A multi-level compound intention prediction method for fighter aircraft based on a knowledge inference engine

Abstract

A kind of fighter multi-level composite intention prediction method based on knowledge inference machine, to solve the technical problems of low precision and single dimension of existing technology for fighter operational intention prediction. The implementation steps are as follows: first, collect multi-source sensor data, preprocess sample data;Then fuse sample data, and extract entities, relationships and attributes in data information, construct data structured knowledge including threat elements, advantage elements and attack index;Use expert knowledge base to obtain enemy aircraft maneuvering action type, and comprehensively consider battlefield situation index to carry out knowledge reasoning, and obtain multi-level composite intention prediction results including single aircraft intention, coordination intention and operational task intention. The present application fully combines the knowledge of fighter identity information, position information and behavior information, adopts multi-source data fusion and multi-level intention composite correlation modeling method, and improves the accuracy of fighter intention prediction.

Description

Technical Field

[0001] The present invention belongs to the field of aviation technology, and in particular relates to a method for predicting multi-level composite intentions of fighter jets based on a knowledge inference engine. Background Art

[0002] Predicting the combat intent of incoming fighter jets is essential in air combat. In complex and ever-changing offensive and defensive environments, comprehensively considering the enemy's combat mission, combat methods, and force deployment, identifying the enemy's behavioral model and combat intent can effectively alleviate the operational pressure on commanders, enabling our side to achieve timely and accurate assessments of the battlefield situation, seize the initiative, and achieve rapid victory.

[0003] Intent prediction refers to the process of identifying, judging, or interpreting the enemy's operational assumptions, intentions, and plans through comprehensive analysis of information perceived by various battlefield sensors. At the decision-making level, multi-source information fusion utilizes established models and related reasoning methods to conduct uncertainty reasoning on the information from various sensors. Intent prediction is based on detected battlefield environmental information and enemy target trajectory data, and is influenced by numerous factors.

[0004] Currently, existing research has used rule-based methods, statistical theory-based methods, and machine learning classification-based methods to conduct intent prediction research. Due to the complexity of intent expression and the hidden nature of intent, existing intent prediction methods have certain applicability and limitations. Among them, the rule-based method constructs an intention recognition expert system through a large amount of data collection and research to achieve intent prediction with limited rules. However, the rule structure of this method is too thin and it is difficult to predict complex intents. The intention prediction method based on statistical theory establishes an intent probability distribution model and estimates the model parameters to achieve intent prediction. However, this method requires a large amount of data statistics in the early stage, over-relies on prior probabilities, and has large deviations. The machine learning classification prediction method relies on feature engineering to establish the association between manually extracted features and intent labels. Due to the limited expressive power of machine learning, the recognition accuracy is relatively low in the process of describing intentions with certain hidden features. Traditional artificial neural network methods have strong representation capabilities and can often utilize richer features when solving complex nonlinear problems. However, this type of method relies on data-driven and lacks reasoning ability, making it difficult to effectively solve the problem of complex intent prediction. In addition, the above methods have limitations such as small target intention space, lack of hierarchy in target intention prediction, lack of correction of prediction results, and low engineering implementation and timeliness. Summary of the Invention

[0005] To address the problems of existing fighter jet intention prediction, which suffers from a single dimension and low accuracy, this paper proposes a multi-level, composite fighter jet intention prediction method based on a knowledge inference engine. Sensor detection data is parsed, processed, and calculated to obtain accurate and effective perception data. Threat factors, advantage factors, and attack indices extracted from knowledge are used to obtain enemy aircraft maneuver types and comprehensive battlefield situation indicators from an expert knowledge base. Finally, a multi-level intention inference unit is used to predict composite intentions at multiple levels, including individual aircraft intentions, collaborative intentions, and combat mission intentions. This method effectively enhances the sensitivity of situational awareness and significantly improves the accuracy of fighter aircraft combat intentions.

[0006] A method for predicting multi-level composite intentions of fighter jets based on a knowledge inference engine comprises the following steps:

[0007] (1) Detection sample data collection

[0008] The multi-sensor detection data of both the enemy and friendly forces within a given time period in a 2V2 simulation environment are collected as a sample data set, including target position, speed information, and relative direction information generated by radar, electronic warfare, and data link detection methods.

[0009] (2) Sample data preprocessing

[0010] Preprocess the data collected in step (1). Use the observed target azimuth, distance, and altitude to group and associate the detected targets. Combined with the changes in the target's motion trajectory, use the Gaussian probability function and the continuity probability function to identify and remove false targets and retain valid data.

[0011] (3) Effective data fusion

[0012] The valid data obtained in step (2) are fused. First, homogeneous sensor data trace fusion is performed, that is, the two points with the smallest distance in the two target groups are selected, and the mean of their spatial position and velocity is used as the fusion result. When performing trace association, the corresponding targets with the smallest distance are regarded as the same targets, and the time series traces are fused using the mean filtering method. When performing heterogeneous sensor trace fusion, the radar, data link, and electronic warfare are assigned coefficient weights ranging from 90% to 10% to obtain the fusion results of multi-sensor target position, velocity and other information.

[0013] (4) Action recognition based on maneuver rules based on expert knowledge base

[0014] The target signal fusion result obtained in step (3) is processed. The aircraft maneuvering action space is constructed, including climb, left climb, right climb, left turn, right turn, dive, left dive, right dive, half roll, level flight deceleration, constant speed level flight, level flight acceleration, upward somersault, and downward somersault. Multi-dimensional action component identification is performed from three dimensions: up and down direction, left and right direction, and acceleration direction, to realize the recognition of the action in the action space. By distinguishing multiple action dimensions, the action recognition is realized.

[0015] (5) Knowledge extraction: Threat assessment index

[0016] The target signal fusion results obtained in step (3) are processed to calculate the threat assessment index, including the angle threat index, speed threat index, distance threat index, altitude threat index, and air combat effectiveness index. The average of the five threat indices is used as the comprehensive threat assessment index to form knowledge with clear entity attribution and relationship attributes.

[0017] (6) Standalone Intention Reasoning Unit

[0018] Based on the target's perception information from radar, data link, and electronic warfare multi-sensors and the results obtained in steps (4) and (5), the target's identity information, location information, behavior information, capability information, relationship information, and target movement knowledge are extracted. The knowledge source is obtained by associating, classifying, expanding attributes, and deeply describing discrete intelligence. A single-machine intention inference unit is constructed by introducing a conditional variational autoencoder network (CVAE) into a long short-term memory network (LSTM). The extracted knowledge source is used to perform intention inference in a single-machine intention space consisting of attack, preparation, defense, and exit. According to different threat levels, a four-level intention warning is given from high to low AD, and the intention label is output.

[0019] (7) Knowledge extraction: advantage index

[0020] The target signal fusion results obtained in step (3) are processed to calculate the advantage index of our aircraft over the enemy aircraft in five dimensions, namely, the angle advantage index, speed advantage index, altitude advantage index, distance advantage index, and energy advantage index. Combined with the near and far distance situations, the advantage evaluation comprehensive index is calculated to participate in the collaborative intention prediction.

[0021] (8) Collaborative Intention Reasoning Unit

[0022] Based on the results obtained in steps (5) and (7), the threat assessment and advantage assessment parameters are normalized to obtain the overall situation. A cascaded CVAE-LSTM collaborative intention reasoning unit is constructed to perform collaborative intention reasoning on the single-machine intention and target knowledge spectrum such as identity information, location information, behavior information, capability information, relationship information, and target movement in step (6). The role of the single machine in the collaborative operation (main attack machine or auxiliary attack machine) is divided, and reasoning is performed in the collaborative intention space of {attack, decoy, prepare attack} and {auxiliary attack, defense, alert, prepare}. Different collaborative intention output results are given according to different situation assessment levels, and the collaborative intention probability of the two enemy aircraft or the target reminder is output.

[0023] (9) Task Intention Reasoning Unit

[0024] Construct a combat scenario library, such as battlefield natural environment, construction fortifications, military equipment, etc., and obtain key military elements in the combat scenario according to the military importance level classification rules. Based on the target signal fusion results obtained in step (3), calculate the enemy aircraft's attack index on our key military elements in three dimensions: weapon attack index, distance attack index, and angle attack index. Based on the coordination intention probability of step (8), calculate the main attack aircraft's comprehensive attack index on different key military elements. Combine the target's identity information, location information, behavior information, capability information, relationship information, and target movement to associate multiple CVAE-LSTM coordination intention reasoning units to obtain a mission intention reasoning unit, and output the fighter's intention label in the mission intention space composed of the name and location of the key military elements.

[0025] Beneficial Effects: This invention fully leverages the fighter's maneuverability and air situational awareness capabilities. Based on the input of multi-source sensor information from the fighter, it analyzes, processes, and calculates the fighter's combat intent, fully reflecting the relationship between the fighter's maneuverability and combat intent. Simulation results demonstrate that this invention can accurately predict the fighter's multi-layered, complex intent. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is an architectural diagram of the present invention.

[0027] Figure 2 This is a flowchart of the multi-level intention prediction method of the present invention.

[0028] Figure 3 This is the architecture diagram of the multi-level intention inference engine of the present invention.

[0029] Figure 4 It is a probability distribution diagram of the intention prediction results of two enemy aircraft at a certain moment in a 2V2 air combat scenario of the present invention. DETAILED DESCRIPTION

[0030] The present invention is described in detail below with reference to the accompanying drawings and combat data of a 2V2 scenario in a certain air combat simulation platform.

[0031] It should be noted that when the present invention is applied to different simulation systems or real-world environments, only the data source will differ; even if the simulation algorithms differ, the model training and prediction methods remain consistent. Furthermore, the described embodiments are intended only to facilitate understanding of the present invention and are not intended to limit any practical applications.

[0032] A multi-level composite intention prediction method for fighter jets based on a knowledge inference engine is implemented through the following steps:

[0033] (1) Detection sample data collection

[0034] The multi-sensor detection data of both the enemy and friendly forces within a given time period in a 2V2 simulation environment are collected as a sample data set, including target position, speed information, and relative direction information generated by radar, electronic warfare, and data link detection methods.

[0035] (2) Sample data preprocessing

[0036] Preprocess the data collected in step (1). Use the observed target azimuth, distance, and altitude to group and associate the detected targets. Combined with the changes in the target's motion trajectory, use the Gaussian probability function and the continuity probability function to identify and remove false targets and retain valid data.

[0037] (3) Effective data fusion

[0038] The valid data obtained in step (2) are fused. First, homogeneous sensor data trace fusion is performed, that is, the two points with the smallest distance in the two target groups are selected, and the mean of their spatial position and velocity is used as the fusion result. When performing trace association, the corresponding targets with the smallest distance are regarded as the same targets, and the time series traces are fused using the mean filtering method. When performing heterogeneous sensor trace fusion, the radar, data link, and electronic warfare are assigned coefficient weights ranging from 90% to 10% to obtain the fusion results of multi-sensor target position, velocity and other information.

[0039] (4) Action recognition based on maneuver rules based on expert knowledge base

[0040] Process the target signal fusion results obtained in step (3). Construct the aircraft maneuvering action space, including climb, left climb, right climb, left turn, right turn, dive, left dive, right dive, half roll, level flight deceleration, constant speed level flight, level flight acceleration, upward somersault, and downward somersault. Use the three dimensions of up and down direction, left and right direction, and acceleration direction to describe the action, namely, north velocity dx, celestial velocity dy, east velocity dz, north acceleration ddx, celestial acceleration ddy, and east acceleration ddz. Among them, dy>0 means flying upward and dy<0 means flying downward; the following conditions are met to judge flying to the left (when (dx>0,dz>0), |dz| becomes larger and |dx| becomes smaller; when (dx>0,dz<0), |dx| becomes larger and |dz| becomes smaller; when (dx<0,dz>0), |dx| becomes larger and |dz| becomes smaller; when (dx>0,dz>0), |dz| becomes larger and |dx| becomes smaller). The following conditions are met to determine rightward flight: (When (dx>0, dz>0), |dz| decreases and |dx| increases; when (dx>0, dz<0), |dx| decreases and |dz| increases; when (dx<0, dz>0), |dx| decreases and |dz| increases; when (dx>0, dz>0), |dz| decreases and |dx| increases). When ddy>0, it is upward acceleration; when ddy<0, it is downward acceleration; when ddx is not equal to 0 or ddz is not equal to 0, it is level flight acceleration. By distinguishing multiple motion dimensions, action recognition is achieved.

[0041] (5) Knowledge extraction: Threat assessment index

[0042] The target signal fusion results obtained in step (3) are processed to calculate the threat assessment index, including the angle threat index, speed threat index, distance threat index, altitude threat index, and air combat effectiveness index.

[0043] (5a) The angle threat index is calculated as follows: q R is the target lead angle, q B is the angle between the target heading and the target line (right deviation is positive), q R ∈(-180°,180°],q B ∈(-180°,180°].

[0044] (5b) The speed threat index calculation formula is: In the formula Indicates the enemy aircraft speed. Indicates the speed of my machine.

[0045] (5c) The distance threat index calculation formula is: Maximum missile launch range fR max , radar air-to-air range AE max, attack zone range fR tr , radar detection target distance fT AE .

[0046] (5d) Calculation formula for high threat index Among them, dy (meters) is our altitude, and du (meters) is the enemy altitude.

[0047] (5e) The air combat effectiveness index calculation formula is STC = [lnMane + ln(∑Fir+1) + ln(∑Dete)]Cont·Viab·Voya·Elec, where STC represents the aircraft's air combat effectiveness index, Mane represents the aircraft's maneuverability, Fir represents the aircraft's firepower parameter, Dete represents the aircraft's detection capability, Cont represents its operational performance parameter, Viab represents the aircraft's survivability coefficient, Voya represents the aircraft's range coefficient, and Elec represents the aircraft's electronic countermeasure capability parameter.

[0048] (5f) Comprehensive threat assessment index, W = (angle index +dis index +V index +h index +STC) / 5, which is the average of the five threat indices and participates in the collaborative intention prediction.

[0049] (6) Standalone Intention Reasoning Unit

[0050] Based on the target's perception information from radar, data link, and electronic warfare multi-sensors and the results obtained in steps (4) and (5), the target's identity information, location information, behavior information, capability information, relationship information, target movement, and other knowledge are extracted. The knowledge source is obtained by associating, classifying, expanding attributes, and deeply describing discrete intelligence. A single-machine intention reasoning unit is constructed by introducing a conditional single variational autoencoder network (CVAE) in a long short-term memory network (LSTM). The extracted knowledge source is used to perform intention reasoning in a single-machine intention space consisting of attack, preparation, defense, and exit. According to the different threat levels, a total of 4 levels of intention warnings are given, ranging from high to low AD, and the intention label is output. When the enemy aircraft has launched a missile, is tracking, approaching, or the comprehensive threat assessment index is [0.85, 1], an A-level (highest) intention warning is output; when the enemy aircraft is searching, or the comprehensive threat assessment index is [0.5, 0.85), a B-level (high) intention warning is output; when the comprehensive threat assessment index is [0.2, 0.5), a C-level (medium) intention warning is output; when the comprehensive threat assessment index is [0, 0.2), a D-level (low) intention warning is output.

[0051] (7) Knowledge extraction: advantage index

[0052] The target signal fusion result obtained in step (3) is processed to calculate the advantage index of our aircraft over the enemy aircraft in five dimensions, namely, angle advantage index, speed advantage index, height advantage index, distance advantage index, and energy advantage index.

[0053] (7a) The formula for calculating the angular advantage index is: When q ji =0°, When S a =1, our aircraft is located directly behind the enemy aircraft, our aircraft is in the best attack position, and the enemy aircraft has the greatest angle advantage over our aircraft.

[0054] (7b) The speed advantage index is calculated as follows: Where v is the carrier speed, v T The speed of the enemy aircraft.

[0055] (7c) The formula for calculating the height advantage index is: Where h is the height of the aircraft, h t The altitude of the enemy aircraft.

[0056] (7d) The distance advantage index is calculated as follows:

[0057] (7e) The energy advantage index is calculated as follows: in , S eb 、S er are the unit energy values ​​of the host and target respectively; H b 、V b are the altitude and speed of the aircraft respectively, H r 、V r are the height and speed of the target respectively, and g is the gravitational acceleration constant.

[0058] (7f) Beyond visual range conditions, That is, the comprehensive advantage evaluation index is the weighted sum of each advantage index. Under the condition of near vision distance, the comprehensive advantage evaluation index S=c1*S a *S d +c2*S e .

[0059] (8) Collaborative Intention Reasoning Unit

[0060] Based on the results obtained in steps (5) and (7), the threat assessment and advantage assessment parameters are normalized to obtain the overall situation, and the situation assessment index is T = αW + βS. A cascaded CVAE-LSTM collaborative intention reasoning unit is constructed to perform collaborative intention reasoning on the single-machine intention and target knowledge spectrum such as identity information, location information, behavior information, capability information, relationship information, and target movement in step (6). The role of the single machine in the collaborative operation (main attack machine or auxiliary attack machine) is divided, and reasoning is performed in the collaborative intention space of {attack, decoy, preparatory attack} and {auxiliary attack, defense, alert, preparatory}. Different collaborative intention output results are given according to different situation assessment levels, and the collaborative intention probability of the two enemy aircraft or the target reminder is output. That is, when the situation assessment index is greater than 0.5, the collaborative intention probability of the two enemy aircraft is output; otherwise, the target reminder is output.

[0061] (9) Task Intention Reasoning Unit

[0062] Construct a combat scenario library, such as the battlefield natural environment, construction fortifications, military equipment, etc., and obtain the key military elements in the combat scenario according to the military importance level classification rules. Based on the target signal fusion results obtained in step (3), calculate the three-dimensional attack index of the enemy aircraft on our key military elements, namely the weapon attack index, the distance attack index, and the angle attack index, as follows:

[0063] (9a) Weapon attack index is Where D is the distance between the attacked key military element and the enemy aircraft, R ob is the center of the far and near boundaries of the attacked area, σ Db It is an adjustment item related to the far and near boundaries of the attacked area;

[0064] (9b) The distance attack index is in R max and R min are the far and near boundaries of the enemy weapon attack area, σ r =0.6(R max -R min ) is the distance adjustment item;

[0065] (9c) The angle of attack index is

[0066] (9d) The comprehensive attack index is F i =F at ×F d ×F q , where i represents different key military elements, and the comprehensive attack index F i The larger it is, the stronger the enemy aircraft's intention to attack key military element i;

[0067] The intention prediction of the enemy aircraft's combat mission is based on the collaborative intention probability of step (8), and the comprehensive attack index of the main attack aircraft on different key military elements is calculated. The multiple CVAE-LSTM collaborative intention reasoning units are associated with the target's identity information, location information, behavior information, capability information, relationship information, target movement and other knowledge to obtain the mission intention reasoning unit, and output the fighter's intention label in the mission intention space composed of the name and position of the key military elements.

Claims

1. A multi-level composite intention prediction method for fighter jets based on a knowledge inference engine, characterized in that: The steps include: (1) Detection sample data collection Collect multi-sensor detection data from both sides of the enemy within a given time period in a 2V2 simulation environment as a sample data set, including target position, speed, and relative direction information generated by radar, electronic warfare, and data link detection methods; (2) Sample data preprocessing Preprocess the data collected in step (1); The detected targets are grouped and associated using the observed target azimuth, distance, and altitude. Combined with the target trajectory changes, Gaussian probability function and continuity probability function are used to identify and remove false targets and retain valid data. (3) Effective data fusion The valid data obtained in step (2) are fused; the homogeneous sensor data traces are fused, the two points with the smallest distance in the two target groups are selected, and the mean of their spatial position and velocity is used as the fusion result. When performing trace association, the corresponding target with the smallest distance is regarded as the same target, and the time series traces are fused using the mean filtering method; When performing heterogeneous sensor point fusion, the radar, data link, and electronic warfare are assigned coefficient weights ranging from 90% to 10% to obtain the fusion results of multi-sensor target position and speed information; (4) Action recognition based on maneuver rules based on expert knowledge base Process the target signal fusion result obtained in step (3); construct the aircraft maneuvering action space, including climbing, left climbing, right climbing, left turn, right turn, diving, left diving, right diving, half roll, level flight deceleration, constant speed level flight, level flight acceleration, upward oblique somersault, downward oblique somersault; perform multi-dimensional action component identification from three dimensions of up and down direction, left and right direction, and acceleration direction to realize the recognition of the action in the action space, and realize the recognition of the action by distinguishing multiple action dimensions; (5) Knowledge extraction: threat assessment index The target signal fusion results obtained in step (3) are processed to calculate the threat assessment index, including the angle threat index, speed threat index, distance threat index, altitude threat index, and air combat effectiveness index; the average of the five threat indices is used as the comprehensive threat assessment index to form knowledge with clear entity attribution and relationship attributes; (6) Standalone Intention Reasoning Unit Based on the target's perception information from radar, data link, and electronic warfare multi-sensors and the results obtained in steps (4) and (5), the target's identity information, location information, behavior information, capability information, relationship information, and target movement knowledge are extracted, and discrete intelligence is associated, graded, attribute expanded, and deeply described to obtain a knowledge source; a conditional single variational autoencoder network (CVAE) is introduced into the long short-term memory network (LSTM) to construct a single-machine intention reasoning unit, and the extracted knowledge source is used to perform intention reasoning in the single-machine intention space composed of attack, preparation, defense, and exit. According to different threat levels, a total of 4 levels of intention warnings are given from high to low AD, and the intention label is output; (7) Knowledge extraction: advantage index Process the target signal fusion results obtained in step (3) to calculate the advantage index of our aircraft over the enemy aircraft in five dimensions, namely, angle advantage index, speed advantage index, altitude advantage index, distance advantage index, and energy advantage index; combine the close and long distance situations to calculate the comprehensive advantage evaluation index and participate in the collaborative intention prediction; (8) Collaborative Intention Reasoning Unit Based on the results obtained in step (5) and step (7), the evaluation parameters of threat assessment and advantage assessment are normalized to obtain the overall situation; a cascaded CVAE-LSTM collaborative intention reasoning unit is constructed to perform collaborative intention reasoning on the single-machine intention and identity information, location information, behavior information, capability information, relationship information, and target movement in step (6), divide the role of the single machine in the collaborative operation, and reason in the collaborative intention space of {attack, lure, prepare for attack} and {auxiliary attack, defense, alert, prepare}. According to different situation assessment levels, different collaborative intention output results are given, and the collaborative intention probability of the two enemy aircraft or the target reminder is output; (9) Task intention reasoning unit Construct a combat scenario library and obtain the key military elements in the combat scenario according to the military importance level classification rules; process the target signal fusion results obtained in step (3) and calculate the three-dimensional attack index of the enemy aircraft on our key military elements, namely the weapon attack index, the distance attack index and the angle attack index; calculate the comprehensive attack index of the main attack aircraft on different key military elements based on the coordination intention probability of step (8), and associate multiple CVAE-LSTM coordination intention reasoning units with the target's identity information, location information, behavior information, capability information, relationship information and target movement to obtain a mission intention reasoning unit, and output the intention label of the fighter in the mission intention space composed of the name and location of the key military elements.

Images