BiLSTM + Attention three-layer air combat target intention prediction method of multi-feature continuous time sequence

Abstract

The invention discloses a BiLSTM + Attention three-layer air combat target intention prediction method of a multi-feature continuous time sequence. The method comprises the following steps: 1, establishing an air combat intention prediction basic data set; 2, performing threat assessment on the air combat intention basic data set to obtain an air combat intention threat data set; 3, constructing a BiLSTM + Attention network model, and inputting the air combat intention threat data set for training; and 4, performing tactical intention prediction on subsequent time series data based on trajectory prediction, threat evaluation and the trained BiLSTM + Attention network model. The method is simple in steps and reasonable in design, intention prediction is carried out by combining forward air combat feature data at multiple continuous moments, backward target aircraft air combat trajectory prediction data and battlefield threat assessment, and the reliability of air combat target intention prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of air combat target intention prediction, and specifically relates to a BiLSTM+Attention three-layer air combat target intention prediction method of multi-feature continuous time series. Background Art

[0002] The modern warfare environment is complex and battlefield situation awareness technology is flexible and changeable. How to use modern technology to mine insights into the information behind real-time battlefield data in complex air combat scenarios, mine valuable information that is conducive to combat decision-making through target status data, and predict the enemy's tactical intentions quickly and intelligently in advance is the key to winning combat.

[0003] Tactical intention prediction refers to extracting characteristic information related to tactical intention from comprehensive analysis of battlefield feature data and battlefield intelligence information, thereby predicting the tactical intention of enemy targets. At present, in the research on tactical intention prediction of air combat, it mostly refers to tactical intentions such as attack, cover, feint, retreat, reconnaissance, and interference. Combat intentions are abstract, confrontational, concealed, and deceptive, and are closely related to their trajectories and the threats they encounter. It is difficult to quickly and accurately identify the tactical intentions of enemy targets by relying solely on manual labor. There is an urgent need to establish an efficient and reliable target intention prediction method to provide auxiliary decision-making support for air combat.

[0004] The current methods for predicting target intentions in air combat mainly include: methods based on knowledge rules and theoretical reasoning, methods based on similarity comparison, and machine learning methods based on neural networks. The knowledge rule method faces problems such as difficulty in knowledge expression, difficulty in establishing a knowledge base, and easy to fall into local optimality; theoretical reasoning methods such as Bayes have a strict mathematical foundation and rigorous theory, and can perform sequential analysis of enemy tactical intentions, but in practical applications, it is usually necessary to determine a large number of prior probabilities and conditional probabilities, which is often difficult; methods based on similarity comparison usually have a large amount of calculation, resulting in great difficulty in storage and search; relatively speaking, machine learning methods based on neural networks have advantages in fitting and approximation capabilities in information warfare situations, and can improve the intelligence and autonomy of tactical intention prediction. However, the actual combat intention is usually hidden in the dynamic and time-series changing feature state information, and the feature state information has forward and backward correlations with data features. Secondly, relying on single-moment feature information to identify the target combat intention is prone to one-sided prediction results.

[0005] Therefore, a multi-feature continuous time series BiLSTM+Attention three-layer air combat target intention prediction method is needed, which combines the forward multiple continuous time air combat feature data and the backward target aircraft air combat trajectory prediction data as well as the battlefield threat assessment to perform intention prediction, thereby improving the reliability of air combat target intention prediction. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a BiLSTM+Attention three-layer air combat target intention prediction method with multiple features and continuous time series in view of the deficiencies in the above-mentioned prior art. The method has simple steps and reasonable design. It combines the forward continuous air combat feature data at multiple moments and the backward target aircraft air combat trajectory prediction data as well as the battlefield threat assessment to perform intention prediction, thereby improving the reliability of air combat target intention prediction.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a BiLSTM+Attention three-layer air combat target intention prediction method of multi-feature continuous time series, characterized in that the method comprises the following steps:

[0008] Step 1: Establish a basic data set for air combat intention prediction:

[0009] Step 101, during the air combat confrontation process, the three-dimensional coordinates, speed, and heading characteristic data of the target aircraft and our aircraft are collected in a continuous time sequence to obtain air combat characteristic data;

[0010] Step 102, according to the air combat feature data, obtaining the time series feature data corresponding to each tactical intention type; wherein the tactical intention types are attack, cover, feint, retreat, reconnaissance, and interference, and the labels of the tactical intention types are 1, 2, 3, 4, 5, and 6 respectively; 1 means attack, 2 means cover, 3 means feint, 4 means retreat, 5 means reconnaissance, and 6 means interference;

[0011] Step 103, judging and correcting the length of the time series feature data corresponding to each tactical intention type to obtain each corrected time series feature data; wherein the length of each corrected time series feature data is the same;

[0012] Step 104, recording each corrected time series feature data and its corresponding tactical intention type as an air combat intention basic data set;

[0013] Step 2: Perform threat assessment on the air combat intention basic data set to obtain the air combat intention threat data set:

[0014] Step 201: Obtain the three-dimensional coordinates of the target aircraft at the jth sampling moment from the single corrected time series feature data, and combine it with the three-dimensional coordinates of our aircraft at the jth sampling moment to obtain the relative distance d between our aircraft and the target aircraft at the jth sampling moment. j ; Wherein, j is a positive integer;

[0015] Step 202: According to Get the angle threat index S at the jth sampling moment aj ; where φ jThe velocity vector of our aircraft and the angle between our aircraft and the target aircraft at the jth sampling moment, q j represents the angle between the velocity vector of the target aircraft and the line connecting our aircraft and the target aircraft at the jth sampling moment;

[0016] Step 203: According to Get the distance threat index S at the jth sampling moment dj ; where d j represents the relative distance between our aircraft and the target aircraft at the jth sampling moment; d l represents the missile range of the target aircraft, d d Indicates the maximum radar detection range of the target aircraft;

[0017] Step 204: According to Get the speed threat index S at the jth sampling moment vj ; Among them, V mj represents the speed of our aircraft at the jth sampling moment, V wj represents the speed of the target aircraft at the jth sampling moment;

[0018] Step 205: Calculate the angle threat index S at the jth sampling moment. aj , the distance threat index S at the jth sampling moment dj , the speed threat index S at the jth sampling moment vj Perform weighted processing to obtain the threat assessment value K of the target aircraft to our aircraft at the jth sampling moment j ;

[0019] Step 206: According to the method of step 201 to step 205, the threat assessment value of the target aircraft to any friendly aircraft at the j-th sampling moment is obtained; wherein the threat assessment value of the target aircraft to the e-th friendly aircraft at the j-th sampling moment is recorded as K j (e); e is the serial number of our aircraft;

[0020] Step 207, select the maximum value from the threat assessment values ​​of the target aircraft to any friendly aircraft at the j-th sampling moment, and record it as the maximum threat value;

[0021] according to Get the total threat F of the target aircraft to our aircraft at the jth sampling moment z ; Where, e and E are both positive integers, and 1≤e≤E, and E represents the total number of our aircraft;

[0022] Step 208, repeating steps 201 to 207 multiple times to obtain the maximum threat value and total threat of each time series feature data;

[0023] Step 209: The three-dimensional coordinates, speed, heading, relative distance, maximum threat value and total threat of the target aircraft are taken as individual optimized time series feature data; and each optimized time series feature data and its corresponding tactical intent type are recorded as an air combat intent threat data set;

[0024] Step 3: Build a BiLSTM+Attention network model and input the air combat intention threat dataset into the training:

[0025] Step 301: Build a BiLSTM+Attention network model;

[0026] Step 302: Input all the air combat intention threat data sets into the BiLSTM+Attention network model for training to obtain a trained BiLSTM+Attention network model;

[0027] Step 4: Subsequent time series data is used to predict tactical intent based on trajectory prediction, threat assessment, and the trained BiLSTM+Attention network model:

[0028] Step 401, in the subsequent air combat confrontation process, the three-dimensional coordinates, speed, and heading characteristic data of the target aircraft and our aircraft in a continuous time sequence are collected to obtain subsequent air combat characteristic data of the target aircraft and subsequent air combat characteristic data of our aircraft;

[0029] Step 402: The three-dimensional coordinates and headings in the subsequent target aircraft air combat feature data are input into the trained LSTM neural network model through standardization, and the subsequent air combat trajectory prediction standardized data that is continuous with the subsequent target aircraft air combat feature data is output; and the subsequent air combat trajectory prediction standardized data is de-standardized to obtain the subsequent target aircraft air combat trajectory prediction data;

[0030] Based on the three-dimensional coordinates in the subsequent target aircraft air combat trajectory prediction data, the subsequent speed prediction data is obtained;

[0031] Step 403, based on the subsequent target aircraft air combat characteristic data and the subsequent our aircraft air combat characteristic data, according to the method of step 2, obtain the subsequent maximum threat value and the subsequent total threat;

[0032] Based on the subsequent target aircraft air combat trajectory prediction data, the subsequent speed prediction data and the air combat feature data of the subsequent our aircraft air combat feature data at the last sampling moment, the predicted maximum threat value and the predicted total threat are obtained according to the method of step 2;

[0033] Step 404: record the three-dimensional coordinates, speed, heading, relative distance, subsequent maximum threat value and subsequent total threat in the subsequent target aircraft air combat feature data as subsequent feature data; record the three-dimensional coordinate predicted value, speed predicted value, heading predicted value, relative distance predicted value, predicted maximum threat value and predicted total threat in the subsequent target aircraft air combat feature data as subsequent predicted feature data; and splice the subsequent feature data and the subsequent predicted feature data to form fusion measurement and prediction time series feature data;

[0034] Step 405: Input the fused measurement and prediction time series feature data into the trained BiLSTM+Attention network model to perform tactical intent prediction and obtain the output probability of each tactical intent;

[0035] Step 406: Select the tactical intention corresponding to the maximum probability from the output probabilities of the various tactical intentions as the predicted tactical intention.

[0036] The above-mentioned BiLSTM+Attention three-layer air combat target intention prediction method of multi-feature continuous time series is characterized in that: step 103, the specific process is as follows:

[0037] Step 1031, obtaining the number of samples of the time series feature data corresponding to the i-th tactical intention type, and recording the number of samples of the time series feature data corresponding to the i-th tactical intention type as the length of the i-th time series feature data; wherein i is a positive integer;

[0038] Step 1032: determine the length of each time series feature data. If the lengths of each time series feature data are different, obtain the mode of the lengths of each time series feature data and record it as B.

[0039] Step 1033: record the length of the i-th time series feature data as Ai. If Ai is less than B, use the cubic spline interpolation method to interpolate the i-th time series feature data until the length of the i-th time series feature data after interpolation is B, thereby obtaining the corrected i-th time series feature data;

[0040] If Ai is greater than B and less than 2B, the part greater than B is deleted until the length of the i-th time series feature data after deletion is B, and the corrected i-th time series feature data is obtained;

[0041] If Ai is greater than αB, the part greater than αB is deleted, and the i-th time series feature data is divided into α corrected i-th time series feature data; wherein α represents a multiple, and α is a positive integer greater than 2;

[0042] Step 1034, repeat step 1033 multiple times to complete the correction of each time series feature data, and obtain each corrected time series feature data.

[0043] The above-mentioned BiLSTM+Attention three-layer air combat target intention prediction method of multi-feature continuous time series is characterized in that: the trained LSTM neural network model is obtained in step 403, and the specific process is as follows:

[0044] Step 4031: Obtain 2 / 3B of the time series feature data from each corrected time series feature data as input data, and use the remaining 1 / 3B of the time series feature data as output label data; wherein the input data and the output label data include three-dimensional coordinates and heading;

[0045] Step 4032: standardize the input data and the output label data using a Z-Score standardization method to obtain input standardized data and output label standardized data;

[0046] Step 4033, repeating steps 4031 and 4032 multiple times to obtain multiple sets of input standardized data and their corresponding output label standardized data, and recording them as trajectory data training sets;

[0047] Step 4034: construct an LSTM neural network model; wherein the LSTM neural network model includes an input layer, an LSTM network layer, a fully connected layer and an output layer;

[0048] Step 4035: input the trajectory data training set into the constructed LSTM neural network model for training to obtain a trained LSTM neural network model.

[0049] The above-mentioned BiLSTM+Attention three-layer air combat target intention prediction method of multi-feature continuous time series is characterized in that: step 205, the specific process is as follows:

[0050] Step 2051: normalize the angle threat index at each sampling moment to obtain the normalized value S of the angle threat index at the jth sampling moment. a ' j ;

[0051] Step 2052: According to K j =k1S a ' j +k2S dj +k3S vj , get the threat assessment value K of the target aircraft to our aircraft at the jth sampling moment j ; Among them, k1 represents the first weight coefficient, and k1 takes the value of 0.35, k2 represents the second weight coefficient, and k2 takes the value of 0.35, k3 represents the third weight coefficient, and k3 takes the value of 0.3.

[0052] The above-mentioned BiLSTM+Attention three-layer air combat target intention prediction method of multi-feature continuous time series is characterized in that: the BiLSTM+Attention network model in step 301 includes a BiLSTM module, an Attention module, a fully connected layer and a Softmax layer; the output of the Softmax layer is 6;

[0053] In step 302, the optimized time series feature data is used as the input of the BiLSTM+Attention network model, and the optimized time series feature data and its corresponding tactical intent type are used as the output of the BiLSTM+Attention network model; wherein, if the optimized time series feature data belongs to its corresponding tactical intent type, the tactical intent type takes a value of 1, otherwise it takes a value of zero.

[0054] The above-mentioned BiLSTM+Attention three-layer air combat target intention prediction method of multi-feature continuous time series is characterized in that: in step 403, the subsequent speed prediction data is obtained based on the three-dimensional coordinates in the subsequent target aircraft air combat trajectory prediction data, as follows:

[0055] The subsequent three-dimensional coordinate data is obtained from the subsequent target aircraft air combat trajectory prediction data, and the subsequent speed prediction data is obtained according to the ratio of the relative distance of the three-dimensional coordinates at two adjacent sampling moments to the sampling time.

[0056] Compared with the prior art, the present invention has the following advantages:

[0057] 1. The method of the present invention has simple steps and reasonable design, and solves the problem that the current prediction of air combat target intention relies on single-time feature information to identify the target combat intention, which easily leads to one-sided prediction results.

[0058] 2. This paper adopts LSTM neural network to predict the target aircraft air combat trajectory, so as to obtain the backward target aircraft air combat trajectory prediction data; and predicts the battlefield threat environment based on the trajectory prediction; finally, the forward continuous air combat feature data at multiple moments and the backward target aircraft air combat trajectory prediction data and battlefield threat environment factors are combined to predict the intention. Through the three-layer air combat target intention prediction, the rationality and accuracy of the intention prediction are improved.

[0059] 3. The present invention corrects the length of each time series feature data based on the mode of the length of each time series feature data, so that the lengths of each subsequent corrected time series feature data are the same, thereby ensuring that the type of each combat intention is suitable for the BiLSTM+Attention network model, thereby improving the rapidity and adaptability of subsequent training.

[0060] 4. When training the BiLSTM+Attention network model, the present invention considers weighting the maximum threat value and the total threat variable based on the angle threat index, distance threat index and speed threat index, and combines different weights to effectively consider the weights of threat environment factors that affect the combat intention type, which can be closer to the performance of tactical intention prediction and improve the effect of intention prediction.

[0061] 5. The present invention introduces a BiLSTM bidirectional loop mechanism into the BiLSTM+Attention network model, which can simultaneously capture the continuous time series feature data of the target aircraft in the past and future, thereby improving the performance of the intention prediction link; the introduction of the attention mechanism can allocate corresponding feature coefficient weights according to the importance of the features learned by the model, thereby achieving the effect of optimizing the performance of the neural network model and improving the target intention prediction effect.

[0062] In summary, the method of the present invention has simple steps and reasonable design. It combines the forward multiple continuous air combat feature data at a time and the backward target aircraft air combat trajectory prediction data as well as the battlefield threat assessment to perform intention prediction, thereby improving the reliability of air combat target intention prediction.

[0063] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The figure is a flowchart of the method of the present invention. DETAILED DESCRIPTION

[0065] like Figure 1 As shown, the present invention provides a BiLSTM+Attention three-layer air combat target intention prediction method with multiple features and continuous time series, comprising the following steps:

[0066] Step 1: Establish a basic data set for air combat intention prediction:

[0067] Step 101, during the air combat confrontation process, the three-dimensional coordinates, speed, and heading characteristic data of the target aircraft and our aircraft are collected in a continuous time sequence to obtain air combat characteristic data;

[0068] Step 102, according to the air combat feature data, obtaining the time series feature data corresponding to each tactical intention type; wherein the tactical intention types are attack, cover, feint, retreat, reconnaissance, and interference, and the labels of the tactical intention types are 1, 2, 3, 4, 5, and 6 respectively; 1 means attack, 2 means cover, 3 means feint, 4 means retreat, 5 means reconnaissance, and 6 means interference;

[0069] Step 103, judging and correcting the length of the time series feature data corresponding to each tactical intention type to obtain each corrected time series feature data; wherein the length of each corrected time series feature data is the same;

[0070] Step 104, recording each corrected time series feature data and its corresponding tactical intention type as an air combat intention basic data set;

[0071] Step 2: Perform threat assessment on the air combat intention basic data set to obtain the air combat intention threat data set:

[0072] Step 201: Obtain the three-dimensional coordinates of the target aircraft at the jth sampling moment from the single corrected time series feature data, and combine it with the three-dimensional coordinates of our aircraft at the jth sampling moment to obtain the relative distance d between our aircraft and the target aircraft at the jth sampling moment. j ; Wherein, j is a positive integer;

[0073] Step 202: According to Get the angle threat index S at the jth sampling moment aj ; where φ j The velocity vector of our aircraft and the angle between our aircraft and the target aircraft at the jth sampling moment, q j represents the angle between the velocity vector of the target aircraft and the line connecting our aircraft and the target aircraft at the jth sampling moment;

[0074] Step 203: According to Get the distance threat index S at the jth sampling moment dj ; where d j represents the relative distance between our aircraft and the target aircraft at the jth sampling moment; d l represents the missile range of the target aircraft, d d Indicates the maximum radar detection range of the target aircraft;

[0075] Step 204: According to Get the speed threat index S at the jth sampling moment vj ; Among them, V mj represents the speed of our aircraft at the jth sampling moment, V wj represents the speed of the target aircraft at the jth sampling moment;

[0076] Step 205: Calculate the angle threat index S at the jth sampling moment. aj , the distance threat index S at the jth sampling moment dj , speed threat index S at the jth sampling moment vj Perform weighted processing to obtain the threat assessment value K of the target aircraft to our aircraft at the jth sampling moment j ;

[0077] Step 206: According to the method of step 201 to step 205, the threat assessment value of the target aircraft to any friendly aircraft at the j-th sampling moment is obtained; wherein the threat assessment value of the target aircraft to the e-th friendly aircraft at the j-th sampling moment is recorded as K j (e); e is the serial number of our aircraft;

[0078] Step 207, select the maximum value from the threat assessment values ​​of the target aircraft to any friendly aircraft at the j-th sampling moment, and record it as the maximum threat value;

[0079] according to Get the total threat F of the target aircraft to our aircraft at the jth sampling moment z ; Where, e and E are both positive integers, and 1≤e≤E, and E represents the total number of our aircraft;

[0080] Step 208, repeating steps 201 to 207 multiple times to obtain the maximum threat value and total threat of each time series feature data;

[0081] Step 209: The three-dimensional coordinates, speed, heading, relative distance, maximum threat value and total threat of the target aircraft are taken as individual optimized time series feature data; and each optimized time series feature data and its corresponding tactical intent type are recorded as an air combat intent threat data set;

[0082] Step 3: Build a BiLSTM+Attention network model and input the air combat intention threat dataset into the training:

[0083] Step 301: Build a BiLSTM+Attention network model;

[0084] Step 302: Input all the air combat intention threat data sets into the BiLSTM+Attention network model for training to obtain a trained BiLSTM+Attention network model;

[0085] Step 4: Subsequent time series data is used to predict tactical intent based on trajectory prediction, threat assessment, and the trained BiLSTM+Attention network model:

[0086] Step 401, in the subsequent air combat confrontation process, the three-dimensional coordinates, speed, and heading characteristic data of the target aircraft and our aircraft in a continuous time sequence are collected to obtain subsequent air combat characteristic data of the target aircraft and subsequent air combat characteristic data of our aircraft;

[0087] Step 402: The three-dimensional coordinates and headings in the subsequent target aircraft air combat feature data are input into the trained LSTM neural network model through standardization, and the subsequent air combat trajectory prediction standardized data that is continuous with the subsequent target aircraft air combat feature data is output; and the subsequent air combat trajectory prediction standardized data is de-standardized to obtain the subsequent target aircraft air combat trajectory prediction data;

[0088] Based on the three-dimensional coordinates in the subsequent target aircraft air combat trajectory prediction data, the subsequent speed prediction data is obtained;

[0089] Step 403, based on the subsequent target aircraft air combat characteristic data and the subsequent our aircraft air combat characteristic data, according to the method of step 2, obtain the subsequent maximum threat value and the subsequent total threat;

[0090] Based on the subsequent target aircraft air combat trajectory prediction data, the subsequent speed prediction data and the air combat feature data of the subsequent our aircraft air combat feature data at the last sampling moment, the predicted maximum threat value and the predicted total threat are obtained according to the method of step 2;

[0091] Step 404: record the three-dimensional coordinates, speed, heading, relative distance, subsequent maximum threat value and subsequent total threat in the subsequent target aircraft air combat feature data as subsequent feature data; record the three-dimensional coordinate predicted value, speed predicted value, heading predicted value, relative distance predicted value, predicted maximum threat value and predicted total threat in the subsequent target aircraft air combat feature data as subsequent predicted feature data; and splice the subsequent feature data and the subsequent predicted feature data to form fusion measurement and prediction time series feature data;

[0092] Step 405: Input the fused measurement and prediction time series feature data into the trained BiLSTM+Attention network model to perform tactical intent prediction and obtain the output probability of each tactical intent;

[0093] Step 406: Select the tactical intention corresponding to the maximum probability from the output probabilities of the various tactical intentions as the predicted tactical intention.

[0094] In this embodiment, step 103, the specific process is as follows:

[0095] Step 1031, obtaining the number of samples of the time series feature data corresponding to the i-th tactical intention type, and recording the number of samples of the time series feature data corresponding to the i-th tactical intention type as the length of the i-th time series feature data; wherein i is a positive integer;

[0096] Step 1032: determine the length of each time series feature data. If the lengths of each time series feature data are different, obtain the mode of the lengths of each time series feature data and record it as B.

[0097] Step 1033: record the length of the i-th time series feature data as Ai. If Ai is less than B, use the cubic spline interpolation method to interpolate the i-th time series feature data until the length of the i-th time series feature data after interpolation is B, thereby obtaining the corrected i-th time series feature data;

[0098] If Ai is greater than B and less than 2B, the part greater than B is deleted until the length of the i-th time series feature data after deletion is B, and the corrected i-th time series feature data is obtained;

[0099] If Ai is greater than αB, the part greater than αB is deleted, and the i-th time series feature data is divided into α corrected i-th time series feature data; wherein α represents a multiple, and α is a positive integer greater than 2;

[0100] Step 1034, repeat step 1033 multiple times to complete the correction of each time series feature data, and obtain each corrected time series feature data.

[0101] In this embodiment, the trained LSTM neural network model is obtained in step 403, and the specific process is as follows:

[0102] Step 4031: Obtain 2 / 3B of the time series feature data from each corrected time series feature data as input data, and use the remaining 1 / 3B of the time series feature data as output label data; wherein the input data and the output label data include three-dimensional coordinates and heading;

[0103] Step 4032: standardize the input data and the output label data using a Z-Score standardization method to obtain input standardized data and output label standardized data;

[0104] Step 4033, repeating steps 4031 and 4032 multiple times to obtain multiple sets of input standardized data and their corresponding output label standardized data, and recording them as trajectory data training sets;

[0105] Step 4034: construct an LSTM neural network model; wherein the LSTM neural network model includes an input layer, an LSTM network layer, a fully connected layer and an output layer;

[0106] Step 4035: input the trajectory data training set into the constructed LSTM neural network model for training to obtain a trained LSTM neural network model.

[0107] In this embodiment, step 205, the specific process is as follows:

[0108] Step 2051: normalize the angle threat index at each sampling moment to obtain the normalized value S of the angle threat index at the jth sampling moment. a 'j ;

[0109] Step 2052: According to K j =k1S a ' j +k2S dj +k3S vj , get the threat assessment value K of the target aircraft to our aircraft at the jth sampling moment j ; Among them, k1 represents the first weight coefficient, and k1 takes the value of 0.35, k2 represents the second weight coefficient, and k2 takes the value of 0.35, k3 represents the third weight coefficient, and k3 takes the value of 0.3.

[0110] In this embodiment, the BiLSTM+Attention network model in step 301 includes a BiLSTM module, an Attention module, a fully connected layer and a Softmax layer; the output of the Softmax layer is 6;

[0111] In step 302, the optimized time series feature data is used as the input of the BiLSTM+Attention network model, and the optimized time series feature data and its corresponding tactical intent type are used as the output of the BiLSTM+Attention network model; wherein, if the optimized time series feature data belongs to its corresponding tactical intent type, the tactical intent type takes a value of 1, otherwise it takes a value of zero.

[0112] In this embodiment, in step 403, the subsequent speed prediction data is obtained based on the three-dimensional coordinates in the subsequent target aircraft air combat trajectory prediction data, as follows:

[0113] The subsequent three-dimensional coordinate data is obtained from the subsequent target aircraft air combat trajectory prediction data, and the subsequent speed prediction data is obtained according to the ratio of the relative distance of the three-dimensional coordinates at two adjacent sampling moments to the sampling time.

[0114] In this embodiment, in actual use, the three-dimensional coordinates of the target aircraft and our aircraft are obtained based on a spatial rectangular coordinate system. The specific spatial rectangular coordinate system is not limited as long as it meets the tactical prediction.

[0115] In this embodiment, in actual use, the value of B is 90, that is, the length of each corrected time series feature data is 90.

[0116] In this embodiment, in actual use, the time series characteristic data in step 102 is characteristic data of a plurality of consecutive sampling moments; the sampling time is 0.025 seconds.

[0117] In this embodiment, in actual use, the tactical intent type corresponding to each time series feature data is a tactical intent type marked based on the knowledge annotation historical data of combat field experts, and is not specifically limited as long as it meets the actual tactical intent.

[0118] In this embodiment, in actual use, the BiLSTM module is a bidirectional long short-term memory network module, and the Attention module is an attention mechanism module.

[0119] In this embodiment, in actual use, when standardization is performed in step 204, the mean and standard deviation need to be recorded. Later, when the LSTM neural network model outputs the predicted trajectory, the mean and standard deviation need to be input for de-standardization to restore the trajectory data to its initial state.

[0120] In this embodiment, when actually used, the tactical intention type is {attack, cover, feint, retreat, reconnaissance, interference}. Among them, considering that in actual combat, the data that can usually be intercepted is trajectory posture data, and it is difficult to know the information of the enemy's switching electronic interference equipment, so the present invention predicts the tactical intention based on trajectory posture data. The "interference" tactical intention of the present invention specifically refers to maneuver interference, the "reconnaissance" tactical intention specifically refers to air combat aircraft to search for targets, and does not include the broad reconnaissance meaning range of other means, and "retreat" includes escape, return, (in exercises) end confrontation and other situations of exiting the battle.

[0121] In this embodiment, in actual use, the input feature length in the input layer is 2, the input sequence length is 2 / 3B, the hidden layer in the LSTM network layer has 128 neurons, the output feature length in the output layer is 2, and the output sequence length is 1 / 3B.

[0122] In this embodiment, in actual use, if there are multiple friendly aircraft, the relative distance is the minimum relative distance between the target aircraft and our aircraft.

[0123] In this embodiment, in actual use, the ADAM optimizer is used in step 4035, and the LSTM neural network model is iteratively optimized using the mean square error loss function until all the trajectory data training sets are trained and one iterative training is completed; it is repeated multiple times until the maximum number of training times of 200 is met, and a trained LSTM neural network model is obtained.

[0124] In this embodiment, in actual use, the ADAM optimizer is used in step 209, and the mean square error loss function is used to iteratively optimize the BiLSTM+Attention network model until the air combat intent threat data set is fully trained and one iterative training is completed; it is repeated multiple times until the maximum number of training times of 300 is met, and a trained BiLSTM+Attention network model is obtained.

[0125] In summary, the method of the present invention has simple steps and reasonable design. It combines the forward multiple continuous air combat feature data at a time and the backward target aircraft air combat trajectory prediction data as well as the battlefield threat assessment to perform intention prediction, thereby improving the reliability of air combat target intention prediction.

[0126] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent structural change made to the above embodiment based on the technical essence of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A BiLSTM+Attention three-layer air combat target intention prediction method based on multi-feature continuous time series, characterized in that: The method comprises the following steps: Step 1: Establish a basic data set for air combat intention prediction: Step 101, during the air combat confrontation process, the three-dimensional coordinates, speed, and heading characteristic data of the target aircraft and our aircraft are collected in a continuous time sequence to obtain air combat characteristic data; Step 102, according to the air combat feature data, obtaining the time series feature data corresponding to each tactical intention type; wherein the tactical intention types are attack, cover, feint, retreat, reconnaissance, and interference, and the labels of the tactical intention types are 1, 2, 3, 4, 5, and 6 respectively; 1 means attack, 2 means cover, 3 means feint, 4 means retreat, 5 means reconnaissance, and 6 means interference; Step 103, judging and correcting the length of the time series feature data corresponding to each tactical intention type to obtain each corrected time series feature data; wherein the length of each corrected time series feature data is the same; Step 104, recording each corrected time series feature data and its corresponding tactical intention type as an air combat intention basic data set; Step 2: Perform threat assessment on the air combat intention basic data set to obtain the air combat intention threat data set: Step 201: Obtain the three-dimensional coordinates of the target aircraft at the jth sampling moment from the single corrected time series feature data, and combine it with the three-dimensional coordinates of our aircraft at the jth sampling moment to obtain the relative distance d between our aircraft and the target aircraft at the jth sampling moment. j ; Wherein, j is a positive integer; Step 202: According to Get the angle threat index S at the jth sampling moment aj ; where φ j The velocity vector of our aircraft and the angle between our aircraft and the target aircraft at the jth sampling moment, q j represents the angle between the velocity vector of the target aircraft and the line connecting our aircraft and the target aircraft at the jth sampling moment; Step 203: According to Get the distance threat index S at the jth sampling moment dj ; where d j represents the relative distance between our aircraft and the target aircraft at the jth sampling moment; d l represents the missile range of the target aircraft, d d Indicates the maximum radar detection range of the target aircraft; Step 204: According to Get the speed threat index S at the jth sampling moment vj ; Among them, V mj represents the speed of our aircraft at the jth sampling moment, V wj represents the speed of the target aircraft at the jth sampling moment; Step 205: Calculate the angle threat index S at the jth sampling moment. aj , the distance threat index S at the jth sampling moment dj , the speed threat index S at the jth sampling moment vj Perform weighted processing to obtain the threat assessment value K of the target aircraft to our aircraft at the jth sampling moment j ; Step 206: According to the method of step 201 to step 205, the threat assessment value of the target aircraft to any friendly aircraft at the j-th sampling moment is obtained; wherein the threat assessment value of the target aircraft to the e-th friendly aircraft at the j-th sampling moment is recorded as K j (e); e is the serial number of our aircraft; Step 207, select the maximum value from the threat assessment values ​​of the target aircraft to any friendly aircraft at the j-th sampling moment, and record it as the maximum threat value; according to Get the total threat F of the target aircraft to our aircraft at the jth sampling moment z ; Where, e and E are both positive integers, and 1≤e≤E, and E represents the total number of our aircraft; Step 208, repeating steps 201 to 207 multiple times to obtain the maximum threat value and total threat of each time series feature data; Step 209: The three-dimensional coordinates, speed, heading, relative distance, maximum threat value and total threat of the target aircraft are taken as individual optimized time series feature data; and each optimized time series feature data and its corresponding tactical intent type are recorded as an air combat intent threat data set; Step 3: Build a BiLSTM+Attention network model and input the air combat intention threat dataset into the training: Step 301: Build a BiLSTM+Attention network model; Step 302: Input all the air combat intention threat data sets into the BiLSTM+Attention network model for training to obtain a trained BiLSTM+Attention network model; Step 4: Subsequent time series data is used to predict tactical intent based on trajectory prediction, threat assessment, and the trained BiLSTM+Attention network model: Step 401, in the subsequent air combat confrontation process, the three-dimensional coordinates, speed, and heading characteristic data of the target aircraft and our aircraft in a continuous time sequence are collected to obtain subsequent air combat characteristic data of the target aircraft and subsequent air combat characteristic data of our aircraft; Step 402: The three-dimensional coordinates and headings in the subsequent target aircraft air combat feature data are input into the trained LSTM neural network model through standardization, and the subsequent air combat trajectory prediction standardized data that is continuous with the subsequent target aircraft air combat feature data is output; and the subsequent air combat trajectory prediction standardized data is de-standardized to obtain the subsequent target aircraft air combat trajectory prediction data; Based on the three-dimensional coordinates in the subsequent target aircraft air combat trajectory prediction data, the subsequent speed prediction data is obtained; Step 403, based on the subsequent target aircraft air combat characteristic data and the subsequent our aircraft air combat characteristic data, according to the method of step 2, obtain the subsequent maximum threat value and the subsequent total threat; Based on the subsequent target aircraft air combat trajectory prediction data, the subsequent speed prediction data and the air combat feature data of the subsequent our aircraft air combat feature data at the last sampling moment, the predicted maximum threat value and the predicted total threat are obtained according to the method of step 2; Step 404: record the three-dimensional coordinates, speed, heading, relative distance, subsequent maximum threat value and subsequent total threat in the subsequent target aircraft air combat feature data as subsequent feature data; record the three-dimensional coordinate predicted value, speed predicted value, heading predicted value, relative distance predicted value, predicted maximum threat value and predicted total threat in the subsequent target aircraft air combat feature data as subsequent predicted feature data; and splice the subsequent feature data and the subsequent predicted feature data to form fusion measurement and prediction time series feature data; Step 405: Input the fused measurement and prediction time series feature data into the trained BiLSTM+Attention network model to perform tactical intent prediction and obtain the output probability of each tactical intent; Step 406: Select the tactical intention corresponding to the maximum probability from the output probabilities of the various tactical intentions as the predicted tactical intention.

2. A BiLSTM+Attention three-layer air combat target intention prediction method based on multi-feature continuous time series according to claim 1, characterized in that: Step 103, the specific process is as follows: Step 1031, obtaining the number of samples of the time series feature data corresponding to the i-th tactical intention type, and recording the number of samples of the time series feature data corresponding to the i-th tactical intention type as the length of the i-th time series feature data; wherein i is a positive integer; Step 1032: determine the length of each time series feature data. If the lengths of each time series feature data are different, obtain the mode of the lengths of each time series feature data and record it as B. Step 1033: record the length of the i-th time series feature data as Ai. If Ai is less than B, use the cubic spline interpolation method to interpolate the i-th time series feature data until the length of the i-th time series feature data after interpolation is B, thereby obtaining the corrected i-th time series feature data; If Ai is greater than B and less than 2B, the part greater than B is deleted until the length of the i-th time series feature data after deletion is B, and the corrected i-th time series feature data is obtained; If Ai is greater than αB, the part greater than αB is deleted, and the i-th time series feature data is divided into α corrected i-th time series feature data; wherein α represents a multiple, and α is a positive integer greater than 2; Step 1034, repeat step 1033 multiple times to complete the correction of each time series feature data, and obtain each corrected time series feature data.

3. A BiLSTM+Attention three-layer air combat target intention prediction method based on multi-feature continuous time series according to claim 2, characterized in that: In step 403, the trained LSTM neural network model is obtained. The specific process is as follows: Step 4031: Obtain 2 / 3B of the time series feature data from each corrected time series feature data as input data, and use the remaining 1 / 3B of the time series feature data as output label data; wherein the input data and the output label data include three-dimensional coordinates and heading; Step 4032: standardize the input data and the output label data using a Z-Score standardization method to obtain input standardized data and output label standardized data; Step 4033, repeating steps 4031 and 4032 multiple times to obtain multiple sets of input standardized data and their corresponding output label standardized data, and recording them as trajectory data training sets; Step 4034: construct an LSTM neural network model; wherein the LSTM neural network model includes an input layer, an LSTM network layer, a fully connected layer and an output layer; Step 4035: input the trajectory data training set into the constructed LSTM neural network model for training to obtain a trained LSTM neural network model.

4. A multi-feature continuous time series BiLSTM+Attention three-layer air combat target intention prediction method according to claim 1, characterized in that: Step 205, the specific process is as follows: Step 2051: normalize the angle threat index at each sampling moment to obtain the normalized value S of the angle threat index at the jth sampling moment. a ' j ; Step 2052: According to K j =k1S a ' j +k2S dj +k3S vj , get the threat assessment value K of the target aircraft to our aircraft at the jth sampling moment j ; Among them, k1 represents the first weight coefficient, and k1 takes the value of 0.35, k2 represents the second weight coefficient, and k2 takes the value of 0.35, k3 represents the third weight coefficient, and k3 takes the value of 0.

3.

5. A BiLSTM+Attention three-layer air combat target intention prediction method based on multi-feature continuous time series according to claim 1, characterized in that: The BiLSTM+Attention network model in step 301 includes a BiLSTM module, an Attention module, a fully connected layer and a Softmax layer; the output of the Softmax layer is 6; In step 302, the optimized time series feature data is used as the input of the BiLSTM+Attention network model, and the optimized time series feature data and its corresponding tactical intent type are used as the output of the BiLSTM+Attention network model; wherein, if the optimized time series feature data belongs to its corresponding tactical intent type, the tactical intent type takes a value of 1, otherwise it takes a value of zero.

6. A multi-feature continuous time series BiLSTM+Attention three-layer air combat target intention prediction method according to claim 1, characterized in that: In step 403, the subsequent speed prediction data is obtained based on the three-dimensional coordinates in the subsequent target aircraft air combat trajectory prediction data, as follows: The subsequent three-dimensional coordinate data is obtained from the subsequent target aircraft air combat trajectory prediction data, and the subsequent speed prediction data is obtained according to the ratio of the relative distance of the three-dimensional coordinates at two adjacent sampling moments to the sampling time.

Images