following technical effects:

Threat factors are selected according to expert experience, and the dynamic Bayesian network (DBN) structure is learned through the discrete firefly algorithm. On this basis, the DBN parameters are determined through a small sample parameter learning algorithm based on data expansion. The obtained model can effectively infer the threat probability value of aerial cluster targets in real time. Based on the reasoning results of the above threat assessment model, the superior and inferior solution distance method is used to convert the probability value into a specific threat value, which improves the intuitiveness of the result and facilitates the threat degree sorting.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a block diagram showing the principle of the aerial cluster target threat assessment method of the present invention.

Figure 2 is a structural diagram of the aerial cluster target threat assessment model.

FIG3 is a schematic diagram showing a comparison of the assessment accuracy of the threat assessment method proposed by the present invention and the traditional method.

Detailed ways

The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

As shown in Figure 1, the present invention mainly includes: an initialization module, a threat assessment model construction module based on a dynamic Bayesian network (DBN), a threat probability distribution numerical module based on the superior and inferior solution distance method, and a cluster target threat value output module. The specific implementation steps are as follows in conjunction with the flowchart:

Step 1: Initialization: The operator inputs the characteristic information of several air cluster targets within a period of time, such as the number of targets, target type, target location, and target carrying load.

Step 2: Formulate the DBN structure by combining expert experience and historical sample information. Before starting to learn the structure, it is necessary to comprehensively consider the characteristics of aerial cluster targets based on expert experience, and select threat assessment indicators from three parts: situation indicators, capability indicators, and cluster indicators. Among them, situation indicators reflect the state of the target during threat assessment, such as speed, location information, etc.; capability indicators reflect the combat capability of the target, such as electronic jamming capability and the type of payload carried; cluster indicators refer to indicators that are unique to cluster targets compared to single targets, such as cluster formation, the number of targets in the cluster, etc. After determining the features that need to be considered for threat assessment, the specific connections of the network are learned according to the discrete firefly algorithm. In the Bayesian network, connections represent causal relationships between nodes. The specific steps are as follows:

First, initialize the parameters: — initial attraction coefficient, — light absorption coefficient, — number of iterations, —Firefly population size. Randomly generated based on the threat assessment indicators mentioned above A Bayesian network structure is constructed and converted into a string as the initial population of fireflies. The fitness of all fireflies in the population is calculated and used express, , Firefly BIC scores of the structures represented.

Next, start iterating, traversing all fireflies, and set Fireflies ,if The fitness of is the maximum value in the population. In order to avoid falling into the local optimal solution, it is mutated, that is, randomly flipped is an element in the structure matrix of If the fitness is not the maximum value, then traverse all fireflies again, and set the traversed fireflies to ,like ,but Towards The position of is moved, the formula is as follows:

(1)

(2)

(3)

in, and Respectively represent and The string obtained by transforming the Bayesian network structure corresponding to each firefly contains elements, is the number of network nodes, is a length of A vector representing the distance moved, Indicates the elements, Respectively and The elements, is the attraction coefficient, is the square of the Euclidean distance between the positions of fireflies. After moving once, if the fitness is better than before the move, the individual retains the result of the move, otherwise the individual remains unchanged.

When all fireflies have finished moving, the population after the move is used as the initial population for the next iteration.

Finally, check whether the number of iterations has been reached. If so, output the threat assessment model structure corresponding to the historical optimal fitness, otherwise continue to iterate.

Step 3: Learn DBN parameters. Considering the problem of insufficient learning samples in the early stage of the battlefield, the parameters are learned through data expansion. First, a constraint set is formulated for DBN parameters based on expert experience. , the original and expanded data sets are scored according to the degree of compliance of the parameters with different constraints. The higher the score, the better the learning effect of the data set. The constraints used in this algorithm and the corresponding scoring formula are listed below. For Node Take the first state, and the parent node set is the The network parameters in the state, that is, the probability of this situation occurring, Represent the scores of the corresponding constraint types. The formula is as follows:

(1) Range constraint: This constraint defines the constraint area of a parameter. and are the left and right endpoints of the interval respectively, and the formula is as follows:

(4)

Rating calculation formula:

(5)

(2) Inequality constraint: This constraint defines the relationship between two parameters. For Different another parameter, the formula is as follows:

(6)

Rating calculation formula:

(7)

(3) Approximate equality constraint: This constraint defines the approximate equality relationship between parameters. is the confidence interval, take 0.1, and the formula is as follows:

(8)

Rating calculation formula:

(9)

After completing the DBN parameter constraints, the maximum likelihood estimation is used as the basic parameter learning algorithm, and the original data set Parameter learning is performed for samples, and the formula is as follows:

(10)

in, The data set satisfies The total number of sample data for the given node value state combination, express middle Total After completing parameter learning, according to the constraint set , calculate the score, set as .

Next, the original data set is randomly sampled M times with replacement using the bootstrap method to obtain the extended data set , respectively use The M groups of sample data in the above example are used for parameter learning and score calculation, and the dataset with the highest score is selected. , the rating is set to .like , then keep the extended data set, otherwise discard the extended data set and resample until a satisfying data set so far.

Finally, based on the data set and Parameters learned The final network parameters are obtained by weighted summation using the following formula ,in and Represents the data sets and The number of samples in .

(11)

Step 4: After determining the structure and parameters of DBN, it is necessary to perform threat value inference based on the target information obtained in step 1. The inference uses a common interface algorithm as the inference algorithm. This method has the advantages of fast inference speed and accurate results. The ultimate goal of threat assessment is to rank the threat level of the target. However, the result of DBN inference is a probability distribution, which is difficult to rank directly. Therefore, the superiority and inferiority distance method is used to quantify the threat probability distribution into a specific value of 0-1, which is convenient for subsequent sorting. Suppose the DBN inference result is , They represent the probability values of the target threat level belonging to low, medium and high levels respectively. Assume that the positive and negative ideal solutions of the threat probability distribution are , , which means the target is completely a high threat target or a low threat target. The threat value calculation formula is as follows:

(12)

in and Distribution represents the target threat probability distribution and and The Euclidean distance of It is the threat value of the target. The larger the value, the greater the threat the target poses to us.

Step 5: Output the threat value and ranking of all air cluster targets in the current time slice.

The structure of the threat assessment model of the present invention is shown in Figure 2, and the experimental results are shown in Figure 3. Among them, the hardware environment is: CPU i7-12700H @ 2.30 GHz, memory 24G, a total of 5 air cluster targets are subjected to real-time threat assessment, and in order to simulate the small sample situation, threat assessment is performed when the sample size is 50, 200, 400, 600, 800, and 1000 respectively. Figure 3 shows the comparison of the threat assessment accuracy of the method proposed by the present invention and the traditional method under different sample sizes. The assessment accuracy refers to the similarity between the threat ranking and the result given according to expert experience. The higher the accuracy, the closer the assessment result is to the human assessment. The traditional method refers to the threat assessment DBN model constructed based entirely on expert experience without relying on data. It can be seen from the results of Figure 3 that compared with the traditional method, the threat assessment method proposed by the present invention can obtain more accurate threat assessment results when samples are scarce.

The established threat assessment model for aerial cluster targets under small sample conditions can build a more reasonable threat assessment model when samples are scarce, thereby improving the accuracy of threat assessment for aerial cluster targets under small sample conditions. When constructing the threat assessment model, the discrete firefly algorithm is used to learn the model structure based on expert experience, and the parameter learning algorithm based on data expansion is used to improve the accuracy of parameter learning under small sample conditions.

The present invention provides a method for assessing the threat of air cluster targets under small sample conditions. There are many methods and ways to implement the technical solution. The above is only a preferred implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims (6)

1. A method for assessing the threat of aerial cluster targets under small sample conditions, characterized in that it comprises the following steps: Step 1: Obtain basic information of enemy air cluster targets, including target position, target speed, target type, target quantity, and cluster formation; Step 2: Select the features of the aerial cluster targets and learn the dynamic Bayesian network model structure of threat assessment through the discrete firefly algorithm. The specific steps are as follows: Step 2-1, formulate the DBN structure, and select threat assessment indicators from three parts: situation indicators, capability indicators, and cluster-specific indicators; Step 2-2, use discrete firefly algorithm to learn the structure; initialization parameters: — initial attraction coefficient, — light absorption coefficient, — number of iterations, —Firefly population size; randomly generated A Bayesian network structure is constructed and converted into a string as the initial population of fireflies. The fitness of all fireflies in the population is calculated and used express, , Firefly BIC score of the structure represented; Step 2-3, start iteration, traverse all fireflies, set Fireflies ,if The fitness of is the maximum value in the population. In order to avoid falling into the local optimal solution, it is mutated, that is, randomly flipped An element in the structure matrix of If the fitness is not the maximum value, then traverse all fireflies again, and set the traversed fireflies to ,like ,but Towards The position of is moved, the formula is as follows: (1) (2) (3) in, and Respectively represent and The string obtained by transforming the Bayesian network structure corresponding to each firefly contains elements, is the number of network nodes, is a length of A vector representing the distance moved, Indicates the elements, Respectively and The elements, is the attraction coefficient, is the square of the Euclidean distance between the positions of fireflies. After moving once, if the fitness is better than before the move, the individual retains the result of the move, otherwise the individual remains unchanged; When all fireflies have finished moving, the population after the move is used as the initial population for the next iteration; Step 2-4, check whether the number of iterations has been reached. If so, output the threat assessment model structure corresponding to the historical optimal fitness, otherwise continue to iterate; Step 3: Based on the model structure in step 2, the model parameters are determined using a small sample parameter learning algorithm based on data expansion. The specific steps are as follows: Step 3-1: Set constraints on DBN parameters , the original and augmented datasets are scored based on how well the parameters conform to different constraints; set up For Node Take the first state, and the parent node set is the The network parameters in the state, that is, the probability of this situation occurring, They represent the scores of the corresponding constraint types respectively; the constraint types are as follows: (1) Range constraint: This constraint defines the constraint area of a parameter. and are the left and right endpoints of the interval respectively, and the formula is as follows: (4) Rating calculation formula: (5) (2) Inequality constraint: This constraint defines the relationship between two parameters. For Different another parameter, the formula is as follows: (6) Rating calculation formula: (7) (3) Approximate equality constraint: This constraint defines the approximate equality relationship between parameters. is the confidence interval, and the formula is as follows: (8) Rating calculation formula: (9) Step 3-2, based on maximum likelihood estimation, the original data set Parameter learning is performed for samples, and the formula is as follows: (10) in, The data set satisfies The total number of sample data for the given node value state combination, express middle Total After completing parameter learning, according to the constraint set , calculate the score, set as ; Step 3-3: Perform M random samplings with replacement on the original data set using the bootstrap method to obtain the extended data set. , respectively use The M groups of sample data in the above example are used for parameter learning and score calculation, and the dataset with the highest score is selected. , the rating is set to ;like , then keep the extended data set, otherwise discard the extended data set and resample until a satisfying So far, Steps 3-4 will be based on the data set and Parameters learned The final network parameters are obtained by weighted summation using the following formula ,in and Represents the data sets and The number of samples in : (11) Step 4: Input the enemy target information in step 1 into the threat assessment model for inference to obtain the threat level probability distribution, and then process it through the superior and inferior solution distance method to obtain the final threat assessment value; Step 5: Output the threat value and ranking of all air cluster targets in the current time slice. 2. The method for assessing air cluster target threats under small sample conditions according to claim 1, wherein step 4 comprises: Step 4-1, performing threat value inference based on the target information obtained in step 1, using the interface algorithm as the inference algorithm; Step 4-2: Use the superior-inferior solution distance method to quantify the threat probability distribution into a specific value of 0-1 for subsequent sorting. Suppose the DBN inference result is , representing the probability of the target threat level belonging to low, medium and high levels respectively. The positive and negative ideal solutions of the threat level are , , which means the target is completely a high threat target or a low threat target. The threat value calculation formula is as follows: (12) in and Distribution represents the target threat probability distribution and and The Euclidean distance of The threat value of the target. The larger the value, the greater the threat of the target to us. Step 4-3: If there are multiple targets for threat assessment, the threat levels of the targets are ranked according to the results of step 4-2. 3. A threat assessment system for air cluster targets under small sample conditions, characterized by comprising: The first module is used to obtain basic information about enemy air cluster targets, including target position, target speed, target type, target quantity, and cluster formation; The second module is used to select the features of aerial cluster targets and learn the dynamic Bayesian network model structure of threat assessment through the discrete firefly algorithm. Specifically: 1) Formulate the DBN structure and select threat assessment indicators from three parts: situation indicators, capability indicators, and cluster-specific indicators; 2) Use discrete firefly algorithm to learn the structure; initialization parameters: — initial attraction coefficient, — light absorption coefficient, — number of iterations, —Firefly population size; randomly generated A Bayesian network structure is constructed and converted into a string as the initial population of fireflies. The fitness of all fireflies in the population is calculated and used express, , Firefly BIC score of the structure represented; 3) Start iteration, traverse all fireflies, set Fireflies ,if The fitness of is the maximum value in the population. In order to avoid falling into the local optimal solution, it is mutated, that is, randomly flipped An element in the structure matrix of If the fitness is not the maximum value, then traverse all fireflies again, and set the traversed fireflies to ,like ,but Towards The position of is moved, the formula is as follows: (1) (2) (3) in, and Respectively represent and The string obtained by transforming the Bayesian network structure corresponding to each firefly contains elements, is the number of network nodes, is a length of A vector representing the distance moved, Indicates the elements, Respectively and The elements, is the attraction coefficient, is the square of the Euclidean distance between the positions of fireflies. After moving once, if the fitness is better than before the move, the individual retains the result of the move, otherwise the individual remains unchanged; When all fireflies have finished moving, the population after the move is used as the initial population for the next iteration; 4) Check whether the number of iterations has been reached. If so, output the threat assessment model structure corresponding to the historical optimal fitness, otherwise continue to iterate; The third module, based on the model structure in the second module, uses a small sample parameter learning algorithm based on data expansion to determine the model parameters, specifically: (1) Establishing constraints on DBN parameters , the original and augmented datasets are scored based on how well the parameters conform to different constraints; set up For Node Take the first state, and the parent node set is the The network parameters in the state, that is, the probability of this situation occurring, They represent the scores of the corresponding constraint types respectively; the constraint types are as follows: 1) Range constraint: This constraint defines the constraint area of a parameter. and are the left and right endpoints of the interval respectively, and the formula is as follows: (4) Rating calculation formula: (5) 2) Inequality constraint: This constraint defines the relationship between two parameters. For Different another parameter, the formula is as follows: (6) Rating calculation formula: (7) 3) Approximate equality constraint: This constraint defines the approximate equality relationship between parameters. is the confidence interval, and the formula is as follows: (8) Rating calculation formula: (9) (2) Based on the maximum likelihood estimation parameter learning algorithm, the original data set Parameter learning is performed for samples, and the formula is as follows: (10) in, The data set satisfies The total number of sample data for the given node value state combination, express middle Total After completing parameter learning, according to the constraint set , calculate the score, set as ; (3) The original data set is randomly sampled M times with replacement using the bootstrap method to obtain the extended data set , respectively use The M groups of sample data in the above example are used for parameter learning and score calculation, and the dataset with the highest score is selected. , the rating is set to ;like , then keep the extended data set, otherwise discard the extended data set and resample until a satisfying So far, (4) Based on the data sets and Parameters learned The final network parameters are obtained by weighted summation using the following formula ,in and Represents the data sets and The number of samples in : (11) The fourth module is used to input the enemy target information in the first module into the threat assessment model for inference, obtain the threat level probability distribution, and then obtain the final threat assessment value through the superior and inferior solution distance method; The fifth module is used to output the threat value and ranking of all air cluster targets in the current time slice. 4. The air cluster target threat assessment system under small sample conditions according to claim 3 is characterized in that the specific implementation method of the fourth module is: (1) Perform threat value inference based on the target information obtained in the first module, using the interface algorithm as the inference algorithm; (2) Use the superior-inferior solution distance method to quantify the threat probability distribution into a specific value between 0 and 1, which is convenient for subsequent sorting. Suppose the DBN inference result is , representing the probability of the target threat level belonging to low, medium and high levels respectively. The positive and negative ideal solutions of the threat level are , , which means the target is completely a high threat target or a low threat target. The threat value calculation formula is as follows: (12) in and Distribution represents the target threat probability distribution and and The Euclidean distance of The threat value of the target. The larger the value, the greater the threat of the target to us. (3) If there are multiple targets for threat assessment, the threat levels of the targets are ranked according to the results of the previous step. 5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for assessing the threat of aerial cluster targets under small sample conditions as described in any one of claims 1 to 2 is implemented. 6. A computer-readable storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the method for assessing the threat of aerial cluster targets under small sample conditions as described in any one of claims 1-2 is implemented.