Target tactical intention recognition method based on PCA-GTN

Abstract

The invention discloses a target tactical intention recognition method based on PCA-GTN, relates to the technical field of tactical intention recognition, and provides a target tactical intention recognition method based on PCA-GTN, which is characterized in that on the basis of air combat time sequence data, dimension reduction processing of target data feature dimensions is realized by using a principal component analysis (PCA) method, and the target tactical intention recognition method based on PCA-GTN is realized. According to the method, a target tactical intention recognition model based on GTN (Generic TransformerNetworks) is constructed, and a target tactical intention recognition model based on GTN (Generic TransformerNetworks) is constructed. According to the model, a gating module is added on the basis of a basic Transform model, the feature dimension and the time dimension of an air combat target data set are organically combined, and the classification function of multivariate time series data is effectively achieved. Through comparative analysis with other traditional intention recognition models, it is verified that the model has high competitiveness in the field of intention recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of tactical intention recognition, and in particular to a target tactical intention recognition method based on PCA-GTN. Background Art

[0002] Target tactical intent identification belongs to the second stage of the situational awareness model—the situation understanding layer. As a key component of situational understanding, it not only extracts and classifies features from the preprocessed data from the first stage but also lays a solid foundation for subsequent battlefield threat assessment and situation prediction. During intense air combat, real-time identification of the target's tactical intent is essential for understanding enemy movements and assisting our own tactical decision-making. To maintain an active advantage in the air battlefield and facilitate the rapid completion of our "OODA (Observation, Orientation, Decision, Action)" cycle, accurate and efficient identification of the target's intent is essential.

[0003] Air combat target tactical intention recognition is a comprehensive analysis of the state of enemy aerial targets based on a thorough understanding of data sources. By acquiring real-time information on enemy target position, velocity, angle, and other status information, it is possible to quickly and accurately identify enemy target tactical maneuvers, predict target state trends, and determine the enemy pilot's tactical intentions. This in turn provides effective decision support to our pilots, accelerates our "OODA" loop during air combat, and significantly enhances our positional advantage during air combat confrontations.

[0004] In the current information age, modern air combat is characterized by high dynamics, uncertainty in battlefield intelligence, and a lack of expert knowledge. This makes it difficult for pilots to accurately and efficiently judge the target's behavioral intentions in real time during actual combat, easily putting them in a passive position. Therefore, in actual air combat, selecting an appropriate recognition mechanism is necessary to effectively utilize observation data, which is also a key factor in ensuring the accuracy of subsequent threat assessment and situation prediction. Summary of the Invention

[0005] In response to the above-mentioned problems, the present invention aims to provide a target tactical intention recognition method based on PCA-GTN. The data set processed by PCA can enable the model to have higher learning efficiency and convergence speed, effectively improving its recognition accuracy, and at the same time effectively accelerating its recognition speed in actual combat.

[0006] The main idea of the technical solution adopted by this invention is to encapsulate expert experience into a labeling system to address the problem of identifying the tactical intentions of air combat targets. However, considering the extremely complex battlefield environment in actual air combat confrontations, it is difficult to quickly and accurately identify the enemy's tactical intentions based solely on human experience. Moreover, judging the enemy's tactical intentions based solely on current information is unscientific and has a low recognition rate. Therefore, a GTN algorithm model is introduced. To further improve the training efficiency and recognition accuracy of the model, the PCA method is introduced to reduce the dimension of air combat data, and a PCA-GTN air combat target tactical intention recognition model is proposed. Through theoretical analysis and simulation, the effectiveness and timeliness of the model are verified, providing a theoretical reference for air combat target tactical intention recognition technology and providing a theoretical basis and data support for subsequent air combat threat perception.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] The target tactical intention recognition method based on PCA-GTN is characterized by comprising the following steps:

[0009] Step 1: Obtain data on hostile targets during air combat confrontation;

[0010] Step 2: Use the PCA method to reduce the dimension of the data obtained in step 1;

[0011] Step 3: Extract and classify the data after dimensionality reduction in step 2 according to different target features to obtain target feature data;

[0012] Step 4: Build an intent tag library based on the target feature data obtained in step 3;

[0013] Step 5: Improve the encoder-decoder in the Transformer model and use the improved Transformer model as the target tactical recognition model;

[0014] Step 6: Based on the target tactical recognition model in step 5, the correlation between the feature dimension and the time dimension of the data after dimensionality reduction in step 2 is extracted. The extracted information is then merged using a gating module. After passing through a linear transformation layer and a softmax layer, the classification result is output.

[0015] Step 7: Match the classification results output in step 6 with the tactical intent in the intent label library in step 4 to achieve tactical intent recognition of the hostile target.

[0016] Furthermore, step 3 includes the following sub-steps:

[0017] Step 3-1: Extract the target feature time series information from the data after dimensionality reduction in step 2 using the following formula:

[0018] X=f(V,Q,D,…)

[0019] In the formula, V, Q, D, ... are the target speed, angle, distance and other state information, among which, is the set of speed information of n time points, is the set of angle information of n time points, is the set of distance information of n time series points; f(*) is the mapping process of feature extraction; X is the extracted target feature time series information;

[0020] Step 3-2: Further mine and analyze the target feature time series information extracted in step 3-1 to obtain the mapping relationship between target tactical intent and target features:

[0021] Y=f(X)=f(X E ,X A ,X S ,…)()

[0022] Where Y is the target's possible tactical intention set; f(*) is the mapping relationship between the feature set and the intention set; is a set of characteristics of the battlefield environment; is the feature set of the target state; A feature set for friendly-enemy interactions;

[0023] Step 3-3: Based on the mapping relationship in step 3-2, the target features are classified to obtain target feature data.

[0024] Furthermore, step 4 includes the following steps:

[0025] Step 4-1: Based on the target feature data obtained in step 3, select multiple tactical intentions, set corresponding types of labels to encode the multiple tactical intentions, and construct a label space;

[0026] Step 4-2: Decode each label in the label space constructed in step 4-1 to obtain multiple tactical intent recognition results. The collection of multiple tactical intent recognition results is the intent space;

[0027] Step 4-3: Construct an intent tag library through the label space in step 4-1 and the intent space in step 4-2, as well as the encoding and decoding mechanism between the label space and the intent space.

[0028] Furthermore, the improved Transformer model in step 5 is obtained by changing the embedding layer of the Transformer model to a fully connected layer and adding a nonlinear activation function instead of the linear transformation; the expression of the nonlinear activation function is:

[0029]

[0030] Among them, tanh(x) is a typical hyperbolic tangent function activation function of a neural network, x∈R, -1<tanh(x)<1.

[0031] Furthermore, step 6 includes the following:

[0032] Step 6-1: Introduce two mechanisms to extract the correlation of feature dimension and time dimension respectively in the improved Transformer model in step 5. The encoder in each mechanism captures the correlation of time dimension and feature dimension through self-attention mechanism and mask attention respectively to obtain extracted information;

[0033] Step 6-2: The extracted information obtained in step 6-1 is merged through the gating module, and the weight of each mechanism is learned through the gating module to obtain the gating weight of each mechanism;

[0034] Step 6-3: The gating weights of each mechanism in step 6-2 are set to vector h using the nonlinear activation C and S of the fully connected layer. After a linear transformation layer and a softmax layer, the output weights g1 and g2 of each encoder are obtained. Then, each gating weight participates in the output of the corresponding mechanism and outputs the final feature vector y. The feature vector y is the final classification result. The expression of the feature vector y is:

[0035] h=W·Concat(C,S)+b

[0036] g1,g2=Softmax(h)

[0037] y=Concat(C·g1,S·g2)

[0038] Where C is the output matrix of the feature dimension encoder, S is the output matrix of the time dimension encoder, Concat(*) is the matrix concatenation operation, W is the linear transformation weight matrix, b is the linear transformation bias vector, h is the linear transformation output, g1 and g2 are the weight matrices of the two encoders respectively, and y is the output vector of the gating module.

[0039] The present invention also discloses a target tactical intention recognition system based on PCA-GTN, which is characterized by comprising:

[0040] Data acquisition module: used to obtain data on hostile targets during air combat confrontation;

[0041] PCA dimensionality reduction module: Use PCA to reduce the dimensionality of the data obtained in the data acquisition module;

[0042] Feature extraction and classification module: used to extract and classify the data after dimensionality reduction in the PCA dimensionality reduction module according to different target features to obtain target feature data;

[0043] Build intent tag library module: used to build intent tag library based on target feature data obtained by feature extraction and classification module;

[0044] Constructing a target tactical recognition model module: Improve the encoder-decoder in the Transformer model and use the improved Transformer model as the target tactical recognition model;

[0045] Output classification result module: Based on the target tactical recognition model in the target tactical recognition model construction module, the module extracts the correlation between the feature dimension and the time dimension of the data after dimensionality reduction in the PCA dimensionality reduction module, then uses the gating module to merge the extracted information, and outputs the classification result after a linear transformation layer and a softmax layer;

[0046] Tactical intention recognition module: used to match the classification results output by the output classification result module with the tactical intentions in the intent label library in the intention label library construction module, so as to realize the tactical intention recognition of the hostile target.

[0047] Furthermore, the feature extraction and classification module includes the following:

[0048] Extracting target feature time series information submodule: used to extract target feature time series information from the data after dimensionality reduction in the PCA dimensionality reduction module;

[0049] Mining and analysis submodule: further mines and analyzes the target feature time series information extracted in the target feature time series information extraction submodule to obtain the mapping relationship between the target tactical intention and the target feature;

[0050] Classification submodule: Based on the mapping relationship of the mining and analysis submodule, the target features are classified to obtain target feature data.

[0051] Furthermore, building the intent tag library module includes the following:

[0052] Constructing the label space submodule: Based on the target feature data obtained in the feature extraction and classification module, select multiple tactical intentions, set corresponding types of labels to encode the multiple tactical intentions, and construct the label space;

[0053] Constructing the intention space submodule: Decode each label in the label space constructed in the constructing label space submodule to obtain multiple tactical intent recognition results. The collection of multiple tactical intent recognition results is the intention space;

[0054] Construct the intent tag library sub-module: Construct the intent tag library by constructing the label space in the label space sub-module and the intent space in the intent space sub-module, as well as the encoding and decoding mechanism between the label space and the intent space.

[0055] The beneficial effects of the present invention are as follows: compared with the prior art, the improvement of the present invention is that:

[0056] (1) The PCA-GTN model proposed in this application can effectively fit the selected air combat dataset, and the dataset processed by PCA can enable the model to have higher learning efficiency and convergence speed, effectively improving its recognition accuracy, and at the same time effectively accelerating its recognition speed in actual combat conditions.

[0057] (2) Through comparative analysis of the feature and time dimensions of the air combat dataset, different tactical intentions are sensitive to changes in various feature information. This application extracts the feature weights of various tactical intentions through the self-attention mechanism. The extraction of feature weights can accurately assess the current situation information in real time and make decision-making actions, which is a key step in preserving oneself and defeating powerful enemies. It can also provide a reference for performance and threat assessment of subsequent research objectives. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of the "OODA" loop feedback of the present invention.

[0059] Figure 2 A graph of factors influencing the present invention is identified.

[0060] Figure 3 This is a diagram of the tactical intent identification process of the present invention.

[0061] Figure 4 It is the target feature space composition diagram of the present invention.

[0062] Figure 5 This is the tactical intent tag library of the present invention.

[0063] Figure 6 This is the basic Transformer model structure diagram of the present invention.

[0064] Figure 7 This is the structural diagram of the self-attention mechanism of the present invention.

[0065] Figure 8 This is the structural diagram of the multi-head attention mechanism of the present invention.

[0066] Figure 9 This is the structure diagram of the target intention recognition model based on GTN in the present invention.

[0067] Figure 10 Schematic diagrams of six typical tactics of the present invention.

[0068] Figure 11 This is a comparison chart of the accuracy of each model of the present invention.

[0069] Figure 12 This is a comparison chart of the loss values of each algorithm in the present invention.

[0070] Figure 13 This is a comparison chart of the training time of each model of the present invention.

[0071] Figure 14 This is a comparison chart of the test time of each model of the present invention. DETAILED DESCRIPTION

[0072] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0073] Air combat target tactical intention recognition is a comprehensive analysis process of the enemy's air target status based on a full grasp of the data source. Through real-time acquisition of enemy target position, speed, angle and other status information, it can quickly and accurately identify the enemy's target tactical actions, predict the changing trend of the target's status, and judge the enemy pilot's tactical intention, thereby providing effective auxiliary decision-making for our pilots, accelerating the "OODA" cycle process during our air combat, and greatly helping to enhance our positional advantage in the process of air combat confrontation. The position of target tactical intention recognition in the "OODA" cycle is as follows: Figure 1 shown.

[0074] In the current information age, modern air combat is characterized by high dynamics of confrontation, uncertainty of battlefield intelligence, and lack of expert knowledge. This makes it difficult for pilots to accurately and efficiently judge the target's behavioral intentions in real time during actual confrontations, and they are easily put into a passive position. Therefore, in the actual air combat confrontation process, it is necessary to select a suitable recognition mechanism to effectively utilize the observation data, which is also a key factor in ensuring the accuracy of subsequent threat assessment and situation prediction. Factors affecting the intention recognition problem include Figure 2 shown.

[0075] This application is based on air combat time series data, uses principal component analysis (PCA) to achieve dimensionality reduction of target data feature dimensions, and constructs a target tactical intention recognition method based on GTN (Gated Transformer Networks). Figures 1-14 .

[0076] The process of tactical intent recognition can also be regarded as a multivariate time series classification problem. For this problem, traditional analysis methods include algorithms based on Euclidean distance and Dynamic Time Warping (DTW), which directly measure and classify the original time series data. The DTW algorithm with K-nearest neighbor as the classifier has achieved good results in this field and has been the gold standard in recent decades. The traditional Transformer network completes the sequence generation and prediction tasks through the superposition of encoders and decoders and position embedding. In combination with the characteristics of multivariate time series classification tasks, this application proposes a multivariate time series classification method based on PCA-based Gated Transformer Networks (GTN). First, PCA is used to reduce the dimension of the feature data, and then the encoder and decoder in the basic Transformer model are slightly modified so that the two modules extract the correlation between the feature dimension and the time dimension respectively, which is called the "dual tower" mechanism. Then, a gating module is added to combine the information of the two towers to output the classification results. The GTN model structure based on the dual tower is as follows: Figure 9 shown.

[0077] The target tactical intention recognition method based on PCA-GTN disclosed in this application includes the following steps:

[0078] The target tactical intention recognition method based on PCA-GTN includes:

[0079] Step 1: Obtain data on hostile targets during air combat confrontation;

[0080] Specifically, most of the data obtained during air combat confrontation is time series data, so the tactical intention recognition of the target is the process of extracting features from a period of time series data and classifying them. Therefore, in the actual air combat confrontation process, the more complete and accurate the enemy aircraft information obtained, the more significant and easy to distinguish its intention features will be. By mining and analyzing the correlation and differences between time series data and matching and comparing them with the samples in the intent library, the accuracy of intent recognition can be effectively improved, thereby improving the fault tolerance of the battlefield. The target tactical intention recognition process based on time series data is as follows: Figure 3 shown.

[0081] Step 2: Use the PCA method to reduce the dimension of the data obtained in step 1;

[0082] Specifically, based on the characteristics of the collected air combat target data, it can be divided into two dimensions. The first is the temporal dimension. Since the information collected by sensors during air combat is time series data, its temporal characteristics can be used to identify the target's motion patterns and predict its next state. The second is the feature dimension. Different sensors collect various target state data, forming a multidimensional time series dataset. By classifying this multidimensional time series data, the target's state characteristics can be extracted and its combat intent can be identified, providing sufficient data support for battlefield situational awareness. Considering that the original air combat dataset contains too many and complex feature dimensions, and some features are correlated, which can easily lead to feature redundancy and affect training and recognition efficiency, PCA is used to improve the GTN model. PCA is used to reduce the feature dimensions of the air combat dataset and extract the features with the highest correlation with various tactical intents. These features are then used in the training and recognition of the PCA-GTN tactical intent recognition model. The main features extracted after PCA processing that are highly correlated with various tactical intents are summarized in Table 1.

[0083] Table 1 Correlation between target intention and main characteristics

[0084]

[0085]

[0086] It should be noted that the use of PCA method to reduce the dimensionality of data is an existing technology, which comes from "Air Combat Target Threat Assessment Based on PCA-LM" [J] Firepower and Command Control, 2024, 49(2): 63-68. "High-level Reconstruction and Evaluation of Air Combat Maneuvers Based on PCA", the specific dimensionality reduction process will not be repeated here.

[0087] Step 3: Extract and classify the data after dimensionality reduction in step 2 according to different target features to obtain target feature data;

[0088] Wherein, step 3 includes:

[0089] Step 3-1: Extract the target feature time series information from the data after dimensionality reduction in step 2 using the following formula:

[0090] X=f(V,Q,D,…)

[0091] In the formula, V, Q, D, ... are the target speed, angle, distance and other state information, among which, is the set of speed information of n time points, is the set of angle information of n time points, is the set of distance information of n time series points; f(*) is the mapping process of feature extraction; X is the extracted target feature time series information;

[0092] Step 3-2: Further mine and analyze the target feature time series information extracted in step 3-1 to obtain the mapping relationship between target tactical intent and target features:

[0093] Y=f(X)=f(X E ,X A ,X S ,…)()

[0094] Where Y is the target's possible tactical intention set; f(*) is the mapping relationship between the feature set and the intention set; is a set of characteristics of the battlefield environment; is the feature set of the target state; A feature set for friendly-enemy interactions;

[0095] Step 3-3: Based on the mapping relationship in step 3-2, the target features are classified to obtain target feature data.

[0096] It should be noted that in the actual air combat confrontation process, the target feature space mainly involved is as follows Figure 4 As shown in the figure, the battlefield environment characteristics in air combat mainly include the information combat environment, meteorological environment, electromagnetic environment, and terrain environment. Among them, the information combat environment includes a large amount of battlefield information shared by land-based, sea-based, air-based, and space-based information nodes; the meteorological environment affects tactics and weapon use; the electromagnetic environment includes active and passive interference in the air battlefield; and the terrain environment includes the undulations of the terrain in certain areas of the battlefield.

[0097] Target characteristics mainly include the target's speed, acceleration, altitude, weapon status, radar status, etc. Among them, the target's speed, acceleration, and altitude are numerical time series data, while the weapon status and radar status are non-numerical data. Generally, the radar status of the target carrier aircraft will vary when performing different missions. For example, a fighter jet may activate its air-to-air radar during a dogfight, while a reconnaissance aircraft may activate its air-to-air and opposite-direction radars during a reconnaissance mission.

[0098] The main characteristics of enemy-friendly interaction include enemy-friendly distance, azimuth, altitude difference, and enemy-friendly equipment type. These characteristics can all be calculated by the fire control system, while enemy-friendly equipment type mainly includes information such as reconnaissance aircraft, bombers, and fighter jets, as well as their models.

[0099] Different target feature data may contain different tactical intent information, and the target will also have restrictions on each state to achieve its tactical intent. The target state and its corresponding tactical intent information are shown in Figure 2.

[0100] Table 2 Different target states correspond to possible target intentions

[0101]

[0102]

[0103] Target tactical intent recognition can be considered a form of pattern recognition, and the proper selection of recognition parameters is a crucial step in this process. First, intent classification is the prerequisite for intent recognition and also its output. Based on the acquired target data, the target's tactical intent is categorized into six types: reconnaissance, assault, jamming, attack, evasion, and escape.

[0104] (1) Reconnaissance Intention

[0105] From the perspective of spatial scope, reconnaissance methods can be divided into aerial reconnaissance and space reconnaissance. This application focuses on the study of enemy aerial reconnaissance methods. The enemy's methods for carrying out aerial reconnaissance can be basically divided into visual reconnaissance, electronic reconnaissance, television imaging reconnaissance, and nuclear radiation reconnaissance. They mainly target our ground, sea, and air combat targets, aiming to ascertain the location of important military targets such as our political and economic centers, transportation hubs, and their weapon configuration, troop deployment, and troop camouflage. When carrying out reconnaissance, fighter jets need to detect targets within the airspace, so they need to keep both radars powered on and the range of changes in each state attribute is small.

[0106] (2) Intent to Assault

[0107] After completing rigorous enemy reconnaissance and search missions, the enemy will formulate a comprehensive strike strategy based on strategic and tactical requirements and launch a surprise attack. The goal is usually to destroy key areas of our ground or maritime air defense systems, or to shoot down large carriers such as supply transport aircraft or bombers in the air. When carrying out surprise attacks, rapid approach to the target is necessary, often involving switching between medium and short ranges. When attacking ground targets, enemy aircraft will also rapidly descend in altitude and significantly increase in speed to avoid detection by ground radar.

[0108] (3) Interference Intention

[0109] Enemy aircraft use electromagnetic interference and other methods to reduce the effectiveness of our electronic equipment and detection systems, also known as electronic jamming. This is intended to ensure mission success and enhance the survivability of enemy aircraft. Jamming operations are often coordinated with other combat units. A typical example in air combat is during an air force penetration or assault mission, where the aircraft or its wingman maintains the jamming source, maneuvering within a defined area to provide cover for their own operations or those of the attacking group.

[0110] (4) Attack Intent

[0111] The purpose of air combat is to preserve oneself and destroy the enemy. When enemy aircraft have the necessary positioning, they will accelerate the OODA loop closing process to launch missiles first. This puts our aircraft in a tactical passive position, forcing them to employ evasive maneuvers. When executing an attack, we usually have a clear positioning advantage and maintain radar lock.

[0112] (5) Avoidance intention

[0113] When an enemy aircraft is locked on by a missile, it usually performs evasive maneuvers. When both sides are in the beyond-visual-range phase, our missile is launched at maximum attack range. Until the missile enters terminal guidance, the enemy aircraft can perform lateral maneuvers or activate electromagnetic interference to escape the missile's tracking. When both sides are in the within-visual-range combat phase, it is necessary to avoid being locked on by the enemy's radar or missiles as much as possible. At this time, high-maneuverability maneuvers such as rolling left and right, flying towards the sun, and somersaults may be performed, and various state attributes will change dramatically.

[0114] (6) Intent to escape

[0115] Because the g-forces a pilot can withstand are limited, generally no more than 8g, while missiles can reach a maximum g-force of 30g, aircraft maneuverability is far inferior to missiles. When conventional evasive maneuvers are insufficient to ensure safety, or when an aircraft enters a no-escape zone, it is necessary to maintain high-speed maneuvers while deploying infrared decoys or employing electronic jamming to confuse missiles.

[0116] Step 4: Build an intent tag library based on the target feature data obtained in step 3;

[0117] Specifically, building the intent tag library described in step 4 includes:

[0118] Step 4-1: Based on the target feature data obtained in step 3, select multiple tactical intentions, set corresponding types of labels to encode the multiple tactical intentions, and construct a label space;

[0119] Step 4-2: Decode each label in the label space constructed in step 4-1 to obtain multiple tactical intent recognition results. The collection of multiple tactical intent recognition results is the intent space;

[0120] Step 4-3: Construct an intent tag library through the label space in step 4-1 and the intent space in step 4-2, as well as the encoding and decoding mechanism between the label space and the intent space.

[0121] More specifically, due to the wide variety of target features and their non-uniform units, it is necessary to organize the feature data separately before performing intent recognition to eliminate the influence of different dimensions. The data used in this application includes both numerical data and non-numerical data. For numerical data such as speed, altitude, and distance, normalization is adopted; for non-numerical data such as radar status and aircraft type, a digitization method is adopted. The specific construction method is as follows:

[0122] (1) Normalization of numerical data. The purpose is to eliminate the influence of different dimensions and accelerate network convergence. The target's speed, height, position and other data are normalized as follows:

[0123]

[0124] Where, X i is a time series set of a certain feature of the target; x t is the characteristic value of the target at a certain time point; x t ' is the characteristic value of the time series point obtained after normalization.

[0125] (2) Numericalization of non-numeric data. The purpose is to concretize abstract feature data into data that can be used for network learning. For example, non-numeric data such as battlefield weather conditions and target radar status can be digitized as follows:

[0126]

[0127] In the formula, g(*) is the numerical mapping process of non-numeric data, X i is a non-numerical feature, X i ' is a numerical feature, such as the target radar state feature X i ={'off','on'} becomes X after the mapping process i '={0,1}, meteorological environment characteristics X i ={'clear', 'thunderstorm', 'fog', 'strong wind', 'snow'} is mapped to X i '={0,1,2,3,4};x i ' is the label corresponding to a certain attribute, x" i ∈[0,1].

[0128] Tactical Intent Space Labeling

[0129] Based on the acquired target feature data, this application selects six tactical intentions, including reconnaissance, assault, interference, attack, avoidance, and escape, and sets six types of labels {1, 2, 3, 4, 5, 6} to encode these intentions. The encoded intent library can be regarded as the output value of the tactical intent recognition model, and the tactical intent recognition result is obtained after decoding. This combination of intent space, intent label space and the encoding and decoding mechanism between them is called the intent label library. Its internal mechanism is as follows Figure 5 shown.

[0130] Step 5: Improve the encoder-decoder in the Transformer model and use the improved Transformer model as the target tactical recognition model;

[0131] Specifically, first explain the basic principles of Transformer

[0132] Over the past few years, the Transformer has made remarkable innovations in deep learning. Initially proposed for natural language processing (NLP), the model has since been extended to other fields, including computer vision (CV). Due to its strong generalization capabilities, numerous Transformer variants have emerged, improving upon the Transformer from various perspectives. These have achieved impressive results in various fields, demonstrating excellent performance across a wide range of tasks.

[0133] The basic structure of the Transformer model is as follows Figure 6 As shown. The classic Transformer model generally has a combination of N layers of encoders and decoders. Each encoder layer in the left box has two sublayers. The first sublayer is the Multi-Head Attention (MHA) mechanism, which is used to build the internal relationship of the input data; the second sublayer is a basic feedforward network layer that is fully connected by position. It linearly transforms and activates the input vector at each position and passes the encoder output to the decoder. The N encoders use parallel computing, which greatly improves the operating efficiency. Residual connection [Deep residual learning for image recognition] is used between the two sublayers, and then layer normalization [layernormalization] is used.

[0134] Assume that the input vector is x=(x1,x2,…x n ), after passing through the input embedding layer and the position encoding layer, the input vector passed to the self-attention layer is:

[0135] X=Embedding(x)+Positional_Encoding

[0136] Among them, the input embedding layer Embedding(x) mainly converts each element in the input into a vector form, which can be encoded in the form of one-hot encoding. The positional encoding layer Positional_Encoding mainly uses the sine and cosine functions to encode the position of the input elements. The formula is:

[0137]

[0138] In the formula, pos represents the position, 2i represents the even dimension, 2i+1 represents the odd dimension, and d model Indicates the input feature dimension, PE (pos,2i) Indicates the position information when the dimension is even, PE (pos,2i+1) Indicates the position information when the dimension is an odd number.

[0139] The output of the input vector after passing through the multi-head attention layer is:

[0140] MHA(Q,K,V)=Concat(h1,…,h H )W O

[0141] Where Q, K, and V represent the query matrix, key matrix, and value matrix of the input vector attention respectively; Concat(*) represents the concatenation operation on the vector; h i is the feature vector of the i-th attention head, H is the number of heads of the attention mechanism, h i The calculation formula is:

[0142] h i =Attention(QW i Q ,KW i K ,VW i V )

[0143] Where W i Q 、W i K 、W i V They represent the weight matrices corresponding to Q, K, and V of the i-th head input respectively.

[0144] The Feed-Forward Network (FFN) layer is a fully connected layer consisting of two linear transformations and a ReLU activation function. Its main function is to prevent the degradation of the model output [Attention is not all you need: pure attention loses rank doubly exponentially with depth]. Its calculation formula is:

[0145] FFN(x)=max(0,xW1+b1)W2+b2

[0146] Where W1 and W2 are weight matrices, b1 and b2 are bias terms, x is the input vector of FFN, and max is the maximum value operation.

[0147] A residual module is connected to the output of each sub-layer, i.e. Figure 6 The "Sum & Normalization" module in . The main function of this part is to solve the problem of gradient disappearance and weight matrix degradation. The calculation formula of its residual connection is:

[0148] X attention =X+MHA(Q,K,V)

[0149] X' attention =LayerNorm(X attention )

[0150] X FFN =X' attention +FFN(X' attention )

[0151] X' FFN =LayerNorm(X FFN )

[0152] Among them, Xattention is the attention time series set (non-numeric), X is the original input vector, MHA multi-head attention vector, Q, K, V are the query matrix, key matrix and value matrix of the input vector attention respectively, X'attention is the attention numerical feature set, XFFN feedforward network layer vector,

[0153] The function of FFN(*) is to process the input vector in the feedforward network layer. X'FFN is the feedforward network feature set. The function of LayerNorm(*) is to normalize the hidden layer in the neural network to make it into a standard normal distribution. This can speed up the training speed and accelerate the network convergence. The specific calculation process is as follows:

[0154]

[0155] Formula (1) means finding the mean of the jth layer; μ j represents the mean value of the jth layer, i represents the vector dimension (1 to m);

[0156] Formula (2) expresses the variance of the jth layer; σ 2 j represents the variance of the jth layer, and xij represents the value of the jth layer in the i-th dimension;

[0157] The function of LayerNorm(*) in formula (3) is to normalize the hidden layer in the neural network to make it into a standard normal distribution. This can speed up the training speed, accelerate the convergence of the network, and obtain the normalized value. The function of ε is to prevent the denominator from being 0.

[0158] The Transformer decoder is largely identical to the encoder, differing in its masked self-attention mechanism. Because it abandons the recurrent mechanism of traditional RNN models and introduces a self-attention mechanism that parallelizes input, a mask mechanism is introduced to mask some of the decoder's inputs to prevent mutual influence between input vectors.

[0159] The Mask mechanism used by the basic Transformer model is mainly divided into padding mask (PM) and sequence mask (SM).

[0160] (1) PM exists in both the encoder and decoder modules. Since the lengths of the input sequences in the two modules are different, there will be deviations when calculating attention. To ensure that the lengths of the input sequences are the same, the shorter sequences need to be padded. Since filling these positions with 0 is meaningless and will affect the global probability value after the softmax operation, in order to ensure that the padded positions do not affect the subsequent calculation process, they need to be padded with negative infinity. After passing through the softmax layer, these positions will become 0 and will not affect the calculation of the global probability.

[0161] (2) SM only exists in the first multi-head attention module in the decoder, so this module is called the Masked Multi-Head Attention (MMHA) mechanism. In the test and verification phase, since the data is input in parallel, when the model decodes the i-th input vector, it must completely rely on the output information of i and before i. Therefore, the data after i must be masked to prevent it from affecting the model's prediction effect. Therefore, during the training phase of the model, the subsequent information should be masked. This can not only save training time, but also reduce the risk of overfitting and improve the generalization ability of the model.

[0162] The core of the Transformer algorithm lies in the application of its self-attention mechanism. Its principle is to use the characteristics of human observation to find the correlation between data, obtain the attention weights between input vectors, and construct its feature matrix. Its basic structure is as follows Figure 7 shown.

[0163] The process is mainly divided into the following steps:

[0164] Step 1: Input vector x i Multiply by W Q 、W K 、W V Weight matrix, mapped to three subspaces - Q i (Query), K i (Key), V i (Value), representing the input vector x i The query vector, key vector and value vector of

[0165] Step 2: According to Q i With all K, calculate x i Correlation with other input vectors;

[0166] Step 3: Normalize it through the softmax function and enhance the differences between the data to facilitate the calculation of its self-attention weight;

[0167] Step 4: Combine the weight coefficient with V i Perform weighted summation to obtain its attention value.

[0168] This process is called scaled dot-product attention and is calculated as:

[0169]

[0170] Where Q and K are the query vector and key vector of the input vector attention respectively, V is the value vector of attention; softmax is the activation function; The role of K is to scale the dot product to avoid it being too large. T represents the transpose of the weight matrix.

[0171] The multi-head attention mechanism of the basic Transformer model is to generate multiple groups of W Q 、W K 、W V The weight matrix performs h linear projections on the Q, K, and V of the input vector x, respectively, and learns the d k d k and d vThen, this application calculates the self-attention value of each group Q, K, V in parallel to generate d v These values are concatenated and projected again to obtain the final value. The specific process is as follows: Figure 8 shown.

[0172] Using the improved Transformer model as the target tactical recognition model

[0173] Specifically, by changing the embedding layer of the Transformer model to a fully connected layer and adding a nonlinear activation function tanh to replace the linear transformation, the improved Transformer model is obtained, and its activation function is:

[0174]

[0175] Among them, tanh(x) is a typical hyperbolic tangent function activation function of a neural network, x∈R, -1<tanh(x)<1.

[0176] Step 6: Based on the target tactical recognition model in step 5, the correlation between the feature dimension and the time dimension of the data after dimensionality reduction in step 2 is extracted. The extracted information is then merged using a gating module. After passing through a linear transformation layer and a softmax layer, the classification result is output.

[0177] Step 6 includes:

[0178] Step 6-1: Introduce two mechanisms to extract the correlation of feature dimension and time dimension respectively in the improved Transformer model in step 5. The encoder in each mechanism captures the correlation of time dimension and feature dimension through self-attention mechanism and mask attention respectively to obtain extracted information;

[0179] Step 6-2: The extracted information obtained in step 6-1 is merged through the gating module, and the weight of each mechanism is learned through the gating module to obtain the gating weight of each mechanism;

[0180] Step 6-3: The gating weights of each mechanism in step 6-2 are set to vector h using the nonlinear activation C and S of the fully connected layer. After a linear transformation layer and a softmax layer, the output weights g1 and g2 of each encoder are obtained. Then, each gating weight participates in the output of the corresponding mechanism and outputs the final feature vector y. The feature vector y is the final classification result. The expression of the feature vector y is:

[0181] h=W·Concat(C,S)+b

[0182] g1,g2=Softmax(h)

[0183] y=Concat(C·g1,S·g2)

[0184] Where C is the output matrix of the feature dimension encoder, S is the output matrix of the time dimension encoder, Concat(*) is the matrix concatenation operation, W is the linear transformation weight matrix, b is the linear transformation bias vector, h is the linear transformation output, g1 and g2 are the weight matrices of the two encoders respectively, and y is the output vector of the gating module.

[0185] Specifically, the two mechanisms in step 6-1 can be called the "twin towers" mechanism

[0186] Multivariate time series has multiple features, each of which can be regarded as a univariate time series. The general assumption is that there are hidden correlations between different features in the case of general or distorted time steps. Acquiring information in both the time dimension and the feature dimension at the same time is the key to multivariate time series research. Unlike other works that use the basic Transformer model for time series classification and prediction, this application introduces a simple "dual tower" structure, in which the encoder in each tower captures the correlation between the time dimension and the feature dimension through self-attention mechanism and mask attention, such as Figure 7 shown.

[0187] The feature dimension encoder is the part within the wireframe on the left side of the model, which calculates the attention weights between different feature vectors across all time dimensions. In the multivariate time series classification task, there is no relative or absolute correlation between the positions of the feature vectors in the multivariate time series. If this application switches the order of the features, it will have no effect on the training results of the model. Therefore, only position encoding needs to be added to the time dimension encoder. The attention layer with masks on all feature dimensions is expected to accurately calculate the correlation between features across all time steps. When inputting time series data to each encoder, the two encoders can be directly implemented by simply transposing the features and time axis.

[0188] The wireframe on the right side of the model contains the time dimension encoder. To encode temporal features, attention weights are calculated for each pair of features across all time steps, using a masked self-attention mechanism to calculate the attention value for each point. In the multi-head self-attention layer, the attention matrix for all time steps is constructed using scaled dot-product attention. As with the original Transformer architecture, positionally fully connected feedforward layers are stacked on top of each multi-head attention layer to enhance feature extraction. At the same time, residual connections are maintained around each sub-layer to guide the flow of information and gradients, followed by layer normalization.

[0189] The gating module in step 6-2 is specifically

[0190] To combine the feature matrices output by the two towers, a simple gating module is used to learn weights for each tower. After obtaining the output of each tower, a fully connected layer nonlinearly activates C and S, setting the vector h. After passing through a linear transformation layer and a softmax layer, the output weights g1 and g2 of each encoder are obtained. Each gating weight is then incorporated into the output of the corresponding tower, resulting in the final feature vector y.

[0191] h=W·Concat(C,S)+b

[0192] g1,g2=Softmax(h)

[0193] y=Concat(C·g1,S·g2)

[0194] Where C is the output matrix of the feature dimension encoder, S is the output matrix of the time dimension encoder, Concat(*) is the matrix concatenation operation, W is the linear transformation weight matrix, b is the linear transformation bias vector, h is the linear transformation output, g1 and g2 are the weight matrices of the two encoders respectively, and y is the output vector of the gating module.

[0195] Step 7: Match the output results in step 6 with the tactical intent in the intent tag library in step 4 to achieve tactical intent recognition of the hostile target.

[0196] Example

[0197] The data used in this simulation is simulated confrontation data, and LOCKON air combat software is used for online confrontation. The relevant data in the air combat confrontation process is extracted and analyzed. Through multiple online confrontations, a large amount of air combat data is generated, which contains a variety of intention data. Among the large amount of sample data, experts in the field of air combat select and refine the typical tactical intention process, and extract and package the corresponding group or section of data, and generate intention labels as part of the tactical intention label library. Thus, all tactical intention data and their labels are synthesized into a tactical intention data set. The tactical intentions used in this application are divided into six types: reconnaissance, assault, interference, attack, avoidance, and escape:

[0198] The six tactical intentions and their common trajectory types selected in this application are as follows: Figure 10As shown in the figure, reconnaissance intent is mainly manifested in circular maneuvers in combat airspace, while air, ground, and surface radars remain active, maintaining search and tracking at all times to fully understand battlefield information; assault intent is mainly manifested in downward dives to reduce altitude and increase speed, in order to achieve a surprise attack effect; jamming intent is mainly manifested in figure-eight maneuvers in the air, while keeping the jammer active, to expand the jamming range while ensuring its own survivability; attack intent is mainly manifested in leaps from low to high, increasing its own altitude advantage and posing a greater threat to the enemy; evasion intent is mainly manifested in emergency turns or altitude pull-ups, which are emergency avoidance measures when missiles attack, and are also commonly used to evade enemy radar detection and improve its own survivability; escape intent is mainly manifested in U-turns or altitude pull-ups to increase speed to shake off missiles. When entering a no-escape zone, jammers will be released or jammers will be activated to confuse the missile.

[0199] This application uses 240,000 sets of sample data and divides them into training and test sets in a ratio of 3:1. Among them, reconnaissance intention accounts for 18.5%, assault intention accounts for 15.3%, interference intention accounts for 16.2%, attack intention accounts for 17.4%, evasion intention accounts for 19.6%, and escape intention accounts for 13.0%. The sample space composed of various tactical intentions and their corresponding eigenvalues is shown in Table 3.

[0200] Table 3 Summary of sample space composed of six tactical intentions

[0201]

[0202] Simulation results analysis

[0203] This application uses a comparative analysis method to verify the effectiveness and timeliness of PCA-GTN. In addition to the PCA-GTN and GTN tactical intent recognition models proposed in this application, five common models, including long short-term memory networks (LSTMs), BP neural networks, and support vector machines (SVMs), were selected for comparative analysis. The effectiveness and practicality of this method were verified mainly in terms of recognition accuracy, loss value, and timeliness.

[0204] The ratio of correct samples in the recognition results to the total number of samples is called recognition accuracy. The recognition accuracy in each iteration of training is used as the evaluation index. The recognition accuracy of each model is plotted as follows: Figure 11 As shown. Select classification cross entropy as the loss function, calculate the loss function value of each model respectively, and draw the loss value of each model as shown in Figure 12 shown.

[0205] like Figure 11As shown in the figure, with the increase of the number of iterations, the recognition accuracy of various models is increasing, among which the LSTM, SVM and BPNN models have the highest accuracy of 96.99%, 96.13% and 88.80% respectively. The recognition accuracy of the PCA-GTN intent recognition model proposed in this application and its basic model GTN is as high as 98.38%. Compared with the other three models, the accuracy is improved by 1.39%, 2.25% and 9.58% respectively. It can be seen that in terms of intent recognition accuracy, the PCA-GTN intent recognition model proposed in this application has better recognition accuracy than traditional models. In addition, through Figure 12 From the comparison of the loss values of each model, it can be seen that with the increase in the number of iterations, the loss values of each model are constantly decreasing, and the loss value of the PCA-GTN model always has a significant advantage over the other models. When training to the first few generations, the loss value of the PCA-GTN model can reach a low level and has good robustness.

[0206] Preliminary analysis and research show that the GTN model and PCA-GTN model proposed in this application have significant advantages in recognition accuracy and loss value. The PCA-GTN model has faster convergence speed and better robustness than the GTN model. This shows that PCA can effectively improve training results, enabling GTN to complete training faster under the same sample size, number of iterations, and learning rate conditions, thereby improving model training efficiency.

[0207] To further verify the recognition performance of the PCA-GTN model, the test set sample data with each tactical intent label are verified separately. When one of the tactical intents is selected as the research object, this intent is regarded as a positive sample, and the other intents are regarded as negative samples. Therefore, the confusion matrix can be introduced and the classification results of the test set data are substituted into the confusion matrix.

[0208] Table 4 Confusion matrix of PCA-GTN model recognition results

[0209]

[0210] Table 4 shows that the PCA-GTN model achieves a high recognition rate. The test datasets used for various tactical intents are also relatively balanced, meeting the requirements of subsequent analysis. Three evaluation metrics, precision (also known as P), recall (also known as R), and macro-average F1 score, were introduced to evaluate the recognition model. Precision (P) refers to the proportion of all recognition results for a given intent that agree with the true value; recall (R) refers to the proportion of all true-value examples of a given intent that are identified as belonging to that class; and F1 is a metric that balances precision and recall. The test results are shown in Tables 5-6.

[0211] Table 5 Intent recognition performance comparison

[0212]

[0213] Table 6 Comparison of F1 values of test results of each model

[0214]

[0215] In Tables 5-6, A, B, C, D, and E represent the PCA-GTN, GTN, LSTM, SVM, and BPNN tactical intent recognition models, respectively. From the test results of these three categories of indicators, it can be seen that various models have good adaptability for the recognition of different tactical intents. The gaps in various indicators between the PCA-GTN model and the GTN model are relatively small, but they are significantly higher than the other models, demonstrating the effectiveness of the GTN and PCA-GTN models proposed in this application. After PCA dimensionality reduction, although the feature dimension of the data set has decreased, the PCA-GTN model still has extremely high recognition accuracy and is superior to other models in all aspects. Therefore, based on the recognition results of the six tactical intents and the three evaluation indicators, it can be determined that the PCA-GTN model has a higher recognition rate and better robustness than other traditional recognition models. PCA can effectively improve the recognition accuracy and robustness of the GTN model to a certain extent.

[0216] Timeliness analysis

[0217] In order to verify the timeliness of the PCA-GTN tactical intention recognition model proposed in this application, the training and testing time of the five recognition models, PCA-GTN, GTN, LSTM, SVM, and BPNN, were compared and analyzed under the same training set and test set data conditions. The simulation results are shown in the figure below. Figure 13-14 shown.

[0218] Depend on Figure 13 It can be seen that the training time of various models is in the thousands of seconds, but the training time of the PCA-GTN model is relatively shorter than that of other models, which is more than 192 seconds shorter than that of other models. Compared with the basic GTN model, the training time of PCA-GTN is reduced by 241 seconds, which shows that PCA can effectively improve the learning speed of GTN by reducing the feature dimension of the air combat dataset, thereby effectively shortening the training time of the GTN tactical recognition model. Figure 14As can be seen, the test time of various models is in the second level, with the SVM model taking the longest, approximately 6.27 seconds, and the PCA-GTN model taking the shortest, approximately 2.68 seconds. The test time of the PCA-GTN model proposed in this application is significantly shorter than that of other models, and its test time is shortened by approximately 40% compared to the GTN model. This shows that PCA can also effectively improve the recognition speed of the GTN model and shorten the recognition time.

[0219] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The target tactical intention recognition method based on PCA-GTN is characterized by: The following steps are involved: Step 1: Obtain data on hostile targets during air combat confrontation; Step 2: Use the PCA method to reduce the dimension of the data obtained in step 1; Step 3: Extract and classify the data after dimensionality reduction in step 2 according to different target features to obtain target feature data; Step 4: Build an intent tag library based on the target feature data obtained in step 3; Step 5: Improve the encoder-decoder in the Transformer model and use the improved Transformer model as the target tactical recognition model; Step 6: Based on the target tactical recognition model in step 5, the correlation between the feature dimension and the time dimension of the data after dimensionality reduction in step 2 is extracted. The extracted information is then merged using a gating module, and the classification result is output after a linear transformation layer and a softmax layer. Step 7: Match the classification results output in step 6 with the tactical intent in the intent label library in step 4 to achieve tactical intent recognition of the hostile target.

2. The target tactical intention recognition method based on PCA-GTN according to claim 1 is characterized in that: Step 3 includes the following sub-steps: Step 3-1: Extract the target feature time series information from the data after dimensionality reduction in step 2 using the following formula: X=f(V,Q,D,…) In the formula, V, Q, D, ... are the target speed, angle, distance and other state information, among which, is the set of speed information of n time points, is the set of angle information of n time points, is the set of distance information of n time series points; f(*) is the mapping process of feature extraction; X is the extracted target feature time series information; Step 3-2: Further mine and analyze the target feature time series information extracted in step 3-1 to obtain the mapping relationship between target tactical intent and target features: Y=f(X)=f(X E ,X A ,X S ,…)() Where Y is the target's possible tactical intention set; f(*) is the mapping relationship between the feature set and the intention set; is a set of characteristics of the battlefield environment; is the feature set of the target state; A feature set for friendly-enemy interactions; Step 3-3: Based on the mapping relationship in step 3-2, the target features are classified to obtain target feature data.

3. The target tactical intention recognition method based on PCA-GTN according to claim 1 is characterized in that: Step 4 includes the following steps: Step 4-1: Based on the target feature data obtained in step 3, select multiple tactical intentions, set corresponding types of labels to encode the multiple tactical intentions, and construct a label space; Step 4-2: Decode each label in the label space constructed in step 4-1 to obtain multiple tactical intent recognition results. The collection of multiple tactical intent recognition results is the intent space; Step 4-3: Construct an intent tag library through the label space in step 4-1 and the intent space in step 4-2, as well as the encoding and decoding mechanism between the label space and the intent space.

4. The target tactical intention recognition method based on PCA-GTN according to claim 1 is characterized in that: The improved Transformer model in step 5 is obtained by changing the embedding layer of the Transformer model to a fully connected layer and adding a nonlinear activation function instead of the linear transformation; the expression of the nonlinear activation function is: Among them, tanh(x) is a typical hyperbolic tangent function activation function of a neural network, x∈R, -1<tanh(x)<1.

5. The target tactical intention recognition method based on PCA-GTN according to claim 1 is characterized in that: The step 6 includes the following contents: Step 6-1: Introduce two mechanisms to extract the correlation of feature dimension and time dimension respectively in the improved Transformer model in step 5. The encoder in each mechanism captures the correlation of time dimension and feature dimension through self-attention mechanism and mask attention respectively to obtain extracted information; Step 6-2: The extracted information obtained in step 6-1 is merged through the gating module, and the weight of each mechanism is learned through the gating module to obtain the gating weight of each mechanism; Step 6-3: The gating weights of each mechanism in step 6-2 are set to vector h using the nonlinear activation C and S of the fully connected layer. After a linear transformation layer and a softmax layer, the output weights g1 and g2 of each encoder are obtained. Then, each gating weight participates in the output of the corresponding mechanism and outputs the final feature vector y. The feature vector y is the final classification result. The expression of the feature vector y is: h=W·Concat(C,S)+b g1,g2=Softmax(h) y=Concat(C·g1,S·g2) Where C is the output matrix of the feature dimension encoder, S is the output matrix of the time dimension encoder, Concat(*) is the matrix concatenation operation, W is the linear transformation weight matrix, b is the linear transformation bias vector, h is the linear transformation output, g1 and g2 are the weight matrices of the two encoders respectively, and y is the output vector of the gating module.

6. The target tactical intention recognition system based on PCA-GTN is characterized by: include: Data acquisition module: used to obtain data on hostile targets during air combat confrontation; PCA dimensionality reduction module: Use PCA to reduce the dimensionality of the data obtained in the data acquisition module; Feature extraction and classification module: used to extract and classify the data after dimensionality reduction in the PCA dimensionality reduction module according to different target features to obtain target feature data; Build intent tag library module: used to build intent tag library based on target feature data obtained by feature extraction and classification module; Constructing a target tactical recognition model module: Improve the encoder-decoder in the Transformer model and use the improved Transformer model as the target tactical recognition model; Output classification result module: Based on the target tactical recognition model in the target tactical recognition model construction module, the module extracts the correlation between the feature dimension and the time dimension of the data after dimensionality reduction in the PCA dimensionality reduction module, then uses the gating module to merge the extracted information, and outputs the classification result after a linear transformation layer and a softmax layer; Tactical intention recognition module: used to match the classification results output by the output classification result module with the tactical intentions in the intent label library in the intention label library construction module, so as to realize the tactical intention recognition of the hostile target.

7. The target tactical intention recognition system based on PCA-GTN is characterized by: The feature extraction and classification module includes the following: Extracting target feature time series information submodule: used to extract target feature time series information from the data after dimensionality reduction in the PCA dimensionality reduction module; Mining and analysis submodule: further mines and analyzes the target feature time series information extracted in the target feature time series information extraction submodule to obtain the mapping relationship between the target tactical intention and the target feature; Classification submodule: Based on the mapping relationship of the mining and analysis submodule, the target features are classified to obtain target feature data.

8. The target tactical intention recognition system based on PCA-GTN is characterized by: Building the intent tag library module includes the following: Constructing the label space submodule: Based on the target feature data obtained in the feature extraction and classification module, select multiple tactical intentions, set corresponding types of labels to encode the multiple tactical intentions, and construct the label space; Constructing the intention space submodule: Decode each label in the label space constructed in the constructing label space submodule to obtain multiple tactical intent recognition results. The collection of multiple tactical intent recognition results is the intention space; Construct the intent tag library sub-module: Construct the intent tag library by constructing the label space in the label space sub-module and the intent space in the intent space sub-module, as well as the encoding and decoding mechanism between the label space and the intent space.

Images