Aerial target intention recognition model based on deep learning

Abstract

An air target intention recognition model based on deep learning belongs to the field of intention recognition and comprises an input module, a storage module, a data generation module, an intention recognition module and an output module. The method comprises the following steps: firstly, constructing an air target intention space and an intention feature set, performing unified coding, and packaging experience knowledge of a decision maker into an intention label; and secondly, constructing an intention data generation model based on a denoising diffusion model, adding Gaussian noise to intention data in a forward process, and denoising in a reverse process so as to generate air target intention data. An improved knowledge distillation method is applied to a de-noising diffusion model, and the sampling process of the model is accelerated. And finally, constructing a time module based on expansion causal convolution, and solving the problem of time sequence feature extraction of the intention data. And a graph attention mechanism is introduced to mine and analyze the relationship between different intention features. According to the intention recognition model, the lack of intention data can be made up, and the time sequence feature extraction capability can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of intention recognition, and in particular relates to an aerial target intention recognition model based on deep learning. Background Art

[0002] With the introduction of new operational concepts such as decision-centric warfare and joint all-domain operations, as well as the development of modern weapons and equipment, the intensity, timeliness and complexity of air-space confrontations have increased significantly. This requires the ability to quickly acquire and process battlefield information, infer the combat intent of enemy targets, achieve clear and effective battlefield situation assessments, and implement reasonable and favorable decisions on this basis. Intent recognition is the key to the transition of modern warfare from the information domain to the cognitive domain. It is the premise and foundation for battlefield cognition and intelligent decision-making. It is of great significance for commanders to understand the battlefield situation and make reasonable decisions. The combat intention recognition of aerial targets in air defense operations specifically refers to the process of extracting the battlefield information in the corresponding time domain and airspace from the real-time, confrontational air defense battlefield environment, the static and real-time dynamic attributes of aerial targets, analyzing these elements, and then judging and inferring the combat intentions of enemy aerial targets. The specific process is as follows: Figure 1 shown.

[0003] With the further development of battlefield informatization and intelligence under the new situation, the uncertainty, confrontation, deception and concealment of war itself have brought great difficulties to the collection of battlefield situation data. At the same time, due to the great asymmetry of military strength and information advantages of the warring parties, there will be great differences in battlefield data collection, resulting in a serious scarcity of air target intention data, which affects the accuracy of identifying enemy target intentions. In particular, the application of artificial intelligence technology in intention recognition tasks requires a large amount of data as support. In response to the problem of data scarcity, one possible solution is sample generation technology, that is, generating more synthetic data based on existing real intention data. Typical sample generation technologies mainly include generative adversarial networks (GAN), diffusion models (DM), variational autoencoders (VAE), etc.

[0004] With the development of decision support systems, more and more researchers have begun to study the problem of tactical intention recognition of aerial targets. The existing target intention recognition methods mainly include methods based on statistical theory (evidence theory, Bayesian network), methods based on cognitive models (template matching, expert system), and methods based on artificial intelligence.

[0005] The intention recognition methods based on statistical theory mainly include evidence theory and Bayesian network. Xia et al. used grey Markov chain to analyze and predict the speed, angle, attack and detection of enemy drones. By combining the prediction factors with the rules provided by rough sets, the intention of enemy drones in the short term in the future can be inferred. Bao et al. used high-dimensional space similarity to measure the support degree of the target state for its intention, and combined with evidence theory to identify the tactical intention of the target in turn. Zhao et al. combined the confidence rule base and evidence theory to identify the intention of aerial targets. Xu et al. applied the dynamic sequence Bayesian network to the problem of target intention recognition, optimized the algorithm through information entropy theory, and achieved good results. Yang et al. proposed an improved algorithm based on data classification confidence on the basis of semi-supervised naive Bayes classifier, which effectively identified the intention of air combat targets. Jin et al. used network nodes and conditional probability tables to establish an initial Bayesian network model with a high model recognition accuracy. The advantages of this method are the ability to handle uncertainty and integrate multi-source information, while the disadvantages are limited ability to handle conflicting evidence and difficulty in constructing basic probability assignment functions.

[0006] Intention recognition methods based on cognitive models mainly include template matching and expert systems. Chen et al. first constructed a situation template based on expert experience and used DS evidence theory to build a reasoning model for intention recognition. Yin et al. used statistical methods to obtain target combat intention knowledge and rule bases for prior knowledge and proposed an intention recognition method based on discriminant analysis. The advantage of this method is that it has strong causal probabilistic reasoning capabilities, but the disadvantage is that parameter estimation is difficult and it has poor adaptability to dynamically changing battlefield environments.

[0007] Intent recognition methods based on artificial intelligence mainly include neural networks and deep learning. Zhou et al. proposed an aerial target intent recognition method based on deep neural networks, using the ReLU (Rectified Linear Unit) function and the adaptive moment estimation (Adam) algorithm to optimize the back propagation algorithm and improve the model recognition effect. In view of the temporal nature of target intent features, Teng et al. combined the temporal convolutional network (TCN) and the bidirectional gated recurrent unit (BiGRU) to extract the temporal characteristics of the features, and introduced the attention mechanism to assign different weights to the features, thereby identifying the tactical intent of the target. In view of the limitations of existing algorithms such as reliance on empirical knowledge, difficulty in extracting full temporal features, and inability to meet actual combat requirements, Yang et al. proposed a target tactical intent recognition algorithm based on bidirectional long short-term memory (BiLSTM), with a recognition accuracy of 92%. In view of the shortcomings of existing methods in terms of temporality and interpretability, Wang et al. designed an aerial target intent recognition model based on BiGRU and conditional random field (CRF), achieving a high recognition accuracy. Qu et al. proposed an aerial target intention recognition method based on fully connected neural networks, convolutional neural networks, and recurrent neural networks, realizing the target intention recognition function based on real-time situation information. The proposed method has good robustness. The advantage of this method is that it has a strong ability to learn data features and does not require the organization of a large number of generation rules. However, the disadvantage is that the model is complex and the interpretability is poor.

[0008] Although the above-mentioned research methods have their own advantages, they still have shortcomings: First, although the existing intention recognition model has achieved a high recognition accuracy, the war costs caused by misjudgment of intention in actual combat are unaffordable for both parties, and the performance and interpretability of the model still need to be improved. Second, the existing intentions are all labeled by experts based on experience and knowledge, while in actual combat, there are great similarities between intentions, and it is difficult to strictly distinguish different intentions. Third, there are few battlefield data collected in actual combat, and the existing models are more dependent on data support. How to obtain more high-quality intention data is also a future research direction. Fourth, from the perspective of the OODA loop, if the combat intention of the enemy target can be predicted in advance, it can lay the foundation for accelerating the closure of the OODA loop, and then it can obtain the prerequisite for victory in the war. In addition, the current method only judges the tactical intention of the target based on its state at a single moment. In actual combat, the tactical intention of the enemy target usually needs to be reflected by the change of state in multiple continuous time periods. At the same time, due to the scarcity of air target intention data, the model training is insufficient, affecting its evaluation accuracy and applicability. Summary of the invention

[0009] The purpose of this invention is to address the problems of data scarcity and insufficient temporal feature extraction in the process of aerial target intention recognition, and propose an aerial target intention recognition model based on deep learning, called IDGTIR, where ID represents the improved diffusion probability model, G represents the graph attention mechanism, T represents the time module based on dilated causal convolution, and IR represents intention recognition. This recognition model realizes the sampling process of the accelerated diffusion model and greatly improves the problem of temporal feature extraction of intent data.

[0010] The technical solution adopted by the present invention is:

[0011] An aerial target intention recognition model based on deep learning includes an input module, a storage module, a data generation module, an intention recognition module, and an output module; the aerial target intention recognition model is established based on a denoising diffusion model and a gated recurrent unit;

[0012] The input module is used to collect battlefield situation data, and then perform unified coding and standardization operations on the collected data before inputting it into the intention recognition module; the battlefield situation data is multimodal data, including the movement state and entity attributes of the target; the battlefield situation data is collected through underlying sensors, radars, etc.;

[0013] The storage module is used to store training data and knowledge information. When the intention recognition module performs modeling and training, the training data and knowledge information are retrieved. At the same time, the storage module transmits the real intention data (i.e., training data and knowledge information) to the data generation module for data expansion; the training data and knowledge information include historical intention data, expert knowledge, domain knowledge, and rule knowledge;

[0014] The data generation module is used to expand the historical intention data, so as to provide sufficient training data sets for the deep learning model, and specifically uses a data generation model based on a denoising diffusion probability model to generate data;

[0015] The intention recognition module uses a deep learning-based intention recognition model to train the model according to the historical intention data, knowledge information in the storage module and the generated data in the data generation module, adjusts the various hyperparameters of the model to achieve the optimal effect of the model, and then processes the data input in the input module to obtain the tactical intention of the target in the current state;

[0016] The output module outputs the target's current tactical intention recognition result obtained by the intention recognition module.

[0017] Furthermore, in the input module, battlefield situation data is collected, and uniform coding and standardization operations are performed on the collected data, including

[0018] S11, taking enemy air targets in air defense operations as the research object, and then determining the intention space of air targets;

[0019] For different battlefield environments, different combat styles, and different combat objectives, the intention space is quite different, so it needs to be defined in combination with the specific combat background. This invention mainly focuses on aerial targets in air defense operations, such as reconnaissance aircraft, fighter jets, early warning aircraft, etc.

[0020] Specifically, we first define the temporal feature set of the target intent I t From time t1 to time t n The real-time characteristics of continuous air combat at the moment; redefine Y t As the tactical intention space of the target, the problem of air combat target intention recognition is described as: t To Y t The mapping process is:

[0021]

[0022] The problem of air combat intention recognition is the specific mapping of target intention recognition features to target intention recognition types. Considering that target intention recognition in air combat occurs in a real-time, complex, and highly confrontational battlefield environment, and that there are certain limitations in the ability to obtain and analyze information, it is usually difficult for our commanders to accurately obtain the combat intentions of enemy targets. Therefore, the probability of the intention occurring is used to represent the result of intention recognition at that moment. In an actual combat environment, enemy targets may conceal their tactical intentions through various means. Therefore, judging the target's tactical intentions based solely on battlefield data at a single moment is somewhat one-sided and false. Therefore, it is chosen to extract target features from battlefield data obtained at continuous moments to identify its tactical intentions.

[0023] S12, establishing an encoding and decoding mechanism for the determined intent space to perform model training;

[0024] After determining the target intention space, the feature information required by the model can be determined based on the relationship between the target attribute characteristics and the intention. From the perspective of the enemy's combat mission, the enemy's aircraft in the air also have certain differences in the feature information displayed when performing different combat missions. In addition, considering the limitations of information technology, there are target features that cannot be directly obtained, such as target type, shape, etc., and this embodiment chooses to use radar one-dimensional range image and radar reflection cross-sectional area instead.

[0025] S13, constructing an air defense combat target intention recognition feature set, including numerical feature data and non-numerical feature data, and then performing a standardization operation on the feature data of the feature set;

[0026] For numerical feature data, the Min-Max standardization method is used to map it to the interval [0,1]. The specific calculation method is:

[0027]

[0028] In the formula, x represents the eigenvalue of a numerical feature, x′ Represents the normalized result of the numerical feature, where min and max correspond to the minimum and maximum values ​​of the feature in the set respectively;

[0029] The non-numerical feature data are all classified data. To facilitate the learning of the neural network, the non-numerical feature data are numerically processed and then mapped to the interval [0,1]. The specific calculation method is:

[0030]

[0031] In the formula, j represents the size of the classification space, and y is the value of the original i-th non-numeric feature after being mapped to the interval [0,1].

[0032] Furthermore, the historical intent data is labeled data or unlabeled data, which can provide a certain training data set for the deep learning model. The expert knowledge, domain knowledge, and rule knowledge are information extracted from battlefield intelligence, historical information, and expert experience, providing relevant knowledge information and constraint rules for the intent recognition module.

[0033] Furthermore, the denoising diffusion model is used to generate intent data. The specific process is as follows:

[0034] Denoising Diffusion Probabilistic Models (DDPM) mainly consists of two processes: a forward process that perturbs data into noise and a reverse process that converts noise into data.

[0035] Given the original data distribution x0~q(x0), Gaussian noise is gradually added to the data during the diffusion process;

[0036] Assume that a total of T steps of diffusion are performed to generate a series of noisy samples x, and the noise parameter in the diffusion process is an increasing sequence β 1:T ∈(0,1] T Determine, then the forward diffusion process gradually transforms the original data distribution q(x0) into the latent variable distribution q(x T ), as shown in formula (4):

[0037]

[0038] in, N represents Gaussian distribution, represents the mean of the Gaussian distribution, (1-α t ) Ι represents the covariance matrix of Gaussian distribution, I is the identity matrix;

[0039] Let α t :=1-β t , but

[0040]

[0041] Where N represents Gaussian distribution, represents the mean of the Gaussian distribution, β t Ι represents the covariance matrix of Gaussian distribution, I is the identity matrix;

[0042] By deriving and transforming formula (5), we can calculate the data distribution q(x t |x0):

[0043]

[0044] Among them, ε∈Ν(0,Ι), N represents Gaussian distribution, represents the mean of the Gaussian distribution, (1-α t ) Ι represents the covariance matrix of Gaussian distribution, I is the identity matrix;

[0045] As t increases, the proportion of noise increases and the proportion of original data decreases; When x T It is almost Gaussian distributed, and the diffusion process of the model is considered to be completed;

[0046] The reverse diffusion process of DDPM uses a Markov chain to transform the distribution of latent variables p θ (x T ) is transformed into data distribution p θ (x0); Considering that the noise added in each step of the forward process is very small, assuming that p θ (x t-1 |x t ) is Gaussian distribution, p θ (x t-1 |x t ) is fitted using a neural network, where θ represents the parameters of the neural network;

[0047] When β t When it is close enough to 1, for all x0, q(x t |x0) approaches the standard normal distribution, so q(x t|x0) is set to the standard normal distribution, then the joint probability distribution of the reverse diffusion process is as follows:

[0048]

[0049] Among them, p θ (x t-1 |x t ):=Ν(x t-1 ;μ θ (x t ,t),σ θ (x t ,t) 2 Ι); N represents Gaussian distribution, μ θ (x t ,t) represents the mean of Gaussian distribution, σ θ (x t ,t) 2 Ι represents the covariance matrix of Gaussian distribution, I is the identity matrix;

[0050] μ θ Decomposed into noise and x t , we can get the approximate value of the mean as follows (9):

[0051]

[0052] Settings and Beta t Related constants is the variance, and the trainable parameters only exist in the mean, then the generation process is expressed as:

[0053]

[0054] Among them, ε θ represents a neural network with the same input and output, ε θ The noise predicted at each step is used in the back-diffusion process;

[0055] The goal of DDPM is to find a data distribution that can make the dual-objective data distribution p θ (x0) maximizes the parameter θ, at the target data p θ Negative log-likelihood function of (x0) - logp θ A KL divergence is added to (x0) to form an upper bound for the negative likelihood logarithm.

[0056] The denoising loss function of DDPM is expressed as formula (11):

[0057]

[0058] in, represents the predicted noise; E qIt means averaging the variables under distribution q; represents the actual noise; C is a constant;

[0059] To further simplify the expression, the following loss function is minimized during training:

[0060]

[0061] In the calculation process, the latent variable x is first sampled from the standard normal distribution T ~N(x T ; 0, Ι), and then get x again from the sampling t-1 When p θ When (x0) is calculated, the iteration process ends, where Formula (12) is a further simplification of formula (11).

[0062] Furthermore, considering that the reverse diffusion process of DDPM often requires hundreds or thousands of steps, the sampling speed is greatly slowed down. In order to speed up the model sampling speed, an improved knowledge distillation method is introduced. In order to meet the needs of future information-based intelligent warfare, multiple teacher models and a single student model are deployed in the cloud computing center and combat terminal respectively. The knowledge distillation method is designed as follows:

[0063] First, design teacher models: design seven teacher models based on the seven combat intentions of air targets The difference between different teacher models lies in their different inputs, which are only single intent data;

[0064] Second, design the student model: input the data of the seven intentions into the teacher model at the same time, and select the model with the best performance through multiple experiments, denoted as At this time, the student model N S Use a relatively small U-Net network instead The U-Net network of the model; the number of channels of the U-Net network input layer used in the student model remains unchanged, and the number of channels of the convolutional neural network in other positions is halved. At the same time, a neural network with a convolution kernel of 1 is added to solve the problem of matching the feature size of the U-Net network after the reduction;

[0065] Third, the design of the loss function: the data of the seven combat intentions are simultaneously input into the seven teacher models and the student model. Given the original input data set With θ S The student model with parameters N S and The teacher model with parameters The loss function designed for the knowledge distillation method mainly consists of three parts: First, calculate the cross entropy L between the true label and the student modellabel ; The second is to calculate the average cross entropy between the teacher model output and the student model output L TS ; The third is to calculate the loss value L of the middle layer inter :

[0066] L total =αL label +βL TS +γL inter (13)

[0067] L label =H(y i ,N S (x i )) (14)

[0068]

[0069] in, and N T and N S Soft output; τ>1 indicates the temperature parameter of the softening operation; represents the output feature map of a certain intermediate layer of the t-th teacher model; F S (·) represents the output feature map of a middle layer of the student model; MSE(·) represents the mean square error between different feature maps; α, β, and γ represent the weights of the three loss values, respectively, and their sum is 1.

[0070] Furthermore, in the intention recognition module, the intention recognition process includes the following steps:

[0071] (1) Data preprocessing: convert non-numerical features into numerical features, and then normalize the features to form a standard data set;

[0072] (2) Feature extraction: dilated causal convolution is introduced to model the connection between inputs at different times and mine the information embedded in the intent data;

[0073] Dilated Causal Convolution (DCC) is a network structure that can process time series data proposed by Shaojie Bai et al. in 2018 in the Temporal Convolution Network (TCN). Compared with CNN, causal convolution can process time series data according to {x1,x2,...,x t} and {y1,y2,...,y t-1 Prediction y t , so that y t Close to the true value.

[0074] In order to fully extract the temporal features of situation data, the idea of ​​Inception is introduced to construct a temporal block. The Temporal Block consists of 4 parts. The size of the dilated causal convolution kernel of each part is the same, but the dilation factor is different.

[0075] Given the processed input data V = {V1, V2, ..., V m}, after the Temporal Block, it is activated by the sigmoid function and the tanh function respectively to obtain the output feature S a and S b .

[0076] S a =tanh(F(V,s)) (17)

[0077] S b =σ(F(V,s)) (18)

[0078] S=S a +S b (19)

[0079] Where tanh(·) represents the tanh function, σ(·) represents the sigmoid function, and F represents the feature vector output by the Temporal Block;

[0080] (3) Enhanced feature representation capability: Introducing the graph attention mechanism to capture the connections between multiple features and dynamically adjust the attention weights of different nodes;

[0081] Graph Attention Network (GAT) is an attention-based node classification network proposed by Petar Velickovic et al. The basic idea is to update the node representation according to the attention of each node on its neighboring nodes.

[0082] At the input layer, GAT receives the node features of the graph data and the topological structure of the graph. For the constructed IDGTIR model, the node features of the graph are the feature vectors extracted by fusing Temporal Block and Conv1d, and the topological structure is the adjacency matrix constructed by the similarity between the feature vectors. The input is the node features. Where M represents the number of nodes, and F represents the feature dimension of the nodes;

[0083] In order to obtain sufficient expressive power, GAT uses a shared linear transformation parameterized by a weight matrix to transform input features into higher-level features. At the same time, a shared attention mechanism a is introduced to calculate the importance of the feature of node j relative to node i. ij :

[0084]

[0085] In the formula, the shared attention mechanism a:R F′ ×R F →R represents a single-layer feedforward neural network, W∈R F′×F represents the weight applied to each node of the graph;

[0086] In order to make the coefficients between different nodes easier to compare, the softmax function is used for normalization to obtain the normalized attention coefficient:

[0087]

[0088] In the formula, is the initialization weight vector of the attention mechanism a, LeakyReLU is used for nonlinear operations, N i is a field of node i in the graph, and || represents a concatenation operation;

[0089] Finally, the normalized attention coefficient is used to calculate the linear combination of the corresponding features, and the output features of the node are obtained through the nonlinear activation function σ:

[0090]

[0091] in, represents the output feature of node i; α ij represents the attention coefficient from node i to node j, W represents the weight matrix, represents the input features of node j;

[0092] In order to stabilize the learning process of the self-attention mechanism, the attention mechanism is extended to a multi-head attention mechanism. K attention mechanisms run independently, and the output features obtained are concatenated to obtain the final features:

[0093]

[0094] in, represents the final output feature of node i; α ij k represents the attention coefficient from node i to node j in the kth attention mechanism, W k represents the weight matrix, represents the input features of node j;

[0095] (4) Intent recognition: The output of the graph attention mechanism will be sent to the softmax layer for classification, and finally the target intent recognition result at the current moment is obtained.

[0096] Beneficial effects of the present invention:

[0097] (1) The problem of aerial target intention recognition was analyzed in detail, and the aerial target intention space and intention feature set were constructed and uniformly encoded, and the decision maker's experience knowledge was encapsulated into intention labels. (2) An intention data generation model was constructed based on the denoising diffusion model. Gaussian noise was added to the intention data in the forward process, and denoised in the reverse process to generate the aerial target intention data. The improved knowledge distillation method was applied to the denoising diffusion model to accelerate the sampling process of the model. Seven teacher models were designed according to the seven combat intentions of aerial targets, and the loss function in the training process of the teacher model and the student model was optimized to accelerate the sampling process of the diffusion model. (3) A time module based on dilated causal convolution was constructed to solve the problem of temporal feature extraction of intent data; at the same time, a graph attention mechanism was introduced to mine and analyze the relationship between different intent features. Combining the two, the feature extraction capability of the model has been greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 Intention reasoning process for aerial targets;

[0099] Figure 2 It is the IDGTIR model framework;

[0100] Figure 3 Encoding and decoding process for target intent;

[0101] Figure 4 Generate models for intent data;

[0102] Figure 5 for improved knowledge distillation methods;

[0103] Figure 6 Identify model structures for intent;

[0104] Figure 7 It is the specific framework of Temporal Block;

[0105] Figure 8 t-SNE graph of the original intent data;

[0106] Fig. 9 The t-SNE graph is a merging of the original intent data and the generated intent data, where (a) is a t-SNE graph of the generated intent data, and (b) is a t-SNE graph after the generated intent data and the original data are merged;

[0107] Fig.10 The curves of accuracy and loss value change, where (a) the curve of accuracy changing with the number of iterations, (b) the curve of loss value changing with the number of iterations;

[0108] Fig.11 is the confusion matrix;

[0109] Fig.12 The accuracy and loss value change curves of different data sets, where (a) the accuracy changes with the number of iterations, (b) the loss changes with the number of iterations;

[0110] Fig.13 The accuracy and loss value change curves of different data sets, where (a) the accuracy changes with the number of iterations, and (b) the loss changes with the number of iterations. DETAILED DESCRIPTION

[0111] The experimental data of this embodiment comes from Air Combat Maneuvering Generator (ACMG). During the data collection process, a variety of air combat scenarios are set, such as air defense operations in different terrains such as plains and mountains. The time characteristic data of the target and the initially set intention data are output through the system interface, and then the target intention is revised by experts in the field of air combat. The data under different scenarios are fused to obtain a total of 3520 samples, with a time distribution of 6 sampling cycles. The classification of different categories of test training is shown in Table 1, of which training samples account for 80% and test samples account for 20%.

[0112] Table 1 Classification of sample data sets

[0113]

[0114] In order to verify the effect of the proposed model, four evaluation indicators are mainly selected: accuracy, recall, precision and F1-score. The tactical intention recognition of aerial targets is essentially a multi-classification problem. Accuracy, recall, precision and F1-score can reflect the effect of model classification. The specific calculation formula is as follows:

[0115]

[0116] Among them, TP is the true positive class, FN is the false negative class, FP is the false positive class, and TN is the true negative class.

[0117] The invention discloses an air target intention recognition model based on deep learning, called IDGTIR model, which includes an input module, a storage module, a data generation module, an intention recognition module, and an output module. The air target intention recognition model is established based on a denoising diffusion model and a gated recurrent unit, such as Figure 2 As shown;

[0118] The input module is used to collect battlefield situation data, and then perform unified coding and standardization operations on the collected data before inputting it into the intention recognition module; the battlefield situation data is multimodal data, including the movement state and entity attributes of the target; the battlefield situation data is collected through underlying sensors, radars, etc.;

[0119] In the input module, battlefield situation data is collected, and unified coding and standardization operations are performed on the collected data, including:

[0120] S11, taking enemy air targets in air defense operations as the research object, and then determining the intention space of air targets;

[0121] For different battlefield environments, different combat styles, and different combat objectives, the intention space is quite different, so it needs to be defined in combination with the specific combat background. This invention mainly focuses on aerial targets in air defense operations, such as reconnaissance aircraft, fighter jets, early warning aircraft, etc.

[0122] Specifically, we first define the temporal feature set of the target intent I t From time t1 to time t n The real-time characteristics of continuous air combat at the moment; redefine Y t As the tactical intention space of the target, the problem of air combat target intention recognition is described as: t To Y t The mapping process is:

[0123]

[0124] The problem of air combat intention recognition is the specific mapping of target intention recognition features to target intention recognition types. Considering that target intention recognition in air combat occurs in a real-time, complex, and highly confrontational battlefield environment, and that there are certain limitations in the ability to obtain and analyze information, it is usually difficult for our commanders to accurately obtain the combat intentions of enemy targets. Therefore, the probability of the intention occurring is used to represent the result of intention recognition at that moment. In an actual combat environment, enemy targets may conceal their tactical intentions through various means. Therefore, judging the target's tactical intentions based solely on battlefield data at a single moment is somewhat one-sided and false. Therefore, it is chosen to extract target features from battlefield data obtained at continuous moments to identify its tactical intentions.

[0125] The intention space of the aerial target determined in this embodiment is 7 intentions: {attack, penetration, interference, feint, surveillance, reconnaissance, retreat}. The detailed description of different intentions is shown in Table 2:

[0126] Table 2 Detailed description of the combat intent of air targets

[0127]

[0128] S12, establishing an encoding and decoding mechanism for the determined intent space to perform model training;

[0129] After determining the target intention space, the feature information required by the model can be determined based on the relationship between the target attribute characteristics and the intention. From the perspective of the enemy's combat mission, the enemy's aircraft in the air also have certain differences in the feature information displayed when performing different combat missions. In addition, considering the limitations of information technology, there are target features that cannot be directly obtained, such as target type, shape, etc., and this embodiment chooses to use radar one-dimensional range image and radar reflection cross-sectional area instead.

[0130] The problem of intent recognition is actually a multi-classification problem. Figure 3 The encoding and decoding mechanism shown is used to train the model by encoding and decoding the enemy target intentions.

[0131] S13, after determining the target intention space, the feature information required by the model can be determined based on the relationship between the target attribute characteristics and the intention. From the perspective of the enemy's combat mission, when the enemy's aircraft perform different combat missions, the feature information displayed by the aircraft also has certain differences.

[0132] (I) As for the target's flight speed, when fighter jets are engaged in air combat, their flight speed is generally 735 to 1470 km / h; when bombers and transport aircraft are performing missions, their flight speed is generally 600 to 850 km / h; when early warning aircraft are performing surveillance missions, their flight speed is generally 750 to 950 km / h.

[0133] (ii) As far as the target's flight altitude is concerned, when enemy aircraft are performing penetration missions, the low-altitude penetration altitude is 50 to 200 meters, and the high-altitude penetration altitude is 10,000 to 11,000 meters; the altitude of fighter jets during air combat is generally 1,000 to 6,000 meters; when enemy aircraft are performing reconnaissance missions, the low-altitude reconnaissance flight altitude is 100 to 1,000 meters, and the ultra-high-altitude reconnaissance flight altitude is above 15,000 meters.

[0134] (3) As far as the target's radar status is concerned, fighter jets usually keep their air radars turned on when engaging in air combat; transport aircraft generally keep their radars silent when performing transport missions; bombers keep their air radars or sea radars turned on only when performing bombing missions; and when aircraft are performing reconnaissance missions, both their air radars and sea radars are turned on.

[0135] In addition, considering the limitations of information technology, there are target features that cannot be directly obtained, such as target type, shape, etc. We choose to use radar one-dimensional range image and radar reflection cross-sectional area instead.

[0136] In summary, the feature set for air defense target intention recognition is constructed as {altitude, speed, acceleration, heading angle, azimuth, distance, radar one-dimensional range image, radar reflection cross-sectional area, air radar status, ground radar status, interference status, friend-or-foe identification response} 12-dimensional features, of which the first 8 are numerical features and the last 4 are non-numerical features. The specific description of the features is shown in Table 3.

[0137] Table 3 Aerial target intention recognition feature set information

[0138]

[0139]

[0140] Standardize the feature data of the feature set:

[0141] For numerical feature data, the Min-Max standardization method is used to map it to the interval [0,1]. The specific calculation method is:

[0142]

[0143] In the formula, x represents the eigenvalue of a numerical feature, x′ Represents the normalized result of the numerical feature, where min and max correspond to the minimum and maximum values ​​of the feature in the set respectively;

[0144] The non-numerical feature data are all classified data. To facilitate the learning of the neural network, the non-numerical feature data are numerically processed and then mapped to the interval [0,1]. The specific calculation method is:

[0145]

[0146] In the formula, j represents the size of the classification space, and x is the value of the original i-th non-numerical feature after being mapped to the interval [0,1].

[0147] The historical intention data is labeled data or unlabeled data, which can provide a certain training data set for the deep learning model. The expert knowledge, domain knowledge, and rule knowledge are information extracted from battlefield intelligence, historical information, and expert experience, providing relevant knowledge information and constraint rules for the intention recognition module.

[0148] The storage module is used to store training data and knowledge information. When the intention recognition module performs modeling and training, the training data and knowledge information are retrieved. At the same time, the storage module transmits the real intention data (i.e., training data and knowledge information) to the data generation module for data expansion; the training data and knowledge information include historical intention data, expert knowledge, domain knowledge, and rule knowledge;

[0149] The data generation module is used to expand the historical intention data, so as to provide sufficient training data sets for the deep learning model. Specifically, a data generation model based on a denoising diffusion probability model is used to generate data; the model is as follows Figure 4 As shown, the specific process is:

[0150] Denoising Diffusion Probabilistic Models (DDPM) mainly consists of two processes: a forward process that perturbs data into noise and a reverse process that converts noise into data.

[0151] Given the original data distribution x0~q(x0), Gaussian noise is gradually added to the data during the diffusion process;

[0152] Assume that a total of T steps of diffusion are performed to generate a series of noisy samples x, and the noise parameter in the diffusion process is an increasing sequence β 1:T ∈(0,1] T Determine, then the forward diffusion process gradually transforms the original data distribution q(x0) into the latent variable distribution q(x T ), as shown in formula (4):

[0153]

[0154] in, N represents Gaussian distribution, represents the mean of the Gaussian distribution, (1-α t )Ι represents the covariance matrix of Gaussian distribution, and I is the identity matrix.

[0155] Let α t :=1-β t , but

[0156]

[0157] Where N represents Gaussian distribution, represents the mean of the Gaussian distribution, β t Ι represents the covariance matrix of the Gaussian distribution, and I is the identity matrix.

[0158] By deriving and transforming formula (5), we can calculate the data distribution q(xt |x0):

[0159]

[0160] Among them, ε∈Ν(0,Ι), N represents Gaussian distribution, represents the mean of the Gaussian distribution, (1-α t )Ι represents the covariance matrix of Gaussian distribution, and I is the identity matrix.

[0161] As t increases, the proportion of noise increases and the proportion of original data decreases; When x T It is almost Gaussian distributed, and the diffusion process of the model is considered to be completed;

[0162] The reverse diffusion process of DDPM uses a Markov chain to transform the distribution of latent variables p θ (x T ) is transformed into data distribution p θ (x0). Considering that the noise added in each step of the forward process is small, we assume that p θ (x t-1 |x t ) is also a Gaussian distribution. θ (x t-1 |x t ) is fitted using a neural network, where θ represents the parameters of the neural network;

[0163] When β t When it is close enough to 1, for all x0, q(x t |x0) approaches the standard normal distribution, so q(x t |x0) is set to the standard normal distribution, then the joint probability distribution of the reverse diffusion process is as follows:

[0164]

[0165] Among them, p θ (x t-1 |x t ):=Ν(x t-1 ;μ θ (x t ,t),σ θ (x t ,t) 2 Ι); N represents Gaussian distribution, μ θ (x t ,t) represents the mean of Gaussian distribution, σ θ (x t ,t) 2 Ι represents the covariance matrix of the Gaussian distribution, and I is the identity matrix.

[0166] μ θ Decomposed into noise and x t , we can get the approximate value of the mean as follows (9):

[0167]

[0168] Settings and Beta t Related constants is the variance, and the trainable parameters only exist in the mean, then the generation process is expressed as:

[0169]

[0170] Among them, ε θ represents a neural network with the same input and output, ε θ The noise predicted at each step is used in the back-diffusion process;

[0171] The goal of DDPM is to find a data distribution that can make the dual-objective data distribution p θ (x0) maximizes the parameter θ, at the target data p θ Negative log-likelihood function of (x0) - logp θ A KL divergence is added to (x0) to form an upper bound for the negative likelihood logarithm.

[0172] The denoising loss function of DDPM is expressed as formula (11):

[0173]

[0174] in, represents the predicted noise; E q It means averaging the variables under distribution q; represents the actual noise; C is a constant.

[0175] To further simplify the expression, the following loss function is minimized during training:

[0176]

[0177] In the calculation process, the latent variable x is first sampled from the standard normal distribution T ~N(x T ; 0, Ι), and then get x again from the sampling t-1 When p θ When (x0) is calculated, the iteration process ends. Indicates average.

[0178] Considering that the reverse diffusion process of DDPM often requires hundreds or thousands of steps, the sampling speed is greatly slowed down. In order to speed up the model sampling speed, an improved knowledge distillation method is introduced, such as Figure 5 As shown. In order to meet the needs of future information-based intelligent warfare, multiple teacher models and a single student model are deployed in the cloud computing center and combat terminal respectively. The knowledge distillation method is designed as follows:

[0179] First, design teacher models: design seven teacher models based on the seven combat intentions of air targets The difference between different teacher models lies in their different inputs, which are only single intent data, e.g. The input is the attack intention data;

[0180] Second, design the student model: input the data of the seven intentions into the teacher model at the same time, and select the model with the best performance through multiple experiments, denoted as At this time, the student model N S Use a relatively small U-Net network instead The U-Net network of the model; the number of channels of the U-Net network input layer used in the student model remains unchanged, and the number of channels of the convolutional neural network in other positions is halved. At the same time, a neural network with a convolution kernel of 1 is added to solve the problem of matching the feature size of the U-Net network after the reduction;

[0181] Third, the design of the loss function: the data of the seven combat intentions are simultaneously input into the seven teacher models and the student model. Given the original input data set With θ S The student model with parameters N S and The teacher model with parameters like Figure 5 As shown in the figure, the loss function designed for the knowledge distillation method mainly consists of three parts: First, calculate the cross entropy L between the true label and the student model label ; The second is to calculate the average cross entropy between the teacher model output and the student model output L TS ; The third is to calculate the loss value L of the middle layer inter :

[0182] L total =αL label +βL TS +γL inter (13)

[0183] L label =H(y i ,N S (x i )) (14)

[0184]

[0185] in, and N T and N S Soft output; τ>1 indicates the temperature parameter of the softening operation; represents the output feature map of a certain intermediate layer of the t-th teacher model; F S (·) represents the output feature map of a middle layer of the student model; MSE(·) represents the mean square error between different feature maps; α, β, and γ represent the weights of the three loss values, respectively, and their sum is 1.

[0186] The intention recognition module uses a deep learning-based intention recognition model to train the model according to the historical intention data, knowledge information in the storage module and the generated data in the data generation module, adjusts the various hyperparameters of the model to achieve the optimal effect of the model, and then processes the data input in the input module to obtain the tactical intention of the target in the current state;

[0187] The process of intention recognition is as follows Figure 6 As shown, the following steps are included:

[0188] (1) Data preprocessing: convert non-numerical features into numerical features, and then normalize the features to form a standard data set;

[0189] (2) Feature extraction: dilated causal convolution is introduced to model the connection between inputs at different times and mine the information embedded in the intent data;

[0190] Dilated Causal Convolution (DCC) is a network structure that can process time series data proposed by Shaojie Bai et al. in 2018 in the Temporal Convolution Network (TCN). Compared with CNN, causal convolution can process time series data according to {x1,x2,...,x t} and {y1,y2,...,y t-1 Prediction y t , so that y t Close to the true value.

[0191] In order to fully extract the temporal features of situation data, the idea of ​​Inception is introduced to construct Figure 7 The temporal block shown in Fig. 2 is composed of four parts, each of which has the same dilated causal convolution kernel size but different dilation factors.

[0192] Given the processed input data V = {V1, V2, ..., V m}, after the Temporal Block, it is activated by the sigmoid function and the tanh function respectively to obtain the output feature S a and S b .

[0193] S a =tanh(F(V,s)) (17)

[0194] S b =σ(F(V,s)) (18)

[0195] S=S a +S b (19)

[0196] Among them, tanh(·) represents the tanh function, σ(·) represents the sigmoid function, and F represents the feature vector output by the Temporal Block.

[0197] (3) Enhanced feature representation capability: Introducing the graph attention mechanism to capture the connections between multiple features and dynamically adjust the attention weights of different nodes;

[0198] Graph Attention Network (GAT) is an attention-based node classification network proposed by Petar Velickovic et al. The basic idea is to update the node representation according to the attention of each node on its neighboring nodes.

[0199] At the input layer, GAT receives the node features of the graph data and the topological structure of the graph. For the constructed IDGTIR model, the node features of the graph are the feature vectors extracted by fusing Temporal Block and Conv1d, and the topological structure is the adjacency matrix constructed by the similarity between the feature vectors. The input is the node features. Where M represents the number of nodes, and F represents the feature dimension of the nodes;

[0200] In order to obtain sufficient expressive power, GAT uses a shared linear transformation parameterized by a weight matrix to transform input features into higher-level features. At the same time, a shared attention mechanism a is introduced to calculate the importance of the feature of node j relative to node i. ij :

[0201]

[0202] In the formula, the shared attention mechanism a:R F′ ×RF →R represents a single-layer feedforward neural network, W∈R F′×F represents the weight applied to each node of the graph;

[0203] In order to make the coefficients between different nodes easier to compare, the softmax function is used for normalization to obtain the normalized attention coefficient:

[0204]

[0205] In the formula, is the initialization weight vector of the attention mechanism a, LeakyReLU is used for nonlinear operations, N i is a field of node i in the graph, and || represents a concatenation operation;

[0206] Finally, the normalized attention coefficient is used to calculate the linear combination of the corresponding features, and the output features of the node are obtained through the nonlinear activation function σ:

[0207]

[0208] in, represents the output feature of node i; α ij represents the attention coefficient from node i to node j, W represents the weight matrix, Represents the input features of node j.

[0209] In order to stabilize the learning process of the self-attention mechanism, the attention mechanism is extended to a multi-head attention mechanism. K attention mechanisms run independently, and the output features obtained are concatenated to obtain the final features:

[0210]

[0211] in, represents the final output feature of node i; α ij k represents the attention coefficient from node i to node j in the kth attention mechanism, W k represents the weight matrix, Represents the input features of node j.

[0212] (4) Intent recognition: The output of the graph attention mechanism will be sent to the softmax layer for classification, and finally the target intent recognition result at the current moment is obtained.

[0213] The output module outputs the target's current tactical intention recognition result obtained by the intention recognition module.

[0214] Among them, in the data generation module, Unet uses Adam to update its model parameters, and the number of steps of the forward and backward diffusion processes of DDPM is set to 1000. In the intent recognition module, the model parameters mainly include the optimizer, batch size, training cycle, learning rate, convolution kernel size and dilation factor of the dilated causal convolution in the Temporal Block, and the number of nodes of the graph attention mechanism. For the optimizer, five algorithms are compared: Stochastic Gradient Descent (SGD), Root Mean Square propagation (RMSprop), Adaptive Moment Estimation (Adam), NAdam and Adamax. The test set accuracy is shown in Table 4. In addition, the accuracy of the model on the test set under different batch sizes and learning rates is compared, and the experimental results are shown in Table 5. The optimizer, batch size and learning rate corresponding to the maximum accuracy are selected as the parameters of IDGTIR. Temporal Block module comparison experiments are also added, and the results are shown in Table 6. The experimental accuracy and model parameters of different numbers of modules are compared, and finally the number of Temporal Block modules is set to 2. The model hyperparameter settings of IDGTIR are shown in Table 7. The experiments conducted in this paper use Python 3.8 and Pytorch 1.12.1, accelerated by NVIDIA GeForce RTX2080 Ti GPU and CUDA 12.2.

[0215] Table 4 Test set accuracy obtained by different optimizers

[0216]

[0217] Table 5 Test set accuracy obtained with different parameters

[0218]

[0219] Table 6 Temporal Block module quantity comparison experiment

[0220]

[0221]

[0222] Table 7 IDGTIR model hyperparameter settings

[0223]

[0224] In order to verify the performance of the proposed IDGTIR model, multiple sets of experiments are set up focusing on the following three questions:

[0225] (1) Can the intent data generation module generate high-quality time series data?

[0226] (2) Can the generated intent data be applied to the intent recognition module to obtain more accurate recognition results?

[0227] (3) Can the intention recognition model effectively extract the temporal characteristics of the intention data, and can the recognition process convince the commander?

[0228] 1. Generate intent data visualization

[0229] In order to verify the quality and diversity of the generated intent data, the T-SNE algorithm is introduced to visualize the generated samples. The T-SNE algorithm can intuitively display the sample distribution of the data in the low-dimensional space while maintaining the similarity relationship between the data in the high-dimensional space. The visualization results are as follows: Figure 8 and Fig. 9 As shown, different colors represent different intent data. Figure 3 The proposed method decodes the intent, i.e., intent 0 corresponds to retreat, Figure 1 Corresponding to the breakthrough, Figure 2 Corresponding attack, meaning Figure 3 Corresponding to interference, meaning Figure 4 Corresponding monitoring, Figure 5 Corresponding reconnaissance, meaning Figure 6 Corresponding feint.

[0230] Figure 8 This is the visualization result of the original intention data. After dimensionality reduction and clustering, the sample data of different intentions are cross-distributed, especially the data of attack intention and feint intention are mixed together. This is consistent with the actual combat situation. At the same time, it also brings certain difficulties to the intention recognition task and it is easy to misjudge the two.

[0231] Fig. 9 (a) t-SNE graph of intent data generated by IDDPM, VAE, and GAN. Fig. 9 (b) is the t-SNE graph after the generated intention data is merged with the original data. Fig. 9 It was found that the distribution of intent data generated by IDDPM in the T-SNE graph is better than that of GAN and VAE. Fig. 9 The intention data generated by the three in (a) can be well clustered. The point sets representing the three intentions of interference, surveillance, and reconnaissance in IDDPM are distributed close to each other, indicating that the characteristics of these three intention data are relatively similar. In addition, the point sets representing the two intentions of attack and feint are also distributed close to each other. Fig. 9In (b), after adding the original intent data, the point set of the original data appears in the upper right corner of the t-SNE graph of the GAN-generated data, while the point set of the original data appears in the upper left corner of the VAE. In contrast, the t-SNE graph of IDDPM does not change significantly after adding the original data, proving that it can better learn the feature relationship between intent data and generate high-quality intent data.

[0232] 2. Intent Recognition Experiment

[0233] 1. Analysis of intent recognition results

[0234] We conducted intent recognition experiments based on original data and generated data, and denoted the original data and IDDPM generated data as D and D respectively. IDDPM (GAN generates data, VAE generates data D GAN , D VAE ), the experimental results are shown in Table 8. From Table 8, it can be found that the data generated by IDDPM has the best evaluation indicators in the experimental results obtained by the intention recognition model.

[0235] Table 8 Comparison of different intent datasets

[0236]

[0237] Fig.10 The accuracy and loss curves of DIDDPM during the training of the intent recognition model are given. It can be found that the model has converged after 30 epochs, and the accuracy and loss values ​​have not changed significantly. The accuracy varies around 98%, with the highest being 98.61%, and the loss value is about 0.05.

[0238] The accuracy only reflects the overall recognition effect of the model on the 7 different intent data. Therefore, we further analyzed the recognition accuracy of each intent and generated the confusion matrix of the test set as follows: Fig.11 . Fig.11The depth of the color in the figure represents the accuracy rate. The darker the color, the higher the accuracy rate. The diagonal line represents the recognition accuracy rate of the intention. From the confusion matrix, it can be found that the IDGTIR model has a high recognition accuracy rate for the seven different intentions, especially for the retreat intention, which has a recognition accuracy rate of up to 100%. The reason for the analysis is that the retreat intention is significantly different from the other six intentions in terms of distance, azimuth, heading angle and other features, so it is easier to identify. In addition, we also found that there is a certain probability of misclassification for the three intentions of interference, reconnaissance and surveillance, and there is also a certain probability of misclassification for the two intentions of attack and feint, which is consistent with actual combat. The target of the feint usually takes a variety of tactical actions to confuse the enemy and cover the target of the attack. The characteristics of the two are relatively similar, so it is easy to misjudge. The targets of the surveillance and reconnaissance intentions have certain similarities in the characteristics of height, speed, acceleration, heading angle, azimuth, distance, etc., so it is also easy to misjudge.

[0239] 2. Ablation experiment

[0240] In order to further verify the effectiveness of the IDGTIR model for intent recognition tasks, ablation experiments of the dataset and ablation experiments of the model were designed respectively.

[0241] (1) Ablation experiments on the dataset.

[0242] Experiments were conducted on the original data set, the generated data set, and the data set after the original data and the generated data were combined. The experimental results are shown in Table 9. The change curves of accuracy and loss value are shown in Fig.12 shown.

[0243] Table 9 Dataset ablation experiment results

[0244]

[0245] From Table 9 and Fig.12 It can be seen that the dataset after merging the original dataset D and the generated data DIDDPM has the best effect, with an accuracy increase of 0.12% compared to DIDDPM and an accuracy increase of 1.85% compared to D. The experimental results show that the intent recognition model has a good feature extraction capability for the generated data. With the increase in the number of training times, the accuracy is significantly improved and the loss value is significantly reduced. The accuracy of the merged dataset has not been greatly improved, which is affected by the feature extraction capability of the model. Combined with the analysis of the experimental results, the generated intent data is relatively consistent with the original intent data in feature distribution, indicating that the generated samples are of high quality and can meet the needs of the intent recognition task.

[0246] (2) Model ablation experiment. Different models are trained on the same dataset. The experimental results are shown in Table 10. The curves of accuracy and loss are shown in Table 10. Fig.13 shown.

[0247] Table 10 Model ablation experiment results

[0248]

[0249] From Table 10 and Fig.13 It can be seen that the accuracy, precision, recall, and F1_score of IDGTIR are higher than those of the other two models. Compared with TCN, the accuracy is increased by 2.21%, and compared with TemporalBlock, it is increased by 0.82%, indicating that the introduction of the graph attention mechanism and Inception improves the model's recognition performance of the target tactical intent. With the increase in the number of training times, the model accuracy generally improves and the loss value generally decreases. The effect of IDGTIR is more obvious. In addition, the IDGTIR model has the least number of parameters. To further compare the recognition performance of the model, the recall rate and F1_score are used to reflect the recognition accuracy of different models for the seven intents. The results are shown in Table 11.

[0250] Table 11 Evaluation indicators for different intentions

[0251]

[0252] As can be seen from Table 11, for seven different intents, IDGTIR has the highest recall rate and F1_score. The comprehensive analysis shows that TCN can better extract the temporal features of intent data. After the introduction of the Inception idea, the stacking of multiple TCN networks improves the feature extraction capability and efficiency. The introduction of the graph attention mechanism helps to explore the feature relationship between different intents, thus improving the accuracy.

[0253] 3. Comparative experiments with other methods

[0254] It is compared with seven existing aerial target intention recognition models, including LSTM-Attention, DBP, Attention-TCN-BiGRU, STABC-IR, BiGRU-Attention, PCLSTM, and SVM. The intention recognition accuracy and loss value of the models are compared under the same dataset, as shown in Table 12.

[0255] Table 12 Experimental results of different intent recognition models

[0256]

[0257] The comparison shows that the recognition accuracy and loss value of the IDGTIR model are much better than those of several other intent recognition models. The recognition accuracy and loss value of the IDGTIR model are 98.73% and 0.038 respectively. These further prove that the IDGTIR model has a good effect in the application of aerial target intent recognition tasks, can meet the needs of data scarcity, and better extract time series features.

[0258] In summary, the visualized intent data shows that the samples generated by the model are of high quality. Comparative experiments with other advanced intent recognition methods show that IDGTIR has higher intent recognition accuracy and performance. Ablation experiments also verify the effectiveness of the model.

Claims

1. A deep learning-based aerial target intention recognition model, characterized in that: It includes input module, storage module, data generation module, intention recognition module and output module; The input module is used to collect battlefield situation data, and then perform unified coding and standardization operations on the collected data before inputting it into the intention recognition module; the battlefield situation data is multimodal data, including the movement state and entity attributes of the target; The storage module is used to store training data and knowledge information. When the intention recognition module performs modeling and training, the training data and knowledge information are retrieved. At the same time, the storage module transmits the real intention data to the data generation module for data expansion; the training data and knowledge information include historical intention data, expert knowledge, domain knowledge, and rule knowledge; The data generation module is used to expand the historical intention data, so as to provide sufficient training data sets for the deep learning model, and specifically uses a data generation model based on a denoising diffusion probability model to generate data; The intention recognition module uses a deep learning-based intention recognition model to train the model according to the historical intention data, knowledge information in the storage module and the generated data in the data generation module, adjusts the various hyperparameters of the model to achieve the optimal effect of the model, and then processes the data input in the input module to obtain the tactical intention of the target in the current state; The output module outputs the target's current tactical intention recognition result obtained by the intention recognition module.

2. The deep learning-based aerial target intention recognition model according to claim 1, characterized in that: In the input module, battlefield situation data is collected, and unified coding and standardization operations are performed on the collected data, including S11, taking enemy air targets in air defense operations as the research object, and then determining the intention space of air targets; Specifically, we first define the temporal feature set of the target intent I t From time t1 to time t n The real-time characteristics of continuous air combat at the moment; redefine Y t As the tactical intention space of the target, the problem of air combat target intention recognition is described as: t To Y t The mapping process is: S12, establishing an encoding and decoding mechanism for the determined intent space to perform model training; S13, constructing an air defense combat target intention recognition feature set, including numerical feature data and non-numerical feature data, and then performing a standardization operation on the feature data of the feature set; For numerical feature data, the Min-Max standardization method is used to map it to the interval [0,1]. The specific calculation method is: In the formula, x represents the eigenvalue of a numerical feature, x′ represents the normalized result of the numerical feature, and min and max correspond to the minimum and maximum values ​​of the eigenvalue in the set respectively; The non-numerical feature data are all classified data. To facilitate the learning of the neural network, the non-numerical feature data are numerically processed and then mapped to the interval [0,1]. The specific calculation method is: In the formula, j represents the size of the classification space, and y is the value of the original i-th non-numeric feature after being mapped to the interval [0,1].

3. The deep learning-based aerial target intention recognition model according to claim 1, characterized in that: The historical intent data is labeled data or unlabeled data, and the expert knowledge, domain knowledge, and rule knowledge are information extracted from battlefield intelligence, historical information, and expert experience.

4. The deep learning-based aerial target intention recognition model according to claim 1, characterized in that: The intention data is generated using a denoising diffusion model, and the specific process is as follows: Given the original data distribution x0~q(x0), Gaussian noise is gradually added to the data during the diffusion process; Assume that a total of T steps of diffusion are performed to generate a series of noisy samples x, and the noise parameter in the diffusion process is an increasing sequence β 1:T ∈(0,1] T Determine, then the forward diffusion process gradually transforms the original data distribution q(x0) into the latent variable distribution q(x T ), as shown in formula (4): in, N represents Gaussian distribution, represents the mean of the Gaussian distribution, (1-α t ) Ι represents the covariance matrix of Gaussian distribution, I is the identity matrix; Let α t :=1-β t , but Where N represents Gaussian distribution, represents the mean of the Gaussian distribution, β t Ι represents the covariance matrix of Gaussian distribution, I is the identity matrix; By deriving and transforming formula (5), we can calculate the data distribution q(x t |x0): Among them, ε∈Ν(0,Ι), N represents Gaussian distribution, represents the mean of the Gaussian distribution, (1-α t ) Ι represents the covariance matrix of Gaussian distribution, I is the identity matrix; As t increases, the proportion of noise increases and the proportion of original data decreases; When x T It is almost Gaussian distributed, and the diffusion process of the model is considered to be completed; The reverse diffusion process of DDPM uses a Markov chain to transform the distribution of latent variables p θ (x T ) is transformed into data distribution p θ (x0); Considering that the noise added in each step of the forward process is very small, assuming that p θ (x t-1 |x t ) is Gaussian distribution, p θ (x t-1 |x t ) is fitted using a neural network, where θ represents the parameters of the neural network; When β t When it is close enough to 1, for all x0, q(x t |x0) approaches the standard normal distribution, so q(x t |x0) is set to the standard normal distribution, then the joint probability distribution of the reverse diffusion process is as follows: Among them, p θ (x t-1 |x t ):=Ν(x t-1 ;μ θ (x t ,t),σ θ (x t ,t) 2 Ι); N represents Gaussian distribution, μ θ (x t ,t) represents the mean of Gaussian distribution, σ θ (x t ,t) 2 Ι represents the covariance matrix of Gaussian distribution, I is the identity matrix; μ θ Decomposed into noise and x t , we can get the approximate value of the mean as follows (9): Settings and Beta t Related constants is the variance, and the trainable parameters only exist in the mean, then the generation process is expressed as: Among them, ε θ represents a neural network with the same input and output, and the noise predicted by εθ at each step is used in the back diffusion process; The denoising loss function of DDPM is expressed as formula (11): in, represents the predicted noise; E q It means averaging the variables under distribution q; represents the actual noise; C is a constant; To further simplify the expression, the following loss function is minimized during training: In the calculation process, the latent variable x is first sampled from the standard normal distribution T ~N(x T ; 0, Ι), and then get x again from the sampling t-1 When p θ When (x0) is calculated, the iteration process ends, where Indicates average.

5. The deep learning-based aerial target intention recognition model according to claim 4, characterized in that: An improved knowledge distillation method is introduced in the reverse diffusion process of DDPM to speed up the model sampling speed. The improved knowledge distillation method is designed as follows: First, design teacher models: design seven teacher models based on the seven combat intentions of air targets The difference between different teacher models lies in their different inputs, which are only single intent data; Second, design the student model: input the data of the seven intentions into the teacher model at the same time, and select the model with the best performance through multiple experiments, denoted as At this time, the student model N S Use a relatively small U-Net network instead The U-Net network of the model; the number of channels of the U-Net network input layer used in the student model remains unchanged, and the number of channels of the convolutional neural network in other positions is halved. At the same time, a neural network with a convolution kernel of 1 is added to solve the problem of matching the feature size of the U-Net network after the reduction; Third, the design of the loss function: the data of the seven combat intentions are simultaneously input into the seven teacher models and the student model. Given the original input data set With θ S The student model with parameters N S and θ Tt The teacher model with parameters The loss function designed for the knowledge distillation method mainly consists of three parts: First, calculate the cross entropy L between the true label and the student model label ; The second is to calculate the average cross entropy between the teacher model output and the student model output L TS ; The third is to calculate the loss value L of the middle layer inter : L total =αL label +βL TS +γL inter (13) L label =H(y i ,N S (x i )) (14) in, and N T and N S Soft output; τ>1 indicates the temperature parameter of the softening operation; represents the output feature map of a certain intermediate layer of the t-th teacher model; F S (·) represents the output feature map of a middle layer of the student model; MSE(·) represents the mean square error between different feature maps; α, β, and γ represent the weights of the three loss values, respectively, and their sum is 1.

6. The deep learning-based aerial target intention recognition model according to claim 1, characterized in that: In the intention recognition module, the intention recognition process includes the following steps: (1) Data preprocessing: convert non-numerical features into numerical features, and then normalize the features to form a standard data set; (2) Feature extraction: dilated causal convolution is introduced to model the connection between inputs at different times and mine the information embedded in the intent data; Construct the time module, which consists of 4 parts. The dilated causal convolution kernel size of each part is the same, but the dilation factor is different; Given the processed input data V = {V1, V2, ..., V m }, after the time module, it is activated by the sigmoid function and the tanh function respectively, and the output feature S is obtained a and S b : S a =tanh(F(V,s)) (17) S b =σ(F(V,s)) (18) S=S a +S b (19) Where tanh(·) represents the tanh function, σ(·) represents the sigmoid function, and F represents the feature vector output by the time module; (3) Enhanced feature representation capability: Introducing the graph attention mechanism to capture the connections between multiple features and dynamically adjust the attention weights of different nodes; At the input layer, GAT receives the node features of the graph data and the topological structure of the graph; for the constructed IDGTIR model, the node features of the graph are the feature vectors extracted by the fusion time module and Conv1d, and the topological structure is the adjacency matrix constructed by the similarity between the feature vectors; the input is the node features Where M represents the number of nodes, and F represents the feature dimension of the nodes; GAT uses a shared linear transformation parameterized by a weight matrix to transform input features into higher-level features. At the same time, it also introduces a shared attention mechanism a to calculate the importance of the feature of node j relative to node i. ij : In the formula, the shared attention mechanism a:R F′ ×R F →R represents a single-layer feedforward neural network, W∈R F′×F represents the weight applied to each node of the graph; Use the softmax function for normalization to get the normalized attention coefficient: In the formula, is the initialization weight vector of the attention mechanism a, LeakyReLU is used for nonlinear operations, N i is a field of node i in the graph, and || represents a concatenation operation; Finally, the normalized attention coefficient is used to calculate the linear combination of the corresponding features, and the output features of the node are obtained through the nonlinear activation function σ: in, represents the output feature of node i; α ij represents the attention coefficient from node i to node j, W represents the weight matrix, represents the input features of node j; In order to stabilize the learning process of the self-attention mechanism, the attention mechanism is extended to a multi-head attention mechanism. K attention mechanisms run independently, and the output features obtained are concatenated to obtain the final features: in, represents the final output feature of node i; α ij k represents the attention coefficient from node i to node j in the kth attention mechanism, W k represents the weight matrix, represents the input features of node j; (4) Intent recognition: The output of the graph attention mechanism will be sent to the softmax layer for classification, and finally the target intent recognition result at the current moment is obtained.

Images