Dynamic dimension air combat situation information fusion method based on Set Transform

Abstract

The invention discloses a dynamic dimension air combat situation information fusion method based on Set Transform, and belongs to the field of air combat situation information fusion, and the method comprises the following steps: respectively defining and dividing a fixed dimension air combat situation information category and a dynamic dimension air combat situation information category, respectively expressing the dynamic dimension information category in the air combat situation information as set data, defining a Set Transform neural network structure, and carrying out the fusion of the dynamic dimension air combat situation information. Respectively inputting the set data into a Set Transform for processing to obtain set features with a fixed number and dimensions, and splicing fixed dimension information in the air combat situation information with the set features with the fixed number and dimensions output by the Set Transform to serve as input of a subsequent machine learning technology; according to the air combat situation information fusion method provided by the invention, the defect that a traditional feedforward neural network can only process fixed dimension input is avoided, lossless and efficient processing of dynamic dimension input is realized, and an upper-layer algorithm can be conveniently expanded among any number of formation combats.

Description

Technical Field

[0001] The present invention belongs to the field of air combat situation information fusion. Specifically, it relates to a method for fusing dynamic dimension air combat situation information based on Set Transformer. Background Art

[0002] In recent years, with the continuous development of machine learning technology, more and more research has started to process air combat related tasks based on machine learning technology, such as providing decision support and situation assessment. Modern machine learning technologies mostly use deep neural networks to fuse and extract features from high-dimensional inputs. For example, reinforcement learning technology uses a deep neural network to directly learn and establish a mapping from air combat situation to fighter behavior. Therefore, the reasonable representation and fusion of air combat situation information is an important prerequisite for the success of machine learning technology in different top-level air combat tasks.

[0003] Air combat situation information has the characteristics of high dimension and dynamic change. For any fighter in air combat, the situation information includes the state information of its own aircraft, the state information of friendly aircraft obtained through the inter-aircraft data link, the missile information launched by the formation obtained through the aircraft-missile data link, the target information detected by the formation radar obtained through the inter-aircraft data link, and the enemy missile information detected by the missile approach warning system. Among them, except for the state information of its own aircraft which is of a fixed dimension, the dimensions of other information will change dynamically according to the battlefield situation. For example, the dimension of the target information detected by the radar is proportional to the number of targets detected by the radar.

[0004] For high-dimensional, especially dynamically changing air combat situation information, traditional feedforward neural networks that can only input information with a fixed dimension show many deficiencies. Previous studies usually fixed the dimension of the network input information to the maximum possible information dimension or a given input dimension. This way will increase the difficulty of network fitting, cause a certain degree of information loss, and it is difficult for a network with fixed-dimension information input to expand between formations with different numbers of fighters.

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

[0006] In view of the above defects, the present invention provides a method for fusing dynamic dimension air combat situation information based on Set Transformer, including the following steps:

[0007] S1. Define and divide the categories of fixed-dimension and dynamic-dimension air combat situation information (i.e., invariant information and variable information) respectively;

[0008] S101. The fixed-dimension situation information in the air combat situation information category is the state information V of its own aircraft ego, including the local number, location, speed, attitude, radar sensor status, remaining missile quantity, etc., and its dimension is denoted as D ego ;

[0009] S102. The dynamic dimension situation information in the air combat situation information category includes the friendly aircraft status information V obtained through the inter-aircraft data link fri , and its single dimension is denoted as D fri , the missile information V launched by the formation obtained through the aircraft-missile data link fm , and its single dimension is denoted as D fm , the target information V detected by the formation radar obtained through the inter-aircraft data link radar , and its single dimension is denoted as D radar , the enemy missile information V detected by the missile approach warning system mw , and its single dimension is denoted as D mw ;

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

[0011] The friendly aircraft status information obtained through the inter-aircraft data link is expressed as a set where is the i-th valid information;

[0012] The missile information launched by the formation obtained through the aircraft-missile data link is expressed as a set

[0013] The target information detected by the formation radar obtained through the inter-aircraft data link is expressed as a set

[0014] The enemy missile information detected by the missile approach warning system is expressed as a set

[0015] S3. Define the Set Transformer neural network structure, and input the set data in S2 into the SetTransformer respectively to obtain set features with fixed numbers and dimensions;

[0016] The output set features with fixed dimensions are:

[0017]

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

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

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

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

[0022] Among them, the concatenated dimension is: D ego + 4D output = D ego + 64.

[0023] Furthermore, the friendly aircraft status information obtained through the inter-aircraft data link in step S2 is expressed as a set where is the i-th valid information;

[0024] The missile information of the formation launched obtained through the aircraft-missile data link is expressed as a set

[0025] The target information detected by the formation radar obtained through the inter-aircraft data link is expressed as a set

[0026] The enemy missile information detected by the missile approach warning system is expressed as a set

[0027] Furthermore, the SetTransformer in step S3 fri , the input is calculated by the expression of the SetTransformer neural network structure and then output:

[0028] The parameter is θ fri ;

[0029] SetTransformer in step S3 fm , input After being calculated by the expression of the Set Transformer neural network structure, the output is:

[0030] The parameter is θ fm ;

[0031] SetTransformer in step S3 radar , input After being calculated by the expression of the Set Transformer neural network structure, the output is:

[0032] The parameter is θ radar ;

[0033] SetTransformer in step S3 mw ,

[0034] input After being calculated by the expression of the Set Transformer neural network structure, the output is:

[0035] The parameter is θ mw .

[0036] The present invention has the following beneficial effects compared with the prior art:

[0037] (1) The air combat situation information fusion method proposed by the present invention can avoid the defect that the traditional feedforward neural network can only process fixed-dimension inputs, and realizes lossless processing of dynamic-dimension inputs;

[0038] (2) The air combat situation information fusion method proposed by the present invention provides a unified neural network information processing interface for any number of formation operations, and can realize the extension of upper-layer machine learning algorithms among different numbers of formation operations;

[0039] (3) Compared with the traditional feedforward neural network, the Set Transformer neural network used in the present invention has invariance to different permutations inside the air combat dynamic-dimension situation information, and can accelerate the situation information fusion efficiency to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is the situation information fusion process of the method proposed by the present invention;

[0041] Figure 2The change curve of the win rate in the 4v4 combat scenario with the training time steps;

[0042] Figure 3 The change curve of the win rate in multiple combat scenarios with the training time steps. Specific implementation manners

[0043] To facilitate the understanding of the present invention, the device of the present invention will be described more comprehensively below with reference to the relevant drawings. Embodiments of the device are shown in the drawings. However, the device can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0044] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "installed", "connected", and "set" should be understood in a broad sense. For example, it can be fixedly connected and set, or detachably connected and set, or integrally connected and set. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0045] Embodiment

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

[0047] S1. Define and divide the fixed dimension and dynamic dimension air combat situation information categories (i.e., invariant information and variable information) respectively;

[0048] S101. The fixed dimension situation information in the air combat situation information category is the own aircraft state information V ego , including the own aircraft number, position, speed, attitude, radar sensor status, remaining missile quantity, etc., and its dimension is denoted as D ego ;

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

[0050] (1) The friendly aircraft state information V fri obtained through the inter-aircraft data link, specifically including its number, status information relative to the own aircraft, number of radar sensors, remaining missile quantity, etc., and its single information dimension is denoted as D fri ;

[0051] (2) The missile information V fm launched by the formation obtained through the aircraft-missile data link, specifically including its status information relative to the own aircraft, the attack zone of the own aircraft on this target, etc., and its single information dimension is denoted as D fm ;

[0052] (3) Target information V detected by the formation radar obtained through the inter-aircraft data link radar , specifically including its status information relative to the own aircraft, the attack area of the own aircraft on this target, etc., and the single information dimension is denoted as D radar ;

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

[0054] (4) Enemy missile information V detected by the missile approach warning system mw , specifically including the warning time of its status information relative to the own aircraft, etc., and the single information dimension is denoted as D mw ;

[0055] The overall dimension of the above information is the number of effective information multiplied by the single information dimension, and the number of effective information will change dynamically with the change of the battlefield situation, which results in the dynamic change of the overall information dimension.

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

[0057] The dimension of the dynamic dimension information is the number of effective information multiplied by the single information dimension, and it can be expressed as set data with single information as elements and the number of elements being the number of effective information.

[0058] (1) The status information of friendly aircraft obtained through the inter-aircraft data link is expressed as a set where is the i-th effective information;

[0059] (2) The missile information launched by the formation obtained through the aircraft-missile data link is expressed as a set

[0060] (3) The target information detected by the formation radar obtained through the inter-aircraft data link is expressed as a set

[0061] (4) The enemy missile information detected by the missile approach warning system is expressed as a set

[0062] In the traditional situation information fusion method based on the feedforward neural network (FNN), since the network can only accept inputs with fixed dimensions, for dynamic dimension situation information (such as friendly aircraft information, missile information, target information, etc.), the following processing methods are usually adopted:

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

[0064] The concatenated friendly aircraft information vector TV fri , with a dimension of

[0065] The concatenated missile information vector TV launched by the formation fm , with a dimension of

[0066] The concatenated detected target information vector TV radar , with a dimension of

[0067] The concatenated enemy missile information vector TV mw , with a dimension of

[0068] When the actual number of dynamic information is less than the set maximum number , the missing data will be filled with zeros. Although this method solves the problem of inputting dynamic information into a fixed structure, it also brings the following problems:

[0069] When the actual number of information is much less than , a large number of filled values (zeros) will cause waste of computing resources;

[0070] The maximum possible number is usually set depending on the scale of the opponent's fighter formation. When the formation scale changes (such as the increase or decrease in the number of enemy aircraft), the network structure needs to be readjusted, and it is difficult to flexibly adapt to different battlefield scenarios;

[0071] Although the filled values do not carry any useful information, they will still be involved in the calculation, increasing unnecessary computational overhead.

[0072] S3. Define the Set Transformer neural network structure, and input the set data in S2 into the Set Transformer respectively to obtain set features with fixed numbers and dimensions;

[0073] The Set Transformer neural network structure can map set data with arbitrary numbers and dimensions to set outputs with fixed numbers and dimensions. The mathematical expression is:

[0074]

[0075] Among them, SV represents the information V processed by the Set Transformer, and the same applies hereinafter.

[0076] The parameters that need to be specified when using the Set Transformer include the dimension D of the elements in the input set input , the number n of elements in the output set output , the dimension D of the elements in the output set output and the neural network related parameter θ.

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

[0078] (1) For the friendly aircraft status information obtained through the inter-aircraft data link, it is defined as SetTransformer fri , and the input is calculated by the expression of the Set Transformer neural network structure and then output:

[0079] The parameter is θ fri ;

[0080] The specific expression is:

[0081]

[0082] (2) For the missile information launched by the formation obtained through the aircraft-missile data link, it is defined as SetTransformer fm , and the input is calculated by the expression of the Set Transformer neural network structure and then output:

[0083] The parameter is θ fm ;

[0084] The specific expression is:

[0085]

[0086] (3) For the target information detected by the formation radar obtained through the inter-aircraft data link, it is defined as SetTransformer radar , and the input is calculated by the expression of the Set Transformer neural network structure and then output:

[0087] The parameter is θ radar ;

[0088] The specific expression is:

[0089]

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

[0091] Input to the Set Transformer neural network structure and output after expression calculation:

[0092] The parameter is θ mw ;

[0093] The specific expression is:

[0094]

[0095] The Set Transformer output of each of the above types of information is a set of data with a fixed number and dimension, and since the output is insensitive to different input set permutations, 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 output = 1 of the output set, and the dimension D output = 16.

[0097] It should be noted that the Set Transformer neural network structure includes an encoder and a decoder, and the working process is as follows:

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

[0099] The decoder takes a learnable set of parameters with the number of elements n output and dimension D hidden as the input, and performs multi-head attention operations between the decoder input set and the encoder output set to aggregate features; several layers of encoders can be optionally stacked after the decoder; finally, the fully connected layer maps D hidden to D output ;

[0100] In this embodiment, n output = 1 is set, that is, a global feature vector is output, so there is only a fully connected layer after the decoder, and D hidden= 32, that is, the hidden layer dimension is 32, and the number of heads of both the encoder and the decoder is 2.

[0101] S4. Concatenate the fixed-dimension information in the air combat situation information with the set features of fixed number and dimension output by the Set Transformer as the input for subsequent machine learning techniques, specifically as follows:

[0102] The output dimensions of the Set Transformer neural network structure are all D output = 16 vector SV fri 、SV fm 、SV radar 、SV mw .

[0103] Further, the four vectors SV output by the Set Transformer fri 、SV fm 、SV radar 、SV mw are concatenated with the local state information V ego to obtain the output feature of fixed dimension ST :

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

[0105] The dimension after concatenation is:

[0106] D ego + 4D output = D ego + 64, and this concatenated feature feature ST is used as the input for the top-level machine learning algorithm.

[0107] For the traditional situation information fusion method based on the feedforward neural network, instead of using the Set Transformer, the input feature vectors TV of dynamic dimension fri 、TV fm 、TV radar 、TV mw are directly concatenated with the local state information V ego :

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

[0109] The dimension after splicing is Used as the input of the top-level machine learning algorithm.

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

[0111] The combat scenario of this task is that the fighter formations of both sides encounter in a certain airspace, and the combat goal is to annihilate all the vital forces of the enemy. The number and performance of the combat platform and all weapons and avionics systems of both formations are the same. The combat platform uses F-16 aircraft, and each aircraft is equipped with semi-active radar-guided medium and long-range air-to-air missiles (4 pieces), airborne fire control radar, electronic support measures, and airborne / aircraft-missile data link systems. This embodiment uses a typical enemy strategy as the opponent and trains the combat strategy of our formation using the multi-agent reinforcement learning MAPPO (Multi-Agent Proximal Policy Optimization) algorithm, including maneuvering strategies, missile launch strategies, and radar switch strategies.

[0112] In order to compare the effects of the present invention and the traditional situation information fusion method based on the feedforward neural network, this embodiment uses these two methods to build the Actor network and Critic network of the MAPPO algorithm respectively.

[0113] Specifically, for the situation information fusion method based on Set Transformer of the present invention, both the Actor network and the Critic network include a Set Transformer layer in S3 (used to process dynamic dimension situation information to obtain feature ST ) and a fully connected layer with 128 neurons (used to further map feature ST to action output or value function); for the traditional situation information fusion method based on the feedforward neural network, both the Actor network and the Critic network include two fully connected layers with 128 neurons (directly map feature FNN to action output or value function).

[0114] First, taking the 4v4 combat scenario as an example, this embodiment uses both the Set Transformer and the traditional feedforward neural network method to build the MAPPO reinforcement learning network to build the reinforcement learning neural network and train it. Figure 2 The change curve of the winning rate of our side with the training time step is shown. It can be seen that the convergence value of the winning rate of the present invention is larger and the convergence speed is faster than that of the traditional feedforward neural network processing method, which improves the situation fusion efficiency.

[0115] Furthermore, in order to verify the advantage that the method proposed by the present invention can achieve the expansion of the upper-layer algorithm among different numbers of formation operations, in this embodiment, random extraction training is carried out in multiple combat scenarios of 2v2, 3v3, 4v4, and 5v5. The change curve of the winning rate of our side with the training time step is shown in Figure 3 , and it can be seen that the method proposed by the present invention can achieve a winning rate of more than 80% in four training scenarios of 2v2, 3v3, 4v4, and 5v5 with the same set of neural networks.

[0116] It should be noted that the structure described in the present invention can be implemented in many different forms and is not limited to the embodiments. Any equivalent transformation made by those of ordinary skill in the art using the description and drawings of the present invention, or directly or indirectly applied in other related technical fields, such as the loading and unloading of other items, is included in the protection scope of the present invention.

Claims

1. A dynamic dimension air combat situation information fusion method based on Set Transformer, characterized in that: The specific steps include: S1. Define the categories of air combat situation information, including fixed dimension information and dynamic dimension information; S2, expressing the dynamic dimension information categories in the air combat situation information as set data respectively; S3, define the Set Transformer neural network structure, input the set data of step S2 into the Set Transformer to obtain the set features of fixed number and dimension; S4. Concatenate the fixed-dimensional information in the air combat situation information with the fixed number and dimension of set features output by Set Transformer as the input for subsequent machine learning techniques.

2. The method for fusion of dynamic dimensional air combat situation information based on Set Transformer as claimed in claim 1, characterized in that: In step S1, the fixed dimension situation information in the air combat situation information category is the aircraft state information V ego , including the aircraft number, position, speed, attitude, radar sensor status and number of remaining missiles, etc., and its dimension is recorded as D ego The dynamic dimension situation information in the air combat situation information category includes the friendly aircraft status information V obtained through the inter-aircraft data link fri , whose single dimension is denoted as D fri , missile information launched by the formation obtained through the aircraft-missile data link V fm , whose single dimension is denoted as D fm , target information detected by the formation radar obtained through the inter-machine data link V radar , whose single dimension is denoted as D radar , Enemy missile information detected by the missile approach warning system V mw , whose single dimension is denoted as D mw .

3. The method for fusion of dynamic dimension air combat situation information based on Set Transformer as claimed in claim 1, characterized in that: The aggregate data in step S2 includes: The friendly aircraft status information obtained through the inter-aircraft data link is expressed as a set in is the i-th valid information; The missile information launched by the formation obtained through the aircraft missile data link is expressed as a collection The target information detected by the formation radar obtained through the inter-machine data link is expressed as a set The enemy missile information detected by the missile approach warning system is expressed as a set 4. The method for fusion of dynamic dimension air combat situation information based on Set Transformer as claimed in claim 3, characterized in that: The output set features of the fixed dimension in step S3 are: For the above four types of dynamic information, the corresponding Set Transformer structures are defined respectively: For the friendly aircraft status information obtained through the inter-machine data link, it is defined as SetTransformer fri , For the missile information launched by the formation obtained through the aircraft-missile data link, it is defined as SetTransformer fm , For the target information detected by the formation radar obtained through the inter-machine data link, it is defined as SetTransformer radar , For the enemy missile information detected by the missile approach warning system, it is defined as SetTransformer mw , and set the number of elements in its output set to n output =1, output set element dimension D output =16.

5. The method for fusion of dynamic dimension air combat situation information based on Set Transformer as claimed in claim 1, characterized in that: The fixed-dimensional output in step S4 is: feature ST =V ego +NE fri +NE fm +NE radar +NE mw ; The dimension after splicing is: D ego +4D output =D ego +64; Among them SV fri SV fm SV radar SV mw The output dimension of the Set Transformer neural network structure is D output =4 vectors of 16.

Images