Unmanned aerial vehicle combat intention judgment model training method, equipment and medium

Abstract

The invention provides an unmanned aerial vehicle combat intention judgment model training method, a use method, equipment and a medium, and belongs to the technical field of unmanned aerial vehicles. The problem that the judgment accuracy of a certain type of combat intention of an unmanned aerial vehicle in a complex battlefield environment is low is solved. According to the technical scheme, the training method comprises the steps of acquiring initial situation information of unmanned aerial vehicles of two parties through real-time data acquisition, analyzing causal association among multiple factors in combination with a causal inference model, and generating initial intention judgment accuracy. By optimizing model parameters, the prediction precision and adaptability of the model are further improved, an optimal unmanned aerial vehicle combat intention judgment model is obtained, and the autonomous decision-making ability and combat efficiency of the unmanned aerial vehicle in a complex battlefield environment are remarkably improved through the model.

Description

Technical Field

[0001] The present invention relates to the technical field, and in particular to a method, device, and medium for training an unmanned aerial vehicle (UAV) combat intention judgment model. Background Art

[0002] As an efficient and flexible intelligent device, UAVs have been widely used in military, reconnaissance, communication, monitoring, and rescue fields. In the modern battlefield environment, UAVs play a key role in battlefield information collection, target tracking, tactical support, and combat mission execution. Their low cost, high efficiency, and high flexibility have gradually made them an indispensable core equipment in modern warfare. In addition, in the case of rapidly changing battlefield situations, UAVs can provide accurate situation information support for command decision-making through real-time data collection and transmission, significantly improving combat efficiency and decision-making quality.

[0003] Although UAV technology has developed rapidly, there are still significant deficiencies in its combat intention judgment ability. Traditional UAV combat intention judgment methods are difficult to cope with the dynamic and changing battlefield environment, mainly manifested as: lacking the ability to deeply analyze multi-dimensional situation information and unable to accurately capture the complex relationships between variables; and in terms of human-machine collaboration, existing technologies cannot fully utilize the decision-making experience and real-time feedback of human experts, resulting in the lack of flexibility in the judgment results of UAVs; the adaptability and generalization ability of existing models are limited, and it is difficult to achieve stable intention judgment in different scenarios. These problems directly limit the decision-making ability of UAVs in complex combat environments, thereby affecting their execution efficiency and mission completion quality.

[0004] In view of the above problems, how to improve the ability of the UAV combat intention judgment model through scientific and efficient training methods has become the research topic of the present invention. Summary of the Invention

[0005] The purpose of the present invention is to provide a method, device, and medium for training a UAV combat intention judgment model. By training and applying the UAV combat intention judgment model, the autonomous decision-making ability and judgment accuracy of UAVs in complex combat environments can be improved, rapid and accurate judgment of combat intentions can be achieved, thereby optimizing the tactical execution efficiency and enhancing the adaptability and mission completion ability of UAVs in dynamic battlefields.

[0006] To achieve the above invention purpose, the technical solutions adopted by the present invention are specifically as follows:

[0007] In a first aspect, the present invention provides a method for training a UAV combat intention judgment model, including:

[0008] A first UAV obtains its own situation information and simultaneously collects relevant situation data of a second UAV, providing basic support for subsequent intention judgment;

[0009] Process and fuse the collected information to generate comprehensive situation information;

[0010] Input the comprehensive situation information into the causal inference model, and generate the initial accuracy rate of combat intention judgment by analyzing the causal relationships among various elements;

[0011] Execute one training cycle, calculate the loss and perform backpropagation to update the network parameters. After each training cycle is completed, evaluate the loss and accuracy rate of the model on the validation set, and store the model with the best performance.

[0012] Input the comprehensive situation information into the causal inference model, and generate the initial accuracy rate of combat intention judgment by analyzing the causal relationships among various elements, including:

[0013] According to the initial comprehensive situation information, select a sample from the sample set in sequence. Then, among the samples with the same label as this sample, use the Euclidean distance to screen out the 10 samples closest to this sample, and randomly select one from them to generate augmented data;

[0014] Extract features from the initial data and the augmented data, generate high-dimensional feature representations through the MLP neural network encoding module. At the same time, standardize the feature space through normalization operations to ensure the distribution consistency of the input features, and separate causal and non-causal factors;

[0015] Based on the separated causal and non-causal factors, further strengthen the important causal features and suppress the non-critical features by combining the masking mechanism;

[0016] Input the masked feature representation into the classifier to generate the accuracy rate of combat intention judgment;

[0017] The intention judgment model executes one training cycle, calculates the loss and performs backpropagation to update the network parameters.

[0018] Based on the separated causal and non-causal factors, further strengthen the important causal features and suppress the non-critical features by combining the masking mechanism, including:

[0019] Design a masker, which consists of an attention mechanism layer and a mask generation layer. This masker receives causal factors and non-causal factors as inputs, and ensures the strengthening of causal features and the gradual suppression of non-causal features by dynamically adjusting the feature selection process;

[0020] The masker dynamically adjusts the number of feature selections according to the training progress. Retain more features at the initial stage of training to learn the variable relationships, gradually reduce the number of selections as the training progresses, and finally only retain the key features that make significant contributions to the target judgment;

[0021] During the process of dynamic feature selection, the attention mechanism assigns importance weights to each feature to measure the contribution of the feature to the target judgment. Features with higher weights are significantly enhanced, and weighted feature representations are generated through element-wise weighting;

[0022] Based on the weighted feature representation, the mask generation layer screens features by dynamically generating a mask matrix, and gradually sparsifies the mask using the Gumbel Softmax method. More features are retained for learning at the beginning of training, and later it focuses on key features, strengthens the role of causal features, and suppresses the interference of non-key features;

[0023] The masker dynamically updates the mask according to each round of training, and generates the final feature set through feature selection and optimization. After the mask result is fused with the input features, optimized features are generated to enhance the model's focusing ability on key causal features.

[0024] The masked feature representation is input into the classifier to generate the accuracy of combat intention judgment, including:

[0025] Using the mask m and the anti-mask 1 - m to divide the features into good dimensions and bad dimensions;

[0026] The good-dimension features are input into the good classifier, and the bad-dimension features are input into the bad classifier to learn the classification capabilities of different features respectively, and the losses of the two classifiers are obtained. The formula is as follows:

[0027]

[0028] Among them, the good classifier is f 1 (·), the bad classifier is f 2 (·), Z o is the feature of the original data, Z a is the feature of the augmented data, y is the corresponding label,

[0029] Strengthen key features through the loss of the good classifier, and reduce the interference of non-key features through the loss of the bad classifier;

[0030] Introduce contrastive loss to optimize the discrimination ability between similar samples and different samples in the feature space;

[0031] Combining the outputs of the good classifier and the bad classifier, generate the accuracy of combat intention judgment. Among them, a total loss function needs to be defined to optimize the encoder and the two classifiers. The formula is:

[0032]

[0033] λ is the weight; and, add L1 regularization L2 regularization to optimize the masker. The formula is:

[0034]

[0035] Optionally, after each training cycle is completed, evaluate the loss and accuracy of the model on the validation set, and store the model with the best performance. This also includes:

[0036] After each training cycle is completed, evaluate the loss and accuracy of the model on the validation set, and store the model with the best performance;

[0037] Load the above-mentioned model, that is, the model based on causal inference, and design and define the classification model;

[0038] Use the model based on causal inference to extract the features of the machine and humans, and fuse these two features as input data for the human-machine hybrid intelligence model based on causal inference;

[0039] Input the fused features into the model for multiple forward propagations to generate the classification prediction results for each time. After all forward propagations are completed, calculate the mean value of the first dimension of the results to obtain the final classification score, and evaluate the classification accuracy and loss based on this;

[0040] During training, if the accuracy of the validation set improves, immediately save the current model. And record the training loss, validation loss, and accuracy for future analysis and model optimization.

[0041] Optionally, load the above-mentioned model, that is, the model based on causal inference, and design and define the classification model, including:

[0042] Load the model based on causal inference, which includes the machine intelligence model and human intelligence model based on causal inference. The human intelligence model is constructed by adding noise to the machine intelligence model;

[0043] The network structure of the classification model consists of two standard feedforward layers and a Bayesian linear layer;

[0044] During the forward propagation process, first calculate the log-likelihood, and use the KL divergence to measure the difference between the posterior distribution and the prior distribution. Then, define the evidence lower bound (ELBO) as the optimization objective, and perform a weighted combination of the log-likelihood and the KL divergence, which is calculated through multiple Monte Carlo samplings to finally obtain the loss of the model.

[0045] In a second aspect, the present invention also provides a method for using a UAV combat intention judgment model, including:

[0046] Receive the comprehensive situation information and call the UAV combat intention judgment model according to the comprehensive situation information, where the UAV combat intention judgment model is obtained by using the training method described in any one of claims 1 or 5-6;

[0047] The first unmanned aerial vehicle (UAV) obtains the first situational information collected by itself and simultaneously receives the second situational information from the second UAV, where the situational information includes combat-related variables such as position, speed, and direction angle.

[0048] Fuse the received first situational information and second situational information to generate the current comprehensive situational information, providing support for subsequent model input.

[0049] Input the comprehensive situational information into the combat intention judgment model, and the model analyzes the combat situation of both UAVs based on the input information and generates a combat intention judgment result.

[0050] Verify and adjust the preliminary combat intention judgment result according to preset rules to ensure that the output result matches the actual combat environment.

[0051] Output the final combat intention judgment result for supporting subsequent mission planning and decision-making.

[0052] Thirdly, the present invention also provides a computer processing device, which includes:

[0053] A memory for storing an operating system, application programs, and various instructions or data to be executed.

[0054] A processor for executing the instructions stored in the memory to complete corresponding functions, implement the training method of the UAV combat intention judgment model, or implement the usage method of the UAV combat intention judgment model.

[0055] Fourthly, the present invention also provides a computer-readable storage medium, which enables the processor to implement the training method or usage method of the UAV combat intention judgment model when the computer program is executed by the processor.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] The embodiments of the present invention provide a training method, device, and medium for a UAV combat intention judgment model. By collecting the situational information of the first UAV and the second UAV in real time and combining the causal inference model to analyze the causal relationship between multiple factors, a preliminary intention judgment accuracy rate is generated; by optimizing the model parameters, the prediction accuracy and adaptability of the model are further improved until the model training is completed, obtaining a UAV combat intention judgment model. Through this model, the autonomous decision-making ability and combat efficiency of UAVs in complex battlefield environments are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings are used to provide a further understanding of the present invention and form a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention.

[0059] Figure 1 It is a schematic diagram of the confrontation of the UAV combat intention judgment model of the present invention. Figure 2 It is a structural block diagram of the computer processing device in the present invention.

[0060] Figure 3 It is a schematic flowchart of the method for training the UAV combat intention judgment model in the present invention;

[0061] Figure 4 It is a schematic flowchart of the model process of the human-machine hybrid model in the present invention;

[0062] Figure 5 It is a schematic flowchart of the method for using the UAV combat intention judgment model in the present invention;

[0063] Figure 6 It is a box plot for comparing the accuracy rates of the UAV combat intention judgment models in the present invention;

[0064] Reference numerals:

[0065] 10. First UAV; 20. Second UAV. Detailed implementation manners

[0066] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to 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.

[0067] Embodiment 1

[0068] See Figure 1 And Figure 6 , this embodiment provides its technical solution as a method, device and medium for training a UAV combat intention judgment model, which can improve the accuracy of intention recognition in UAV combat by constructing and training an efficient combat intention judgment model, thereby enhancing the intelligence and precision of combat decision-making and improving combat efficiency and response capabilities.

[0069] As Figure 1 shown, Figure 1 It is a schematic diagram of the confrontation of the UAV combat intention judgment model.

[0070] As Figure 1As shown in the figure, the first unmanned aerial vehicle (UAV) 10 includes a fuselage, a flight control module, an information collection module, a communication module, an energy supply module, etc. Specifically, the fuselage is used to carry each functional component and protect the internal structure; the flight control module realizes autonomous or remotely controlled flight actions by sensing the attitude of the UAV and executing flight algorithms; the information collection module (such as cameras, sensors, etc.) is used to obtain relevant data of the surrounding environment or target object in real time; the energy supply module (such as a battery) provides the required energy support for functions such as the flight and information collection of the UAV; the communication module is used to perform data interaction with external systems or devices. The computer processing device 30 is communicatively connected to the first UAV 10 and is used to receive the information collected by the first UAV 10, including the situation information such as the position, speed, and direction angle of the first UAV 10 and the second UAV 20, process and integrate the situation information of the first UAV 10 and the second UAV 20, generate comprehensive situation information and input it into the model. The model analyzes the combat situation of both UAVs based on the input information and generates a judgment result of the combat intention.

[0071] Taking the application of the UAV combat intention judgment model training method and usage method described in this embodiment to UAV intention judgment as an example, our UAV can be regarded as the first UAV 10, and the enemy UAV can be regarded as the second UAV 20. During the process of model training and usage, the situation information generated by the first UAV 10 and the second UAV 20 respectively during flight, such as data on position, speed, direction angle, etc., is mainly collected, and analysis and annotation are carried out based on this, so as to complete the training and usage of the UAV combat intention judgment model. At the same time, the computing and processing device 30 can serve as a controller communicatively connected to the first UAV 10, and is used to output the combat intention judgment result to the first UAV 10, so that the first UAV 10 receives the corresponding judgment result and executes the corresponding operation, thereby realizing the accurate recognition and response to the UAV combat intention.

[0072] As Figure 2 shown, Figure 2 is a structural block diagram of the computing and processing device.

[0073] The processor is the core, the memory provides data and instruction storage, the I / O interacts with the outside, and the bus acts as a "bridge" for data and signal transmission between the processor, the memory, and the I / O device.

[0074] Specifically, the processor can be a central processing unit (CPU), a graphics processing unit (GPU), or other forms of integrated processors, which are used to execute the instructions stored in the memory and control the input / output devices. The memory can include read-only memory (ROM), random access memory (RAM), flash memory, or long-term or short-term storage media such as hard disk drives and solid-state drives (SSDs), which are used to store the operating system, application programs, and various types of data. When the processor executes the relevant instructions, it can implement the specific execution and control of the system functions.

[0075] In this embodiment, the memory stores instructions for implementing the functions or methods described in this embodiment. The processor reads and executes these instructions to complete operations such as data calculation, task scheduling, and logical control. When the processor needs to exchange data with peripherals or the network, it interacts with external devices through the I / O module, such as reading data from sensors or outputting information to display devices. In addition, the computing processing device usually communicates with a remote server or other computing devices through a network interface to obtain or send more data.

[0076] The bus provides a transmission channel for data and control signals between the processor, the memory, and the I / O devices. Different buses can be designed differently according to bandwidth, latency, or parallelism requirements. For example, they can be based on PCI, USB, I2C, SPI, or other high-speed serial bus technologies.

[0077] Through the organic combination of the above functional modules, the computing processing device can meet the requirements in multiple aspects such as data processing, algorithm execution, and external interaction, providing reliable hardware support for the related technical solutions of this embodiment. It should be noted that Figure 2 This is only an exemplary structural schematic diagram. In actual applications, some functional units can be added or omitted according to specific requirements, and an architecture with multiple buses or multiple processors can also be adopted to improve the parallel processing ability and redundancy reliability of the system.

[0078] In an exemplary embodiment, the processor executes a computer program stored in the memory and, when executing the program, implements the following steps:

[0079] The first unmanned aerial vehicle (UAV) obtains its own situational information and simultaneously collects the relevant situational data of the second UAV, providing basic support for subsequent intention judgment.

[0080] Process and fuse the collected information to generate comprehensive situational information.

[0081] Input the comprehensive situational information into the causal inference model, and generate a preliminary accuracy rate of combat intention judgment by analyzing the causal relationships between various elements.

[0082] Execute one training cycle, calculate the loss and perform backpropagation to update the network parameters. After each training cycle is completed, evaluate the loss and accuracy of the model on the validation set and store the model with the best performance.

[0083] In some embodiments, the processor inputs the situation synthesis information into the causal inference model, and generates a preliminary combat intention judgment accuracy by analyzing the causal relationships between various elements, including:

[0084] According to the initial situation synthesis information, select a sample from the sample set in sequence. Then, among the samples with the same label as this sample, use the Euclidean distance to screen out the 10 samples closest to this sample, and randomly select one from them to generate augmented data;

[0085] Extract features from the initial data and the augmented data, generate high-dimensional feature representations through the MLP neural network encoding module. At the same time, standardize the feature space through normalization operations to ensure the distribution consistency of the input features, and separate causal and non-causal factors;

[0086] Based on the separated causal and non-causal factors, further strengthen the important causal features and suppress the non-critical features in combination with the masking mechanism;

[0087] Input the masked feature representation into the classifier to generate the combat intention judgment accuracy;

[0088] The intention judgment model executes one training cycle, calculates the loss and performs backpropagation to update the network parameters.

[0089] In some embodiments, the processor further strengthens the important causal features and suppresses the non-critical features in combination with the masking mechanism based on the separated causal and non-causal factors, including:

[0090] Design a masker, which consists of an attention mechanism layer and a mask generation layer. This masker receives causal factors and non-causal factors as inputs, and ensures the strengthening of causal features and the gradual suppression of non-causal features by dynamically adjusting the feature selection process;

[0091] The masker dynamically adjusts the number of feature selections according to the training progress. Retain more features at the initial stage of training to learn the variable relationships, gradually reduce the number of selections as the training progresses, and finally only retain the key features that make significant contributions to the target judgment;

[0092] In the process of dynamic feature selection, the attention mechanism assigns importance weights to each feature to measure the contribution degree of the feature to the target judgment. The features with higher weights are significantly strengthened, and the weighted feature representations are generated by the element-wise weighting method;

[0093] Based on the weighted feature representation, the mask generation layer screens features by dynamically generating a mask matrix and gradually sparsifies the mask using the Gumbel Softmax method. More features are retained for learning in the initial stage of training, and later it focuses on key features, strengthens the role of causal features, and suppresses the interference of non-key features;

[0094] The masker dynamically updates the mask according to each round of training and generates the final feature set through feature selection and optimization. After the mask result is fused with the input features, optimized features are generated to enhance the model's focusing ability on key causal features.

[0095] In some embodiments, when the processor inputs the masked feature representation into the classifier to generate the accuracy of combat intention judgment, it includes:

[0096] Using the mask and the anti-mask to divide the features into superior dimensions (important causal features) and inferior dimensions (non-key features);

[0097] The features in the superior dimensions are input into the superior classifier, and the features in the inferior dimensions are input into the inferior classifier to learn the classification capabilities of different features respectively;

[0098] Strengthen key features through the loss of the superior classifier and reduce the interference of non-key features through the loss of the inferior classifier;

[0099] Introduce contrastive loss to optimize the discrimination ability between similar and different samples in the feature space;

[0100] Combine the outputs of the superior classifier and the inferior classifier to generate the accuracy of combat intention judgment.

[0101] In some embodiments, after each training cycle is completed, the processor evaluates the loss and accuracy of the model on the validation set and stores the model with the best performance. It also includes:

[0102] After each training cycle is completed, evaluate the loss and accuracy of the model on the validation set and store the model with the best performance;

[0103] Load the above model, that is, the model based on causal inference, and design and define the classification model;

[0104] Use the model based on causal inference to extract the features of machines and humans, and fuse these two features as input data for the human-machine hybrid intelligent model based on causal inference;

[0105] Input the fused features into the model for multiple forward propagations to generate the classification prediction results each time. After all forward propagations are completed, calculate the mean of the first dimension of the results to obtain the final classification score, and evaluate the classification accuracy and loss based on this;

[0106] During training, if the accuracy of the validation set improves, immediately save the current model. Also record the training loss, validation loss, and accuracy for future analysis and model optimization.

[0107] In some embodiments, the processor loads the above-mentioned model, i.e., the model based on causal inference, and designs and defines a classification model, including:

[0108] Load a model based on causal inference, which includes a machine intelligence model and a human intelligence model based on causal inference. The human intelligence model is constructed by adding noise to the machine intelligence model;

[0109] The network structure of the classification model consists of two standard feedforward layers and a Bayesian linear layer;

[0110] During the forward propagation process, first calculate the log-likelihood and use the KL divergence to measure the difference between the posterior distribution and the prior distribution. Then, define the evidence lower bound (ELBO) as the optimization objective, combine the log-likelihood and the KL divergence through weighted combination, and calculate it through multiple Monte Carlo samplings to finally obtain the loss of the model.

[0111] In one embodiment, when the processor executes the computer program, the following steps can also be implemented:

[0112] After receiving the comprehensive situation information, call the UAV combat intention judgment model according to this information, where the model is obtained by the above-mentioned UAV combat intention judgment model training method;

[0113] The first UAV obtains the first situation information collected by itself and simultaneously receives the second situation information from the second UAV, where the situation information includes combat-related variables such as position, speed, and direction angle;

[0114] Fuse the received first situation information and the second situation information to generate the current comprehensive situation information to support the subsequent model input;

[0115] Input the comprehensive situation information into the combat intention judgment model, and the model analyzes the combat situation of both UAVs according to the input information and generates a combat intention judgment result;

[0116] Verify and adjust the preliminary combat intention judgment result according to the preset rules to ensure that the output result matches the actual combat environment;

[0117] Output the final combat intention judgment result to support subsequent mission planning and decision-making.

[0118] This embodiment also provides a computer-readable storage medium, which stores a computer program executable by a processor. When the computer program is executed, it is used to implement the steps of the method in this embodiment. The storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, or other types of storage devices.

[0119] Next, in combination with the working principle of the computing processing device, the training method and usage method of the UAV combat intention judgment model provided in this embodiment will be elaborated in detail.

[0120] As Figure 3 shown, Figure 3 is a schematic flowchart of the UAV combat intention judgment model training method provided in this embodiment.

[0121] Step S1: The first UAV obtains its own situational information and simultaneously collects the relevant situational data of the second UAV, providing basic support for subsequent intention judgment.

[0122] The first UAV obtains its own first situational information and the relevant situational data of the second UAV, including key parameters such as position, speed, and direction angle. Through GPS, IMU, and camera radar, the first UAV obtains its own status information, and at the same time obtains the motion data of the second UAV through wireless communication. The collected situational information can be used as the basis for subsequent combat intention judgment analysis after preprocessing, providing necessary data support for the intelligent decision-making of the UAV.

[0123] Step 2: Process and fuse the collected information to generate comprehensive situational information.

[0124] The computing processing device of this method first obtains the first situational information R r = [x r , y r , z r , v r , γ r , nx r , nz r , roll r collected by the first UAV and the second situational information R b = [x b , y b , z b , v b , γ b , nx b , nz b , roll b, where the first situation information includes the flight state of the first UAV itself, while the second situation information reflects the flight condition of the second UAV. Specifically, the data includes the position information x of the UAV r ,x b ,y r ,y b ,z r ,z b 、pitch angle γ r ,γ b 、yaw angle control variable nx r ,nx b nz r ,nz b and roll angle roll r ,roll b . After obtaining these two types of situation information, the computing and processing device will process and fuse them to generate comprehensive situation information which contains the relative position information of the UAVs on both sides, expressed as Meanwhile, the velocity component of the red UAV in the axial direction is expressed as v r_x ,v r_y ,v r_z , and the velocity component of the blue UAV in the axial direction is expressed as v b_x ,v b_y ,v b_z .

[0125] Step S3: Input the comprehensive situation information into the causal inference model, and generate a preliminary combat intention judgment accuracy rate by analyzing the causal relationships between various elements.

[0126] In some embodiments, step S3 specifically includes:

[0127] According to the initial comprehensive situation information, select a sample from the sample set in sequence. Then, among the samples with the same label as this sample, use the Euclidean distance to screen out the 10 samples closest to this sample, and randomly select one from them to generate augmented data;

[0128] Extract features from the initial data and the augmented data, generate a high-dimensional feature representation through the MLP neural network encoding module. At the same time, standardize the feature space through a normalization operation to ensure the distribution consistency of the input features, and separate causal and non-causal factors;

[0129] Based on the separated causal and non-causal factors, further strengthen the important causal features and suppress the non-critical features in combination with the masking mechanism;

[0130] Input the masked feature representation into the classifier to generate the combat intention judgment accuracy rate;

[0131] The intention judgment model executes one training cycle, calculates the loss, performs backpropagation, and updates the network parameters.

[0132] Specifically, the processing device calculates according to the comprehensive situation information (initial data) X o ={X i} where where M represents the total number of samples. Then, from the data X with the same label as the sample oo ={X j}, find the 10 samples closest according to the Euclidean distance, and randomly select one of them to generate augmented data X a . Feature extraction is performed on the initial data and the augmented data to obtain Generate a high-dimensional feature representation through the MLP neural network encoding module. At the same time, standardize the feature space through a normalization operation to ensure the distribution consistency of the input features, and separate causal and non-causal factors. Based on the separated causal and non-causal factors, combined with the masking mechanism, further strengthen the important causal features and suppress the non-critical features. Input the masked feature representation into the classifier to generate the accuracy of combat intention judgment. The intention judgment model executes one training cycle, calculates the loss, performs backpropagation, and updates the network parameters.

[0133] In some embodiments, based on the separated causal and non-causal factors, combined with the masking mechanism to further strengthen the important causal features and suppress the non-critical features, including:

[0134] Design a masker, which consists of an attention mechanism layer and a mask generation layer. The masker receives causal factors and non-causal factors as inputs, and through dynamically adjusting the feature selection process, ensures the strengthening of causal features and the gradual suppression of non-causal features;

[0135] The masker dynamically adjusts the number of feature selections according to the training progress. Retain more features at the initial stage of training to learn variable relationships, gradually reduce the number of selections as the training progresses, and finally only retain the key features that make significant contributions to the target judgment;

[0136] In the process of dynamic feature selection, the attention mechanism assigns importance weights to each feature to measure the contribution degree of the feature to the target judgment. The features with higher weights are significantly strengthened, and a weighted feature representation is generated by element-wise weighting;

[0137] Based on the weighted feature representation, the mask generation layer screens features by dynamically generating a mask matrix, and uses the Gumbel-Softmax method to gradually sparsify the mask. Retain more features to participate in learning at the initial stage of training, focus on key features in the later stage, strengthen the role of causal features, and suppress the interference of non-critical features;

[0138] The masker dynamically updates the mask according to each round of training, and generates the final feature set through feature selection and optimization. After the masked result is fused with the input features, optimized features are generated, enhancing the model's focusing ability on key causal features.

[0139] Specifically, the computing device strengthens key causal features and suppresses non-key features based on the separated causal and non-causal factors in combination with the masking mechanism. First, the computing device constructs a masker including an attention mechanism layer and a mask generation layer to dynamically screen the input features. In the initial stage of training, more features are retained to learn the variable relationships. As the training progresses, non-key features are gradually reduced, and only the causal features that have a significant impact on the target judgment are retained:

[0140]

[0141] where k init is the number of feature selections set during initialization, t represents the current training round, and E is the total number of training rounds.

[0142] Meanwhile, the attention mechanism assigns weights to the features, and the mask generation layer uses the Gumbel Softmax method to dynamically adjust the feature selection, making the mask gradually become sparse, and finally obtaining an optimized feature set, thereby enhancing the model's attention to key causal features and improving the accuracy and robustness of intention judgment.

[0143] In some embodiments, the masked feature representation is input into a classifier for further processing, including:

[0144] Using the mask and the inverse mask to divide the features into good dimensions (important causal features) and bad dimensions (non-key features);

[0145] The good-dimension features are input into a good classifier, and the bad-dimension features are input into a bad classifier to learn the classification capabilities of different features respectively;

[0146] Strengthening key features through the loss of the good classifier and reducing the interference of non-key features through the loss of the bad classifier;

[0147] Introducing a contrastive loss to optimize the discrimination ability between similar and different samples in the feature space;

[0148] Combining the outputs of the good classifier and the bad classifier to generate the judgment accuracy of the combat intention.

[0149] Specifically, the computing device uses the mask m and the inverse mask 1 - m to divide the features into good dimensions (important causal features) and bad dimensions (non-key features), and inputs them into the good classifier and the bad classifier respectively for learning:

[0150]

[0151] The superior classifier focuses on the classification ability of key features, while the inferior classifier is used to measure the interference effect of non-key features. By optimizing the loss of the superior classifier, key features are strengthened, and at the same time, the loss of the inferior classifier is reduced to minimize the impact of non-key features.

[0152] In addition, a contrastive loss is introduced to enhance the discrimination ability of different classes of samples in the feature space.

[0153]

[0154] Among them, represents the L2 norm, which is used to measure the distance between two samples; margin is used to set the minimum distance between different classes of samples.

[0155] Finally, by combining the outputs of the superior classifier and the inferior classifier, the judgment accuracy of the combat intention is generated. Among them, a total loss function needs to be defined to optimize the encoder and the two classifiers. The formula is:

[0156]

[0157] λ is the weight; and, L1 regularization is added L2 regularization to optimize the masker. The formula is:

[0158]

[0159] Step S4: Execute one training cycle, calculate the loss and perform backpropagation to update the network parameters. After each training cycle is completed, evaluate the loss and accuracy of the model on the validation set, and store the model with the best performance.

[0160] The computing device executes one training cycle. First, it generates prediction results through forward propagation and calculates the loss value according to the true labels. Then, it uses the backpropagation algorithm to calculate the gradients and adopts the Adam optimizer to update the model parameters to improve the accuracy of combat intention judgment. After each training cycle ends, the computing device evaluates the model performance on the validation set, calculates the loss and accuracy, and dynamically adjusts the training strategy. At the same time, save the current optimal model to ensure that the final model has the best performance during testing.

[0161] As Figure 4 shown, in some embodiments, step S4 further includes:

[0162] After each training cycle is completed, evaluate the loss and accuracy of the model on the validation set, and store the model with the best performance;

[0163] Load the above-mentioned model, i.e., the model based on causal inference, and design and define a classification model;

[0164] Use the model based on causal inference to extract the features of machines and humans, and fuse these two features as input data for the human-machine hybrid intelligent model based on causal inference;

[0165] Input the fused features into the model for multiple forward propagations to generate the classification prediction results for each time. After all forward propagations are completed, calculate the mean of the first dimension of the results to obtain the final classification score, and evaluate the classification accuracy and loss based on this;

[0166] During the training process, whenever the accuracy of the validation set improves, save the state of the current model. At the same time, record the training loss, validation loss, and accuracy for subsequent analysis and optimization.

[0167] In some embodiments, load the above-mentioned model, i.e., the model based on causal inference, and design and define a classification model, including:

[0168] Load the model based on causal inference, which includes a machine intelligence model and a human intelligence model based on causal inference. The human intelligence model is constructed by adding noise to the machine intelligence model;

[0169] The network structure of the classification model consists of two standard feedforward layers and a Bayesian linear layer;

[0170] During the forward propagation process, first calculate the log-likelihood and use the KL divergence to measure the difference between the posterior distribution and the prior distribution. Then, define the evidence lower bound ELBO as the optimization objective, perform a weighted combination of the log-likelihood and the KL divergence, and calculate through multiple Monte Carlo samplings to finally obtain the loss of the model.

[0171] For the trained UAV combat intention judgment model, this embodiment further proposes an application method thereof, which is applicable to terminal devices that can communicate with the first UAV, such as the computing and processing device that executes the model training process or other devices with processing capabilities. It should be noted that the UAV combat intention judgment model is optimized through the above training method.

[0172] Figure 5 It is a schematic flowchart of the method for using the UAV combat intention judgment model.

[0173] As Figure 5 shown, the method for using the UAV combat intention judgment model specifically includes steps S51 to S56:

[0174] Step S51: Receive the comprehensive situation information and call the UAV combat intention judgment model according to the comprehensive situation information. Among them, the UAV combat intention judgment model is obtained by adopting the foregoing training method.

[0175] Step S52: Receive the collected first situation information and second situation information, where the situation information includes combat-related variables such as position, speed, and direction angle;

[0176] Step S53: Perform fusion processing on the received first situation information and second situation information to generate the current comprehensive situation information, providing support for subsequent model input;

[0177] Step S54: Input the comprehensive situation information into the combat intention judgment model, and the model analyzes the combat situation of both UAVs according to the input information and generates a combat intention judgment result;

[0178] Step S55: Verify and adjust the preliminary combat intention judgment result according to the preset rules to ensure that the output result matches the actual combat environment;

[0179] Step S56: Output the final combat intention judgment result for supporting subsequent mission planning and decision-making.

[0180] Specifically, the computing and processing device receives the comprehensive situation information, calls the trained UAV combat intention judgment model, and respectively obtains the situation information collected by the first UAV and the second UAV, including key combat variables such as position, speed, and direction angle. Subsequently, perform fusion processing on the obtained situation information to generate the current comprehensive situation information as the model input. The model analyzes the combat situation of both UAVs according to the input information, generates a combat intention judgment result, and verifies and adjusts the judgment result according to the preset rules, and finally outputs an accurate combat intention judgment result to support subsequent mission planning and decision-making.

[0181] In summary, by collecting and fusing the situation information of the first UAV and the second UAV, and combining the pre-trained combat intention judgment model, the computing and processing device can analyze the combat situation of both UAVs and generate an accurate combat intention judgment result. Further, verify and optimize the judgment result through preset rules to ensure that it conforms to the actual combat environment, thereby effectively supporting subsequent mission planning and decision-making, and realizing the intelligent identification and prediction of UAV combat intentions.

[0182] 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 training a UAV combat intention judgment model, characterized in that: The following steps are involved: S1. The first UAV obtains its own situation information and collects relevant situation data of the second UAV at the same time to provide basic support for subsequent intention judgment; S2, process and integrate the collected information to generate comprehensive situation information; S3. Input the comprehensive situation information into the causal inference model, and generate the accuracy rate of preliminary combat intention judgment by analyzing the causal relationship between various elements; S4. Perform a training cycle, calculate the loss and perform backpropagation, update the network parameters, and after each training cycle, evaluate the loss and accuracy of the model on the validation set and store the model with the best performance.

2. The training method according to claim 1, characterized in that: In step S3, the comprehensive situation information is input into a causal inference model, and a preliminary combat intention judgment accuracy is generated by analyzing the causal relationship between various elements, including the following steps: S31, selecting a sample from the sample set in order according to the initial situation comprehensive information; S32. Among the samples with the same label as the sample, use the Euclidean distance to select the 10 samples closest to the sample, and randomly select one of them to generate enhanced data; S33, extract features from the initial data and enhanced data, generate high-dimensional feature representation through the MLP neural network encoding module, standardize the feature space through normalization operation, ensure the distribution consistency of input features, and separate causal and non-causal factors; S34, based on the separated causal and non-causal factors, combined with the masking mechanism, further strengthen the important causal features and suppress the non-critical features; S35, inputting the masked feature representation into the classifier to generate the accuracy of combat intention judgment; S36. The intent judgment model executes a training cycle, calculates the loss and performs back propagation, and updates the network parameters.

3. The training method according to claim 2, characterized in that: In step S34, based on the separated causal and non-causal factors, the masking mechanism is combined to further strengthen the important causal features and suppress the non-critical features, including: S341. Design a masker consisting of an attention mechanism layer and a mask generation layer. The masker receives causal factors and non-causal factors as input and ensures the strengthening of causal features and the gradual suppression of non-causal features by dynamically adjusting the feature selection process. S342, the masker dynamically adjusts the number of feature selections according to the training progress; S343. In the process of dynamic feature selection, the attention mechanism assigns importance weights to each feature to measure the contribution of the feature to the target judgment, and generates weighted feature representation by weighting each element; S344, based on weighted feature representation, the mask generation layer dynamically generates a mask matrix to filter features, and uses the GumbelSoftmax method to gradually sparse the mask. In the early stage of training, more features are retained for learning, and in the later stage, key features are focused on, the role of causal features is strengthened, and the interference of non-key features is suppressed; S345, the masker dynamically updates the mask according to each round of training, generates the final feature set through feature selection and optimization, and generates optimized features after the mask result is fused with the input features.

4. The training method according to claim 2, characterized in that: In step S35, the masked feature representation is input into the classifier to generate the combat intention judgment accuracy, including: S351, using mask m and inverse mask 1-m to divide the features into superior dimensions and inferior dimensions; S352, the superior dimensional features are input into the superior classifier, and the inferior dimensional features are input into the inferior classifier. The classification capabilities of different features are learned respectively, and the losses of the two classifiers are obtained. The formula is as follows: The best classifier is f1(·), the worst classifier is f2(·), and Z o is the characteristic of the original data, Z a To enhance the features of the data, y is the corresponding label, S353, strengthen key features through superior classifier loss, and reduce interference of non-key features through inferior classifier loss; S354, introduce contrast loss to optimize the ability to distinguish between samples of the same type and samples of different types in feature space; S355. Combine the outputs of the superior classifier and the inferior classifier to generate the accuracy of the judgment of the combat intention. It is necessary to define a total loss function to optimize the encoder and the two classifiers. The formula is: λ is the weight; and, L1 regularization is added L2 Regularization To optimize the masker, the formula is:

5. The training method according to claim 1, characterized in that: In step S4, after each training cycle is completed, the loss and accuracy of the model on the validation set are evaluated, and the model with the best performance is stored, which also includes: S41. After each training cycle is completed, evaluate the model's loss and accuracy on the validation set, and store the best performing model, that is, the model that exhibits the lowest loss on the validation set: S42, loading the above model, that is, the model based on causal inference, and designing and defining a classification model; S43. Extract machine and human features using a causal inference-based model, and fuse the two features as input data for use by a causal inference-based human-machine hybrid intelligent model; S44, input the fused features into the model for multiple forward propagations to generate classification prediction results each time. After all forward propagations are completed, the first dimension of the results is averaged to obtain the final classification score, and the classification accuracy and loss are evaluated accordingly; S45. During the training process, whenever the accuracy of the validation set improves, save the state of the current model. At the same time, record the training loss, validation loss, and accuracy to facilitate subsequent analysis and optimization.

6. The training method according to claim 5, characterized in that: In step S42, the above model, i.e., the model based on causal inference, is loaded, and a classification model is designed and defined, including: S421. Loading a model based on causal inference, which includes a machine intelligence model based on causal inference and a human intelligence model, wherein the human intelligence model is constructed by adding noise to the machine intelligence model; S422, the network structure of the classification model consists of two standard feed-forward layers and one Bayesian linear layer; S423. In the forward propagation process, the log-likelihood is first calculated, and the KL divergence is used to measure the difference between the posterior distribution and the prior distribution. The evidence lower bound ELBO is defined as the optimization target, the log-likelihood and KL divergence are weightedly combined, and calculated through multiple Monte Carlo sampling to finally obtain the model loss.

7. A computer processing device, characterized in that The equipment includes: Memory, used to store operating systems, applications, and various instructions or data that need to be executed; A processor is used to execute instructions in a memory to implement corresponding functions and complete the UAV combat intention judgment model training method described in any one of claims 1 or 5-6.

8. A computer-readable storage medium, characterized in that: storing a computer program; When the processor executes the program, the UAV combat intention judgment model training method described in any one of claims 1, 5 or 6 is implemented.

Images