Unmanned aerial vehicle combat intention judgment method based on Bayesian model and man-machine complementarity

Abstract

The invention provides an unmanned aerial vehicle combat intention judgment method based on a Bayesian model and man-machine complementarity, and the technical scheme is as follows: collecting and analyzing real-time situation information of an enemy unmanned aerial vehicle by constructing a man-machine hybrid intelligent framework combining Bayesian inference and a man-machine complementarity theory; feature extraction is carried out on the situation data of the enemy unmanned aerial vehicle, independent intention judgment is carried out through machine intelligence and human intelligence, and related parameters are dynamically obtained through Bayesian inference; fusing the prediction results of the human and the machine by using a confidence feedback mechanism and combining with a man-machine complementarity coefficient; and through dynamic weight distribution and parameter adjustment, a more robust and accurate intention recognition result is generated. The method has the beneficial effect that the complementarity advantage of intuitionistic judgment of human intelligence and high-speed calculation of machine intelligence is effectively fused in a complex and dynamic battlefield environment.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) intention prediction, and particularly to a method for judging UAV combat intention based on a Bayesian model and human-machine complementarity. Background Art

[0002] In military operations, the judgment and decision-making of human intelligence have always been the key factors in the outcome of the battlefield. For example, commanders need to face complex and changing tactical situations, quickly evaluate the behavior of the enemy, and make accurate decisions. The human-machine hybrid intelligence method based on the Bayesian model effectively captures the changing trend of the behavior of enemy UAVs by introducing a dynamically adjusted reasoning mechanism. At the same time, by combining the global intuitive judgment of human intelligence and the fast computing ability of machine intelligence, a hybrid decision-making system with strong complementarity is formed.

[0003] With the efficient processing of high-dimensional data and the accurate calculation of complex models, big data and artificial intelligence technologies have gradually penetrated into the field of military operations. However, in practical applications, the limitations of the interpretability and generalization ability of artificial intelligence technologies have gradually emerged, resulting in their deficiencies in dealing with the uncertainty, openness, and vulnerability of complex systems. In view of the limitations of artificial intelligence technologies at the present stage and the system's demand for intelligent solutions, some researchers have proposed a comprehensive modeling method that combines computer simulation and expert experience, and thus the concept of human-machine hybrid intelligence has emerged, which provides a new idea for UAV intention recognition. Compared with relying solely on machine or human analysis, combining the advantages of human experience and machine learning helps UAVs better improve their recognition performance when processing data.

[0004] However, most of the existing human-machine hybrid intelligence methods mix the predictions of human intelligence and machine intelligence through the form of an intervention mechanism for human-machine interaction, making it difficult to utilize their complementary advantages and difficult to capture the characteristics of real-time changing data when dealing with complex and dynamic actual environments. For example, the human-machine hybrid model proposed in the article "Bayesian modeling of human–AI complementarity" explores the conditions for achieving complementarity by combining the predictions of human and machine classifiers, revealing that the advantages and disadvantages of humans and machines in tasks are complementary. However, similar methods mostly rely on fixed or preset parameter models, making it difficult to capture the characteristics of real-time changing data and task difficulty when dealing with complex and dynamic actual environments, resulting in a rigid decision-making process, and the traditional human-machine complementarity model does not fully consider the uncertainty in the prediction task. For example, the current method cannot dynamically feedback and adjust parameters, resulting in poor performance of the classifier in high-difficulty tasks or extreme situations, affecting the overall decision-making effect.

[0005] To solve this problem, the present invention proposes a method for judging the combat intention of an unmanned aerial vehicle (UAV) based on a Bayesian model and human-machine complementarity, which is used to improve the accuracy of UAV intention recognition. This method combines the probabilistic outputs of human intelligence and machine intelligence, dynamically adjusts classification parameters through Bayesian inference, and obtains more realistic parameters. In addition, to improve the accuracy of overall decision-making, we calculate the contribution ratios of human intelligence and machine intelligence through a complementarity coefficient, and generate a weighted mixed confidence level. Summary of the Invention

[0006] The object of the present invention is to propose a method for judging the combat intention of an unmanned aerial vehicle (UAV) based on a Bayesian model and human-machine complementarity, which is used to assist in UAV combat intention recognition and decision-making optimization. This method can combine the intuitive judgment ability of human intelligence and the high-speed computing advantage of machine intelligence to achieve efficient and reliable intention prediction and decision-making support in complex and uncertain battlefield environments. The present invention can not only improve the accuracy of battlefield situation awareness and prediction, but also enhance the adaptability of the model in dynamic environments by dynamically adjusting decision weights, thereby improving combat effectiveness in high-risk battlefields and reducing the risk of battle losses caused by decision-making mistakes.

[0007] The inventive concept of the present invention is as follows: First, based on the situation information of friendly and enemy UAVs, by using methods such as Bayesian inference, the prediction results of human intelligence and machine intelligence are fused to obtain a complementary model; subsequently, by analyzing the uncertainty distribution in the prediction results, the human-machine complementarity coefficient is calculated, and based on this, the combined weights of human and machine intelligence are dynamically adjusted to generate a more accurate final prediction result; on this basis, the model parameters are continuously optimized through a dynamic feedback mechanism to adapt to the real-time changes in the battlefield environment. The present invention endows UAVs with the ability to recognize the intentions of enemy UAVs during combat, and enhances decision-making reliability by quantifying uncertainty and complementarity, thereby providing new technical support and practical application value for the field of military intelligence.

[0008] The present invention is implemented by the following measures: A method for judging the combat intention of an unmanned aerial vehicle (UAV) based on a Bayesian model and human-machine complementarity, comprising the following steps:

[0009] S1. Use the sensor system of our UAV to obtain the situation information of the enemy UAV, and then according to the situation information of our UAV, organize the data as training set data, and use this data to obtain the prediction results of human intelligence and machine intelligence.

[0010] S2. Based on the prediction results, a complementary model is constructed through methods such as Bayesian inference, logit score and confidence level, and human-machine complementarity coefficient.

[0011] S3. Realize human-machine hybrid decision-making through a dynamic combined weight allocation and parameter update mechanism to obtain the final intention of the enemy UAV, and provide reliable intelligence support for decision-makers.

[0012] Furthermore, Step 1 includes the following steps:

[0013] 1-1) Collect the current state information of the enemy drones: position information, flight speed, heading angle, roll angle, pitch angle, control amount, etc., in order to predict the future behavior of the enemy drones.

[0014] 1-2) Obtain the situation information of the red and blue drones. The situation information of the red drones is denoted as: The situation information of the blue drones is denoted as: where, where, x r , x b , y r , y b , z r , z b are respectively the position information of the drones, γ r , γ b is the pitch angle, is the yaw angle, nx r , nx b , nz r , nz b is the control amount, roll r , roll b is the roll angle;.

[0015] 1-3) Set the independently completed machine intelligence training model as M, and its output result is: where is the probability that the model M judges the data x as the category y i and the total number of categories is k. Set the independently completed human intelligence training model as H, and the human judgment result of the category of x is Y H (x), and the credibility is C H (x). Based on the probability assignment principle, the performance of humans on x in each category can be expressed using a function as: where is the probability that the model H judges the data x as the category y i and the total number of categories is k, f yc is the probability assignment model constructed based on expert knowledge.

[0016] Furthermore, Step 2 includes the following steps:

[0017] 2-1) Construct the prior distribution of each parameter, μ H1 ~N(1, 2), μ H0 ~N(-1, 2), μ M1 ~N(1, 2), μ M0 ~N(-1, 2), σH ~N(1, 2), σ M ~N(1, 2), ρ HM ~N(0.5, 0.3), where μ H1 is the mean of human intelligence on positive examples, μ H0 is the mean of human intelligence on negative examples, μ M1 is the mean of machine intelligence on positive examples, μ M0 is the mean of machine intelligence on negative examples, σ H is the standard deviation of human intelligence, σ M is the standard deviation of machine intelligence, ρ HM is the correlation coefficient between human intelligence and machine intelligence;

[0018] 2 - 2), Calculate the mean, standard deviation, and correlation for each label. For the mean of human intelligence positive example observations: The mean of negative example observations is: The variance is: The correlation coefficient is: where ph[j] represents the probability output of human intelligence on label j, correct and incorrect represent the correctness of each sample on class j, num_correct and num_incorrect record the number of samples correctly and incorrectly classified on this class, Cov(ph[j], pm[j]) represents the covariance between human intelligence and machine intelligence, σ H [j] represents the variance of human intelligence on label j, σ M [j] represents the variance of machine intelligence on label j;

[0019] 2 - 3), Construct the logit score for each data sample i and each class label j, which reflects the performance of the classifier under different class labels.

[0020] The mean during logit score generation is determined by the matching of the true label z i . During the generation process, when label j matches the true label, the positive class mean and are used, otherwise the negative class mean and are used. Specifically expressed as:

[0021]

[0022] Construct the covariance matrix ∑ to represent the potential correlation between the logit scores of human and machine classifiers on different labels;

[0023]

[0024] Among them, and respectively represent the variance of the logit scores of the human and machine classifiers, while ρ HM [j] represents the correlation between the two.

[0025] Construct the human and machine logit scores, sample the logit scores using the bivariate normal distribution, which is expressed as:

[0026]

[0027] Construct the confidence scores. The logit scores are non-linearly transformed through the sigmoid function. The corresponding confidence score calculation formula is: Among them, γ H [i,j] and γ M [i,j] respectively represent the confidence scores of the human and machine classifiers on label j;

[0028] 2 - 4), construct the human - machine complementarity coefficient r HM , which is a key indicator for measuring the collaborative effect of humans and machines in a hybrid intelligent system, and describes the degree of combination and effectiveness of human intelligence and machine intelligence in a given task. The calculation formula of the human - machine complementarity coefficient is as follows:

[0029]

[0030] Among them, μ H represents the mean of human intelligence, μ M represents the mean of machine intelligence, σ H represents the variance of human intelligence, σ M represents the variance of machine intelligence, and ρ HM is the correlation coefficient between human intelligence and machine intelligence.

[0031] Furthermore, step three includes the following steps:

[0032] 3 - 1), construct a human - machine hybrid intelligent model. Through dynamic combined weight allocation, the system can flexibly adjust the weights of humans and machines in different tasks, making the prediction of the human - machine combination more accurate. The combined weight allocation uses parameters α and β to represent the weights of humans and machines in decision - making. By introducing a complementarity coefficient to balance the confidence of the human classifier and the confidence of the machine classifier, the specific calculation formula is as follows:

[0033]

[0034] β = 1 - α (6)

[0035] The final combined score is: combined_scores = α·γ H +β·γ M , where r HM is the human-machine complementarity coefficient, γ H is the confidence score of human intelligence, γ M is the confidence score of machine intelligence. The clip function restricts the value to be between [0.1, 0.9] to avoid parameter out-of-bounds;

[0036] 3-2), Construct a dynamic update mechanism to adjust parameters in real time to adapt to the performance changes of humans and machines during the task. This mechanism reduces the error rate by continuously adjusting key parameters, thereby improving the stability of the system. The correlation parameter ρ HM is updated based on the difference between the human and machine confidence levels. The formula is: ρ HM = clip(ρ HM + 0.001·sign(confidence_diff - 0.1), -1.0, 1.0), where confidence_dif is the absolute difference between the human and machine confidence levels, and the sign function is used to indicate the direction of increasing or decreasing ρ HM The clip function restricts the value to be between [-1.0, 1.0] to avoid parameter out-of-bounds;

[0037] The variance parameter is updated according to the error rate of each intelligence. When the error rate is high, the corresponding variance parameter will increase, indicating an increase in uncertainty; conversely, the variance parameter decreases. The update formula is: σ H = clip(σ H + 0.001·sign(errors H - 0.5), 1e-6, 10.0), σ M = clip(σ M + 0.001·sign(errors M - 0.5), 1e-6, 10.0).

[0038] where errors H and errors M represent the average error rates of humans and machines respectively;

[0039] 3-3), Use the training set data to train the model, and save the trained model and parameters such as accuracy.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] (1) The present invention dynamically quantifies the complementary coefficient between human intelligence and machine intelligence through a Bayesian model, enabling efficient data fusion and real-time decision optimization in complex and dynamic battlefield environments. Through a mechanism for dynamically updating parameters, the present invention effectively captures the uncertainty in prediction results and environmental changes, making decisions more flexible and reliable, especially suitable for high-risk and high-dynamic scenarios in UAV operations.

[0042] (2) Based on the situational information of friendly and enemy UAVs, the present invention combines the respective advantages of human intelligence and machine intelligence to construct a Bayesian inference model and a multi-level hybrid intelligence architecture. By modeling the combat intentions of enemy UAVs and combining relative position and dynamic situational information, the accuracy of intention recognition and decision-making efficiency are significantly improved, providing more reliable technical support for battlefield situation awareness and action decision-making.

[0043] (3) By introducing a human-machine complementary coefficient and a dynamic weight adjustment mechanism, the present invention overcomes the limitations of traditional systems in high-uncertainty environments. It not only improves the overall accuracy of intention recognition but also optimizes parameter updates through a dynamic feedback mechanism, enabling the system to have stronger robustness and generalization ability during long-term operation, providing a new solution for UAV operations and intelligent decision-making.

[0044] Explanation of the attached drawings

[0045] The accompanying drawings are used to provide a further understanding of the present invention and form a part of the specification. Example 1 of the present invention is used to explain the present invention and does not constitute a limitation to the present invention.

[0046] Figure 1 It is the overall flowchart of the UAV combat intention judgment method based on the Bayesian model and human-machine complementarity provided by the present invention.

[0047] Figure 2 It is the model framework diagram of the UAV combat intention judgment method based on the Bayesian model and human-machine complementarity provided by the present invention.

[0048] Figure 3 It is the connection schematic diagram between each neuron of the machine intelligence neural network provided by the present invention.

[0049] Figure 4 It is the schematic diagram of the K-fold cross-validation results of each model provided by the present invention.

[0050] Figure 5 It is the schematic diagram of the accuracy of four models provided by the present invention under different noises. Detailed implementation manner

[0051] To make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0052] Embodiment 1

[0053] Embodiment 1 provides a method for judging the combat intention of an unmanned aerial vehicle based on a Bayesian model and human-machine complementarity, which includes the following steps:

[0054] Step 1): Use the sensor system of our unmanned aerial vehicle to obtain the situation information of the enemy unmanned aerial vehicle, and then organize the data according to the situation information of our unmanned aerial vehicle as training set data, and use this data to obtain the prediction results of human intelligence and machine intelligence;

[0055] Step 2): Based on the prediction results, construct a complementary model through methods such as Bayesian inference, logit score and confidence, and human-machine complementarity coefficient;

[0056] Step 3): Implement human-machine hybrid decision-making through dynamic combined weight allocation and parameter update mechanism to obtain the final intention of the enemy unmanned aerial vehicle and provide reliable intelligence support for decision-makers;

[0057] Step 4): Compare the models and analyze the accuracy rate, F1 score, recall rate, and precision of each model.

[0058] Step 1): Obtain the situation information of the enemy unmanned aerial vehicle, and then organize the data according to the situation information of our unmanned aerial vehicle as training set data, and use this data to obtain the prediction results of human intelligence and machine intelligence. The specific steps are as follows:

[0059] 1-1): Obtain the situation information of the unmanned aerial vehicles of the red and blue sides. The situation information of the red side unmanned aerial vehicle is recorded as: The situation information of the blue side unmanned aerial vehicle is recorded as: Among them, among them, x r , x b , y r , y b , z r , z b Are the position information of the unmanned aerial vehicle respectively, γ r , γ b Is the pitch angle, Is the yaw angle, nx r , nx b , nz r , nz b Are the control amounts, roll r , roll b Is the roll angle.

[0060] 1-2), set the independently completed machine intelligence training model as M, and its output result is: Among them is the probability that the model M judges the data x as the category y i , and the total number of categories is k. Set the independently completed human intelligence training model as H, and the human judgment result of the category of x is Y H (x), and the confidence level is C H (x). Based on the probability assignment principle, the performance of humans on x in each category can be expressed by a function as: Among them is the probability that the model H judges the data x as the category y i , the total number of categories is k, and f yc is the probability assignment model constructed based on expert knowledge.

[0061] Step 2) Based on the prediction results, a complementary model is constructed through methods such as Bayesian inference, logit score, confidence level, and human-machine complementarity coefficient. The specific steps are as follows:

[0062] 2-1), construct the prior distribution of each parameter, μ H1 ~N(1, 2), μ H0 ~N(-1, 2), μ M1 ~N(1, 2), μ M0 ~N(-1, 2), σ H ~N(1, 2), σ M ~N(1, 2), ρ HM ~N(0.5, 0.3), where μ H1 is the mean of human intelligence on positive examples, μ H0 is the mean of human intelligence on negative examples, μ M1 is the mean of machine intelligence on positive examples, μ M0 is the mean of machine intelligence on negative examples, σ H is the standard deviation of human intelligence, σ M is the standard deviation of machine intelligence, ρ HM is the correlation coefficient between human intelligence and machine intelligence;

[0063] 2-2), calculate the mean, standard deviation, and correlation of each label. For the mean of human intelligence positive example observations:

[0064] The mean of negative example observations is: The variance is: The correlation coefficient is: Among them, ph[j] represents the probability output of human intelligence on label j, correct and incorrect represent the correctness of each sample on class j, num_correct and num_incorrect record the number of samples correctly and incorrectly classified on this class, Cov(ph[j], pm[j]) represents the covariance between human intelligence and machine intelligence, and σ H [j] represents the variance of human intelligence on label j, and σ M [j] represents the variance of machine intelligence on label j;

[0065] 2-3), Construct the logit score based on each data sample i and each class label j, which reflects the performance of the classifier under different class labels.

[0066] The mean value during the generation of the logit score is determined by the matching situation of the true label z i . During the generation process, when label j matches the true label, the positive class mean and are used, otherwise the negative class mean and are used. Specifically expressed as:

[0067]

[0068] Construct the covariance matrix ∑ to represent the potential correlation between the logit scores of the human and machine classifiers on different labels;

[0069]

[0070] Among them, and represent the variances of the logit scores of the human and machine classifiers respectively, and ρ HM [j] represents the correlation between the two.

[0071] Construct the human and machine logit scores, and sample the logit scores using the bivariate normal distribution, expressed as:

[0072]

[0073] Construct the confidence score, the logit score is non-linearly transformed through the sigmoid function, and the corresponding confidence score calculation formula is: Among them, γ H [i, j] and γ M [i, j] represent the confidence scores of the human and machine classifiers on label j respectively;

[0074] 2-4), Construct the human-machine complementarity coefficient rHM , which is a key metric for measuring the collaborative effectiveness of humans and machines in a hybrid intelligent system, describes the degree of combination and effectiveness of human intelligence and machine intelligence in a given task. The calculation formula for the human-machine complementarity coefficient is as follows:

[0075]

[0076] where μ H represents the mean of human intelligence, and μ M represents the mean of machine intelligence, σ H represents the variance of human intelligence, and σ M represents the variance of machine intelligence, and ρ HM is the correlation coefficient between human intelligence and machine intelligence.

[0077] Step 3) Achieve human-machine hybrid decision-making through a dynamic combination of weight allocation and parameter update mechanism to obtain the final enemy UAV intention. The specific steps are as follows:

[0078] 3-1), Construct a human-machine hybrid intelligent model. Through dynamic combination weight allocation, the system can flexibly adjust the weights of humans and machines in different tasks, making the prediction of the combination of humans and machines more accurate. The combination weight allocation uses parameters α and β to represent the weights of humans and machines in decision-making. By introducing a complementarity coefficient, the confidence of the human classifier and the confidence of the machine classifier are balanced. The specific calculation formula is as follows:

[0079]

[0080] β = 1 - α (6)

[0081] The final combined score is: combined_scores = α·γ H + β·γ M , where r HM is the human-machine complementarity coefficient, γ H is the confidence score of human intelligence, γ M is the confidence score of machine intelligence. The clip function restricts the value to be between [0.1, 0.9] to avoid parameter out-of-bounds;

[0082] 3-2), Construct a dynamic update mechanism to adjust parameters in real time to adapt to the performance changes of humans and machines during the task. This mechanism reduces the error rate by continuously adjusting key parameters, thereby improving the stability of the system. The update of the correlation parameter ρ HM is based on the difference between the confidence of humans and machines. The formula is: ρ HM ·clip(ρ HM+0.001·sign(confidence_diff - 0.1), -1.0, 1.0), where confidence_dif is the absolute difference between the human and machine confidence levels, and the sign function is used to indicate whether to increase or decrease ρ HM in the direction, and the clip function restricts the value to be between [-1.0, 1.0] to avoid parameter out-of-bounds;

[0083] The variance parameter is updated based on the error rate of each intelligence. When the error rate is high, the corresponding variance parameter will increase, indicating an increase in uncertainty; conversely, the variance parameter decreases. The update formula is: σ H = clip(σ H + 0.001·sign(errors H - 0.5), 1e - 6, 10.0), σ M = clip(σ M + 0.001·sign(errors M - 0.5), 1e - 6, 10.0).

[0084] where errors H and errors M represent the average error rates of humans and machines respectively;

[0085] 3 - 3), Use the training set data to train the model, and save the trained model and parameters such as accuracy.

[0086] Step 4) Compare the models and analyze the accuracy, F1 score, recall rate, and precision of each model. The specific steps are as follows:

[0087] Table 1 shows the accuracy, precision, recall rate, and F1 score of four models on different categories. The B - HM - C model with human - machine hybrid optimization is generally better than the other three models, demonstrating excellent performance.

[0088] Table 1 is a comparison table of four models

[0089]

[0090] Based on the above comparison results, it can be found that the method for judging the combat intention of drones based on the Bayesian model and human - machine complementarity proposed by the present invention effectively combines the advantages of both humans and machines. Different from the static Bayesian model, this model shows strong accuracy based on the confidence feedback and dynamic parameter update mechanism, providing effective support for drone combat decision - making in complex battlefield environments.

[0091] Embodiment 2

[0092] Embodiment 2 provides a method for judging the combat intention of an unmanned aerial vehicle based on a Bayesian model and human-machine complementarity, which includes the following steps:

[0093] Step 1) Use the sensor system of our unmanned aerial vehicle to obtain the situation information of the enemy unmanned aerial vehicle, and then organize the data according to the situation information of our unmanned aerial vehicle as the training set data, and use this data to obtain the prediction results of human intelligence and machine intelligence;

[0094] Step 2) Based on the prediction results, construct a complementary model through methods such as Bayesian inference, logit score and confidence, and human-machine complementarity coefficient;

[0095] Step 3) Realize human-machine hybrid decision-making through dynamic combined weight allocation and parameter update mechanism to obtain the final intention of the enemy unmanned aerial vehicle, and provide reliable intelligence support for decision-makers;

[0096] Step 4) Compare the models and analyze the K-fold cross-validation of each model.

[0097] Steps 1), 2) are the same as those in Embodiment 1 above.

[0098] Step 4) By evaluating the performance of the model on unseen data, ensure that the model has good generalization ability, reduce the risks of overfitting and underfitting, so as to construct a more reliable and robust model, Figure 4 showing the K-fold cross-validation results of each model. From Figure 4 the results, it can be found that a method for judging the combat intention of an unmanned aerial vehicle based on a Bayesian model and human-machine complementarity proposed by the present invention performs significantly better than other models in the task of unmanned aerial vehicle intention judgment, proving that this method has high generalization ability, can effectively overcome the limitations of traditional methods, adapt to the changeable battlefield environment, and provide more reliable support for unmanned aerial vehicle air combat decision-making.

[0099] Embodiment 3

[0100] Embodiment 3 provides a method for judging the combat intention of an unmanned aerial vehicle based on a Bayesian model and human-machine complementarity, which includes the following steps:

[0101] Step 1) Use the sensor system of our unmanned aerial vehicle to obtain the situation information of the enemy unmanned aerial vehicle, and then organize the data according to the situation information of our unmanned aerial vehicle as the training set data, and use this data to obtain the prediction results of human intelligence and machine intelligence;

[0102] Step 2) Based on the prediction results, construct a complementary model through methods such as Bayesian inference, logit score and confidence, and human-machine complementarity coefficient;

[0103] Step 3) Implement human-machine hybrid decision-making through a dynamic combination of weight assignment and parameter update mechanism to obtain the final intention of the enemy UAV, providing reliable intelligence support for decision-makers;

[0104] Step 4) Compare the models and analyze the robustness and stability of each model.

[0105] Steps 1), 2) are the same as those in Embodiment 1 above.

[0106] Step 4) Test the robustness and stability of the model by setting the noise range at 5%, 10%, 15%, 20% and 25% of the data, Figure 5 showing the accuracy of each model under different noises. From Figure 5 the results, it can be found that the method for judging UAV combat intention based on Bayesian model and human-machine complementarity proposed by the present invention has superiority under all noise levels, and its accuracy drops the least under different noise conditions. It drops from 0.928 at 5% noise to 0.713 at 25% noise. Even in a high-noise environment, the model can relatively stably maintain a high accuracy, being able to better cope with the uncertainties in complex environments, thus providing more reliable and robust support for the air combat intention recognition of UAVs.

[0107] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for judging the combat intention of an unmanned aerial vehicle based on a Bayesian model and human-machine complementarity, characterized in that, It includes the following steps: 1) Obtaining the situation information of both sides' unmanned aerial vehicles (UAVs), and predicting human intelligence and machine intelligence; 2) Constructing a complementary model based on the prediction results through methods such as Bayesian inference, logit score, confidence level, and human-machine complementarity coefficient; 3) Achieving human-machine hybrid decision-making through a dynamic combined weight allocation and parameter update mechanism to obtain the final intention of the enemy UAV.

2. The method for judging the combat intention of an unmanned aerial vehicle based on a Bayesian model and human-machine complementarity according to claim 1, wherein, In step 1), the obtaining of the situation information of both sides' UAVs and the prediction of human intelligence and machine intelligence include the following steps: S21. Obtain the situation information of the red and blue drones. The situation information of the red drones is denoted as: The situation information of the blue drones is denoted as: where, (x r , y r , z r ) represents the position information of the red drones in three-dimensional space, v r represents the speed of the red drones, γ r represents the pitch angle of the red drones, φ r represents the heading angle of the red drones, (nx r , nz r ) represents the control amount of the red drones, roll r represents the roll angle of the red drones, (x b , y b , z b ) represents the position information of the blue drones in three-dimensional space, v b represents the speed of the blue drones, γ b represents the pitch angle of the blue drones, φ b represents the heading angle of the blue drones, (nx b , nz b ) represents the control amount of the blue drones, roll b represents the roll angle of the blue drones; S22. Set the machine intelligence training model independently completed as M, and its output result is: Among them is the probability that the model M judges the data x as the category y i The total number of categories is k. Set the human intelligence training model independently completed as H, and the human's judgment result of the category of x is Y H (x), and the credibility is C H (x). Based on the probability assignment principle, the human's performance of x in each category is represented by a function as: Among them is the probability that the model H judges the data x as the category y i The total number of categories is k, and f yc is a probability assignment model constructed based on expert knowledge.

3. The method for judging the combat intention of an unmanned aerial vehicle based on a Bayesian model and human-machine complementarity according to claim 1, characterized in that In step 2), constructing a complementary model based on the prediction results through methods such as Bayesian inference, logit score, confidence level, and human-machine complementarity coefficient includes the following steps: S31. Construct the prior distributions of each parameter, μ H1 ~N(1, 2), μ H0 ~N(-1, 2), μ M1 ~N(1, 2), μ M0 ~N(-1, 2), σ H ~N(1, 2), σ M ~N(1, 2), ρ HM ~N(0.5, 0.3), where N represents the normal distribution, μ H1 is the mean of human intelligence on positive examples, μ H0 is the mean of human intelligence on negative examples, μ M1 is the mean of machine intelligence on positive examples, μ M0 is the mean of machine intelligence on negative examples, σ H is the standard deviation of human intelligence, σ M is the standard deviation of machine intelligence, ρ HM is the correlation coefficient between human intelligence and machine intelligence; S32. Calculate the mean, standard deviation, and correlation for each label. The mean for positive observations of human intelligence is: The mean for negative observations is: The variance is: The correlation coefficient is: Among them, ph[j] represents the probability output of human intelligence on label j, correct and incorrect represent the correctness of each sample on category j, num_correct and num_incorrect record the number of samples correctly and incorrectly classified on this category, Cov(ph[j], pm[j]) represents the covariance between human intelligence and machine intelligence, and σ H [j] represents the variance of human intelligence on label j, and σ M [j] represents the variance of machine intelligence on label j; S33. Constructing a logit score based on each data sample i and each class label j, which is used to reflect the performance of the classifier under different class labels; The mean value during logit score generation is determined by the matching condition of the true label z i During the generation process, when the label j matches the true label, the positive class mean is adopted and otherwise the negative class mean is adopted and It is expressed as: Constructing a covariance matrix ∑ to represent the potential correlation between the logit scores of human and machine intelligence on different labels; where and represent the variance of the logit scores of the human and machine classifiers, respectively, while ρ HM [j] represents the correlation between the two. The logit scores are constructed and sampled using a bivariate normal distribution, expressed as: Construct a confidence score. The logit score is non-linearly transformed through the sigmoid function. The corresponding confidence score calculation formula is as follows: where γ H [i, j] and γ M [i, j] respectively represent the confidence scores of the human and machine classifiers on label j; S34. Construct the human-machine complementarity coefficient r HM , which is a key indicator for measuring the collaboration effect of humans and machines in a hybrid intelligent system, describes the degree and effectiveness of the combination of human intelligence and machine intelligence in a given task. The calculation formula for the human-machine complementarity coefficient is as follows: Among them, μ H represents the mean of human intelligence, μ M represents the mean of machine intelligence, σ H represents the variance of human intelligence, σ M represents the variance of machine intelligence, ρ HM is the correlation coefficient between human intelligence and machine intelligence.

4. The method for judging the combat intention of an unmanned aerial vehicle based on a Bayesian model and human-machine complementarity according to claim 1, characterized in that In step 3), achieving human-machine hybrid decision-making through a dynamic combined weight allocation and parameter update mechanism to obtain the final intention of the enemy UAV includes the following steps: S41. Constructing a human-machine hybrid intelligence model. Through dynamic combined weight allocation, the system flexibly adjusts the weights of humans and machines in different tasks. The combined weight allocation uses parameters α and β to represent the weights of humans and machines in decision-making. By introducing a complementarity coefficient to balance the confidence levels of the human classifier and the machine classifier, the calculation formula is as follows: β＝1-α (6) The final combined score is: combined_scores = α·γ H +β·γ M , where, r HM is the human-machine complementarity coefficient, γ H is the confidence score of human intelligence, γ M is the confidence score of machine intelligence. The clip function restricts the value to be between [0.1, 0.9] to avoid parameter out-of-bounds; S42. Build a dynamic update mechanism to adjust parameters in real time, and the correlation parameter ρ HM is updated based on the difference between human and machine confidence, and the formula is: ρ HM = clip(ρ HM + 0.001·sign(confidence_diff - 0.1), -1.0, 1.0), where confidence_dif is the absolute difference between human and machine confidence, and the sign function is used to indicate the direction of increasing or decreasing ρ HM . The update of the variance parameter is based on the error rate of each intelligence. When the error rate is high, the corresponding variance parameter will increase, indicating an increase in uncertainty; conversely, the variance parameter decreases. The formula is: σ H = clip(σ H + 0.001·sign(errors H - 0.5), 1e-6, 10.0), σ M = clip(σ M + 0.001·sign(errors M - 0.5), 1e-6, 10.0), where errors H and errors M represent the average error rates of humans and machines respectively; S43. Using the training set data to train the model, and saving the trained model and accuracy parameters.

Images