Man-machine hybrid intelligent unmanned aerial vehicle combat intention judgment method based on causal inference

Abstract

The invention provides a man-machine hybrid intelligent unmanned aerial vehicle combat intention judgment method based on causal inference, belongs to the technical field of unmanned aerial vehicle intention judgment, and solves the technical problem of non-independent identically distributed training data. According to the technical scheme, human expert knowledge and a machine learning algorithm are combined, causal and non-causal factors are separated through a structural causal model (SCM), feature selection is optimized by means of a mask module combined with dynamic feature selection, an attention mechanism and a differentiable sampling method, the causal relationship after unmanned aerial vehicle decision making is revealed, and the unmanned aerial vehicle decision making efficiency is improved. By exploring the internal causal relation between the data and the labels, the recognition capability of the model for a few categories is improved, the generalization capability of the model in different environments is enhanced, and misprediction caused by environmental deviation is avoided. According to the method, in a complex and changeable battlefield environment, accurate judgment of the combat intention of the unmanned aerial vehicle is achieved, and higher flexibility and robustness are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of UAV intention judgment, and in particular to a method for judging the combat intention of a human-machine hybrid intelligent UAV based on causal inference. Background Art

[0002] At present, drones are widely used in reconnaissance, surveillance, communication relay, strike and other fields due to their advantages such as high mobility, low cost, flexibility and real-time performance, especially in the military and security industries, and have achieved remarkable results. The modern battlefield environment is becoming increasingly complex, the enemy and our situation is changing rapidly, and the amount of battlefield information is huge and dynamic. Therefore, accurately judging the intentions of enemy drones plays a vital role in combat. Accurately judging the intentions of enemy drones can help commanders make tactical decisions quickly, thereby gaining the upper hand in air combat.

[0003] In order to accurately judge the intention of drones, it is necessary to comprehensively use multiple data sources, such as sensor data, behavior pattern recognition, and tactical environment analysis. Especially in air combat environments, judging the intention of enemy drones usually depends on accurate analysis of enemy aircraft behavior patterns and real-time perception of the battlefield environment. However, existing intention judgment algorithms focus on the correlation of data and lack in-depth mining of causal relationships. They often lead to overfitting due to lack of understanding of causal relationships, making it difficult to ensure the reliability of decisions. The algorithms based on confidence rule base and evidential reasoning mentioned in the article "Air target intention recognition method based on confidence rule base and evidential reasoning" focus on identifying the behavior patterns of targets from historical data and predicting the future behavior of targets through correlation. In addition, most existing intention judgment algorithms are pure machine intelligence methods, which are not adaptable enough in dynamic environments and cannot flexibly respond to emergencies. In the article "An Air Target Tactical Intention Recognition Model Based on Bidirectional GRU With Attention Mechanism", the bidirectional gated recurrent unit (BiGRU) and the attention mechanism are mainly used to process the time series data of air combat intention features, capture the dynamic changes of target tactical intentions, and require a large amount of high-quality time series data. In the case of insufficient data or incomplete information, they lack the ability to respond flexibly and are prone to misjudgment when facing unknown situations. Therefore, the present invention is to solve the problems of poor adaptability of existing intention judgment algorithms in dynamic air combat environments, lack of deep mining of causal relationships, and lack of flexible response capabilities in the case of insufficient data or incomplete information. Summary of the invention

[0004] The purpose of the present invention is to propose a method for judging the combat intention of a human-machine hybrid intelligent UAV based on causal inference, which is used for decision support and intention judgment of UAVs in complex battlefield environments. By combining causal inference with human-machine hybrid intelligence, it can not only handle the complexity and dynamic changes of data, but also effectively deal with technical problems of small amounts of data and non-independent and identically distributed data, thereby improving the accuracy of intelligent decision-making and intention judgment of UAVs in real-time combat scenarios, improving combat effectiveness and reducing potential losses. The present invention is particularly suitable for responding to emergencies and highly complex mission scenarios, providing accurate and reliable auxiliary support for military decision makers.

[0005] The idea of ​​the present invention is to use the situation information of drones to divide the battlefield situation into four categories, and group the data according to preset classification rules. By establishing a human-machine hybrid intelligent model based on causal inference, the intention of the enemy drone can be accurately judged, and the accuracy of each category of judgment can be calculated. By analyzing the accuracy of different categories of intention judgment, the model effect of the causal inference human-machine hybrid intelligent network is evaluated. The present invention gives drones the ability to judge the enemy's combat intentions and provides an accuracy evaluation of the results.

[0006] The present invention is achieved through the following measures: a method for judging the combat intention of a human-machine hybrid intelligent unmanned aerial vehicle based on causal inference, comprising the following steps:

[0007] S1. Acquisition of situation information of enemy and friendly drones, data processing and data enhancement;

[0008] S2. Build a causal inference-based model for intention judgment to obtain human characteristics and machine characteristics;

[0009] S3. Build a human-machine hybrid intelligent model based on causal inference, integrate human characteristics and machine characteristics, and use this model to obtain the accuracy of judging the final intention of the enemy drone;

[0010] Furthermore, the step 1 comprises the following steps:

[0011] 1-1) Obtain the situation information of the red and blue drones. The situation information of the red and blue drones are: R r =[x r ,y r ,z r ,v r ,γ r ,nx r ,nz r ,roll r ], R b =[x b ,y b ,z b,v b ,γ b ,nx b ,nz b ,roll b ], where x r ,x b ,y r ,y b ,z r ,z b are the location information of the UAV, γ r ,γ b is the pitch angle, is the yaw angle, nx r ,nx b , nz r ,nz b To control the amount, roll r ,roll b is the rolling angle. The acquired information is preprocessed to obtain comprehensive situation information in, is the relative position information of the two UAVs, expressed as 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 are the velocity components of the blue drone's velocity vector in the x, y, and z directions.

[0012] 1-2) Through randomly generated data i=1,2…M, j=1,2…M, M is the total number of samples, which constitutes X o ={X i}, X oo ={X j}. In X o Select a sample in order, and then select a sample from X oo Select the data with the same label as the sample, find the 10 closest data according to the Euclidean distance, randomly select one to implement data enhancement, and the generated enhanced data is represented as X a .

[0013] Furthermore, the step 2 comprises the following steps:

[0014] 2-1) Construct an MLP network for X o and X a Encode and get features

[0015] Z o and Z a Perform normalization to obtain normalized features pass and Get the cross-correlation matrix and define the separation loss as the diagonal loss of the cross-correlation matrix and off-diagonal losses The weighted sum of ensures that the model reduces the correlation between different features while maintaining the high correlation of the same feature. The calculation formula is:

[0016]

[0017] 2-2) Construct a neural network-based masker to detect and enhance feature dimensions that have a significant causal impact on intent judgment. The masker consists of two main parts: an attention mechanism layer and a mask generation layer. First, in order to prevent overfitting, the masker dynamically adjusts the number of selected features according to the training progress, k init is the number of feature selections defined during initialization, t is the current training round, and E is the total round. k dynamic As training progresses, the model will gradually select fewer features to ensure that only one feature is selected at the end of training. The calculation method is as follows:

[0018]

[0019] The attention mechanism is used to calculate the importance of each feature, emphasizing the causal features that contribute to the intention judgment and enhancing the model's attention to important features. First, the attention weight is calculated, and then the attention weight is applied to the input feature using element-by-element multiplication to obtain the weighted feature z′ i , and then the initial mask m is obtained through the mask generation network layer. In the mask generation process, the Gumbel Softmax method is used to dynamically adjust the temperature parameter τ, so that the mask generation is both differentiable and gradually tends to be sparse. Define a tensor q with the same shape as the mask m tensor and zero content for k dynamic After Gumbel Softmax sampling, q can be gradually updated to the maximum feature selected in the iterative process. This is the process of mask updating, which can be expressed as:

[0020]

[0021] 2-3) Multiply the features generated by the encoding network by the mask m and 1-m to obtain the quality dimension of the features. Design two classifiers f 1 (·) and f 2 (·), put the good and bad dimensions into these two classifiers, and get two classification losses,

[0022]

[0023] Defining the contrastive loss function Enhance the model's ability to distinguish samples of the same type from samples of different types in the feature space. And corresponding labels and The contrast loss formula is:

[0024]

[0025] in is the L2 norm (Euclidean distance), which is used to measure the distance between two samples. margin is used to specify the minimum distance between samples of different classes. This method defines a total loss function to optimize the encoder and two classifiers, and the formula is:

[0026]

[0027] Calculate the weight λ, which is used to adjust the overall loss function during training. This weight gradually increases during the training process.

[0028] 2-4) Sample each set of features multiple times and calculate the mean of the output to improve the robustness of the model. In order to impose sparsity constraints on the encoder and reduce the weights of unimportant features, L1 regularization is added And to prevent the model from overfitting, add L2 regularization In addition, the weight coefficients of L1 and L2 regularization are dynamically adjusted with the training cycle.

[0029] Therefore, a loss function is defined to optimize the encoder, and the calculation formula is:

[0030]

[0031] 2-5) Use the training set data to train the model, extract and save the features in the training process. Evaluate the model accuracy on the test data set, and predict the output intent and classification accuracy. Store the model when the best accuracy is achieved in the verification.

[0032] Furthermore, the step three comprises the following steps:

[0033] 3-1) Load the machine intelligence model and human intelligence model based on causal inference, define the classification model, and the network structure contains a common feedforward layer (fc1, fc2) and a Bayesian linear layer (blinear3). In the forward propagation method, the log-likelihood and the KL divergence of the difference between the posterior distribution and the prior distribution of the measurement model are calculated, and the evidence lower bound ELBO is defined as the loss function. The weighted combination of the log-likelihood and the KL divergence is used to calculate the loss through multiple Monte Carlo sampling forward propagation;

[0034] 3-2) Splicing the machine features and human features obtained through the machine intelligence model based on causal inference and the human intelligence model based on causal inference to provide input for the human-machine hybrid intelligence model based on causal inference;

[0035] 3-3) The model performs 20 forward propagations on the input features. Each forward propagation produces an output, which is the classification result of the current sample. After all samples are completed, the output is averaged in the first dimension (20 samples) to obtain the final score, which is used to calculate the classification accuracy and loss;

[0036] 3-4) During the training process, if the accuracy of the validation set improves, save the current model. The training loss, validation loss, and accuracy are stored for subsequent analysis.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] (1) The present invention has the ability to process data that is not independent and identically distributed, which is usually difficult to achieve in traditional models. Through causal inference, the causal relationship between data can be discovered instead of just focusing on the correlation between data, which can improve the accuracy of small class data.

[0039] (2) The present invention can effectively grasp the accuracy and reliability of battlefield situation perception and decision-making. By modeling the combat intention of the drone and combining the situation information and relative position, the enemy's combat intention can be accurately inferred. At the same time, the integration of machine intelligence and human intelligence can further improve the robustness of intention judgment.

[0040] (3) The present invention combines human expert knowledge with machine learning algorithms to separate causal and non-causal factors through a structural causal model (SCM), and optimizes feature selection with the help of a mask module that combines dynamic feature selection, attention mechanism, and differentiable sampling method, thereby revealing the causal relationship behind drone decision-making and solving the problem of non-independent and identically distributed training data.

[0041] (4) By exploring the inherent causal relationship between data and labels, the present invention improves the model's ability to identify minority categories, enhances the model's generalization ability in different environments, and avoids incorrect predictions caused by environmental bias. This method achieves accurate judgment of drone combat intentions in a complex and changeable battlefield environment, with higher flexibility and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] 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.

[0043] Figure 1 This is an overall flow chart of the method for judging the combat intention of a human-machine hybrid intelligent UAV based on causal inference provided by the present invention.

[0044] Figure 2 The overall framework diagram of the model based on causal inference provided by the present invention.

[0045] Figure 3 This is an overall framework diagram of the human-machine hybrid intelligent model based on causal inference provided by the present invention.

[0046] Figure 4 A human-machine hybrid intelligent causal diagram is provided for the present invention. DETAILED DESCRIPTION

[0047] 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.

[0048] Example 1

[0049] This example provides a method for determining the combat intention of a human-machine hybrid intelligent UAV based on causal inference, which includes the following steps:

[0050] Step 1) Acquisition of situation information of enemy and friendly drones, data processing and data enhancement;

[0051] Step 2) construct a causal inference-based model for intention judgment to obtain human characteristics and machine characteristics;

[0052] Step 3) construct a human-machine hybrid intelligent model based on causal inference, integrating human characteristics and machine characteristics. Through this model, the accuracy of judging the final intention of the enemy UAV is obtained;

[0053] Step 4) Compare the models and analyze the accuracy, F1 score, and recall rate of each model;

[0054] Step 1) Obtain the situation information of the enemy and our drones, process the data and enhance the data. The specific steps are as follows:

[0055] 1-1) Obtain the situation information of the red and blue drones. The situation information of the red drone: R r =[x r ,y r ,z r ,v r ,γ r ,nx r ,nz r ,roll r ], the situation information of the blue drone is recorded as: B b =[x b ,y b ,z b ,v b ,γ b ,nx b ,nz b ,roll b ], where x r ,x b ,y r ,y b ,z r ,z b are the location information of the UAV, γ r ,γ b is the pitch angle, is the yaw angle, nx r ,nx b , nz r ,nz b To control the amount, roll r ,roll b is the rolling angle. The acquired information is preprocessed to obtain comprehensive situation information in, is the relative position information of the two UAVs, expressed as 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 are the velocity components of the blue drone's velocity vector in the x, y, and z directions.

[0056] 1-2) Through randomly generated data i=1,2…M, j=1,2…M, M is the total number of samples, which constitutes X o ={X i}, Xoo ={X j}.

[0057] 1-2) In X o Select a sample in order, and then select a sample from X oo Select the data with the same label as the sample, find the 10 closest data according to the Euclidean distance, and randomly select one of them. It is necessary to ensure that the enhanced data corresponds to the original data one by one and the order must not be disrupted. The generated enhanced data is represented by X a , the formula for data enhancement is expressed as:

[0058] X a ={x j ∈X oo ∣y j =y i}

[0059] Where X oo Represents another data set, x j ,y j Represents the data and labels selected from the dataset, y i Represents the labels of the original data.

[0060] Step 2) Build a causal inference model for intention judgment to obtain human characteristics and machine characteristics. The specific steps are as follows:

[0061] 2-1) The overall framework of the model based on causal inference for intention judgment is as follows Figure 2 As shown, construct an MLP network and o and X a Encode and get features Z o and Z a Perform normalization processing, the normalization formula is:

[0062]

[0063] in is the feature of the original data after normalization, is the feature of the normalized enhanced data, μ o To calculate the average value of all features of the original data before normalization, μ a To calculate the average value of all features of the enhanced data before normalization, σ o To calculate the standard deviation of all features of the original data before normalization, σ a It is to calculate the standard deviation of all features of the enhanced data before normalization. The separation loss is defined as the diagonal loss and off-diagonal losses The weighted sum of is used to ensure that the model reduces the correlation between different features while maintaining the high correlation of the same feature. The calculation formula is:

[0064]

[0065] Among them, the diagonal loss Used to measure the correlation between the same features, non-diagonal loss Used to measure the independence between different features, the calculation formulas are:

[0066]

[0067] The related cross-correlation matrix formula can be expressed as:

[0068]

[0069] 2-2) Build a neural network-based masker to detect and enhance feature dimensions that have significant causal effects on the intent judgment task. The masker consists of two main parts: one is the attention mechanism layer, which is used to assign a weight to each input feature to enhance the model's attention to important features. The other is the mask generation layer, which generates the final mask through operations such as fully connected layers, Batch Normalization, ReLU, and Dropout. First, in order to prevent overfitting, the masker dynamically adjusts the number of selected features according to the training progress, k init The number of feature selections defined during initialization is 1024, t is the current training epoch, and E is the total epoch. k dynamic It will gradually decrease with the training rounds. As the training progresses, the model will gradually select fewer features for processing, ensuring that only one feature is selected at the end of training. It is calculated as follows:

[0070]

[0071] The importance of each feature is calculated using the attention mechanism or causal weight, emphasizing the causal features that contribute to the intention judgment. The attention weight a can be calculated as:

[0072] a=Softmax(W 2 tanh(W 1 z i ))

[0073] in is the input feature vector, dimension is N, is the weight matrix of the attention layer. Tanh is the hyperbolic tangent activation function. Softmax is the normalization of each feature dimension. Then the attention weights are applied to the input features using element-by-element multiplication to obtain the weighted feature z′ i , and then the initial mask m is obtained through the mask generation network layer. In the mask generation process, the Gumbel Softmax method is used to dynamically adjust the temperature parameter τ, so that the mask generation is both differentiable and gradually tends to be sparse. The dynamic adjustment formula of the temperature parameter is:

[0074]

[0075] Define a tensor q with the same shape as the mask m tensor and zero content for k dynamic After GumbelSoftmax sampling, q can be gradually updated to the maximum feature selected in the iterative process. This is the mask update process, and the formula is designed as follows:

[0076]

[0077] 2-3) Multiply the representation generated by the encoding network by the mask m and 1-m to obtain the quality dimension of the representation. Design two classifiers f 1 (·) and f 2 (·), put the good and bad dimensions into these two classifiers, and get two classification losses,

[0078]

[0079] Defining the contrastive loss function Enhance the model's ability to distinguish samples of the same type from samples of different types in the feature space. And corresponding labels and The contrast loss formula is:

[0080]

[0081] in is the L2 norm (Euclidean distance), which is used to measure the distance between two samples. Margin is used to specify the minimum distance between samples of different classes. This method defines a total loss function to optimize the encoder and two classifiers, and the formula is:

[0082]

[0083] Calculate the weight λ, which is used to adjust the overall loss function during training. This weight gradually increases during the training process.

[0084] 2-4) Sample each set of features multiple times and calculate the mean of the output to improve the robustness of the model. In order to impose sparsity constraints on the encoder and reduce the weights of unimportant features, L1 regularization is added In addition, in order to prevent the model from overfitting, L2 regularization is designed It improves the generalization ability of the model by suppressing the magnitude of the weights. The calculation formulas for L1 and L2 are:

[0085]

[0086] where ω i is the i-th parameter of the network weight. λ(t) is the weight coefficient of L1 and L2 regularization, which is dynamically adjusted with the training cycle. The calculation formula is as follows:

[0087]

[0088] where λ start is the initial weight. end is the final weight. t is the current training cycle (epoch). T is the total number of training cycles (total number of epochs).

[0089] Therefore, a loss function is designed to optimize the masker, expressed as:

[0090]

[0091] 2-5) Use the training set data to train the causal inference-based model, extract and save the features generated during the training process. Evaluate the accuracy of the model on the test data set, make data predictions, and return the output intent and classification accuracy. Store the model when the best accuracy is achieved in the validation.

[0092] Step 3) Build a human-machine hybrid intelligent model based on causal inference, integrating human characteristics and machine characteristics. Through this model, the accuracy of judging the final intention of the enemy drone is obtained. The specific framework is as follows: Figure 3 As shown in the causal diagram, Figure 4 As shown, the specific steps are as follows:

[0093] 3-1) Load the machine intelligence model based on causal inference and the human intelligence model based on causal inference, define the classification model, and the network structure contains a common feedforward layer (fc1, fc2) and a Bayesian linear layer (blinear3). In the forward propagation method, the log-likelihood and the KL divergence of the difference between the posterior distribution and the prior distribution of the measurement model are calculated, and the evidence lower bound ELBO is defined as the loss function. The weighted combination of the log-likelihood and the KL divergence is used to calculate the loss through multiple Monte Carlo sampling forward propagation;

[0094] 3-2) Splicing the machine features and human features obtained through the machine intelligence model based on causal inference and the human intelligence model based on causal inference to provide input for the human-machine hybrid intelligence model based on causal inference;

[0095] 3-3) The model performs 20 forward propagations on the input features. Each forward propagation produces an output, which is the classification result of the current sample. After all samples are completed, the output is averaged in the first dimension (20 samples) to obtain the final score, which is used to calculate the classification accuracy and loss;

[0096] 3-4) During the training process, if the accuracy of the validation set improves, save the current model. The training loss, validation loss, and accuracy are stored for subsequent analysis.

[0097] Step 5) Compare the models and analyze the accuracy, F1 score, and recall rate of each model. The specific steps are as follows:

[0098] Table 1 compares the precision, recall, and F1 scores of the three models in different categories (such as attack, reconnaissance, penetration, and retreat). The CI_BNN_HM model with human-machine hybrid optimization generally outperforms the other two models, has a lower misclassification rate than other models, and does not perform poorly in a specific category. Even in the "penetration" task with a small number of samples, it shows high precision, recall, and F1 scores, especially a significant improvement in recall. This shows that the CI_BNN_HM model can effectively improve the accuracy of a class of judgments with a small amount of data when dealing with non-independent and identically distributed data.

[0099] Table 1 Comparison of three models

[0100]

[0101] Based on the above comparison, it can be found that the human-machine hybrid model based on causal inference proposed in this embodiment is significantly better than other models in judging the intention of drones, especially in dealing with data imbalance and complex dependencies.

[0102] Example 2

[0103] Step 1) Acquisition of situation information of enemy and friendly drones, data processing and data enhancement;

[0104] Step 2) construct a traditional Bayesian neural network (BNN) and a multi-layer perceptron (MLP), save the trained network model and parameters, and obtain the accuracy of judging the final intention of the enemy UAV;

[0105] Step 3) Compare the judgment results of the traditional neural network with those of the human-machine hybrid intelligent model based on causal inference, and analyze the accuracy, F1 score, and recall rate of each model;

[0106] Step 1) is the same as in Example 1 above.

[0107] Step 2) Construct a traditional Bayesian neural network (BNN) and a multi-layer perceptron (MLP), save the trained network model and parameters, and obtain the accuracy of judging the final intention of the enemy drone. The specific steps are as follows:

[0108] 2-1) The Bayesian neural network (BNN) network structure has three layers. The first one is the ordinary fully connected layer nn.Linear(input_dim,200), the second one is nn.Linear(200,128), and the third one is the Bayesian linear layer BBBLinear(128,output_dim). In the forward propagation method, the log-likelihood and the KL divergence of the posterior distribution and prior distribution difference of the measurement model are calculated. The evidence lower bound ELBO is defined as the loss function. The weighted combination of the log-likelihood and KL divergence is used to calculate the loss through multiple Monte Carlo sampling forward propagation, and the trained parameters and structure are saved.

[0109] 2-2) The multilayer perceptron (MLP) network structure has three layers. The first one is a normal fully connected layer nn.Linear(inputs_dims,150), the second one is nn.Linear(150,64), and the third one is nn.Linear(64,outputs_dims). The cross entropy loss is used as the loss function, and the trained parameters and structure are saved.

[0110] Step 3) Compare the judgment results of the traditional neural network with those of the human-machine hybrid intelligent model based on causal inference, and analyze the accuracy, F1 score, and recall rate of each model. The specific steps are as follows:

[0111] Table 2 compares the precision, recall, and F1 scores of the three models in different categories (such as attack, reconnaissance, penetration, and retreat). The CI_BNN_HM model generally outperforms the two traditional neural network models and performs best overall in various tasks. Its precision, recall, and F1 scores are all improved compared to other models, especially in the penetration category with a small number of models, highlighting its advantages in small sample scenarios. This shows that the CI_BNN_HM model can effectively improve the accuracy of a type of judgment with a small amount of data when dealing with non-independent and identically distributed data.

[0112] Table 2 Comparison with traditional models

[0113]

[0114] Example 3

[0115] Step 1) Acquisition of situation information of enemy and friendly drones, data processing and data enhancement;

[0116] Step 2) construct a causal inference-based model for intention judgment to obtain human characteristics and machine characteristics;

[0117] Step 3) construct a human-machine hybrid intelligent model based on causal inference, integrating human characteristics and machine characteristics. Through this model, the accuracy of judging the final intention of the enemy UAV is obtained;

[0118] Step 4) Compare the models and analyze the accuracy, F1 score, and recall rate of each model;

[0119] The steps are the same as those in Example 1, but the simulation data used are different, as shown in the following table:

[0120] Table 3 records the precision, recall and F1 scores of the three models in different categories (such as attack, reconnaissance, penetration and retreat). The CI_BNN_HM model outperforms the other two models (CI_MLP_M and CI_MLP_H) in overall performance, with higher precision, recall and F1 scores, smaller fluctuations and better stability. In the small sample task penetration, the F1 score of CI_BNN_HM (0.8175) is significantly higher than that of CI_MLP_H (0.7402), slightly better than CI_MLP_M (0.8003), and the recall rate of CI_BNN_HM (0.8881) is slightly higher than that of CI_MLP_M (0.8626), and significantly higher than that of CI_MLP_H (0.8453), reflecting good small sample adaptability and can effectively improve the accuracy of categories with less data.

[0121] Table 3 Model comparison table

[0122]

[0123] 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 method for judging the combat intention of a human-machine hybrid intelligent UAV based on causal inference, comprising the following steps: 1) Acquisition of situational information of enemy and friendly drones, data processing and data enhancement; 2) Build a causal inference model for intention judgment to obtain human characteristics and machine characteristics; 3) Construct a human-machine hybrid intelligent model based on causal inference, through which human characteristics and machine characteristics are integrated to obtain the accuracy of judging the ultimate intention of the enemy drone.

2. The method for judging combat intention of human-machine hybrid intelligent UAV based on causal inference 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 red and blue drones. The situation information of the red and blue drones are: R r =[x r ,y r ,z r ,v r ,γ r ,nx r ,nz r ,roll r ], R b =[x b ,y b ,z b ,v b ,γ b ,nx b ,nz b ,roll b ], where x r ,x b ,y r ,y b ,z r ,z b are the location information of the UAV, γ r ,γ b is the pitch angle, is the yaw angle, nx r ,nx b , nz r ,nz b To control the amount, roll r ,roll b is the rolling angle; the acquired information is preprocessed to obtain the comprehensive situation information in, is the relative position information of the two UAVs, expressed as 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 are the velocity components of the blue drone’s velocity vector in the x, y, and z directions; S22, through randomly generated data M is the total number of samples, which constitutes X o ={X i }, X oo ={X j }, in X o Select a sample in order, and then select a sample from X oo Select the data with the same label as the sample, find the 10 closest data according to the Euclidean distance, randomly select one to implement data enhancement, and the generated enhanced data is represented as X a .

3. The method for judging the combat intention of a human-machine hybrid intelligent UAV based on causal inference according to claim 1 is characterized in that, in the step 2), a model based on causal inference for intention judgment is constructed to obtain human characteristics and machine characteristics, comprising the following steps: S31, build MLP network, for X o and X a Encode and get features Z o and Z a Perform normalization to obtain normalized features pass and Get the cross-correlation matrix and define the separation loss as the diagonal loss of the cross-correlation matrix and off-diagonal losses The weighted sum of ensures that the model reduces the correlation between different features while maintaining the high correlation of the same feature. The calculation formula is: S32. Construct a neural network-based masker to detect and enhance feature dimensions that have significant causal effects on intent judgment. The masker consists of two parts: an attention mechanism layer and a mask generation layer. First, to prevent overfitting, the masker dynamically adjusts the number of selected features according to the training progress, k init is the number of feature selections defined during initialization, t is the current training round, E is the total round, k dynamic As training progresses, the model will gradually select key features to ensure that only one feature is selected at the end of training. The calculation method is as follows: The attention mechanism is used to calculate the importance of each feature, emphasizing the causal features that contribute to the intention judgment and enhancing the model's attention to important features. First, the attention weight is calculated, and then the attention weight is applied to the input feature using element-by-element multiplication to obtain the weighted feature z′. i , and then the initial mask m is obtained through the mask generation network layer. In the mask generation process, the Gumbel Softmax method is used to dynamically adjust the temperature parameter τ, so that the mask generation is both differentiable and gradually tends to be sparse. A tensor q with the same shape as the mask m tensor and zero content is defined for k dynamic After Gumbel Softmax sampling, q can be gradually updated to the maximum feature selected in the iterative process. This is the process of mask updating, which is expressed as: S33, multiply the features generated by the encoding network by the mask m and 1-m to obtain the quality dimension of the features, design two classifiers f1(·) and f2(·), put the quality dimension into these two classifiers, and obtain two classification losses, Defining the contrastive loss function Enhance the model's ability to distinguish samples of the same type from samples of different types in the feature space. And corresponding labels and The contrast loss formula is: in is the L2 norm, which is used to measure the distance between two samples. Margin is used to specify the minimum distance between samples of different classes. A total loss function is defined to optimize the encoder and two classifiers. The formula is: Calculate the weight λ, which is used to adjust the overall loss function during training. This weight gradually increases during the training process. S34. Sample each set of features multiple times and calculate the mean of the output to improve the robustness of the model. In order to impose sparsity constraints on the encoder and reduce the weights of unimportant features, L1 regularization is added. And to prevent the model from overfitting, add L2 regularization Moreover, the weight coefficients of L1 and L2 regularization are dynamically adjusted with the training cycle; Therefore, a loss function is defined to optimize the masker, and the calculation formula is: S35. Train the model using the training set data, extract and save the features in the training process, evaluate the model accuracy on the test data set, predict and return the output intent and classification accuracy, and store the model when the best accuracy is achieved during verification.

4. The method for judging combat intention of human-machine hybrid intelligent UAV based on causal inference according to claim 1 is characterized in that: In step 3), a human-machine hybrid intelligent model based on causal inference is constructed to integrate human characteristics and machine characteristics. Through the model, the accuracy of judging the final intention of the enemy UAV is obtained, including the following steps: S41. Load the machine intelligence model and human intelligence model based on causal inference, define the classification model, and the network structure includes a common feedforward layer (fc1, fc2) and a Bayesian linear layer. In the forward propagation method, the log-likelihood and the KL divergence of the difference between the posterior distribution and the prior distribution of the measurement model are calculated. The evidence lower bound ELBO is defined as the loss function, and the weighted combination of the log-likelihood and the KL divergence is used to calculate the loss through forward propagation through multiple Monte Carlo samplings. S42, splicing the machine features and human features obtained by the machine intelligence model based on causal inference and the human intelligence model based on causal inference to provide input for the human-machine hybrid intelligence model based on causal inference; S43, the model performs 20 forward propagations on the input features, and each forward propagation obtains an output, that is, the classification result of the current sample. After all sampling is completed, the output is averaged in the first dimension to obtain the final score, which is used to calculate the classification accuracy and loss; S44. During the training process, if the accuracy of the validation set is improved, the current model is saved; the training loss, validation loss and accuracy are stored for subsequent analysis.

Images