Unmanned aerial vehicle combat intention recognition method based on causal characterization

Abstract

The invention provides an unmanned aerial vehicle combat intention recognition method based on causal characterization, and aims to solve the technical problems of prediction and decision making under non-independent identically distributed data. And the output of human intelligence and machine intelligence is fused by using the Bayesian neural network, so that the robustness and decision-making ability of the system in a complex environment are improved. The model framework designed by the invention is divided into three levels: the first level is independent processing of machine intelligence and human intelligence, the second level is used for converting features into potential causal variables through CE-VAE, and the third level is used for fusing the potential variables through a Bayesian neural network, generating joint causal representation and outputting a final decision.

Description

Technical Field

[0001] The present invention relates to the technical field of UAV intention prediction, and in particular to a method for identifying UAV combat intention based on causal representation. Background Art

[0002] In machine learning and statistics, the independent and identically distributed (IID) assumption is the cornerstone of many classic algorithms and theories. However, real-world data often does not conform to this assumption. For example, data in sensor networks may be influenced by environmental factors, and data in financial markets often exhibit complex dependencies. This non-IID property poses a significant challenge to model performance and generalization, resulting in the model's inability to effectively capture the inherent structure of the data in real applications. This problem becomes increasingly prominent as data size and complexity increase. Addressing the challenges of non-IID requires re-examining and adjusting traditional algorithm design to ensure the model's robustness and accuracy when processing complex data.

[0003] To address the challenges posed by non-IID, there are currently two main approaches. One approach relies on traditional statistical and machine learning methods, which primarily rely on adjusting the algorithms themselves to accommodate data dependencies. The other approach incorporates causal inference, identifying and leveraging causal relationships in the data to improve model accuracy and robustness.

[0004] In traditional research without causal inference, researchers have improved algorithmic structures and introduced regularization techniques to make models more adaptable to complex data dependencies. They have also utilized a combination of multiple models to improve overall performance and mitigate a single model's sensitivity to changes in data distribution. Data transformation and synthesis techniques have been used to increase data diversity and help models generalize better. Causal inference is used to address the non-IID problem. Currently, researchers are identifying causal relationships in the data to reduce the impact of data bias on the model. By combining causal inference with transfer learning, researchers are improving the model's generalization ability in new environments. Causal models are used to generate or enhance data, alleviating data imbalance.

[0005] Although the above methods have solved the accuracy error of prediction results when the data are not independent and identically distributed to a certain extent, they still lack in interpretability and stability. Summary of the Invention

[0006] The purpose of this invention is to propose a method for identifying UAV combat intentions based on causal representation, aiming to improve the interpretability and stability of military decision-making, solve the result errors when the data is not independent and identically distributed, extract the essential reasons for human and machine intelligence decision-making through the causal representation model, and enhance it, establish classification rules, simulate the judgment process of machine intelligence and human intelligence respectively, and use variational autoencoders and Bayesian networks for data processing and result output, and use Bayesian neural networks to mix the output results of human intelligence, machine intelligence and variational autoencoders to finally obtain the prediction results.

[0007] The inventive concept of the present invention is to construct a three-layer model framework: the first layer processes data independently, the second layer extracts potential causal variables through causal variational autoencoders, and the third layer uses Bayesian neural networks to fuse these variables and output the final decision, aiming to improve the robustness, interpretability and stability of the model in complex environments.

[0008] The present invention is achieved by the following measures: a method for identifying UAV combat intention based on causal representation, comprising the following steps:

[0009] S1. Acquisition of UAV combat intention data;

[0010] S2. Design and train a deep learning-based combat intent judgment model to obtain a machine intelligence model;

[0011] S3, the design of the combat intention judgment model based on human intelligence knowledge base, and the human intelligence-based;

[0012] S4. Extract the causal characteristics of machine intelligence and human intelligence respectively through causal representation;

[0013] S5. Build Bayesian deep learning to complete the causal feature fusion of machine intelligence and human intelligence, and output the final enemy drone intention judgment result.

[0014] Furthermore, the step 1 includes the following steps:

[0015] 1-1) UAV data information acquisition, the drone information of the red and blue sides are R r =[x r ,y r ,z r ,v r ,γ r ,φ r ,nx r ,nz r ,roll r ] and R b =[x b ,y b ,z b ,vb ,γ b ,φ b ,nx b ,nz b ,roll b ]

[0016] 1-2) Extract the position coordinate information of the red and blue drones, normalize the position coordinates, and divide the coordinate values ​​by 1000 to reduce the data level; calculate the velocity direction vectors of the red and blue drones based on the pitch and yaw angles of the drones. The calculation formula for the velocity direction vector is:

[0017] u r =[cos(γ r )·sin(φ r ),cos(γ r )·cos(φ r ),sin(γ r )] (1)

[0018] u b =[cos(γ b )·sin(φ b ),cos(γ b )·cos(φ b ),sin(γ b )] (2)

[0019] 1-3) Calculate the Euclidean distance d between the red drone and the blue drone = || p r -p b ||, where p r =[x r ,y r ,z r ],p b =[x b ,y b ,z b ] are the position information of the red and blue drones respectively, and the unit direction vector d of the red drone pointing to the blue drone is calculated. r→b =(p b -p r ) / (||p b -p r ||), the speed unit direction vector of the red team's drone Calculate the angle between the speed direction of the red drone and the direction the red drone points to the blue drone Calculate the unit direction vector d of the blue drone pointing to the red drone b→r =(p b -p r ) / (||p b -pr ||), the speed unit direction vector of the blue drone Calculate the angle between the speed direction of the blue drone and the direction of the blue drone pointing to the red drone

[0020] 1-4) Record the relative position information of the red and blue drones as follows: , record the obtained information as The feature vector is sliced ​​to filter out a specific number of data samples containing four labels, and then spliced ​​to generate a new dataset; the new dataset is randomly divided into training and test sets, and the data is shuffled during model training to ensure randomness and diversity of the data.

[0021] Furthermore, the step 2 includes the following steps:

[0022] 2-1) Build an MLP neural network for machine intelligence. The input layer receives the input vector inputs_dims. Fully connected layer 1 (fc1) maps the input dimension inputs_dims to dimension 200. Activation function: The softplus activation function is applied to the output of fc1. Fully connected layer 2 (fc2) maps dimension 200 to dimension 128. Activation function: The softplus activation function is applied to the output of fc2. Fully connected layer 3 (fc3) maps dimension 128 to dimension output_dims. Activation function: The sigmoid activation function is applied to the output of fc3 to produce the final output.

[0023] 2-2) Each iteration includes steps such as data loading, forward propagation, loss calculation, gradient calculation, and parameter update. First, a batch of data is extracted from the training set and input into the MLP network to calculate the predicted output. Then, the loss function value is calculated based on the predicted output and the true label, using the cross entropy loss function:

[0024]

[0025] Next, the gradient of the loss function with respect to the model parameters is calculated, and an optimizer is used to update the model parameters based on this gradient information, gradually reducing the loss function value. During training, a validation dataset is used to evaluate model performance, and model parameters are adjusted based on the validation results to prevent overfitting.

[0026] Furthermore, the step three includes the following steps:

[0027] 3-1) Construct human intelligent judgment rules, record θ r→b is θ1, θ b→rLet θ2 be the value, and the distance between the two drones be D. Random noise is added to the judgment boundary to simulate the randomness of human judgment and the fuzziness of boundary perception. The angle noise Δθ is set to U(-5,5) and the distance noise Δd is set to U(-50,50). The following judgment rule is constructed:

[0028]

[0029] 3-2) Human intelligence in the present invention is human judgment of x. Let human judgment result of x be Y. Human (x), the credibility is C Human (x), based on the probability distribution principle, human performance of x in each category can be expressed as:

[0030]

[0031] Furthermore, the step 4 includes the following steps:

[0032] 4-1) Construct two variational autoencoders VAE Machine and VAE Human , extract features from the judgment results of machine intelligence and human intelligence respectively, It is the cause of x and machine intelligence, VAE Machine The potential features extracted from the data, It is the cause of x and human intelligence Human, which is VAE Human Potential features for data extraction. It is VAE Machine The reconstructed output of It is VAE Human The training process is mainly divided into two stages. The first stage is the training and learning of the feature encoding representation part, based on x, P Machine and P Human By encoding the causal variables and For learning, the corresponding loss function is:

[0033]

[0034] 4-2) Construct two causal variational autoencoders CE-VAE Machine and CE-VAE Human , by setting some latent variables output by the encoder to 0, and finally reconstructing the original data through the decoder. If the reconstructed data is similar to the original data, it can be considered that the part set to 0 is an irrelevant factor. It is the cause of x and machine intelligence, which is CE-VAE Machine The latent variables extracted, The cause of x and human intelligence is CE-VAE Human The latent variables extracted, It is CE-VAE Machine The reconstructed output of It is CE-VAE Human The reconstructed output of . Its loss function consists of three parts: L VAE 、L CE-VAE And the mean square error (MSE) between the two reconstructed data.

[0035]

[0036] Furthermore, the step five includes the following steps:

[0037] 5-1) Data and the output P of machine intelligence respectively Machine and the output of human intelligence P Human Splice to get data set X Machine =[X,P Machine ] and X Human =[X,P Human ]. X Machine =[X,P Mac ] and X Human =[X,P Human ] is passed to the causal variational autoencoder CE-VAE Machine and CE-VAE Human , after the encoding layer generates the mean μ and logarithmic variance logσ 2 By reparameterization Get the characteristics Z of machine intelligence and human intelligence Machine and Z Human , mapping both directly to and The spliced ​​data is recorded as

[0038] 5-2) Construct the BBB Linear layer to learn the uncertainty of weights. Including μ w The mean of the weights), σ w (used to calculate weight standard deviation), μ b (biased mean), σ b Used to calculate the bias standard deviation) and prior (prior distribution of weights).

[0039] 5-3) Set up a Bayesian neural network based on BayesianMLP, with an input layer (receives input data of dimension input_dim), BBB Linear layer 1 (maps the input to dimension 360 and uses BBB Linear to learn the uncertainty of the weights, with the prior distribution specified by the priors parameter), a traditional linear layer fc1 (maps the input to dimension 200), a Softplus activation function (applied to the output of fc1), a traditional linear layer fc2 (maps dimension 200 to dimension 128), a Softplus activation function (applied to the output of fc2), and a BBB Linear layer 3 (maps dimension 128 to dimension output_dim, outputting out_1, out_2, and the KL divergence). During the forward propagation, the input data first passes through fc1 and fc2 and the Softplus activation function, then passes through blinear3, and the final output is the KL divergence of out_1 and out_2.

[0040] 5-4) Using Bayesian neural network, the data set The training process is based on the stochastic gradient descent algorithm, which minimizes the loss function in an iterative manner. The Evidence Lower Bound (ELBO) is used as the loss function:

[0041]

[0042] After calculating the loss, the backpropagation algorithm is executed to calculate the gradient of the loss function with respect to the model parameters. The optimizer then updates the model parameters based on this gradient information. After completing training for a batch, the loss for that batch is added to the total loss, and the model's predicted probability is recorded. Ultimately, the training process returns the average total loss, the total error cost, the total KL divergence cost, and the probability values ​​for all batches.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] (1) This paper uses a causal variational autoencoder (CE-VAE) to extract potential causal variables from the data, enabling the model to understand the real reasons behind the decision. This causal representation not only enhances the interpretability of the model, but also improves the robustness and generalization ability of the model under non-independent and identically distributed data.

[0045] (2) The three-layer model framework proposed in this paper places equal emphasis on machine intelligence and human intelligence, and fuses their outputs through a Bayesian neural network. This framework not only fully utilizes the powerful data processing capabilities of machine intelligence and the empirical knowledge of human intelligence, but also compensates for the shortcomings of traditional human-machine hybrid intelligence methods through causal representation. This multi-layered hybrid intelligence framework provides a new approach to solving complex problems.

[0046] (3) This invention uses a Bayesian neural network to fuse the outputs of human and machine intelligence, enabling the model to better handle noise and uncertainty in the data. This uncertainty modeling capability helps improve the reliability and stability of the model, especially in areas with high risk assessment requirements, such as military decision-making, and has important application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings are used to provide 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 of the present invention.

[0048] Figure 1 This is an overall flow chart of a method for identifying UAV combat intention based on causal representation provided by the present invention.

[0049] Figure 2 This is an overall framework diagram of a model based on causal representation in a human-machine hybrid intelligence framework provided by the present invention.

[0050] Figure 3 This is the average causal effect data diagram provided by the present invention.

[0051] Figure 4 The overall schematic diagram of the CE-VAE provided by the present invention

[0052] Figure 5 Schematic diagram of the present invention under non-intervention and intervention conditions. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, 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 intended to limit the present invention.

[0054] Example 1

[0055] This example provides a causal representation-based UAV combat intention recognition method to solve the non-independent and identically distributed problem. The specific steps are as follows.

[0056] Step 1) UAV combat intention data acquisition;

[0057] Step 2) Design and train a deep learning-based combat intent judgment model to obtain a machine intelligence model;

[0058] Step 3) Design of a combat intention judgment model based on human intelligence knowledge base to obtain human intelligence;

[0059] Step 4) Extract the causal features of machine intelligence and human intelligence respectively through causal representation;

[0060] Step 5) Build Bayesian deep learning to complete the causal feature fusion of machine intelligence and human intelligence, and output the final enemy drone intention judgment result.

[0061] Step 1: Obtain and process enemy and friendly drone data

[0062] 1-1) UAV data information acquisition, the drone information of the red and blue sides are R r =[x r ,y r ,z r ,v r ,γ r ,φ r ,nx r ,nz r ,roll r ] and R b =[x b ,y b ,z b ,v b ,γ b ,φ b ,nx b ,nz b ,roll b ], where the x and y coordinates range from -5000 to 5000, and the z coordinate ranges from 2000 to 4000; the velocity (v) is fixed at 150 units; the pitch angle (γ) is randomly generated between -200 and 200 and converted to radians; the heading angle (φ) is randomly generated between 0 and 200 and converted to radians; the acceleration (n x ,n z ) where n x Fixed to 0, n z Fixed to; Roll angle (roll): fixed to 0.

[0063] 1-2) Extract the position coordinate information of the red and blue drones, normalize the position coordinates, and divide the coordinate values ​​by 1000 to reduce the data level; calculate the velocity direction vectors of the red and blue drones based on the pitch and yaw angles of the drones. The calculation formula for the velocity direction vector is:

[0064] u r =[cos(γ r )·sin(φ r ),cos(γ r )·cos(φ r ),sin(γ r )] (1)

[0065] u b =[cos(γ b )·sin(φ b ),cos(γ b )·cos(φ b ),sin(γ b )] (2)

[0066] 1-3) Calculate the Euclidean distance d between the red drone and the blue drone = || p r -p b ||, where p r =[x r ,y r ,z r ],p b =[x b ,y b ,z b ] are the position information of the red and blue drones respectively, and the unit direction vector d of the red drone pointing to the blue drone is calculated. r→b =(p b -p r ) / (||p b -p r ||), the speed unit direction vector of the red team's drone Calculate the angle between the speed direction of the red drone and the direction the red drone points to the blue drone Calculate the unit direction vector d of the blue drone pointing to the red drone b→r =(p b -p r ) / (||p b -p r ||), the speed unit direction vector of the blue drone Calculate the angle between the speed direction of the blue drone and the direction of the blue drone pointing to the red drone

[0067] 1-4) Record the relative position information of the red and blue drones as follows: The information obtained is recorded as The feature vector is sliced ​​to filter out a specific number of data samples containing four labels, and then spliced ​​to generate a new dataset; the new dataset is randomly divided into training and test sets, and the data is shuffled during model training to ensure randomness and diversity of the data.

[0068] Step 2: Build a machine intelligence training model for training

[0069] 2-1) Build an MLP neural network for machine intelligence. The input layer receives the input vector inputs_dims. Fully connected layer 1 (fc1) maps the input dimension inputs_dims to dimension 200. Activation function: The softplus activation function is applied to the output of fc1. Fully connected layer 2 (fc2) maps dimension 200 to dimension 128. Activation function: The softplus activation function is applied to the output of fc2. Fully connected layer 3 (fc3) maps dimension 128 to dimension output_dims. Activation function: The sigmoid activation function is applied to the output of fc3 to produce the final output.

[0070] 2-2) Each iteration includes steps such as data loading, forward propagation, loss calculation, gradient calculation, and parameter update. First, a batch of data is extracted from the training set and input into the MLP network to calculate the predicted output. Then, the loss function value is calculated based on the predicted output and the true label, using the cross entropy loss function:

[0071]

[0072] Next, the gradient of the loss function with respect to the model parameters is calculated, and an optimizer is used to update the model parameters based on this gradient information, gradually reducing the loss function value. During training, a validation dataset is used to evaluate model performance, and model parameters are adjusted based on the validation results to prevent overfitting.

[0073] Step 3: Build human intelligence classification rules for data judgment

[0074] 3-1) Construct human intelligent judgment rules, record θ r→b is θ1, θ b→r Let θ2 be θ2, and the distance between the two drones be D. Add random noise to the judgment boundary to simulate the randomness of human judgment and the ambiguity of boundary perception. Set the angle noise Δθ to U(-5,5) and the distance noise Δd to U(-50,50). Construct the following judgment rule:

[0075]

[0076] 3-2) Human intelligence in the present invention is human judgment of x. Let human judgment result of x be Y. Human (x), the credibility is C Human (x), based on the probability distribution principle, human performance of x in each category can be expressed as:

[0077]

[0078] Step 4: Build CE-VAE to extract features from the dataset

[0079] 4-1) Characteristics It is the cause of x and machine intelligence, VAE Machine The potential features extracted from the data, It is the cause of x and human intelligence Human, which is VAE Human Potential features for data extraction. It is VAE Machine The reconstructed output of It is VAE Human The training process is mainly divided into two stages. The first stage is the training and learning of the feature encoding representation part, based on x, P Machine and P Human By encoding the causal variables and For learning, the corresponding loss function is:

[0080]

[0081] 4-2) Construct two causal variational autoencoders CE-VAE Machine and CE-VAE Human , by setting some latent variables output by the encoder to 0, and finally reconstructing the original data through the decoder. If the reconstructed data is similar to the original data, it can be considered that the part set to 0 is an irrelevant factor. It is the cause of x and machine intelligence, which is CE-VAE Machine The latent variables extracted, The cause of x and human intelligence is CE-VAE Human The latent variables extracted, It is CE-VAE Machine The reconstructed output of It is CE-VAE Human The reconstructed output of . Its loss function consists of three parts: L VAE 、L CE-VAE And the mean square error (MSE) between the two reconstructed data.

[0082]

[0083]

[0084] Step 5: Build a human-machine hybrid intelligent model based on causal representation, use the integrated data as input, and obtain the final prediction results

[0085] 5-1) Data and the output P of machine intelligence respectively Machine and the output of human intelligence PHuman Splice to get data set X Machine =[X,P Machine ] and X Human =[X,P Human ]. X Machine =[X,P Mac ] and X Human =[X,P Human ] is passed to the causal variational autoencoder CE-VAE Machine and CE-VAE Human , after the encoding layer generates the mean μ and logarithmic variance logσ 2 By reparameterization Get the characteristics Z of machine intelligence and human intelligence Machine and Z Human , mapping both directly to and The spliced ​​data is recorded as

[0086] 5-2) Construct the BBB Linear layer to learn the uncertainty of weights. Including μ w The mean of the weights), σ w (used to calculate weight standard deviation), μ b (biased mean), σ b Used to calculate the bias standard deviation) and prior (prior distribution of weights).

[0087] 5-3) Set up a Bayesian neural network based on BayesianMLP, with an input layer (receives input data of dimension input_dim), BBB Linear layer 1 (maps the input to dimension 360 and uses BBB Linear to learn the uncertainty of the weights, with the prior distribution specified by the priors parameter), a traditional linear layer fc1 (maps the input to dimension 200), a Softplus activation function (applied to the output of fc1), a traditional linear layer fc2 (maps dimension 200 to dimension 128), a Softplus activation function (applied to the output of fc2), and a BBB Linear layer 3 (maps dimension 128 to dimension output_dim, outputting out_1, out_2, and the KL divergence). During the forward propagation, the input data first passes through fc1 and fc2 and the Softplus activation function, then passes through blinear3, and the final output is the KL divergence of out_1 and out_2.

[0088] 5-4) Using Bayesian neural network, the data set The training process is based on the stochastic gradient descent algorithm, which iteratively minimizes the loss function. The loss function is as follows:

[0089]

[0090] After calculating the loss, the backpropagation algorithm is executed to calculate the gradient of the loss function with respect to the model parameters. The optimizer then updates the model parameters based on this gradient information. After completing training for a batch, the loss for that batch is added to the total loss, and the model's predicted probability is recorded. Ultimately, the training process returns the average total loss, the total error cost, the total KL divergence cost, and the probability values ​​for all batches.

[0091] Step 6) Compare the output results obtained by human intelligence, machine intelligence, and hybrid intelligence, as shown in Table 1 below:

[0092] Table 1 compares the accuracy, recall, precision and F1 score of the three models.

[0093]

[0094] Table 1 compares the accuracy, recall, precision, and F1 scores of the three models. C1, C2, C3, and C4 represent attack, reconnaissance, penetration, and retreat, respectively. Table 1 shows that the human-machine hybrid intelligence model based on causal representation performs better in most metrics. Specifically, the hybrid intelligence model achieves higher accuracy, recall, precision, and F1 scores across all categories than the individual human and machine intelligence models, and also outperforms the overall performance. This demonstrates that the hybrid intelligence model leverages the strengths of both human and machine intelligence, achieving superior predictive performance and explanatory power.

[0095] Example 2

[0096] This example provides a causal variational autoencoder (CE-VAE) to extract features of machine intelligence and human intelligence judgment of data. It includes the following steps:

[0097] Step 1) UAV data acquisition

[0098] Step 2) Build and train a deep learning model to achieve preliminary intelligent judgment of the drone's intentions.

[0099] Step 3) Extract expert knowledge, design human intelligence rules, and assist in intention judgment.

[0100] Step 4) Extract feature representations with causal relationships from machine intelligence and human intelligence respectively.

[0101] Step 5) Verify the effect of CE-VAE by using the Average Causal Effect (ACE)

[0102] Step 1), step 2), step 3) and step 4) are the same as those in the above embodiment 1

[0103] Step 5) Construct the average causal effect formula and conduct the generation experiment to observe the ELBO value

[0104] 5-1) If Figure 5 , remember Z n is the extracted random latent variable, U Z Represents the function of intervening on random latent variables, U X Represents the function that generates the data set, through P(Y=1|do(Z n =1)) represents the probability that the outcome variable Y is 1 under the treatment (intervention) condition, P(Y=1|do(Z n =0)) represents the probability that the outcome variable Y is 1 under the untreated (no intervention) condition, do(Z n =1) means to intervene in the variable X so that its value is 1, do(Z n =0) means intervening in the variable X so that its value is 0. The following formula is obtained:

[0105] ACE=P(Y=1|do(Z n =1))-P(Y=1|do(Z n =0)) (1)

[0106] The formula for calculating the probability after intervention is as follows:

[0107]

[0108] The latent variable extracted by CE-VAE is recorded as Z n , set Z n =z n and Z n =0 into formula (2), (1). Record the result as Figure 3 .

[0109] from Figure 3 As can be observed in the figure, the ACE value fluctuates somewhat under different numbers of training rounds. This fluctuation may be caused by a variety of factors, including the model's adaptability to the data at different training stages, the randomness of the training data, and the adjustment of model parameters. However, despite these fluctuations, the ACE value remains consistently between 0.004 and 0.007, indicating that the causal effect of the CE-VAE model varies within a small range, without experiencing drastic fluctuations or instability. The stability of the ACE value indicates that the causal effect of CE-VAE on the outcome variable Y is reliable under both treated (intervention) and untreated (no intervention) conditions. An ACE value close to 0 indicates that the latent variable is irrelevant and does not affect the causal effect.

[0110] 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 principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying UAV combat intention based on causal representation, characterized by: The following steps are involved: 1) Acquisition of UAV combat intention data; 2) Design and train a deep learning-based combat intent judgment model to obtain a machine intelligence model; 3) Design of a combat intention judgment model based on human intelligence knowledge base; 4) Extracting causal features of machine intelligence and human intelligence respectively through causal representation; 5) Construct Bayesian deep learning to complete the causal feature fusion of machine intelligence and human intelligence, and output the final enemy drone intention judgment result.

2. The method for identifying UAV combat intention based on causal representation according to claim 1 is characterized in that: In step 1), the drone data acquisition includes the following steps: S21, UAV data information acquisition, the red and blue UAV information are R r =[x r ,y r ,z r ,v r ,γ r ,φ r ,nx r ,nz r ,roll r ] and R b =[x b ,y b ,z b ,v b ,γ b ,φ b ,nx b ,nz b ,roll b ], where (x r ,y r ,z r ) represents the position information of the red team’s UAV in three-dimensional space, v r represents the speed of the red drone, γ r represents the pitch angle of the red drone, φ r Indicates the heading angle of the red team’s drone, (nx r ,nz r ) represents the control amount of the red drone, roll r Indicates the roll angle of the red team’s drone; (x b ,y b ,z b ) represents the position information of the blue UAV in three-dimensional space, v b represents the speed of the blue drone, γ b represents the pitch angle of the blue drone, φ b Indicates the heading angle of the blue drone, (nx b ,nz b ) represents the control amount of the blue drone, roll b Indicates the roll angle of the blue team's drone; S22. Calculate the velocity direction vectors of the red drone and the blue drone based on the pitch angle and yaw angle of the drone. The velocity direction vectors of the red and blue drones are u r and u b , the calculation formula of the velocity direction vector is: you r =[cos(γ r )·sin(φ r ),cos(γ r )·cos(φ r ),sin(γ r )] (1) you b =[cos(γ b )·sin(φ b ),cos(γ b )·cos(φ b ),sin(γ b )] (2) S23. Calculate the Euclidean distance d between the red drone and the blue drone = || p r -p b ||, where p r =[x r ,y r ,z r ],p b =[x b ,y b ,z b ] are the position information of the red and blue drones respectively, and the unit direction vector d of the red drone pointing to the blue drone is calculated. r→b =(p b -p r ) / (||p b -p r ||), the speed unit direction vector of the red team's drone Calculate the angle between the speed direction of the red drone and the direction the red drone points to the blue drone Calculate the unit direction vector d of the blue drone pointing to the red drone b→r =(p b -p r ) / (||p b -p r ||), the speed unit direction vector of the blue drone Calculate the angle between the speed direction of the blue drone and the direction of the blue drone pointing to the red drone S24. Record the relative position information of the red and blue drones as follows: The information obtained is recorded as The feature vector is sliced ​​to filter out a specific number of data samples containing four labels, and then spliced ​​to generate a new dataset; the new dataset is randomly divided into training and test sets, and the data is shuffled during model training to ensure randomness and diversity of the data.

3. The method for identifying UAV combat intention based on causal representation according to claim 1 is characterized in that: In step 2), the machine intelligence model construction and training includes the following steps: S31. Build an MLP neural network for machine intelligence. Input layer: receives input vector inputs_dims. Fully connected layer 1 (fc1): maps input dimension inputs_dims to dimension 200. Activation function: Softplus activation function is applied to the output of fc1. Fully connected layer 2 (fc2): maps dimension 200 to dimension 128. Activation function: Softplus activation function is applied to the output of fc2. Fully connected layer 3 (fc3): maps dimension 128 to dimension output_dims. Activation function: Sigmoid activation function is applied to the output of fc3 to obtain the final output. S32. Each iteration includes data loading, forward propagation, loss calculation, gradient calculation, and parameter update. First, a batch of data is extracted from the training set and input into the MLP network to calculate the predicted output. Then, the loss function value is calculated based on the predicted output and the true label, using the cross entropy loss function: Among them, N represents the batch size, C represents the number of categories, and y n,c represents the true label of the nth sample, P Machine (y n,c |X n ,w) represents the probability of the machine intelligence prediction model, given the input X n Under the conditions of and model parameters w, predict the probability that the nth sample belongs to c categories; Calculate the gradient of the loss function with respect to the model parameters, and use the optimizer to update the model parameters based on the gradient information so that the loss function value gradually decreases. During the training process, use the validation dataset to evaluate the model performance, and adjust the model parameters based on the validation results to prevent overfitting.

4. The method for identifying UAV combat intention based on causal representation according to claim 1 is characterized in that: In step 3), the design and judgment of human intelligent rules include the following steps: S41. Construct human intelligence judgment rules, record θ r→b is θ1, θ b→r Let θ2 be the value, and the distance between the two drones be D. Random noise is added to the judgment boundary to simulate the randomness of human judgment and the fuzziness of boundary perception. The angle noise Δθ is set to U(-5,5) and the distance noise Δd is set to U(-50,50). The following judgment rule is constructed: S42. Human intelligence is the human judgment of x. Let the human judgment result of x be Y. Human (x), the credibility is C Human (x), based on the probability distribution principle, human performance of x in each category can be expressed as: where f yc is a probability distribution model based on expert knowledge. is the probability predicted by human intelligence for the kth model.

5. The method for identifying UAV combat intention based on causal representation according to claim 1 is characterized in that: In step 4), causal feature extraction includes the following steps: S51. Build two variational autoencoders VAE Machine and VAE Human , extract features from the judgment results of machine intelligence and human intelligence respectively, It is the cause of x and machine intelligence, which is VAE Machine The potential features extracted from the data, It is the cause of x and human intelligence Human, which is VAE Human Potential features of data extraction, It is VAE Machine The reconstructed output of It is VAE Human The reconstructed output of x, P is mainly divided into two stages. The first stage is the training and learning of the feature encoding representation part. Machine and P Human By encoding the causal variables and For learning, the corresponding loss function is: Among them, E qφ(Z|x,P) [logP θ (X|Z)] represents the expectation of the reconstruction loss in the VAE loss function, which is the posterior distribution q of the latent variable Z given the input x and the machine / human intelligence prior P. φ Find logP θ The expectation of (X|Z), p θ (X|Z) is the conditional probability distribution learned by the decoder, which represents the probability of reconstructing the output X under the condition of the given latent variable Z, P θ (P|Z) represents the probability of predicting machine / human intelligence prior information P under the condition of the latent variable Z, KL(q φ (Z|X,P)||p(Z)) is the KL divergence, which measures the posterior distribution q learned by the encoder φ and the difference between the prior distribution P(Z); S52. Construct two causal variational autoencoders CE-VAE Machine and CE-VAE Human , by setting some latent variables output by the encoder to 0, and finally reconstructing the original data through the decoder. If the reconstructed data is similar to the original data, it is considered that the part set to 0 is an irrelevant factor. It is the cause of x and machine intelligence, which is CE-VAE Machine The latent variables extracted, The cause of x and human intelligence is CE-VAE Human The latent variables extracted, It is CE-VAE Machine The reconstructed output of It is CE-VAE Human The reconstruction output of the loss function consists of three parts: L VAE , L CE-VAE And the mean square error MSE between the two reconstructed data; Where α and γ are weight coefficients.

6. The method for identifying UAV combat intention based on causal representation according to claim 1 is characterized in that: In step 5), the construction and integration of the human-machine hybrid intelligent model includes the following steps: S61, the data and the output P of machine intelligence respectively Machine and the output of human intelligence P Human Splice to get data set X Machine =[X,P Machine ] and X Human =[X,P Human ], respectively and X Human =[X,P Human ] is passed to the causal variational autoencoder CE-VAE Machine and CE-VAE Human , after the encoding layer generates the mean μ and logarithmic variance logσ 2 , by reparameterization Get the characteristics Z of machine intelligence and human intelligence Machine and Z Human , mapping both directly to and The spliced ​​data is recorded as S62, build BBBLinear layer to learn the uncertainty of weights, including μ w The mean of the weights, σ w Used to calculate weight standard deviation, μ b The mean of the bias, σ b Used to calculate the prior distribution of bias standard deviation and prior weight; S63. A Bayesian neural network is established based on BayesianMLP, which includes the input layer, BBB Linear layer 1, traditional linear layer fc1, Softplus activation function, traditional linear layer fc2, Softplus activation function, and BBB Linear layer 3. During the forward propagation, the input data first passes through fc1 and fc2 and the Softplus activation function, then passes through blinear3, and the final output is the KL divergence of out_1 and out_2. S64. Using Bayesian neural network, the data set The training process is based on the stochastic gradient descent algorithm, which minimizes the loss function in an iterative manner. The evidence lower bound ELBO is used as the loss function: Among them, q(ω) represents the posterior distribution of the model parameters, p(ω) represents the prior distribution of the model parameters, KL[q(ω)||p(ω)] represents the difference between q(ω) and p(ω), and E q(ω) [lnp(D|ω)] represents the expectation of the likelihood function lnp(D|ω) under the approximate posterior distribution q(ω), n and γ represent weight functions. After calculating the loss, the backpropagation algorithm is executed to calculate the gradient of the loss function with respect to the model parameters, and the optimizer is used to update the model parameters according to the gradient information. After completing the training of a batch, the loss value of the batch is added to the total loss, and the probability value predicted by the model is recorded. Finally, the training process returns the average total loss, total error cost, total KL divergence cost, and the probability value of all batches.

Images