A tactical intention recognition method based on BiLSTM-Attention

Abstract

The present invention discloses a tactical intent recognition method based on BiLSTM-Attention, which comprises the following steps: Step S1: describing aerial target intent recognition as a mapping of air combat intent recognition features to air combat intent types; Step S2: establishing an air combat intent recognition model based on BiLSTM-Attention. The beneficial effects of the present invention are: introducing a bidirectional (Bidirectional) loop mechanism and an attention (Attention) mechanism based on an LSTM network, thereby simulating the decision-maker's reasoning process regarding air combat situations. Inputting the encoded air combat feature vector into the BiLSTM layer can fully utilize historical and future moment information to make a comprehensive judgment compared to putting it into the LSTM layer; putting the output vector of the BiLSTM layer into the Attention mechanism layer can further highlight the key information that affects the intention and improve the accuracy of intention recognition.

Description

Technical Field

[0001] The present invention relates to a tactical intention recognition method, specifically a tactical intention recognition method based on BiLSTM-Attention Background Art

[0002] In modern information-based air combat, the booming development of aviation science and military technology has led to increasingly severe threats to aerial targets. At the same time, the continuous application of high technology has led to the increasing complexity and information asymmetry of battlefield environments. Relying solely on expert experience makes it difficult to accurately identify enemy targets and intentions in real time within complex air combat environments. Therefore, intelligent reasoning methods are urgently needed to overcome the shortcomings of traditional manual methods and help our side gain air superiority and even victory in war.

[0003] In recent years, a large number of intention recognition studies have been carried out in the military field to meet the needs of combat decision-making systems. Existing methods for enemy target intention recognition in complex battlefield environments mainly include evidence theory, template matching, expert systems, Bayesian networks, and neural networks.

[0004] For example, the characteristic information of aerial targets can be measured using the ship's sensors, a confidence rule base can be established, and the target intention can be identified by fusing multi-source information using evidence reasoning.

[0005] Based on the template constructed based on the situation database, an intention recognition reasoning model is designed, and an intention recognition template matching method based on DS evidence theory is proposed.

[0006] The knowledge base is constructed using domain expert knowledge, and then the correspondence between battlefield situation and combat intention is expressed in the form of rules, and finally the inference engine is used to obtain the inference results.

[0007] The Bayesian network parameters are determined based on the knowledge of military experts. Features are represented by nodes, transfer relationships are represented by directed arcs, and relationship strength is represented by conditional probability. New events are used to influence backward propagation to update the network parameters until an intention exceeds a threshold. This intention is then identified.

[0008] Information from the actual battlefield is collected, appropriate features are selected and data preprocessing is performed to obtain a data set, which is then input into a neural network. The adaptive and self-learning capabilities of the neural network are used to obtain combat intention recognition rules, and then the air combat intention recognition rules are used to infer the enemy target's combat intention.

[0009] Existing methods, such as those described above, rely on single-moment feature information for analysis and calculation, making it difficult to effectively extract underlying information from the time-varying target state characteristics. In reality, on the battlefield, target intent is implemented through a series of tactical actions. Therefore, the target's dynamic properties and battlefield environment exhibit dynamic and time-varying characteristics. Furthermore, enemy combat operations can be deceptive and concealed. Therefore, inferring enemy intent based on single-moment feature information is unscientific.

[0010] A tactical intention intelligent recognition model based on a long short-term memory (LSTM) network is proposed to target the temporal variation characteristics. This model is effective in identifying the target's combat intentions and conforms to the temporal characteristics and logical relationships in battlefield situation information. However, it can only use historical information to make judgments on current information and cannot use future information. Moreover, its accuracy still has much room for improvement. Summary of the Invention

[0011] The purpose of the present invention is to provide a tactical intention recognition method based on BiLSTM-Attention, which can accurately recognize the tactical intention of aerial targets.

[0012] The technical solution of the present invention is as follows: a tactical intention recognition method based on BiLSTM-Attention, comprising the following steps:

[0013] Step S1: describing the air target intention recognition as mapping the air combat intention recognition features to the air combat intention types;

[0014] Step S2: Establish an air combat intention recognition model based on BiLSTM-Attention.

[0015] The step S1 is to determine the tactical intention space I to the temporal feature set V T The mapping function is as follows:

[0016]

[0017] Where I=(i1,i2,…,i n ) is the tactical intention space of air targets, namely, seven types of intentions: {penetration, feint, attack, reconnaissance, retreat, surveillance, and electronic jamming}, V t is the real-time characteristic information of the battlefield at time t, V T From t1 to t T The temporal feature set is composed of feature sets of T consecutive moments, and the function f is a mapping function between an intention type and the real-time feature information of the battlefield collected at each moment.

[0018] The step S1 further includes:

[0019] Step S11: Describing the target tactical intention space

[0020] The tactical intention space for establishing enemy targets includes seven types of intentions: {penetration, feint, attack, reconnaissance, retreat, surveillance, and electronic jamming};

[0021] Step S12: Description of aerial target tactical intention identification features

[0022] The air combat capability factor is also an important factor in determining the degree of target threat. For the air combat capability of a fighter, a single-aircraft air combat capability threat function C is constructed:

[0023] C=[lnε1+ln(ε2+1)+ln(∑ε3+1)]ε4ε5ε6ε7 (2)

[0024] Among them, ε1 is the fighter's maneuverability, ε2 is the performance of onboard weapons, ε3 is the detection capability of onboard equipment, ε4 is the basic flight performance of the fighter, ε5 is the operating performance of the fighter, ε6 is the combat survivability performance of the fighter, and ε7 is the electronic information confrontation performance.

[0025] The step S2 includes: the BiLSTM-Attention model is divided into three parts: an air combat feature vector input layer, a hidden layer and an output layer, and the hidden layer includes a BiLSTM layer, an Attention layer and a Dense layer.

[0026] The step S2 includes:

[0027] Step S21: Calculate the air combat feature vector input layer;

[0028] Step S22: Calculate the hidden layer.

[0029] The step S21 includes:

[0030] Step S211: Read the collected data and perform data cleaning;

[0031] Step S212: normalizing the numerical air combat feature data;

[0032] Normalize 11 numerical air combat feature data types, including enemy aircraft acceleration, enemy aircraft altitude, enemy aircraft speed, enemy aircraft air combat capability factor, heading angle, azimuth angle, our aircraft acceleration, our aircraft altitude, our aircraft speed, our aircraft air combat capability factor, and distance between both sides;

[0033] For the xth numerical data F x =[f x1 ,f x2 ,…f xi ,…,fxn ](x=1,2,…,11), n ​​is the total number of data, the i-th original data value f in the x-th type xi The result of mapping to the interval [0, 1] is f′ xi , the formula is:

[0034]

[0035] Where: minF x is the x-th dimension feature F x The minimum value of maxF x is the x-th dimension feature F x The maximum value of

[0036] Step S213: Encode the non-numeric air combat feature data

[0037] The four attribute data types of air radar status, sea radar status, interference status, and interference status are coded as 0 and 1. For the two attribute data types of maneuver type and enemy aircraft type, the coded data of each non-numeric feature is obtained and then normalized.

[0038] Step S214: Encode the seven enemy target combat intention types into category labels;

[0039] Step S215: randomly initialize the data and divide the training set and test set into 8:2;

[0040] After steps S211 to S215, the collected air combat feature data becomes a feature vector form that can be directly accepted and processed by the hidden layer.

[0041] The step S22 includes:

[0042] Step S221: Calculate the BiLSTM layer

[0043] Calculate using the following formula:

[0044] Γ f =σ(W f [h t-1 ,x t ]+b f ) (4)

[0045] Γ u =σ(W u [h t-1 ,x t ]+b u ) (5)

[0046]

[0047]

[0048] Γ o =σ(W o [h t-1 ,x t +b o ]) (8)

[0049] h t =Γ o *tanh C t (9)

[0050] Formula (4) calculates the value of the forget gate. From the form of formula (4), it can be seen that the value of the forget gate at time t is determined by x t and h t-1 Jointly determined; Formula (5) calculates the activation (W u [h t-1 , x t ]+b u ) cell state value; Formula (6) is calculated by h t-1 and x t The value of the candidate memory unit determined by c t-1 and For Γ f and Γ u After the regulation of t Value; Formulas (8) and (9) calculate the value of h at time t t-1 and x t The output h of the final hidden state of the LSTM after the internal loop and update is determined t ;

[0051] Among them, x t represents the input features at time t; C t-1 represents the neuron before updating; C t represents the updated neuron; h t-1 and h t Represent the output features of the previous moment and the current moment respectively; Γ f , Γ u and Γ o Represent the forget gate, update gate and output gate respectively; is a candidate neuron; σ is the Sigmoid function. The operation process is as follows, where w f 、w u 、w c 、w o and b f 、b u 、b c 、b o is the weight coefficient matrix and offset vector of each corresponding part;

[0052] The hidden state of BiLSTM at the current moment is determined by the current input x t , the output of the previous moment's forward propagation hidden layer state And the output of the backpropagation hidden layer state at the next moment The three parts jointly determine that BiLSTM is composed of two unidirectional LSTMs, so the hidden state of BiLSTM at time t can be obtained by forward propagation of the hidden state. and the back-propagated hidden state The two parts are obtained,

[0053] The calculation formula is shown in (10)(11)(12), where w i (i=1, 2, ..., 6) represents the weight from one unit layer to another unit layer.

[0054]

[0055]

[0056]

[0057] Step S222: Calculate the Attention layer

[0058] The hidden state s of each initial input t and the weight coefficient α of each hidden layer t The final output state vector Y is obtained by accumulating the product of .

[0059] e t =tanh(w t s t +b t ) (13)

[0060]

[0061]

[0062] Where: e t The state vector s represents the t-th eigenvector t The energy value determined, w t represents the weight coefficient matrix of the t-th eigenvector, b t represents the offset corresponding to the t-th eigenvector, e i The meaning of e tSimilarly, according to formula (14), the conversion from the initial input state to the new attention state can be achieved, and then the final output state vector Y is obtained through (15). Finally, Y is integrated with the Dense layer as an output value and input into the final output layer;

[0063] Step S223: Calculate the output layer

[0064] The input of the output layer is the output of the Attention mechanism layer in the hidden layer. The multi-classification Softmax function is used to perform corresponding calculations on the input of the output layer to obtain the classification of the tactical intention of the air target. The specific formula is:

[0065] y k =softmax(w1Y+b1) (16)

[0066] Among them: w1 represents the weight coefficient matrix that needs to be trained from the Attention mechanism layer to the output layer; b1 represents the corresponding bias that needs to be trained; y k Predict labels for the output of the output layer.

[0067] The beneficial effects of this invention lie in: introducing a bidirectional loop mechanism and an attention mechanism based on an LSTM network, thereby simulating the decision-maker's reasoning process regarding air combat situations. Inputting the encoded air combat feature vector into the BiLSTM layer allows for comprehensive judgment based on historical and future information, compared to the LSTM layer. Inputting the BiLSTM layer's output vector into the Attention mechanism layer further highlights key information that influences intent, improving the accuracy of intent recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 reasoning process for tactical intent for aerial targets;

[0069] Figure 2 Tactical intention recognition process for air targets;

[0070] Figure 3 Coding and pattern analysis for combat intent;

[0071] Figure 4 Characterize tactical intent for air targets;

[0072] Figure 5 It is the BiLSTM-Attention model;

[0073] Figure 6 It is an LSTM structure;

[0074] Figure 7 It is a BiLSTM structure;

[0075] Figure 8 It is the Attention mechanism model;

[0076] Figure 9 Changes in accuracy of each model in the ablation experiment;

[0077] Figure 10 Changes in various loss values ​​in the ablation experiment. DETAILED DESCRIPTION

[0078] In order to accurately describe the aerial target tactical intention recognition model, the following assumptions are made: (1) The battlefield environmental conditions such as terrain, atmosphere, and climate of the air combat between the enemy and our side are roughly the same; (2) The enemy's aerial target tactical intention does not change in the extracted time series.

[0079] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0080] A tactical intent recognition method based on BiLSTM-Attention includes the following steps:

[0081] Step S1: Describe the air target intention recognition as a mapping from air combat intention recognition features to air combat intention types.

[0082] The tactical intention of air targets is a process of inferring the enemy's combat intention by extracting battlefield environment information in the corresponding time and space domain, static attributes of air combat enemy and friendly targets, and real-time dynamic information from a real-time, confrontational environment, and combining it with the corresponding military field knowledge. The tactical intention inference process of air targets is as follows: Figure 1 shown.

[0083] Specifically, because aerial target tactical intent recognition is conducted under complex, highly contested battlefield conditions, identified targets will attempt to deceive our decision-makers and force them to make incorrect judgments. Therefore, using single-moment features to identify enemy combat intent can significantly differ from actual conditions. Therefore, inferring the combat intent of an identified target from its features at multiple consecutive moments is more scientific.

[0084] Determine the tactical intention space I to the temporal feature set V T The mapping function is as follows:

[0085]

[0086] Where I=(i1,i2,…,i n ) is the tactical intention space of air targets, namely the following seven intention types: {penetration, feint, attack, reconnaissance, retreat, surveillance, and electronic jamming}, V t is the real-time characteristic information of the battlefield at time t.T From t1 to t T The temporal feature set consists of feature sets of T consecutive moments.

[0087] Due to the high confrontation, uncertainty, and complexity of air combat itself, it is difficult to deduce the mapping relationship between tactical intention type and time series feature set through mathematical formula. This paper uses air combat dataset to train BiLSTM-Attention network structure, thereby establishing the mapping relationship between tactical intention type and time series feature set. The whole air combat intention recognition process is as follows: Figure 2 shown.

[0088] like Figure 2 As shown in the figure, when performing the tactical intention recognition process for air targets, the intention type of historical data is first calibrated to obtain a complete training data set. Then, the preprocessed data set is input into the BiLSTM-Attention network training to obtain the mapping relationship between the air combat intention type and the time series feature set. In actual air combat, the sensor collects N consecutive moments (T n ~T n+N ) The target state information is finally integrated and encoded and input into the trained target intention recognition model to obtain the target intention recognition result. The specific process is as follows:

[0089] Step S11: Describing the target tactical intention space

[0090] The target tactical intention space has different intention spaces for different combat forms, different scenarios, and different enemy entities. Therefore, it is necessary to define an appropriate tactical intention space according to the corresponding combat situation.

[0091] The present invention takes UAV combat in a certain airspace as the research object, and establishes the tactical intention space of the enemy target, including seven intention types: {penetration, feint attack, attack, reconnaissance, retreat, surveillance, and electronic interference}.

[0092] After establishing the intent space, the key to applying the BiLSTM-Attention model to tactical intent recognition is converting human cognitive patterns into labels for intelligent model training and mapping them to intent types in the tactical intent space. Analysis of the process by which decision-makers reason about enemy targets' tactical intent reveals that it is difficult for decision-makers to explicitly express their judgment of enemy targets' intent based on battlefield situation information and their own experience. However, human cognitive experience is often implicit in the process of reasoning about enemy targets' operational intent. Therefore, the decision-makers' cognitive experience can be encapsulated as labels to train the BiLSTM-Attention model.

[0093] For the seven enemy target combat intention types established in this invention, the corresponding combat intention type coding and pattern analysis mechanism are as follows: Figure 3 For example, if the BiLSTM-Attention model outputs an intent recognition result of 4, the enemy's combat intent can be considered surveillance. Therefore, encoding the enemy's combat intent can simply and clearly express the decision maker's cognitive experience and make model training easier.

[0094] Step S12: Description of aerial target tactical intention identification features

[0095] The enemy's tactical intent is highly correlated with the threat level and combat mission of both sides. For example, if our threat level to the enemy is significantly greater than the enemy's threat level to us, the likelihood that the enemy's tactical intent is "attack" is significantly reduced. Therefore, to identify different threat levels and combat missions, different air combat features need to be extracted.

[0096] From the perspective of threat level, there are many factors that affect the target threat level. This paper mainly considers the distance between the enemy and the friendly party, speed, angle, and flight acceleration. The air combat capability factor is also an important factor in determining the target threat level. For the air combat capability of the fighter, a single-aircraft air combat capability threat function C is constructed:

[0097] C=[lnε1+ln(ε2+1)+ln(∑ε3+1)]ε4ε5ε6ε7 (2)

[0098] Among them, ε1 to ε7 represent the fighter's maneuverability, onboard weapon performance, onboard equipment detection capability, basic flight performance, operational performance, combat survivability, and electronic information countermeasure performance, respectively.

[0099] The air combat capability threat is the inherent performance of the fighter. Therefore, the air combat capability factors of various fighters of both sides within a certain period can be calculated according to formula (2), stored in the database, and updated in real time according to the equipment development situation mastered by our side.

[0100] From a combat mission perspective, when an enemy aircraft executes a specific mission, certain characteristics must meet certain conditions. For example, fighter jets typically approach enemy targets at high speed during air combat attacks, typically flying at speeds between 735 and 1470 km / h. Penetration missions are divided into low-altitude and high-altitude penetrations, corresponding to altitudes of 50 to 200 meters and 1000 to 11000 meters. The target radar signal status is also linked to the combat mission. For example, during dogfights, air defense radars are typically kept active, while during reconnaissance missions, both air defense and surface defense radars are kept active. Different types of aircraft have different application value and tactical significance. Fighters are more offensive, while reconnaissance aircraft have stronger reconnaissance capabilities. Therefore, enemy aircraft type can also be used as a characteristic for identifying tactical intent.

[0101] In addition, the realization of aerial target tactical intention is closely related to fighter aircraft maneuvering. Common maneuvering action library has two design types: one is "typical tactical action library" based on typical air combat tactical aircraft action as design basis, and the other is "basic maneuvering action library" based on basic air combat control action as design basis. Since the present invention studies the time sequence characteristics, tactical intention recognition is carried out by collecting 12 frames of target feature information, and the control algorithm of "typical tactical action library" is complex to solve, and the exit and conversion time points of action are difficult to determine, so the present invention adopts "basic maneuvering action library". This library is proposed by scholars of the National Advisory Committee for Aeronautics (NASA) of the United States based on the most commonly used maneuvering mode in air combat, and mainly includes 7 maneuvers of {maximum acceleration, maximum deceleration, maximum overload climb, maximum overload dive, maximum overload right turn, maximum overload left turn, stable typing}. However, the maneuvers combined by these seven maneuvering modes are not enough, and all of them take extreme control and obviously do not conform to the reality of air combat. The present invention selects 11 improved basic control actions, including {constant speed forward flight, decelerated forward flight, accelerated forward flight, climb, right climb, left climb, dive, right dive, left dive, right turn, left turn}.

[0102] In summary, the aerial target tactical intention recognition feature set of the present invention is {aircraft radar status, sea radar status, interference status, interference status, maneuver type, enemy aircraft type, enemy aircraft acceleration, enemy aircraft altitude, enemy aircraft speed, enemy aircraft air combat capability factor, heading angle, azimuth angle, our aircraft acceleration, our aircraft altitude, our aircraft speed, our aircraft air combat capability factor, distance between the two sides} 17-dimensional feature vector, the feature description diagram is as follows: Figure 4 As shown, it can be divided into numerical features and non-numerical features.

[0103] Step S2: Establish an air combat intention recognition model based on BiLSTM-Attention

[0104] The BiLSTM-Attention model is divided into three parts: air combat feature vector input layer, hidden layer and output layer. Among them, the hidden layer includes BiLSTM layer, Attention layer and Dense layer. Among them, the BiLSTM-Attention model structure is as follows Figure 5 shown.

[0105] Step S21: Calculate the air combat feature vector input layer

[0106] The air combat feature vector input layer of the present invention mainly preprocesses the collected air combat feature data sets, that is, processes these data sets into feature vector forms that can be directly accepted and processed by the BiLSTM layer. The specific steps of air combat feature vectorization are as follows:

[0107] Step S211: Read the collected data and perform data cleaning

[0108] Step S212: Normalize the numerical air combat feature data

[0109] Data normalization can eliminate the influence of data dimension and improve network convergence efficiency. The present invention normalizes 11 types of numerical air combat feature data, including enemy aircraft acceleration, enemy aircraft altitude, enemy aircraft speed, enemy aircraft air combat capability factor, heading angle, azimuth angle, our aircraft acceleration, our aircraft altitude, our aircraft speed, our aircraft air combat capability factor, and the distance between the two sides.

[0110] For the xth numerical data F x =[f x1 ,f x2 ,…f xi ,…,f xn ](x=1,2,…,11); n is the total number of data. xi The result of mapping to the interval [0, 1] is f′ xi , the formula is:

[0111]

[0112] Where: minF x is the x-th dimension feature F x The minimum value of maxF x is the x-th dimension feature F x The maximum value of .

[0113] Step S213: Encode the non-numeric air combat feature data

[0114] The four attribute data types of air defense radar status, sea defense radar status, jamming status, and jammed status are encoded as 0 and 1. For example, in the air defense radar status, 0 indicates the radar is off, and 1 indicates it is on. For the two attribute data types of maneuver type and enemy aircraft type, the encoded data of each non-numeric feature is obtained and then normalized.

[0115] Step S214: Encode the seven enemy target combat intention types into category labels.

[0116] Step S215: Randomly initialize the data and divide the training set and test set into 8:2 ratios.

[0117] After the above steps S211 to S215 are performed, the collected air combat feature data becomes a feature vector form that can be directly accepted and processed by the hidden layer.

[0118] Step S22: Calculate the hidden layer

[0119] Step S221: Calculate the BiLSTM layer

[0120] As a special type of recurrent neural network (RNN), the Long Short-Term Memory (LSTM) network also has a recursive structure similar to that of RNN. However, unlike simple RNN networks, LSTM simulates the forgetting and memory mechanisms of the human brain by introducing the idea of ​​gated switches, thereby overcoming the problems of gradient vanishing and gradient exploding during long sequence training.

[0121] In the LSTM architecture, information input and output are realized through forget gate, update gate, output gate and memory unit. The single neuron architecture of LSTM is as follows: Figure 6 shown.

[0122] In the figure: x t represents the input features at time t; c t-1 represents the neuron before updating; c t represents the updated neuron; h t-1 and h t Represent the output features of the previous moment and the current moment respectively; Γ f , Γ u and Γ o Represent the forget gate, update gate and output gate respectively; is a candidate neuron; σ is the Sigmoid function. The operation process is as follows, where w f 、w u 、w c 、w o and b f 、b u 、bc 、b o is the weight coefficient matrix and offset vector of each corresponding part. Formula (4) calculates the value of the forget gate to see how much information can be retained. From the form of formula (4), it can be seen that the value of the forget gate at time t is determined by x t and h t-1 Jointly determined; Formula (5) calculates the activation (W u [h t-1 , x t ]+b u ) cell state value; Formula (6) is calculated by h t-1 and x t The value of the candidate memory unit determined by c t-1 and For Γ f and Γ u After the regulation of t Value; Formula (8) and Formula (9) calculate the value of h at time t t-1 and x t The output h of the final hidden state of the LSTM after the internal loop and update is determined t .

[0123] Γ f =σ(W f [h t-1 ,x t ]+b f ) (4)

[0124] Γ u =σ(W u [h t-1 ,x t ]+b u ) (5)

[0125]

[0126]

[0127] Γ o =σ(W o [h t-1 ,x t +b o ]) (8)

[0128] h t =Γ o *tanh C t (9)

[0129] The traditional LSTM network is a unidirectional neural network structure. The information it obtains is historical information before the current moment, which leads to the neglect of future information. The BiLSTM network is composed of a forward LSTM network and a backward LSTM network, which has the function of capturing the characteristics of the previous and next information. Its model structure is as follows: Figure 7 shown.

[0130] from Figure 7 It can be seen that the hidden state O of BiLSTM at time t t By forwarding the hidden state and the backward hidden state The two parts are obtained, the forward hidden layer state The current input x t and the hidden state at time (t-1) Determine the backward hidden state The current input x t and the hidden state at time (t+1) The calculation formula is as follows (10)(11)(12), where w i (i=1,2…,6) represents the weight from one unit layer to another unit layer.

[0131]

[0132]

[0133]

[0134] Step S222: Calculate the Attention layer

[0135] The Attention mechanism is similar to the brain signal processing mechanism unique to human vision. It calculates the weights of the feature vectors output in the BiLSTM network at different times, highlights the features that account for a larger proportion of the prediction results, and thus makes the entire neural network model show better performance. The Attention mechanism performs well in time series data such as machine translation and speech recognition, and also has relatively good results in classification prediction. It can be used alone or as a layer in other hybrid models. In the tactical intention recognition of aerial targets in this article, the neural network uses the Attention mechanism to focus on some key features during the training process. The core is the weight coefficient. First, the importance of each feature is learned, and then the corresponding weight is assigned to each feature according to the importance. For example, when the enemy aircraft has an attack intention, the heading angle, maneuver type and other features will be assigned more weights by the Attention mechanism to deepen the model memory. The basic structure of the Attention mechanism model is as follows: Figure 8 As shown.t The tth feature vector output by the BiLSTM network is input into the hidden layer of the attention mechanism to obtain the initial state vector s t , and then with the weight coefficient α t The corresponding multiplication and cumulative sum are used to obtain the final output state vector Y. The calculation formula is as follows:

[0136] e t =tanh(w t s t +b t ) (13)

[0137]

[0138]

[0139] Where: e t The state vector s represents the t-th eigenvector t The energy value determined; w t represents the weight coefficient matrix of the t-th eigenvector; b t Represents the offset corresponding to the t-th eigenvector. According to formula (14), the transition from the initial input state to the new attention state can be achieved. Then, the final output state vector Y is obtained through formula (15). Finally, Y is integrated with the Dense layer as an output value and input to the final output layer.

[0140] Step S223: Calculate the output layer

[0141] The input of the output layer is the output of the Attention mechanism layer in the hidden layer. The multi-classification Softmax function is used to calculate the input of the output layer to obtain the classification of the tactical intention of the air target. The specific formula is:

[0142] y k =softmax(w1Y+b1) (16)

[0143] Among them: w1 represents the weight coefficient matrix that needs to be trained from the Attention mechanism layer to the output layer; b1 represents the corresponding bias that needs to be trained; y k Predict labels for the output of the output layer.

[0144] Comparative experiments were conducted with various models and the proposed model. Using accuracy, loss, recall, precision, and F1-score as evaluation indicators, the effectiveness of the proposed model in identifying the tactical intentions of aerial targets was demonstrated. The details are as follows:

[0145] Selection of experimental data sets: The experiment is based on the UAV combat in a certain airspace, and the experimental data comes from a combat simulation system. By running the simulation system multiple times, a variety of air combat intention patterns are obtained, from which 10,000 air combat intention samples are randomly extracted, and 12 consecutive frames of information are collected for each sample (each frame of information includes 17 dimensional feature information such as heading angle, flight altitude, interference status, radar status, etc.). Due to the large amount of sample data, air combat experts are selected to write intention recognition rules based on air combat experience to generate intention labels. After that, the air combat intention samples are classified by computer. Finally, air combat experts revise the sample data with ambiguous intention classification based on air combat experience. The data set includes 7 types of target tactical intentions, and the proportion of each tactical intention data is 21.6% for attack intention, 20.0% for penetration intention, 19.8% for reconnaissance intention, 12.9% for surveillance intention, and 10.9% for feint intention. Figure 10 0%, electronic interference intention Figure 9 .25%, intention to withdraw Figure 6 The sample size is 10,000, of which the training set and test set are divided into 8:2, so the training set sample size is 8,000 and the test set sample size is 2,000.

[0146] Experimental setup

[0147] This experiment uses accuracy and loss as evaluation metrics. The experiment was conducted in Python, using a GPU-accelerated environment and the Keras deep learning framework. The computer configuration was Windows 10, a GTX 960M graphics card, and 8GB of RAM. Many hyperparameters required setting and adjustment. After multiple experiments, adjustments were made based on accuracy and loss, resulting in the parameter settings shown in Table 1.

[0148] Table 1 Model experimental parameters

[0149]

[0150] Experimental results analysis

[0151] Analysis of BiLSTM-Attention Model Recognition Results

[0152] After training, the BiLSTM-Attention model was tested on 20% of the samples. The experiment showed that the accuracy of the network model proposed in this paper reached 97.3%. To further observe the relationship between the recognition intents, a confusion matrix of the data set was created. The diagonal line represents the number of samples that were correctly recognized, as shown in Table 2.

[0153] Table 2 Intent recognition confusion matrix

[0154]

[0155] Table 2 shows the test set intention recognition confusion matrix obtained by training a sample using the model of the present invention. As can be seen from the table, the model has a high recognition accuracy for all seven intentions, especially the retreat intention recognition accuracy, which can reach 100%. For attack intention, a small number of them are mistakenly identified as feint intentions, and there are also a small number of mutual recognition errors between reconnaissance intention and surveillance intention. Analysis shows that the air combat characteristics corresponding to the two intentions are highly similar and the intentions are more deceptive. The BiLSTM neural network cannot ensure that the trained model weights have a significant difference in identifying such tactical intentions. Therefore, the final attention mechanism layer cannot accurately perceive the weight difference between the two, resulting in a small number of mutual recognition errors, which is consistent with the actual situation.

[0156] Comparative experiments with LSTM, SAE, BP, MLP, and SVM models

[0157] This experiment selected the highest accuracy on the test set during 100 iterations as the accuracy of the model, and the corresponding loss value was the loss value of the model. The BiLSTM-Attention model proposed in this invention was used to compare with the stacked autoencoder (SAE) tactical intention intelligent recognition model; the LSTM-based battlefield enemy target tactical intention recognition model; the BP neural network aerial target combat intention recognition model optimized using the ReLU function and the Adam algorithm; and the traditional multi-classification models support vector machine (SVM) and multi-layer perceptron (MLP). The specific results are shown in Table 3:

[0158] Table 3 Intent recognition results of different models

[0159]

[0160] Table 3 shows that the proposed BiLSTM-Attention model outperforms the other five models in both accuracy and loss. It improves accuracy by approximately 30% compared to two traditional machine learning methods, SVM and MLP, by approximately 20% compared to traditional neural network methods, and by 2.8% compared to the basic LSTM model. This demonstrates the effectiveness of the proposed model for identifying tactical intent in aerial targets. Further analysis shows that LSTM and BiLSTM-Attention, as RNN-based temporal feature network models, are more suitable for identifying tactical intent in aerial targets than the other models, further demonstrating the scientific validity of using temporal feature changes to determine intent.

[0161] Model ablation experiment

[0162] Although the comparative experiments of the BiLSTM-Attention model with LSTM, SAE, BP, MLP, and SVM models have fully demonstrated that the BiLSTM-Attention model has the characteristics of high accuracy and low loss rate, and can accurately identify the tactical intentions of aerial targets, it is not a comparison of the same type of mixed experimental models after all, and lacks certain experimental persuasiveness. Therefore, a model ablation experiment was conducted on the same dataset, and the experimental results are shown in Table 4. Figure 9 、 Figure 10 As shown:

[0163] From Table 4, we can see that the accuracy of the proposed model can reach 97.3%, which is 2.8%, 1.5%, and 1.1% higher than that of the LSTM, LSTM-Attention, and BiLSTM models, respectively. The loss value of the proposed model is also lower than that of the other three models. Figure 9 and 10 An analysis of the accuracy and loss changes in ablation experiments shows that, across the four models, accuracy improves and loss decreases with increasing training rounds. The BiLSTM-Attention model consistently outperforms the other three models. Shortly after initial training, the accuracy and loss of BiLSTM-Attention and BiLSTM models significantly outperform the other two models. Analysis shows that the bidirectional propagation mechanism effectively improves training results, enabling neural network models to learn faster with the same batch size, learning rate, and training rounds. The curves for the BiLSTM and LSTM-Attention models are very similar, and both significantly outperform the LSTM model. This demonstrates that the introduction of the bidirectional propagation and attention mechanisms significantly improves the LSTM model.

[0164] Three model evaluation indicators, namely precision (the ratio of the number of positive samples with correct intent recognition to the number of positive samples determined by the intent recognizer), recall (the ratio of the number of positive samples with correct intent recognition to the actual number of positive samples), F1-score (the harmonic mean of precision and recall), and accuracy (the ratio of the number of samples with correct intent recognition to the total number of samples), are introduced to further verify the superiority of the model.

[0165] Table 4 Model ablation experiment results

[0166]

[0167]

[0168] Table 5 Intent recognition performance metrics

[0169]

[0170] Tables 5 (①, ②, ③, and ④) represent the BiLSTM-Attention, BiLSTM, LSTM-Attention, and LSTM models for identifying tactical intentions in air combat targets, respectively. The table shows that the four models have relatively low recognition rates for feint and surveillance intentions, but high recognition rates for retreat intentions. The BiLSTM and LSTM-Attention models show relatively small differences in the three evaluation metrics, but all significantly outperform the LSTM. The BiLSTM-Attention model dominates across all evaluation metrics, demonstrating that the proposed model can accurately identify the tactical intentions of air targets.

[0171] This paper analyzes the characteristics of the problem of identifying tactical intentions of air combat targets. It adopts a hierarchical strategy to select a total of 17 dimensions of air combat features from the perspectives of threat level and combat mission, and divides them into numerical and non-numerical data, which are encoded and normalized separately. The cognitive experience of decision makers is encapsulated as labels. At the same time, this paper proposes a BiLSTM-Attention air combat target tactical intention identification model that utilizes the changes in the time series characteristics of air targets. This model uses a BiLSTM neural network to fully learn 12 consecutive frames of air combat feature information, extracts deeper features, and then uses the attention mechanism to assign different weights to the features to achieve the purpose of more accurate intention identification. Comparative experiments with other models show that the proposed model has a faster learning speed and higher recognition accuracy. However, the proposed model still has deficiencies in identifying intentions with high similarity and strong deception in air combat features. The next focus of research will be how to identify intentions that change in the detected time series.

Claims

1. A tactical intention recognition method based on BiLSTM-Attention, characterized by: The steps include: Step S1: describing the air target intention recognition as mapping the air combat intention recognition features to the air combat intention types; Step S2: Establish an air combat intention recognition model based on BiLSTM-Attention; The step S2 includes: the BiLSTM-Attention model is divided into three parts: an air combat feature vector input layer, a hidden layer, and an output layer, wherein the hidden layer includes a BiLSTM layer, an Attention layer, and a Dense layer; include: Step S21: Calculate the air combat feature vector input layer; Step S22: Calculate the hidden layer; The step S21 includes: Step S211: Read the collected data and perform data cleaning; Step S212: normalizing the numerical air combat feature data; Normalize 11 numerical air combat feature data types, including enemy aircraft acceleration, enemy aircraft altitude, enemy aircraft speed, enemy aircraft air combat capability factor, heading angle, azimuth angle, our aircraft acceleration, our aircraft altitude, our aircraft speed, our aircraft air combat capability factor, and distance between both sides; For the xth numerical data F x =[f x1 ,f x2 ,…f xi ,…,f xn ](x=1,2,…,11), n ​​is the total number of data, the i-th original data value f in the x-th type xi The result of mapping to the interval [0, 1] is f′ xi , the formula is: Where: minF x is the x-th dimension feature F x The minimum value of maxF x is the x-th dimension feature F x The maximum value of Step S213: Encode the non-numeric air combat feature data The four attribute data types of air radar status, sea radar status, interference status, and interference status are coded as 0 and 1. For the two attribute data types of maneuver type and enemy aircraft type, the coded data of each non-numeric feature is obtained and then normalized. Step S214: Encode the seven enemy target combat intention types into category labels; Step S215: randomly initialize the data and divide the training set and test set into 8:2; After steps S211 to S215, the collected air combat feature data becomes a feature vector form that can be directly accepted and processed by the hidden layer.

2. The tactical intention recognition method based on BiLSTM-Attention according to claim 1, characterized in that: The step S1 is to determine the tactical intention space I to the temporal feature set V T The mapping function is as follows: Where I=(i1,i2,…,i n ) is the tactical intention space of air targets, namely, seven types of intentions: {penetration, feint, attack, reconnaissance, retreat, surveillance, and electronic jamming}, V T From t1 to t T The temporal feature set is composed of feature sets of T consecutive moments, and the function f is a mapping function between an intention type and the real-time feature information of the battlefield collected at each moment.

3. The tactical intention recognition method based on BiLSTM-Attention according to claim 2, characterized in that: The step S1 further includes: Step S11: Describing the target tactical intention space The tactical intention space for establishing enemy targets includes seven types of intentions: {penetration, feint, attack, reconnaissance, retreat, surveillance, and electronic jamming}; Step S12: Description of aerial target tactical intention identification features The air combat capability factor is also an important factor in determining the degree of target threat. For the air combat capability of a fighter, a single-aircraft air combat capability threat function C is constructed: C=[lnε1+ln(ε2+1)+ln(∑ε3+1)]ε4ε5ε6ε7 (2) Among them, ε2 airborne weapon performance, ε3 airborne equipment detection capability, ε4 basic flight performance of fighter, ε5 fighter operation performance, ε6 fighter combat survivability performance, and ε7 electronic information confrontation performance.

4. The tactical intention recognition method based on BiLSTM-Attention according to claim 1, characterized in that: The step S22 includes: Step S221: Calculate the BiLSTM layer Calculate using the following formula: C f =σ(W f [h t-1 ,x t ]+b f ) (4) C u =σ(W u [h t-1 ,x t ]+b u ) (5) C t =tanh(W c [h t-1 ,x t ]+b c ) (6) C t =C u *C t +C f *C t-1 (7) C o =σ(W o [h t-1 ,x t +b o ]) (8) h t =Γ o *tanhC t (9) Formula (4) calculates the value of the forget gate. From the form of formula (4), it can be seen that the value of the forget gate at time t is determined by x t and h t-1 Jointly determined; Formula (5) calculates the activation (W u [h t-1 , x t ]+b u ) cell state value; Formula (6) is calculated by h t-1 and x t The value of the candidate memory unit determined by c t-1 and C t For Γ f and Γ u After the regulation of t Value; Formulas (8) and (9) calculate the value of h at time t t-1 and x t The output h of the final hidden state of the LSTM after the internal loop and update is determined t ; Among them, x t represents the input features at time t; C t-1 represents the neuron before updating; C t represents the updated neuron; h t-1 and h t Represent the output features of the previous moment and the current moment respectively; Γ f , Γ u and Γ o Represent the forget gate, update gate and output gate respectively; C t is a candidate neuron; σ is a Sigmoid function; its operation process is as follows, where w f 、w u 、w c 、w o and b f 、b u 、b c 、b o is the weight coefficient matrix and offset vector of each corresponding part; The hidden state of BiLSTM at the current moment is determined by the current input x t , the output of the previous moment's forward propagation hidden layer state And the output of the backpropagation hidden layer state at the next moment The three parts jointly determine that BiLSTM is composed of two unidirectional LSTMs, so the hidden state of BiLSTM at time t can be obtained by forward propagation of the hidden state. and the back-propagated hidden state The two parts are obtained, The calculation formula is shown in (10)(11)(12), where w i Represents the weight from one unit layer to another unit layer, i = 1, 2..., 6; Step S222: Calculate the Attention layer The hidden state s of each initial input t and the weight coefficient α of each hidden layer t The final output state vector Y is obtained by accumulating the product of . e t =tanh(w t S t +b t ) (13) Where: e t The state vector s represents the t-th eigenvector t The energy value determined, w t represents the weight coefficient matrix of the t-th eigenvector, b t represents the offset corresponding to the t-th eigenvector, e i The meaning of e t Similarly, according to formula (14), the conversion from the initial input state to the new attention state can be achieved, and then the final output state vector Y is obtained through (15). Finally, Y is integrated with the Dense layer as an output value and input into the final output layer; Step S223: Calculate the output layer The input of the output layer is the output of the Attention mechanism layer in the hidden layer. The multi-classification Softmax function is used to perform corresponding calculations on the input of the output layer to obtain the classification of the tactical intention of the air target. The specific formula is: and k =softmax(w1Y+b1) (16) Among them: w1 represents the weight coefficient matrix that needs to be trained from the Attention mechanism layer to the output layer; b1 represents the corresponding bias that needs to be trained; y k Predict labels for the output of the output layer.

Images