Dynamic dimension air combat situation information fusion method based on set transformer

Abstract

The application discloses a dynamic dimension air combat situation information fusion method based on a Set Transformer, belongs to the field of air combat situation information fusion, and comprises the following steps: fixed dimension and dynamic dimension air combat situation information categories are respectively defined and divided, dynamic dimension information categories in the air combat situation information are respectively expressed as set data, a Set Transformer neural network structure is defined, set data is respectively input into the Set Transformer to obtain set features with a fixed number and dimension, fixed dimension information in the air combat situation information is spliced with the set features with the fixed number and dimension output by the Set Transformer, and the spliced set features are taken as the input of subsequent machine learning technology; the air combat situation information fusion method provided by the application avoids the defect that a traditional feedforward neural network can only process fixed dimension input, realizes lossless and efficient processing of dynamic dimension input, and facilitates expansion of an upper algorithm between any number of formation battles.

Description

Technical Field

[0001] This invention belongs to the field of air combat situation information fusion, and specifically, it relates to a dynamic dimension air combat situation information fusion method based on Set Transformer. Background Technology

[0002] In recent years, with the continuous development of machine learning technology, more and more research has begun to use machine learning to handle air combat-related tasks, such as providing decision support and situation assessment. Modern machine learning techniques often employ deep neural networks to fuse and extract features from high-dimensional inputs. For example, reinforcement learning techniques use deep neural networks to directly learn and establish a mapping between air combat situation and aircraft behavior. Therefore, the reasonable representation and fusion of air combat situation information is an important prerequisite for the success of machine learning technology in various high-level air combat tasks.

[0003] Air combat situational information is characterized by its high dimensionality and dynamic changes. For any aircraft in air combat, situational information includes its own status information, friendly aircraft status information obtained through inter-aircraft data links, missile launch information obtained through air-to-missile data links, target information detected by formation radar obtained through inter-aircraft data links, and enemy missile information detected by missile approach warning systems. Except for its own status information, which has a fixed dimension, the dimensions of other information change dynamically according to the battlefield situation. For example, the dimension of target information detected by radar is proportional to the number of targets detected by radar.

[0004] Traditional feedforward neural networks, which can only input fixed-dimensional information, exhibit many shortcomings when dealing with high-dimensional, especially dynamically changing, air combat situational information. Previous studies have typically fixed the network input dimension to the maximum possible information dimension or a given input dimension. This approach increases the difficulty of network fitting, causes a certain degree of information loss, and makes it difficult for networks with fixed-dimensional input to scale across formations with varying numbers of aircraft.

[0005] Therefore, based on the above issues, how to represent and fuse dynamic air combat situation information is a problem that needs to be solved. Summary of the Invention

[0006] To address the above shortcomings, this invention provides a dynamic dimensional air combat situation information fusion method based on Set Transformer, comprising the following steps:

[0007] S1. Define and classify fixed-dimensional and dynamic-dimensional air combat situation information categories (i.e., unchanging information and changing information);

[0008] S101, The fixed-dimensional situational information in the air combat situational information category is the machine's status information V. egoThis includes the aircraft's serial number, location, speed, attitude, radar sensor status, and remaining missile count, and its dimension is denoted as D. ego ;

[0009] S102, Dynamic Dimension Situation Information in the Air Combat Situation Information category includes friendly aircraft status information obtained through inter-aircraft data links. fri Its single dimension is denoted as D. fri Information on missiles launched in formation obtained via the missile data link. fm Its single dimension is denoted as D. fm Target information detected by formation radar obtained through inter-aircraft data link V radar Its single dimension is denoted as D. radar Enemy missile information detected by the missile approach warning system V mw Its single dimension is denoted as D. mw ;

[0010] S2. The dynamic dimension information categories in the air combat situation information are respectively represented as set data, including:

[0011] The status information of friendly machines obtained through the inter-machine data link is represented as a set. in This is the i-th valid piece of information;

[0012] Missile information obtained via the machine-missile data link for formation-launched missiles is represented as a set.

[0013] Target information detected by formation radar and obtained via inter-machine data link is represented as a set.

[0014] Information about enemy missiles detected by the missile approach warning system is represented as a set.

[0015] S3. Define the Set Transformer neural network structure, and input the set data in S2 into the SetTransformer to obtain a fixed number and dimension of set features;

[0016] The fixed-dimensional output set features are:

[0017]

[0018] For the four types of dynamic information mentioned above, define the corresponding Set Transformer structure:

[0019] For the status information of friendly machines obtained through inter-machine data links, it is defined as SetTransformer friFor missile formation launch information obtained through the missile data link, it is defined as SetTransformer. fm For target information detected by formation radar obtained through inter-machine data link, it is defined as SetTransformer. radar For enemy missile information detected by the missile approach warning system, it is defined as SetTransformer. mw And set the number of elements n in its output set. output =1, output the dimension D of the set elements. output =16;

[0020] S4. Concatenate the fixed-dimensional information from the air combat situation information with the fixed number and dimensions of set features output by Set Transformer to obtain the final fixed-dimensional output feature. ST As input for subsequent machine learning techniques:

[0021] feature ST =V ego +SV fri +SV fm +SV radar +SV mw ;

[0022] The concatenated dimension is: D ego +4D output =D ego +64.

[0023] Furthermore, the friendly machine status information obtained through the inter-machine data link in step S2 is represented as a set. in This is the i-th valid piece of information;

[0024] Missile information obtained via the machine-missile data link for formation-launched missiles is represented as a set.

[0025] Target information detected by formation radar and obtained via inter-machine data link is represented as a set.

[0026] Information about enemy missiles detected by the missile approach warning system is represented as a set.

[0027] Furthermore, in step S3, SetTransformer fri ,enter The output after calculating the expression of the SetTransformer neural network structure is:

[0028] The parameter is θ fri ;

[0029] SetTransformer in step S3 fm ,enter The output after calculating the expression of the Set Transformer neural network structure is:

[0030] The parameter is θ fm ;

[0031] SetTransformer in step S3 radar ,enter The output after calculating the expression of the Set Transformer neural network structure is:

[0032] The parameter is θ radar ;

[0033] SetTransformer in step S3 mw ,

[0034] enter The output after calculating the expression of the Set Transformer neural network structure is:

[0035] The parameter is θ mw .

[0036] Compared with the prior art, the present invention has the following advantages:

[0037] (1) The air combat situation information fusion method proposed in this invention can avoid the defect that traditional feedforward neural networks can only process fixed-dimensional inputs and achieve lossless processing of dynamic-dimensional inputs;

[0038] (2) The air combat situation information fusion method proposed in this invention provides a unified neural network information processing interface for any number of formations to operate, and can realize the expansion of upper-level machine learning algorithms between different numbers of formations to operate;

[0039] (3) Compared with traditional feedforward neural networks, the Set Transformer neural network used in this invention has invariance to different arrangements within the dynamic dimension situation information of air combat, which can accelerate the efficiency of situation information fusion to a certain extent. Attached Figure Description

[0040] Figure 1 This is the situation information fusion process of the method proposed in this invention;

[0041] Figure 2The curve showing the change in win rate over training time steps in a 4v4 combat scenario.

[0042] Figure 3 The curves showing the change in win rate over training time steps in multiple combat scenarios. Detailed Implementation

[0043] To facilitate understanding of the present invention, the apparatus of the present invention will now be described more fully with reference to the accompanying drawings. Embodiments of the apparatus are shown in the drawings. However, the apparatus can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the present invention more thorough and complete.

[0044] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "configuration" should be interpreted broadly. For example, they can refer to a fixed connection or configuration, a detachable connection or configuration, or an integral connection or configuration. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0045] Example

[0046] like Figure 1 As shown, this embodiment provides a dynamic dimensional air combat situation information fusion method based on Set Transformer, which specifically includes the following steps:

[0047] S1. Define and classify fixed-dimensional and dynamic-dimensional air combat situation information categories (i.e., unchanging information and changing information);

[0048] S101, The fixed-dimensional situational information in the air combat situational information category is the machine's status information V. ego This includes the aircraft's serial number, location, speed, attitude, radar sensor status, and remaining missile count, and its dimension is denoted as D. ego ;

[0049] S102. The dynamic dimension of situational information in the air combat situational information category includes:

[0050] (1) The status information of the friendly machine obtained through the inter-machine data link V fri Specifically, this includes its serial number, status information relative to the local machine, number of radar sensors, and number of remaining missiles, with each individual information dimension denoted as D. fri ;

[0051] (2) Missile information V obtained through the machine-missile data link for formation launch fm Specifically, this includes its state information relative to the local machine, the attack zone of the local machine against the target, etc., and its individual information dimension is denoted as D. fm ;

[0052] (3) Target information V detected by formation radar obtained through inter-machine data link radar Specifically, this includes its state information relative to the local machine, the attack zone of the local machine against the target, etc., and its individual information dimension is denoted as D. radar ;

[0053] It should be noted that if the formation radar does not detect a target, but the electronic support measures detect the target, the target information detected by the electronic support measures will be incorporated into the radar information.

[0054] (4) Enemy missile information detected by the missile approach warning system V mw Specifically, this includes alarm times relative to the local machine's status information, with a single information dimension denoted as D. mw ;

[0055] The overall dimension of the above information is the number of valid information items multiplied by the dimension of a single information item. The number of valid information items changes dynamically with the battlefield situation, which makes the overall information dimension dynamic.

[0056] S2. Represent the dynamic dimension information categories in the air combat situation information as set data;

[0057] The dimension of dynamic dimensional information is the number of valid information items multiplied by the dimension of a single information item. It can be expressed as a set of data where each element is a single piece of information and the number of elements is the number of valid information items.

[0058] (1) The status information of friendly machines obtained through the inter-machine data link is represented as a set. in This is the i-th valid piece of information;

[0059] (2) The missile information obtained through the machine-missile data link for formation launch is represented as a set.

[0060] (3) Target information detected by formation radar obtained through inter-machine data link is represented as a set

[0061] (4) Enemy missile information detected by the missile approach warning system is represented as a set.

[0062] In traditional situation information fusion methods based on feedforward neural networks (FNNs), since the network can only accept inputs of fixed dimensions, dynamic situation information (such as friendly aircraft information, missile information, target information, etc.) is usually processed in the following way:

[0063] First, the maximum possible quantity of each type of dynamic information (such as the number of friendly aircraft, the number of missiles, etc.) is set to a fixed upper limit. Then, all the information from the dynamic dimensions is concatenated into a fixed-length vector, for example:

[0064] spliced ​​friend machine information vector TV fri , dimension

[0065] The spliced ​​missile information vector TV of the formation launch fm , dimension

[0066] The spliced ​​target information vector TV radar , dimension

[0067] The stitched enemy missile information vector TV mw , dimension

[0068] When the actual amount of dynamic information is less than the set maximum amount In this case, missing data will be filled with zeros. While this method solves the problem of fixed structure for dynamic information input, it also brings the following issues:

[0069] When the actual amount of information is much smaller At that time, a large number of padding values ​​(zeros) will waste computing resources;

[0070] Maximum possible number The configuration usually depends on the size of the enemy's fighter formation. When the formation size changes (such as the number of enemy aircraft increases or decreases), the network structure needs to be readjusted, making it difficult to flexibly adapt to different battlefield scenarios.

[0071] Although the padding values ​​do not carry any useful information, they are still involved in the calculations, increasing unnecessary computational overhead.

[0072] S3. Define the Set Transformer neural network structure, and input the set data in S2 into the SetTransformer to obtain a fixed number and dimension of set features;

[0073] The Set Transformer neural network structure can map a set of data of any number and dimension to a set of outputs of a fixed number and dimension. The mathematical expression is:

[0074]

[0075] SV represents the information V after being processed by SetTransformer, and the same applies below.

[0076] When using Set Transformer, the parameters that need to be specified include the dimension D of the input collection elements. input Output the number of elements in the set, n. output Output the dimension D of the set elements. output And the neural network-related parameter θ.

[0077] Define a Set Transformer neural network structure for each type of dynamic information:

[0078] (1) For the status information of the friendly machine obtained through the inter-machine data link, it is defined as SetTransformer fri ,enter The output after calculating the expression of the Set Transformer neural network structure is:

[0079] The parameter is θ fri ;

[0080] The specific expression is:

[0081]

[0082] (2) For missile information launched in formation obtained through the missile data link, it is defined as SetTransformer fm ,enter The output after calculating the expression of the Set Transformer neural network structure is:

[0083] The parameter is θ fm ;

[0084] The specific expression is:

[0085]

[0086] (3) For target information detected by formation radar obtained through inter-machine data link, it is defined as SetTransformer radar ,enter The output after calculating the expression of the Set Transformer neural network structure is:

[0087] The parameter is θ radar ;

[0088] The specific expression is:

[0089]

[0090] (4) For enemy missile information detected by the missile approach warning system, it is defined as SetTransformer mw ,

[0091] enter The output after calculating the expression of the Set Transformer neural network structure is:

[0092] The parameter is θ mw ;

[0093] The specific expression is:

[0094]

[0095] The Set Transformer output for each type of information is a set of data with a fixed number and dimensions. Since the output is not sensitive to different arrangements of the input sets, the output remains unchanged (i.e., the input order does not affect the output result).

[0096] For SetTransformer fri SetTransformer fm SetTransformer radar and SetTransformer mw Set the number of elements n in its output set. output =1. Output the dimension D of the set elements. output =16.

[0097] It should be noted that the Set Transformer neural network structure consists of two parts: an encoder and a decoder. The workflow is as follows:

[0098] The encoder performs multi-head self-attention operations on the elements of the input set, and the output set has the same size as the input set, with dimension D. hidden ;

[0099] The decoder uses n elements output Dimension D hidden The learnable parameter set is taken as input, and a multi-head attention operation is performed between the decoder input set and the encoder output set to aggregate features; several encoder layers can be stacked after the decoder; finally, a fully connected layer merges the D... hidden Mapped to D output ;

[0100] In this embodiment, n is set. output =1, meaning it outputs a global feature vector, so there is only a fully connected layer after the decoder. Additionally, D is set... hidden=32, meaning the hidden layer dimension is 32, and the encoder and decoder each have 2 heads.

[0101] S4. The fixed-dimensional information in the air combat situation information is concatenated with the set features of a fixed number and dimensions output by Set Transformer, and used as the input for subsequent machine learning techniques, as follows:

[0102] The output dimension of the Set Transformer neural network structure is always D. output =16 vector SV fri SV fm SV radar SV mw .

[0103] Further, the four vectors SV output by Set Transformer are... fri SV fm SV radar SV mw With local status information V ego Concatenation yields a fixed-dimensional output feature. ST :

[0104] feature ST =V ego +SV fri +SV fm +SV radar +SV mw ;

[0105] The dimensions after splicing are:

[0106] D ego +4D output =D ego +64, this splicing feature ST As input to top-level machine learning algorithms.

[0107] Traditional situation information fusion methods based on feedforward neural networks do not use Set Transformer, but instead directly input the dynamic dimension feature vector TV. fri TV fm TV radar TV mw With local status information V ego splicing:

[0108] feature FNN =V ego +TV fri +TV fm +TV radar +TV mw,

[0109] The dimensions after splicing are As input to top-level machine learning algorithms.

[0110] The following embodiment uses a beyond-visual-range formation air combat strategy generation task based on multi-agent reinforcement learning to verify the effectiveness of the dynamic dimension air combat situation information fusion method proposed in this invention.

[0111] The operational scenario for this mission involves an encounter between friendly and enemy aircraft formations in a certain airspace, with the objective of eliminating all enemy personnel. Both formations share identical combat platforms, weapons, and avionics systems in terms of quantity and performance. The combat platforms are F-16 aircraft, each equipped with four semi-active radar-guided medium-to-long-range air-to-air missiles, an onboard fire control radar, electronic support measures, and an onboard / missile data link system. This embodiment uses a typical enemy strategy as the adversary and employs the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm to train the friendly formation's operational strategy, including maneuvering strategies, missile launch strategies, and radar on / off strategies.

[0112] To compare the effects of this invention and the traditional situation information fusion method based on feedforward neural networks, this embodiment uses both methods to build the Actor network and Critic network of the MAPPO algorithm.

[0113] Specifically, for the situation information fusion method based on Set Transformer in this invention, both the Actor network and the Critic network include a Set Transformer layer in S3 (used to process dynamic dimensional situation information to obtain features). ST ) and a fully connected layer with 128 neurons (used to process the feature) ST Further mapping to action output or value function); For traditional situation information fusion methods based on feedforward neural networks, both Actor networks and Critic networks include two fully connected layers with 128 neurons each (directly mapping the feature output to the action output or value function); FNN Mapped to action output or value function).

[0114] First, this embodiment takes a 4v4 combat scenario as an example, and uses two methods—Set Transformer and traditional feedforward neural network—to build and train the MAPPO reinforcement learning network. Figure 2 The curve showing the change of our win rate over the training time step demonstrates that, compared with the traditional feedforward neural network processing method, the present invention has a larger win rate convergence value, a faster convergence speed, and accelerates the efficiency of situation fusion.

[0115] Furthermore, to verify the advantage of the proposed method in enabling the upper-level algorithm to be extended across different numbers of formations in combat, this embodiment conducted random sampling training in multiple combat scenarios including 2v2, 3v3, 4v4, and 5v5. The curves showing the change in our win rate over training time steps are shown below. Figure 3 It can be seen that the method proposed in this invention can achieve a win rate of over 80% in four training scenarios: 2v2, 3v3, 4v4, and 5v5, using the same neural network.

[0116] It should be noted that the structure described in this invention can be implemented in many different forms and is not limited to the embodiments described. Any equivalent transformations made by those skilled in the art based on the description and drawings of this invention, or direct or indirect applications in other related technical fields, such as the loading and unloading of other items, are included within the protection scope of this invention.

Claims

1. A dynamic dimensional air combat situation information fusion method based on Set Transformer, characterized in that, Specifically, the following steps are included: S1. Define the categories of air combat situation information, including fixed-dimensional information and dynamic-dimensional information. Fixed-dimensional situation information is the local machine's status information. This includes the aircraft's serial number, location, speed, attitude, radar sensor status, and remaining missile count, denoted by the dimension denoted as . Dynamic situational information includes the status information of friendly machines obtained through inter-machine data links. Its single dimension is denoted as Information on missiles launched in formation obtained through the missile data link Its single dimension is denoted as Target information detected by formation radar obtained through inter-aircraft data link Its single dimension is denoted as Enemy missile information detected by the missile approach warning system Its single dimension is denoted as ; S2. The dynamic dimension information categories in air combat situational information are respectively represented as set data, which includes: The status information of friendly machines obtained through the inter-machine data link is represented as a set. ,in For the first One piece of valid information; Missile information obtained via the machine-missile data link for formation-launched missiles is represented as a set. ; Target information detected by formation radar and obtained via inter-machine data link is represented as a set. ; Information about enemy missiles detected by the missile approach warning system is represented as a set. ; S3. Define the Set Transformer neural network structure, and input the set data from step S2 into the Set Transformer to obtain a fixed number and dimension of set features; S4. Concatenate the fixed-dimensional information from the air combat situation information with the fixed number and dimensions of features output by the Set Transformer, and use this as input for subsequent machine learning techniques. The fixed-dimensional output is: ； The concatenated dimensions are: ; in , , , The output dimension of the Set Transformer neural network structure is 1. The four vectors.

2. The dynamic dimensional air combat situation information fusion method based on Set Transformer as described in claim 1, characterized in that: The fixed-dimensional output set features in step S3 are: ,in For parameters related to the neural network; For the four types of dynamic information mentioned above, corresponding Set Transformer structures are defined respectively: For the status information of friendly machines obtained through inter-machine data links, it is defined as follows: Information on missiles launched in formation obtained through the missile data link is defined as follows: For target information detected by formation radar obtained through inter-aircraft data links, it is defined as... Information about enemy missiles detected by a missile approach warning system is defined as follows: And set the number of elements in its output set. Output the dimension of the set elements .

Images