A human intelligence assessment method based on Bayesian deep learning

Abstract

The present invention provides a human intelligence assessment method based on Bayesian deep learning, which belongs to the technical field of UAV intention recognition; it solves the technical problem that the intention of the enemy UAV cannot be accurately judged in an uncertain environment. Its technical solution is: establish a Bayesian deep learning network suitable for intention recognition and collect situation information of the enemy UAV; use the situation information of the enemy UAV as input, and use the established Bayesian deep learning network and human intelligence to independently judge the intention of the enemy UAV; integrate the judgment results of machine intelligence through the reasoning mechanism, so as to infer and verify the judgment results of human intelligence. The beneficial effects of the present invention are: in a complex and uncertain environment, combining the computational advantages of machine intelligence and the empirical judgment of human intelligence, providing more accurate enemy UAV intention recognition, and enhancing the decision-making ability and tactical advantage of our UAV on the battlefield.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle intention prediction, and in particular to a human intelligence evaluation method based on Bayesian deep learning. Background Art

[0002] In recent years, with the rapid development of artificial intelligence technology, deep learning has shown great application potential in many fields. Especially in complex decision-making and prediction tasks, deep learning algorithms have gradually become mainstream technologies due to their powerful feature extraction and pattern recognition capabilities. However, conventional deep learning algorithms, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have significant deficiencies in dealing with uncertainty and model interpretability. This has brought application challenges in some key areas, such as medical diagnosis, financial risk assessment, and military decision-making.

[0003] To make up for this shortcoming, Bayesian deep learning networks, as an algorithm that integrates Bayesian statistical methods and deep learning technology, have gradually attracted widespread attention. Bayesian deep learning networks can not only quantify uncertainty in the prediction process, but also perform effective model training through methods such as variational inference. These characteristics make Bayesian deep learning networks have obvious advantages in scenarios that require high reliability and interpretability.

[0004] In military operations, human intelligence assessment and decision-making play a vital role. For example, in a battlefield environment, commanders need to accurately assess complex and ever-changing tactical situations and make decisions quickly. In this case, Bayesian deep learning algorithms can combine a large amount of battlefield data to provide high-confidence predictions of enemy behavior through modeling and reasoning.

[0005] The application of traditional deep learning models in tactical assessment often relies on a large amount of historical data and fixed patterns. This method is powerless when faced with dynamic and complex battlefield environments. In addition, the prediction results of conventional models lack uncertainty assessment, which may lead to overconfident decisions. For example, in the article "Bayesian Deep Learning for Robust Human Motion Prediction", researchers used Bayesian neural networks to predict human motion and successfully achieved high-precision and high-reliability predictions in complex environments. This successful application has inspired the military field to use similar Bayesian deep learning methods to conduct real-time evaluation and prediction of battlefield environments, thereby assisting decision-making. Summary of the invention

[0006] The purpose of the present invention is to propose a human intelligence assessment algorithm based on Bayesian deep learning, which is used to evaluate human intelligence and assist decision-making. This method can not only process complex and uncertain battlefield data, but also provide more reliable decision-making support by quantifying and predicting uncertainty. The present invention can effectively improve the accuracy and reliability of battlefield situation perception and decision-making, thereby reducing the battle loss ratio and improving combat effectiveness in high-risk environments.

[0007] The inventive concept of the present invention is: first, the combat intention of the drone is modeled, and then the battlefield situation is divided into four categories through the situation information of the drones of both sides. According to the relative azimuth and distance, the most likely intention of the enemy drone in each area is inferred, and then the intention of the enemy drone is predicted using the established Bayesian deep learning network, and the uncertainty of each intention is obtained. By analyzing the uncertainty of different intentions, the human intelligent recognition process is evaluated; the present invention gives the drone the ability to predict the combat intention of the enemy drone and gives the uncertainty of the prediction result.

[0008] The present invention is achieved by the following measures: a method for evaluating human intelligence based on Bayesian deep learning, comprising the following steps:

[0009] S1. Use the sensor system of our drone to obtain the situation information of the enemy drone, and then organize the data according to the situation information of our drone as the training set data.

[0010] S2. Human intelligence is used to independently judge the intentions of enemy drones. Rules are designed to divide the battlefield space into several areas, and the possible intentions of enemy drones in each area are inferred based on experience and the current battlefield situation.

[0011] S3. Establish a Bayesian deep learning network for judging the intention of enemy drones and save the trained network model and parameters.

[0012] S4. Comprehensively analyze the judgment results of the Bayesian deep learning network and the judgment results of human intelligence, infer the judgment process of human intelligence, further understand the decision-making logic of human intelligence, and verify the prediction results of the Bayesian deep learning network. Finally, by integrating the advantages of the two methods, the most credible combat intentions of enemy drones are obtained, providing reliable intelligence support for decision makers.

[0013] Furthermore, the step 1 comprises the following steps:

[0014] 1-1) Collect the current status information of the enemy UAV: ​​location information, flight speed, heading angle, roll angle, pitch angle, control amount, etc., in order to predict the future behavior of the enemy UAV.

[0015] 1-2) Calculate the position information difference of the red drone relative to the blue drone, and get the velocity components of the velocity vector in the x, y, and z directions. Organize the information into Slice X1, select a specific number of data samples of four labels, and splice them into a new data set X2. Then randomly divide the data set into training and test sets, and shuffle the data during model training.

[0016] Furthermore, the step 2 comprises the following steps:

[0017] 2-1) Define azimuth parameters: determine the azimuth θ1 of our combat aircraft based on the enemy combat aircraft and determine the azimuth θ2 of the enemy combat aircraft based on our combat aircraft; define parameter d: the distance between the enemy UAV and our UAV.

[0018] 2-2) Divide the battlefield space into several areas according to the azimuth angle θ and the distance d, and set the possible intentions of the enemy drone in each area.

[0019] Balance of power between the two sides:

[0020]

[0021] When θ1≤θ2, d>1000m, the enemy's intention is inferred to be reconnaissance; d<1000m, θ2 is subdivided, θ2∈(0,45°), distance d<300m, it is inferred to be attack intention, when the distance d is between 300-1000m, it is inferred to be penetration intention. θ2∈(45°,90°), d<300m, it is inferred to be penetration intention, d>300m, it is inferred to be retreat intention.

[0022] When θ1≥θ2, d>1000m, it is inferred as reconnaissance intention, and d<1000m, it is inferred as attack intention.

[0023] Our advantages:

[0024]

[0025] θ1<θ2, the enemy is at a disadvantage. When θ2∈(90°, 135°), d>1000m, it is inferred to be a reconnaissance intention, and d<1000m, it is a penetration intention. When θ2∈(135°, 180°), it is mainly based on the flight direction of the enemy UAV, and it is inferred to be a retreat intention.

[0026] Enemy Advantages:

[0027]

[0028] θ1>θ2, we are at a disadvantage. When θ1∈(90°,180°), θ2∈(0°,90°), d>1000m, it is inferred to be a reconnaissance intention, and d<1000m is an attack intention.

[0029] Both parties are neutral:

[0030]

[0031] When θ1<θ2, θ2∈(135°, 180°), it is inferred as reconnaissance intention; when θ2∈(90°, 135°), d>1000m, the enemy's intention is inferred as reconnaissance, and when d<1000m, it is inferred as attack.

[0032] When θ1>θ2, θ2∈(135°, 180°), the enemy's intention is inferred to be reconnaissance; when θ2∈(90°, 135°), d>1000m, the enemy's intention is inferred to be reconnaissance, and when d<1000m, it is inferred to be attack.

[0033] Furthermore, the step three comprises the following steps:

[0034] 3-1) Establish a Bayesian linear layer, which defines the weights and biases of the nodes as random variables sampled from a normal distribution. The mean matrix shape of the weights in the Bayesian linear layer is [out_features, in_features], and the variance matrix shape is [in_features, out_features]. The mean and variance of the bias are both one-dimensional matrices of [out_features].

[0035] 3-2) Construct a deep learning network including traditional linear layers and Bayesian linear layers. The first Bayesian linear layer has a feature dimension of [input_dim,360]; the second Bayesian linear layer has a feature dimension of [360,128], the first traditional linear layer has a feature dimension of [input_dim,200], the second traditional linear layer has a feature dimension of [200,128], and the third Bayesian linear layer has a feature dimension of [128,output_dim]. In the forward propagation method, the Bayesian deep learning network calculates the logarithmic prior distribution, logarithmic posterior distribution, logarithmic likelihood, and KL divergence composed of the difference between the prior distribution and the posterior distribution. Define the evidence lower bound ELBO (Evidence Lower Bound) as the loss function, which is composed of the weighted sum of the logarithmic likelihood and the KL divergence, and calculate the loss by forward propagation multiple times through Monte Carlo sampling.

[0036] 3-3) Use the training set data to train the Bayesian deep learning network and save the trained network model and parameters.

[0037] Furthermore, the step 4 comprises the following steps:

[0038] 4-1) Obtain the enemy drone intention judgment result P through the Bayesian deep learning network M , the intention judgment result Y of human intelligence is obtained through the intention rules of the drone H ;

[0039] 4-2) Integrate the two judgment results, infer the logic of human intelligence to judge the enemy drone, and obtain the final intention of the enemy drone;

[0040] 4-3) Feedback the final intention judgment results to our drones to improve battlefield decision-making capabilities.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] (1) The present invention has the ability to process complex and uncertain data. The Bayesian deep learning network can explain the uncertainty in the prediction, making the decision more reliable. By modeling the distribution of neural network weights and biases, it can cope with complex and uncertain battlefield environments.

[0043] (2) The present invention can grasp the accuracy and reliability of battlefield situation perception and decision-making, accurately infer the enemy's intention by modeling the combat intention of the UAV, and using situation information and relative position. Combined with human intelligence judgment, the accuracy of combat intention recognition is improved.

[0044] (3) The present invention comprehensively analyzes the prediction results of the Bayesian deep learning network and the results of human intelligent judgment, thereby improving the overall accuracy of intention recognition and providing decision makers with more reliable intelligence support by analyzing the uncertainty of different intentions.

[0045] Description of the accompanying tables and graphs

[0046] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with Example 1 of the present invention to explain the present invention and do not constitute a limitation to the present invention.

[0047] Figure 1 This is an overall flow chart of the human intelligence assessment method based on Bayesian deep learning provided by the present invention.

[0048] Figure 2 The present invention completes the action mode prediction flow chart of the enemy aircraft based on the air combat situation and maneuver strategy information.

[0049] Figure 3 Schematic diagram of different air combat situations of the present invention.

[0050] Figure 4The present invention provides the determined distance and angle information between the two combat aircrafts, so as to facilitate the formulation of the combat aircraft intention rules.

[0051] Figure 5 The present invention provides various spatial area division diagrams for different air combat situations; wherein, (a) is a division diagram for the area of ​​equilibrium between the two sides; (b) is a division diagram for the area of ​​advantage for our side; (c) is a division diagram for the area of ​​advantage for the enemy; and (d) is a division diagram for the area of ​​neutrality for both sides.

[0052] Figure 6 The present invention provides a Bayesian deep learning-based prediction of the probability of each intention when the distance between the two drones is determined and the azimuth of the enemy drone changes.

[0053] Figure 7 The human intelligence assessment method based on Bayesian deep learning provided by the present invention is a graph representing the probability and probability density of each intention at different azimuth angles. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0055] Example 1

[0056] This embodiment 1 provides a method for evaluating human intelligence based on Bayesian deep learning, which includes the following steps:

[0057] Step 1) Use the sensor system of our drone to obtain the situation information of the enemy drone, and then organize the data according to the situation information of our drone as the training set data;

[0058] Step 2) Design of rules for human intelligence to independently judge the intention of enemy drones;

[0059] Step 3) Establish a Bayesian deep learning network for judging the intention of the enemy UAV, and save the trained network model and parameters;

[0060] Step 4) Comprehensively analyze the prediction results of the Bayesian deep learning network and the judgment results of human intelligence, infer the judgment process of human intelligence, and further understand the decision-making logic of human intelligence. Finally, by combining the advantages of the two methods, the most credible combat intentions of enemy drones are obtained, providing reliable intelligence support for decision makers.

[0061] Step 1) Obtain the situation information of the enemy UAV, and then organize the data according to the situation information of our UAV as the training set data. The specific steps are as follows:

[0062] 1-1) Obtain the situation information of the red and blue drones. The situation information of the red drone: The blue side’s UAV situation information is recorded as: Among them, x, y, z are the position information of the drone, γ is the pitch angle of the drone, is the yaw angle of the drone, nx,nz is the control amount of the drone, and roll is the roll angle of the drone.

[0063] 1-2) Arrange the situation information as follows in, is the relative position information of the red and blue drones, v r_x ,v r_y ,v r_z is the velocity component of the Red Team UAV’s velocity vector in the x, y, and z directions, v b_x ,v b_y ,v b_z is the velocity component of the blue drone's velocity vector in the x, y, and z directions. Slice the data X1 and randomly select a specific number of data samples with four labels, 4,000 data samples of attack intent, 20,000 data samples of reconnaissance intent, 1,500 data samples of penetration intent, and 8,000 data samples of retreat intent, and splice them to form a new data set X2. Then, 6,000 data in X2 are divided into test sets, and the rest are training sets. The data are randomly shuffled for model training. The obtained training set data dimension information is [33500,9].

[0064] Step 2) The specific steps for designing rules for human intelligence to independently judge the intention of the enemy drone are as follows:

[0065] 2-1) Determine the azimuth angle θ1 to be -90°; the azimuth angle θ2 varies in the range of [0°180°]; the distance d between the enemy UAV and our UAV is determined to be 424m.

[0066] 2-2) Divide the battlefield space according to the relative azimuth angle θ and distance d, and set the possible intentions of the enemy drone in each area.

[0067] Balance of power between the two sides:

[0068]

[0069] When θ1≤θ2, d>1000m, the enemy's intention is inferred to be reconnaissance; d<1000m, θ2 is subdivided, θ2∈(0,45°), distance d<300m, it is inferred to be attack intention, when the distance d is between 300-1000m, it is inferred to be penetration intention. θ2∈(45°,90°), d<300m, it is inferred to be penetration intention, d>300m, it is inferred to be retreat intention.

[0070] When θ1≥θ2, d>1000m, it is inferred as reconnaissance intention, and d<1000m, it is inferred as attack intention.

[0071] Our advantages:

[0072]

[0073] θ1>θ2, the enemy is at a disadvantage. When θ2∈(90°, 135°), d>1000m, it is inferred to be a reconnaissance intention, and d<1000m, it is a penetration intention. When θ2∈(135°, 180°), it is mainly based on the flight direction of the enemy UAV, and it is inferred to be a retreat intention.

[0074] Enemy Advantages:

[0075]

[0076] θ11>θ2, we are at a disadvantage. When θ1∈(90°,180°), θ2∈(0°,90°), d>1000m, it is inferred to be a reconnaissance intention, and d<1000m is an attack intention.

[0077] Both parties are neutral:

[0078]

[0079] When θ1<θ2, θ2∈(135°, 180°), it is inferred as reconnaissance intention; when θ2∈(90°, 135°), d>1000m, the enemy's intention is inferred as reconnaissance, and when d<1000m, it is inferred as attack.

[0080] When θ11>θ2, θ2∈(135°, 180°), the enemy's intention is inferred to be reconnaissance; when θ2∈(90°, 135°), d>1000m, the enemy's intention is inferred to be reconnaissance, and d<1000m is inferred to be attack.

[0081] Step 3) Establish a Bayesian deep learning network, save the trained network model and parameters, and pass the situation information to the Bayesian deep learning network to predict the enemy UAV combat intention. The specific steps are as follows:

[0082] 3-1) Establish a Bayesian linear layer. The weights and biases of the nodes defined in this layer are random variables sampled from a normal distribution with a mean of 0 and a variance of 1. The mean matrix shape of the weights in this Bayesian linear layer is [out_features, in_features], and the variance matrix shape is [in_features, out_features]. The mean and variance of the bias are both one-dimensional matrices of [out_features].

[0083] 3-2) Construct the Bayesian deep learning network defined above, which includes a traditional linear layer and a Bayesian linear layer. The first Bayesian linear layer has a feature dimension of [input_dim,360]; the second Bayesian linear layer has a feature dimension of [360,128], the first traditional linear layer has a feature dimension of [input_dim,200], the second traditional linear layer has a feature dimension of [200,128], and the third Bayesian linear layer has a feature dimension of [128,output_dim]. In the forward propagation method, the Bayesian deep learning network calculates the logarithmic prior distribution log_prior, the logarithmic posterior distribution log_post, the logarithmic likelihood log_like, and the KL divergence formed by the difference between the prior distribution and the posterior distribution. The evidence lower bound ELBO (Evidence Lower Bound) is defined as the loss function, which consists of the weighted sum of log likelihood and KL divergence, and the loss is calculated by Monte Carlo sampling multiple times through forward propagation. The loss function calculation formula of the defined network is:

[0084]

[0085] in:

[0086] pre_loss=loss_KL×complexity_cost_weight(6)

[0087]

[0088] In this loss function, sample_nbr is the number of times adopted, pre_loss is the complexity loss of the model, which reflects the degree of deviation of the parameters from the prior distribution; like_loss is the likelihood loss, which is also the performance loss, reflecting the predictive performance of the model on the training data; k is set to 1 to control the weight, a is the mediation coefficient, which is used to balance between pre_loss and like_loss, and is set to 0.001; complexity_cost_weight is the weight for weighing the KL divergence, which is set to 1.

[0089] 3-3) Use the training set data to train the Bayesian deep learning network, set the training batch epoch, and save the model and network parameters after the training is completed.

[0090] Step 4) Comprehensively analyze the prediction results of the Bayesian deep learning network and the judgment results of human intelligence, infer the judgment process of human intelligence, and further understand the decision-making logic of human intelligence. Finally, by integrating the advantages of the two methods, the most credible combat intentions of enemy drones are obtained, providing reliable intelligence support for decision makers. The specific steps are as follows:

[0091] 4-1) Input the situation information of both drones into the Bayesian deep learning network to obtain the intention P of the enemy drone M The distance between the two drones is selected as d = 424m, the azimuth angle θ1 is fixed at -90°, and the azimuth angle θ2 is selected as 0°, 15°, 30°, 45°, 60°, 75°, 90°, 105°, 120°, 135°, 150°, 165°, 180°. The distance between the two drones is determined, and when the azimuth angle changes, the most likely intention of the enemy drone is as follows: Figure 6 As shown;

[0092] When the angle and distance between the two drones are determined, the normal distribution functions of each intention are plotted when the azimuth angle θ2 is 0°, 30°, 60°, 90°, 120°, and 150°, as shown in Figure 7 As shown in the figure, as the azimuth angle θ increases, the probability of attack intention increases significantly, the peak value of probability density also increases, the enemy UAV's attack intention is more obvious, and the probability of reconnaissance intention does not change much at different azimuths, but shows a decreasing trend. The probability of penetration intention gradually decreases with the increase of azimuth. The change in probability density shows that at the azimuth angle (150°), the penetration intention is the least obvious, while at 0°, the penetration intention is more obvious. The probability of retreat intention is also decreasing. Therefore, when the distance between the two UAVs is d=424m, the enemy UAV's azimuth angle θ2 changes, 0°, 30°, and 60° are judged as penetration intentions, and at 90°, 120°, and 150°, it is inferred that the enemy UAV's intention is to attack.

[0093] According to different air combat situations, the battlefield area is divided, so that human intelligence can obtain the intention of enemy drones Y H The inference is that when the distance between the two drones is d = 424m and θ2 = 0°, it is judged as a penetration intention; when θ2 = 30°, it is judged as a penetration intention; when θ2 = 60°, it is judged as a penetration intention; when θ2 = 90°, it is judged as an attack intention; when θ2 = 120°, it is judged as an attack intention; when θ2 = 150°, it is judged as an attack intention;

[0094] 4-2) Prediction results P by integrating and analyzing the Bayesian deep learning network M And human intelligence inference result Y H It can be seen that when θ2 is 0°, 30°, 60°, 90°, 120°, and 150°, the human intelligence reasoning idea is that the air combat situation of the two drones changes from our advantage to a balance between the two sides. Considering the distance and azimuth information of the two drones, it is concluded that the enemy drone's intention is a penetration-attack change. When θ2 = 90° and 120°, the Bayesian deep learning network judges the intention to attack, and the human intelligence judges it to attack. It can be concluded from 7(b) and 7(c) that the probability of attack and penetration intentions is similar, and the peak value of the probability density is not much different, so both intentions are possible. However, because the drones of both sides are in a balance and the distance is 424m, it is inferred that the intention of the drone at this time is an attack intention.

[0095] 4-3) Feedback the final intention judgment results to our drones to improve battlefield decision-making capabilities.

[0096] Example 2

[0097] Referring to Table 1 below, the technical solution provided in this embodiment is a method for predicting intention probability based on a model of traditional deep learning, comprising the following steps:

[0098] Step 1) Use the sensor system of our drone to obtain the situation information of the enemy drone, and then organize the data according to the situation information of our drone as the training set data;

[0099] Step 2) Design of rules for human intelligence to independently judge the intention of enemy drones;

[0100] Step 3) Establish a traditional neural network for judging the intention of the enemy UAV and save the trained network model and parameters;

[0101] Step 4) Compare the prediction results of the traditional neural network with those of the Bayesian deep learning network to derive the most credible combat intentions of the enemy UAV and provide reliable intelligence support for decision makers.

[0102] Table 1d=424 Bayesian deep learning network predicts the probability of each drone intention

[0103]

[0104] Step 1) and step 2) are the same as those in the above-mentioned embodiment 1.

[0105] Step 3) Establish a traditional neural network, save the trained network model and parameters, and pass the situation information to the Bayesian deep learning network to predict the enemy UAV combat intention. The specific steps are as follows:

[0106] 3-1) Establish a three-layer traditional neural network. The feature dimension of the first fully connected layer is [input_dim, 200], the feature dimension of the second fully connected layer is [200, 128], and the feature dimension of the third fully connected layer is [128, output_dim]. Use cross entropy loss as the loss function to save the parameters and structure of the trained traditional neural network.

[0107] Step 4) Compare the prediction results of the traditional neural network with those of the Bayesian deep learning network to obtain the most credible combat intentions of the enemy drone and provide reliable intelligence support for decision makers. The specific steps are as follows:

[0108] 4-1) Input the situation information of both drones into the traditional neural network to obtain the enemy drone intention C M , select the distance between the two drones d = 424, the azimuth angle θ1 is fixed to -90°, and the azimuth angle θ2 is selected to be 0°, 30°, 60°, 90°, 120°, and 150°. The distance between the two drones is determined, and when the azimuth angle changes, the probability of each intention of the enemy drone is obtained, as shown in Table 2 below.

[0109] Table 2d=424 Traditional neural network prediction probability table of each UAV intention

[0110]

[0111] 4-2) Input the situation information of both drones into the Bayesian deep learning network to obtain the enemy drone intention P M The distance and azimuth are the same as those in step 4-1), and the probability of each intention of the enemy drone is shown in Table 1.

[0112] Comparing the probability data of each intention in Table 1 and Table 2, when θ2 = 60°, the most likely intention obtained by the Bayesian deep learning network is penetration, and the most likely intention obtained by the traditional neural network is attack. The intention judgment results obtained by the Bayesian deep learning network are more accurate and consistent with the results of human intelligent judgment. At the same time, it has advantages in dealing with uncertainty, and the prediction probabilities of each intention are more balanced, and the prediction results are more stable. Traditional neural networks show higher confidence in some intentions, but the prediction results fluctuate greatly, are more sensitive to specific features, and have relatively low flexibility and robustness. Therefore, the Bayesian deep learning network has more advantages in dealing with high uncertainty and variability problems.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A human intelligence assessment method based on Bayesian deep learning, characterized in that: The following steps are involved: 1) UAV situation information acquisition; 2) Design of rules for human intelligence to independently judge the intentions of enemy drones; In step 2), the design of rules for human intelligence to independently judge the intention of the enemy drone includes the following steps: S31, define parameter θ1 as the azimuth of our combat aircraft determined based on the enemy combat aircraft, parameter θ2 as the azimuth of the enemy combat aircraft determined based on our combat aircraft, and parameter d as the distance between the enemy UAV and our UAV; S32, dividing the battlefield space into several areas according to the relative azimuth angle θ and the distance d, and setting the possible intentions of the enemy UAV in each area; When the enemy and us are in balance, when θ1≤θ2 and d is greater than 1000m, it is inferred that the enemy's intention is reconnaissance; when d is less than 1000m, θ2 is subdivided, θ2∈(0, 45°), when the distance d is less than 300m, it is inferred to be an attack intention, and when the distance d is between 300-1000m, it is inferred to be a penetration intention, θ2∈(45°, 90°θ, when d is less than 300m, it is inferred to be a penetration intention, and when it is greater than 300m, it is inferred to be a retreat intention; when θ1≥θ2 and d is greater than 1000m, it is inferred to be a reconnaissance intention, and when it is less than 1000m, it is inferred to be an attack intention; Our side has an advantage and the enemy has a disadvantage. θ1∈(0°, 90°). When θ2∈(90°, 135°), d is greater than 1000m, it is inferred to be a reconnaissance intention. If it is less than 1000m, it is a penetration intention. θ2∈(135°, 180°θ, mainly based on the flight direction of the enemy UAV, it is inferred to be a retreat intention. The enemy has an advantage, our side has a disadvantage, θ1∈(90°,180°), θ2∈(0°,90°), d is greater than 1000m, it is inferred to be a reconnaissance intention, less than 1000m is an attack intention; Both parties are neutral, θ1∈(90°, 180°), θ2∈(90°, 180°), when θ1<θ2, θ2∈(135°, 180°), it is inferred that the intention is reconnaissance; θ2∈(90°, 135°), d is greater than 1000m, the enemy's intention is inferred to be reconnaissance, and less than 1000m is inferred to be attack; when θ1>θ2θ2∈(135°, 180°), the enemy's intention is inferred to be reconnaissance; θ2∈(90°, 135°), d is greater than 1000m, the enemy's intention is inferred to be reconnaissance, and less than 1000m is inferred to be attack; 3) Build a Bayesian deep learning network for intent recognition; In step 3), constructing a Bayesian deep learning network for intent recognition includes the following steps: S41. Establish a Bayesian linear layer, which defines the weights and biases of the nodes as random variables sampled from a normal distribution. The mean matrix shape of the weights in the Bayesian linear layer is [out_features, in_features], the variance matrix shape is [in_features, out_features], and the mean and variance of the bias are both one-dimensional matrices of [out_features]. S42. Construct a deep learning network including traditional linear layers and Bayesian linear layers. The first Bayesian linear layer has a feature dimension of [input_dim,360]; the second Bayesian linear layer has a feature dimension of [360,128], the first traditional linear layer has a feature dimension of [input_dim,200], the second traditional linear layer has a feature dimension of [200,128], and the third Bayesian linear layer has a feature dimension of [128,output_dim]. In the forward propagation method, the Bayesian deep learning network calculates the logarithmic prior distribution, the logarithmic posterior distribution, the logarithmic likelihood, and the KL divergence formed by the difference between the prior distribution and the posterior distribution. The evidence lower bound ELBO is defined as the loss function, which is composed of the weighted sum of the logarithmic likelihood and the KL divergence. The loss is calculated by forward propagation multiple times through Monte Carlo sampling. S43, using the training set data to train the Bayesian deep learning network, and saving the trained network model and parameters; 4) Integrate the intention judgment results of the Bayesian deep learning network and human intelligence, infer the human intelligence judgment process, and obtain the final intention of the enemy UAV.

2. The human intelligence assessment method based on Bayesian deep learning according to claim 1 is characterized in that: In the step 1), the acquisition of drone situation information includes the following steps: S21. Obtain the situation information of the drones of both sides. The situation information of the drone of the red side is recorded as: The blue side’s UAV situation information is recorded as: Among them, x, y, z are the position information of the drone, γ is the pitch angle of the drone, is the yaw angle of the drone, nx,nz is the control amount of the drone, and roll is the roll angle of the drone; S22. Calculate the position information difference of the red drone relative to the blue drone, obtain the velocity components of the velocity vector in the x, y, and z axis directions, and organize the information into Slice X1, select a specific number of data samples of four labels, and splice them into a new data set X2. Then randomly divide the data set into training and test sets, and shuffle the data during model training.

3. The human intelligence assessment method based on Bayesian deep learning according to claim 1 is characterized in that: In step 4), the judgment results of the Bayesian deep learning network and human intelligence are integrated to infer the human intelligence judgment process to obtain the final intention of the enemy drone, including the following steps: S51. Obtain the enemy drone intention judgment result P through the Bayesian deep learning network M , the intention judgment result Y of human intelligence is obtained through the intention rules of the drone H ; S52, integrating the two judgment results, inferring the process of human intelligence judging the enemy UAV, and obtaining the final intention of the enemy UAV; S53. Feedback the final intention judgment results to our drones to improve battlefield decision-making capabilities.

Images