LSTM-based target maneuver prediction method, electronic device and storage medium

Through the LSTM-based maneuver prediction method, the spatial situation characteristics during the drone battle are extracted and the LSTM network model is established, which solves the problem of inaccurate prediction in the existing technology and achieves accurate prediction of target maneuver actions.

CN115480582BActive Publication Date: 2024-12-03XIDIAN UNIV +1
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202210986452.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-17
Publication Date
2024-12-03
Estimated Expiration
2042-08-17

AI Technical Summary

Technical Problem

The prior art fails to fully consider the spatial situation characteristics and timing when predicting the maneuvering operation of the target drone, resulting in inaccurate prediction.

Method used

Using the LSTM-based maneuver prediction method, the target maneuver selection data set is constructed in the drone battle process, the spatial situation characteristics are extracted, and the LSTM network model with two hidden layers is established, and the cross entropy loss function is trained to predict the target. maneuver control parameters of the machine.

Benefits of technology

It achieves a more accurate prediction of target maneuvering actions, fully explores the maneuver selection rules in air combat confrontation data, and achieves better prediction results compared with other network models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115480582B_ABST
    Figure CN115480582B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for predicting the maneuverability of a target based on LSTM, an electronic device and a storage medium, comprising the following steps: step 1: constructing a target maneuverability selection data set in a UAV battle process; step 2: extracting spatial situation characteristics of the target maneuverability confrontation process; step 3: segmenting and preprocessing the data set; step 4: establishing a maneuverability prediction network model for a target based on LSTM; step 5: using a training set to train the maneuverability prediction network model for a target based on LSTM, and using a test set to detect the accuracy. The present invention can achieve good results in predicting a single maneuvering control parameter or predicting all maneuvering control parameters correctly at the same time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of aircraft flight control and artificial intelligence technology, and in particular to a maneuver prediction method, electronic equipment and storage medium of a target based on LSTM. Background Art

[0002] In recent years, with the rapid development of drone-related technologies, drones are increasingly used in the military field, and the problem of predicting the next stage of maneuvering actions of target drones has gradually attracted people's attention. In the process of air combat confrontation, the drones of both sides, as intelligent agents, strategically select the next stage of maneuvering actions, so that the drones of the own side are in an advantageous state in space occupation. How to predict the maneuvers that the target drone will take in the next stage in advance is the key issue of the maneuvering action selection strategy of the own side.

[0003] At present, the method of predicting the maneuvering action of the target UAV regards the maneuvering action selection of the target aircraft as the weighted sum of several fixed maneuvers, and the weight of each maneuver is determined by expert knowledge. However, the maneuvering action selection of the UAV is a continuous and dynamic process of optimizing the spatial situation. The prediction method of the fixed weighted sum of maneuvers does not consider the spatial situation characteristics of the target maneuvering action selection, ignores the dynamic and temporal characteristics of the target maneuvering action, and does not consider the target aircraft's perception of our UAV's maneuvering. Therefore, it is difficult to accurately describe the law of the target maneuvering action selection, resulting in inaccurate prediction of the target aircraft's future maneuvering action. Summary of the invention

[0004] In order to overcome the shortcomings of the above technologies, the purpose of the present invention is to provide a maneuvering prediction method, electronic device and storage medium based on LSTM targets, which can achieve good results whether it is the prediction of a single maneuvering control parameter or the simultaneous prediction of all maneuvering control parameters.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A maneuver prediction method of a target based on LSTM, comprising the following steps;

[0007] Step 1: Construct a target maneuver selection dataset S of the UAV combat process, where the target refers to the target UAV that is fighting against our UAV;

[0008] Step 2: Extract the spatial situation characteristics of the target maneuver confrontation process;

[0009] Step 3: Segment and preprocess the target maneuver selection dataset S in step 1;

[0010] Step 4: Establish a maneuver prediction network model based on LSTM target;

[0011] Step 5: Using the training set Train the LSTM-based target maneuver prediction network model and use the test set Perform accuracy test.

[0012] In step 1, a one-on-one air combat simulation system is used to set different initial states for the two sides to conduct multiple combat experiments. The state information of the drones of the two sides is recorded every 20ms in each round of combat. t and maneuver control information c t , forming air combat confrontation data, and each battle process ends when one side wins or the battle lasts for more than 2 minutes.

[0013] 280 sets of air combat confrontation data are selected as the target maneuver prediction data set S. At any time t, the state information s of the drones of the two sides in the battle t and maneuver control information c t Contains information in the following multiple dimensions, as shown in the formula:

[0014] s t =[X t ,Y t ,Z t ,EX t ,EY t ,EZ t ,AX t ,AY t ,AZ t ,V t ]

[0015] c t =[a t ,p t ,y t ]

[0016] Where, X t ,Y t ,Z t They represent the three-dimensional spatial position coordinates of the drone at time t, EX t ,EY t ,EZ t They represent the Euler angles of the drone’s attitude at time t, (AX t ,AY t ,AZ t ) constitutes the nose heading vector of the drone at time t, V t Indicates the speed of the drone at time t, s t Contains status information of 10 dimensions, a t ,p t ,y tThey represent the acceleration information, pitch angle information and yaw angle information for controlling the speed and direction of the drone, respectively. t Contains 3-dimensional maneuvering control information. The state information of our drone at any time t is used Indicates that the maneuver control information is used Indicates; the status information of the target UAV is indicated by Indicates that the maneuver control information is used Represents; Data set S = [s u ,c u ,s e ,c e ].

[0017] In the process of close-range air combat confrontation, the key to the victory of the UAV in step 2 is that when the target aircraft is within the attack range of the own UAV, the angle of the own UAV is in an advantageous state. At this time, the azimuth angle p of the target aircraft and the entry angle q of the target aircraft are close to 0. When predicting the maneuver of the target aircraft, it is necessary to consider the distance d between the target aircraft and our aircraft and the angle of the target aircraft T. a In this case, the distance feature d and angle situation feature T of the target maneuver are extracted from the target maneuver prediction data set S obtained in step 1. a :

[0018]

[0019]

[0020] Where, X u ,Y u ,Z u They represent the three-dimensional spatial coordinates of our drone, X e ,Y e ,Z e They respectively represent the three-dimensional spatial position coordinates of the target UAV.

[0021] The extracted distance features and angle situation features are incorporated into the target maneuver prediction dataset S:

[0022] S=[s u ,c u ,s e ,c e ,d,T a ]

[0023] In the formula, s u ,c u They represent the state information and maneuver control information of our UAV in the target maneuver prediction dataset S, s e ,c e They respectively represent the state information and maneuver control information of the target UAV in the target maneuver prediction dataset S.

[0024] The step 3 comprises:

[0025] 3a, divide the target maneuver prediction dataset S in step 1 into a training set S according to the ratio of training set to test set of 7:3 train and the test set S test ;

[0026] 3b, extract the training set S respectively train and the test set S test The various dimensions of maneuver control information are pre-processed and labeled to obtain labeled training set control information data

[0027] and test set status information data The UAV controls its speed and direction through acceleration, pitch velocity and yaw velocity. The magnitude of acceleration will affect the change of the UAV speed. The existence of pitch velocity and yaw velocity controls the direction of the UAV. In order to simulate the different maneuvers of the UAV in air combat, the simulation system sets the acceleration magnitude to 4 values ​​{-30, -10, 40, 60} (unit: m / s2), each value represents a different acceleration or deceleration maneuver of the UAV, and the pitch velocity has 9 values ​​{-30, -21, -12, -5, 0, 5, 12, 21, 30} (unit: degrees / second ), each value represents that the drone performs different pitch maneuvers at a certain angular velocity, and the yaw angular velocity has 9 value cases {-60, -34, -18, -8, 0, 8, 18, 34, 60} (unit: degree / second), each value represents that the drone performs different maneuvers that deviate from the original route at a certain angular velocity, so the drone maneuvering control parameters have a total of 4*9*9=324 value combinations, each combination represents a maneuvering action of the drone, and the prediction of the maneuvering action is the prediction of the value of the maneuvering control parameter combination, and the values ​​of acceleration, pitch angular velocity, and yaw angular velocity are mapped into the form of category labels;

[0028] 3c, respectively for the training set S train and the test set S test The data of each dimension is normalized and preprocessed to obtain the normalized training set data And the test set data

[0029]

[0030] In the formula, x i represents the data of the i-th dimension in the training set or test set, represents the minimum value of the i-th dimension data, represents the maximum value of the i-th dimension data, is the result after normalization of the i-th dimension data;

[0031] Normalized training set data And test set data They are all stored in the dimensional format of (batch, seq, input), where batch represents the dimension of the network model batch processing, seq represents the length of the time series used to predict the target's next maneuver, and input represents the dimension of each data in the normalized dataset.

[0032] The step 4 is specifically as follows:

[0033] The LSTM-based target maneuver prediction network model includes an LSTM network with 2 hidden layers, 3 fully connected layers and 3 cross entropy loss functions. The input of the LSTM network model is the normalized training set data sequence of consecutive seq moments, and the input format is (batch, seq, input); the output is the encapsulated hidden state of the target machine maneuver within the near seq time period The output format is (batch, seq, hidden);

[0034] Hidden state of the target machine maneuvering within the time period of seq Input 3 fully connected layers respectively, combined with 3 cross entropy loss functions L a , L p , L y , respectively obtain the probability of acceleration, pitch angular velocity and yaw angular velocity taking each type of value at the time seq+1:

[0035]

[0036]

[0037] In the formula, L represents the calculation formula of the cross entropy function, N represents the number of samples, M represents the number of categories, and y ic is a symbolic function. If the true category of sample i is c, y ic Take 1, otherwise y ic Take 0; L a , L p , L y They represent the acceleration cross entropy loss, pitch angle cross entropy loss and yaw angle cross entropy loss calculated using the cross entropy loss function L for acceleration, pitch angular velocity and yaw angular velocity respectively; P a Indicates the probability of the target acceleration taking each type of value at the time seq+1, It represents the probability of the target acceleration taking the mth value at the time seq+1, P pIndicates the probability of the target pitch velocity taking each type of value at the time seq+1, It represents the probability of the target pitch velocity taking the nth value at the time seq+1, P y Indicates the probability of the target yaw angular velocity taking each type of value at the time seq+1, Indicates the probability that the target yaw angular velocity takes the nth value at the time seq+1.

[0038] The category values ​​with the highest probability of target acceleration a, pitch angular velocity p, and yaw angular velocity y are taken as the predicted category values ​​of the target machine acceleration a, pitch angular velocity p, and yaw angular velocity y at the time seq+1, P a pre , P p pre , P y pre :

[0039]

[0040] The predicted category values ​​of acceleration a, pitch angular velocity p, and yaw angular velocity y are combined to obtain the predicted category label value of the maneuvering action. At this point, the maneuvering prediction problem of the drone is transformed into a classification prediction problem of target acceleration a, pitch angular velocity p, and yaw angular velocity y.

[0041] In step 5, during model training, the acceleration cross entropy loss L a , pitch angle cross entropy loss L p and yaw angle cross entropy loss L y Perform weighted fusion and use the fused loss function value as the training loss L of the entire network model train :

[0042] L train =w a L a +w p L p +w y L y

[0043] In the formula, w a , w p , w y Represents the acceleration cross entropy loss L a , pitch angle cross entropy loss L p and yaw angle cross entropy loss L y The weight of .

[0044] An electronic device comprises a processor, a memory and a communication bus, wherein the processor and the memory communicate with each other via the communication bus;

[0045] Memory, used to store computer programs;

[0046] The processor is used to implement the above-mentioned LSTM-based target maneuvering prediction method when executing the program stored in the memory.

[0047] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned LSTM-based target maneuvering prediction method is implemented.

[0048] Beneficial effects of the present invention:

[0049] When predicting the maneuver of the target aircraft, the present invention extracts the spatial situational position characteristics of the target aircraft, and uses them together with the state and maneuver information of our UAV and the state and maneuver information of the target aircraft as the basis for predicting the maneuver of the target aircraft. This method can fully explore the law of maneuver selection of the target aircraft in the air combat confrontation data, making the prediction of the maneuver of the target aircraft more accurate.

[0050] The present invention analyzes the relationship between the maneuvering action of the UAV and the acceleration, pitch angular velocity and yaw angular velocity of the UAV, labels the values ​​of acceleration, pitch angular velocity and yaw angular velocity, and converts the prediction problem of the maneuvering action selection of the target aircraft into a classification prediction problem of the three maneuvering control variables of acceleration, pitch angular velocity and yaw angular velocity. It can accurately predict the future maneuvering action of the target aircraft and provide a new idea for the research of the maneuver prediction method of the target aircraft.

[0051] The present invention starts from the characteristic that the maneuver of the target aircraft is time-series, establishes a target aircraft maneuver prediction network model based on LSTM, and utilizes the powerful prediction ability of LSTM for time series data. According to the state information and maneuver information of our UAV, the state information and maneuver information of the target aircraft and the spatial situation occupancy characteristics of the target aircraft within a period of time, the maneuvering action of the target aircraft at the next moment is predicted. Compared with other network models, better prediction effect is achieved, and accurate prediction of the maneuvering action of the target aircraft is realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a schematic diagram of the azimuth angles between our UAV and the target aircraft during air combat confrontation.

[0053] Figure 2 It is a schematic diagram of labeling processing of UAV maneuver control parameters.

[0054] Figure 3 It is a schematic diagram of the maneuver prediction method of the LSTM-based target. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below in conjunction with embodiments.

[0056] The LSTM-based target maneuver prediction method includes the following steps:

[0057] 1. Construct a dataset of target maneuver selection in the UAV combat process;

[0058] Using a one-on-one air combat simulation system, the two sides set different initial states to conduct multiple combat experiments. In each round of combat, the status information of the drones of the two sides was recorded every 20ms. t and maneuver control information c t , forming air combat confrontation data, and each battle process ends when one side wins or the battle lasts for more than 2 minutes. In this example, 280 groups of battle experiments are selected, including a total of 14,397,712 battle maneuver data as the target maneuver prediction data set S. At any time t, the state information s of the drones of the two battle parties t and maneuver control information c t Contains information in the following multiple dimensions, as shown in the formula:

[0059] s t =[X t ,Y t ,Z t ,EX t ,EY t ,EZ t ,AX t ,AY t ,AZ t ,V t ]

[0060] c t =[a t ,p t ,y t ]

[0061] Where, X t ,Y t ,Z t They represent the three-dimensional spatial position coordinates of the drone at time t, EX t ,EY t ,EZ t They represent the Euler angles of the drone’s attitude at time t, (AX t ,AY t ,AZ t ) constitutes the nose heading vector of the drone at time t, V t Indicates the speed of the drone at time t, s t Contains status information of 10 dimensions. t ,p t ,y tThey represent the acceleration information, pitch angle information and yaw angle information for controlling the speed and direction of the drone, respectively. t Contains 3-dimensional maneuver control information. In this example, the state information of our drone at any time t is expressed as Indicates that the maneuver control information is used Indicates; the status information of the target UAV is indicated by Indicates that the maneuver control information is used Represents; Data set S = [s u ,c u ,s e ,c e ].

[0062] 2. Extract the spatial situation characteristics of the target maneuver confrontation process;

[0063] In the process of air combat, the purpose of the drone's continuous maneuvering is to compete with the target drone, so as to put its own space position in an advantageous state and achieve the purpose of attacking the target aircraft. Therefore, spatial situation information is a key factor in the drone's maneuver selection. In the process of close-range air combat, the key to the drone's victory is that when the target aircraft is within the attack range of its own drone, the angle situation of its own drone is in an advantageous state. At this time, the azimuth angle p of the target aircraft and the entry angle q of the target aircraft are close to 0, such as Figure 1 Therefore, when predicting the maneuver of the target aircraft, it is necessary to consider the distance d between the target aircraft and our aircraft and the angle of the target aircraft T. a In this case, the distance feature d and angle situation feature T of the target maneuver are extracted from the target maneuver prediction dataset S. a :

[0064]

[0065]

[0066] Where, X u ,Y u ,Z u They represent the three-dimensional spatial coordinates of our drone, X e ,Y e ,Z e They respectively represent the three-dimensional spatial position coordinates of the target UAV.

[0067] The extracted distance features and angle situation features are incorporated into the target maneuver prediction dataset S:

[0068] S=[s u ,c u ,s e ,c e ,d,T a ]

[0069] In the formula, s u ,c u They represent the state information and maneuver control information of our UAV in the target maneuver prediction dataset S, s e ,c e They respectively represent the state information and maneuver control information of the target UAV in the target maneuver prediction dataset S.

[0070] 3. Segment and preprocess the dataset.

[0071] 3a, the target maneuver prediction dataset S is divided into training set S and test set S according to the ratio of 7:3. train and the test set S test ;

[0072] 3b, extract the training set S respectively train and the test set S test The various dimensions of maneuver control information are pre-processed and labeled to obtain labeled training set control information data

[0073] and test set status information data The drone controls its own speed and direction through acceleration, pitch velocity and yaw velocity. The magnitude of acceleration will affect the change of the drone's speed, and the existence of pitch velocity and yaw velocity controls the direction of the drone. In order to simulate the different maneuvers of drones in air combat, the simulation system sets the acceleration size to 4 values ​​{-30, -10, 40, 60} (unit: m / s squared), each value represents that the drone performs different acceleration or deceleration maneuvers, the pitch angular velocity has 9 values ​​{-30, -21, -12, -5, 0, 5, 12, 21, 30} (unit: degrees / second), each value represents that the drone performs different pitch maneuvers at a certain angular velocity, and the yaw angular velocity has 9 values ​​{-60, -34, -18, -8, 0, 8, 18, 34, 60} (unit: degrees / second), each value represents that the drone performs different maneuvers that deviate from the original route at a certain angular velocity. Therefore, there are 4*9*9=324 combinations of values ​​for the drone maneuver control parameters, each combination represents a maneuver of the drone, and the prediction of the maneuver is the prediction of the combination value of the maneuver control parameters. The values ​​of acceleration, pitch angular velocity, and yaw angular velocity are mapped into category labels, such as Figure 2 As shown;

[0074] 3c, respectively for the training set S train and the test set S test The data of each dimension is normalized and preprocessed to obtain the normalized training set data

[0075] And the test set data

[0076]

[0077] In the formula, x i represents the data of the i-th dimension in the training set or test set, represents the minimum value of the i-th dimension data, represents the maximum value of the i-th dimension data, is the result after normalization of the i-th dimension data;

[0078] Normalized training set data And the test set data They are stored in the format of (batch, seq, input), where batch represents the dimension of the network model batch processing, seq represents the length of the time series used to predict the next maneuver of the target, and input represents the dimension of each data in the normalized data set. In this example, batch is set to 256, seq is set to 14, and input is 28.

[0079] 4. Establish a maneuver prediction network model based on LSTM;

[0080] like Figure 3 As shown in the figure, the LSTM-based target maneuver prediction network model includes an LSTM network with 2 hidden layers, 3 fully connected layers, and 3 cross entropy loss functions. The input of the LSTM network model is a normalized training set data sequence of consecutive seq moments, and the input format is (batch, seq, input); the output is the encapsulated hidden state of the target machine maneuver within the seq time period The output format is (batch, seq, hidden). In this example, hidden is set to 128.

[0081] Hidden state of the target machine maneuvering within the time period of seq Input 3 fully connected layers respectively, combined with 3 cross entropy loss functions L a , L p , L y , respectively obtain the probability of acceleration, pitch angular velocity and yaw angular velocity taking each type of value at the time seq+1:

[0082]

[0083]

[0084] In the formula, L represents the calculation formula of the cross entropy function, N represents the number of samples, M represents the number of categories, and y icis a symbolic function. If the true category of sample i is c, y ic Take 1, otherwise y ic Take 0; L a , L p , L y They represent the acceleration cross entropy loss, pitch angle cross entropy loss and yaw angle cross entropy loss calculated using the cross entropy loss function L for acceleration, pitch angular velocity and yaw angular velocity respectively; P a Indicates the probability of the target acceleration taking each type of value at the time seq+1, It represents the probability of the target acceleration taking the mth value at the time seq+1, P p Indicates the probability of the target pitch velocity taking each type of value at the time seq+1, It represents the probability of the target pitch velocity taking the nth value at the time seq+1, P y Indicates the probability of the target yaw angular velocity taking each type of value at the time seq+1, Indicates the probability that the target yaw angular velocity takes the nth value at the time seq+1.

[0085] The category values ​​with the highest probability of target acceleration a, pitch angular velocity p, and yaw angular velocity y are taken as the predicted category values ​​of the target machine acceleration a, pitch angular velocity p, and yaw angular velocity y at the time seq+1, P a pre , P p pre , P y pre :

[0086]

[0087] The predicted category values ​​of acceleration a, pitch angular velocity p, and yaw angular velocity y are combined to obtain the predicted category label value of the maneuver. Figure 2 The mapping relationship between the category label and the maneuvering control parameter value shown in the figure can obtain the corresponding maneuvering control variable value. At this point, the maneuvering prediction problem of the UAV is transformed into a classification prediction problem of the target acceleration a, pitch angular velocity p and yaw angular velocity y.

[0088] 5. Using the training set Train the LSTM-based target maneuver prediction network model and use the test set Conduct accuracy testing;

[0089] When training the model, the acceleration cross entropy loss L a , pitch angle cross entropy loss L p and yaw angle cross entropy loss L yPerform weighted fusion and use the fused loss function value as the training loss L of the entire network model train :

[0090] L train =w a L a +w p L p +w y L y

[0091] In the formula, w a , w p , w y Represent the acceleration cross entropy loss L a , pitch angle cross entropy loss L p and yaw angle cross entropy loss L y The weight of w is determined by experiment in this example. a , w p , w y The values ​​are 0.2, 0.4, and 0.4 respectively.

[0092] In order to prevent overfitting during model training and make the network model more robust, a dropout mechanism is added to the LSTM network layer, and the dropout value is set to 0.4 in this example; the learning rate hyperparameter of the network model is set to 0.0001, and the number of training epochs is set to 500. The network parameters of the model are saved for each round of training.

[0093] The test set data is tested and verified using the saved network model parameters. The evaluation index used in this model is the accuracy. The closer the accuracy is to 1, the more accurate the classification prediction of the model. The present invention respectively counts the prediction accuracy of acceleration, pitch angular velocity and yaw angular velocity in the test set Acc(a), Acc(p), Acc(y) and the accuracy of the acceleration, pitch angular velocity and yaw angular velocity predicted correctly at the same time Acc(a, p, y). At the same time, this model is compared with the BP neural network and the RNN neural network, and multiple independent experiments are performed on each compared network model. The experimental results are shown in Table 1:

[0094] Table 1 Comparison of target maneuver prediction accuracy of different network models

[0095]

[0096] It can be seen from the results in Table 1 that, in this example, compared with the BP neural network and the RNN, the network model of the present invention has achieved good results both in predicting a single maneuvering control parameter and in correctly predicting all maneuvering control parameters simultaneously, wherein correctly predicting all maneuvering control parameters simultaneously means that the combination value of the maneuvering control parameters is correctly predicted, that is, the target maneuver prediction at a certain moment is correct.

[0097] The present invention analyzes the existing methods for predicting target maneuvers in air combat confrontation processes, and finds that the existing methods for predicting target maneuvers only start from the target itself, and divide the maneuvers of the target aircraft into a weighted combination of several fixed actions, and the weight of each action represents the probability of the target aircraft taking this maneuver. The relationship between its maneuvering action and the situation of the space it occupies is not considered; and the influence of the state of our UAV on the maneuvering of the target aircraft is not considered; therefore, when predicting the maneuvering of the target aircraft, the present invention extracts the spatial situation occupancy characteristics of the target aircraft, and uses them together with the state and maneuvering information of our UAV and the state and maneuvering information of the target aircraft as the basis for predicting the maneuvering of the target aircraft. This can fully explore the rules of maneuvering selection of the target aircraft in the air combat confrontation data, making the prediction of the maneuvering of the target aircraft more accurate.

[0098] The present invention analyzes the relationship between the maneuvering action of the UAV and the acceleration, pitch angular velocity and yaw angular velocity of the UAV, labels the values ​​of acceleration, pitch angular velocity and yaw angular velocity, and converts the maneuvering prediction problem of the target aircraft into a classification prediction problem of the three maneuvering control variables of acceleration, pitch angular velocity and yaw angular velocity. It can accurately predict the future maneuvering action of the target aircraft and provides a new idea for the research of the maneuvering prediction method of the target aircraft.

[0099] The present invention starts from the characteristic that the maneuver of the target aircraft is time-series, establishes a target aircraft maneuver prediction network model based on LSTM, and uses the powerful prediction ability of LSTM for time series data to predict the maneuvering action of the target aircraft at the next moment according to the state information and maneuvering information of our drone, the state information and maneuvering information of the target aircraft, and the spatial situation occupancy characteristics of the target aircraft during this period of time. Due to space limitations, the present invention shows a prediction example in which the target aircraft actually performs a batch of 256 maneuvers in a certain period of time during a certain air combat confrontation and the model performs a batch of 256 predicted maneuvers accordingly. The prediction effect is shown in Table 2. The experimental results further verify the effectiveness of the method of the present invention in predicting the maneuver of the target aircraft.

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0108] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A maneuver prediction method for a target based on LSTM, characterized in that: The steps include: Step 1: Construct a target maneuver selection dataset S of the UAV combat process, where the target refers to the target UAV that is fighting against our UAV; Step 2: Extract the spatial situation characteristics of the target maneuver confrontation process; Step 3: Segment and preprocess the target maneuver selection dataset S in step 1; Step 4: Establish a maneuver prediction network model based on LSTM target; Step 5: Using the training set Train the LSTM-based target maneuver prediction network model and use the test set Conduct prediction accuracy test; In step 1, a one-on-one air combat simulation system is used to set different initial states for the two sides to conduct multiple combat experiments. The state information of the drones of the two sides is recorded every 20ms in each round of combat. t and maneuver control information c t , forming air combat confrontation data, and each round of combat process ends when one side wins or the combat duration exceeds 2 minutes; 280 sets of air combat confrontation data are selected as the target maneuver prediction data set S. At any time t, the state information s of the drones of the two sides in the battle t and maneuver control information c t Contains information in the following multiple dimensions, as shown in the formula: s t [X t ,Y t ,Z t ,AND t ,EY t ,EZ t ,SURE t ,DO YOU t ,AZ t ,V t ] c t =[a t ,p t ,y t ] Where, X t ,Y t ,Z t They represent the three-dimensional spatial position coordinates of the drone at time t, EX t ,EY t ,EZ t They represent the Euler angles of the drone’s attitude at time t, (AX t ,AY t ,AZ t ) constitutes the nose heading vector of the drone at time t, V t Indicates the speed of the drone at time t, s t Contains status information of 10 dimensions, a t ,p t ,y t They represent the acceleration information, pitch angle information and yaw angle information for controlling the speed and direction of the drone, respectively. t Contains 3-dimensional maneuvering control information. The state information of our drone at any time t is used Indicates that the maneuver control information is used Indicates; the status information of the target UAV is indicated by Indicates that the maneuver control information is used Representation; Dataset In the process of close-range air combat confrontation, the key to the victory of the UAV in step 2 is that when the target aircraft is within the attack range of the own UAV, the angle of the own UAV is in an advantageous state. At this time, the azimuth angle p of the target aircraft and the entry angle q of the target aircraft are close to 0. When predicting the maneuver of the target aircraft, it is necessary to consider the distance d between the target aircraft and our aircraft and the angle of the target aircraft T. a In this case, the distance feature d and angle situation feature T of the target maneuver are extracted from the target maneuver prediction data set S obtained in step 1. a : Where, X u ,Y u ,Z u They represent the three-dimensional spatial coordinates of our drone, X e ,Y e ,Z e They respectively represent the three-dimensional spatial position coordinates of the target UAV; The step 4 is specifically as follows: The LSTM-based target maneuver prediction network model includes an LSTM network with 2 hidden layers, 3 fully connected layers and 3 cross entropy loss functions. The input of the LSTM network model is the normalized training set data sequence of consecutive seq moments, and the input format is (batch, seq, input); the output is the encapsulated hidden state of the target machine maneuver within the near seq time period The output format is (batch, seq, hidden); Hidden state of the target machine maneuvering within the time period of seq Input 3 fully connected layers respectively, combined with 3 cross entropy loss functions L a , L p , L y , respectively obtain the probability of acceleration, pitch angular velocity and yaw angular velocity taking each type of value at the time seq+1: In the formula, L represents the calculation formula of the cross entropy function, N represents the number of samples, M represents the number of categories, and y ic is a symbolic function. If the true category of sample i is c, y ic Take 1, otherwise y ic Take 0; L a , L p , L y They represent the acceleration cross entropy loss, pitch angle cross entropy loss and yaw angle cross entropy loss calculated using the cross entropy loss function L for acceleration, pitch angular velocity and yaw angular velocity respectively; P a Indicates the probability of the target acceleration taking each type of value at the time seq+1, It represents the probability of the target acceleration taking the mth value at the time seq+1, P p Indicates the probability of the target pitch velocity taking each type of value at the time seq+1, It represents the probability of the target pitch velocity taking the nth value at the time seq+1, P y Indicates the probability of the target yaw angular velocity taking each type of value at the time seq+1, Indicates the probability that the target yaw angular velocity takes the nth value at the time seq+1.

2. The method for predicting target maneuvers based on LSTM according to claim 1, characterized in that: The extracted distance features and angle situation features are incorporated into the target maneuver prediction dataset S: S=[s u ,c u ,s e ,c e ,d,T a ] In the formula, s u ,c u They represent the state information and maneuver control information of our UAV in the target maneuver prediction dataset S, s e ,c e They respectively represent the state information and maneuver control information of the target UAV in the target maneuver prediction dataset S.

3. The method for predicting the maneuverability of a target based on LSTM according to claim 1, characterized in that: The step 3 comprises: 3a, divide the target maneuver prediction dataset S in step 1 into a training set S according to the ratio of training set to test set of 7:3 train and the test set S test ; 3b, extract the training set S respectively train and the test set S test The various dimensions of maneuver control information are pre-processed and labeled to obtain labeled training set control information data And test set status information data The UAV controls its speed and direction through acceleration, pitch velocity and yaw velocity. The magnitude of acceleration will affect the change of the UAV speed. The existence of pitch velocity and yaw velocity controls the direction of the UAV. In order to simulate the different maneuvers of the UAV in air combat, the simulation system sets the acceleration magnitude to 4 values ​​{-30, -10, 40, 60} (unit: m / s2), each value represents a different acceleration or deceleration maneuver of the UAV, and the pitch velocity has 9 values ​​{-30, -21, -12, -5, 0, 5, 12, 21, 30} (unit: degrees / second ), each value represents that the drone performs different pitch maneuvers at a certain angular velocity, and the yaw angular velocity has 9 value cases {-60, -34, -18, -8, 0, 8, 18, 34, 60} (unit: degree / second), each value represents that the drone performs different maneuvers that deviate from the original route at a certain angular velocity, so the drone maneuvering control parameters have a total of 4*9*9=324 value combinations, each combination represents a maneuvering action of the drone, and the prediction of the maneuvering action is the prediction of the value of the maneuvering control parameter combination, and the values ​​of acceleration, pitch angular velocity, and yaw angular velocity are mapped into the form of category labels; 3c, respectively for the training set S train and the test set S test The data of each dimension is normalized and preprocessed to obtain the normalized training set data And test set data In the formula, x i represents the data of the i-th dimension in the training set or test set, represents the minimum value of the i-th dimension data, represents the maximum value of the i-th dimension data, is the result after normalization of the i-th dimension data; Normalized training set data And the test set data They are all stored in the dimensional format of (batch, seq, input), where batch represents the dimension of the network model batch processing, seq represents the length of the time series used to predict the target's next maneuver, and input represents the dimension of each data in the normalized dataset.

4. The method for predicting target maneuvers based on LSTM according to claim 1, characterized in that: The category values ​​with the highest probability of target acceleration a, pitch angular velocity p, and yaw angular velocity y are taken as the predicted category values ​​of the target machine acceleration a, pitch angular velocity p, and yaw angular velocity y at the time seq+1, P a pre , P p pre , P y pre : The predicted category values ​​of acceleration a, pitch angular velocity p, and yaw angular velocity y are combined to obtain the predicted category label value of the maneuvering action. At this point, the maneuvering prediction problem of the drone is transformed into a classification prediction problem of target acceleration a, pitch angular velocity p, and yaw angular velocity y.

5. The method for predicting target maneuvers based on LSTM according to claim 1, characterized in that: In step 5, during model training, the acceleration cross entropy loss L a , pitch angle cross entropy loss L p and yaw angle cross entropy loss L y Perform weighted fusion and use the fused loss function value as the training loss L of the entire network model train : L train =w a L a +w p L p +w y L y In the formula, w a , w p , w y Represent the acceleration cross entropy loss L a , pitch angle cross entropy loss L p and yaw angle cross entropy loss L y The weight of .

6. An electronic device, characterized in that: It includes a processor, a memory and a communication bus, wherein the processor and the memory communicate with each other via the communication bus; Memory, used to store computer programs; A processor is used to implement a LSTM-based target maneuvering prediction method as described in any one of claims 1 to 5 when executing a program stored in a memory.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for maneuvering prediction of a target based on LSTM as described in any one of claims 1 to 5 is implemented.

Citation Information

Patent Citations

  • Track association method, system and device and readable storage medium

    CN112001433A