CN115130357A - GRU-based air target combat intention prediction system and method - Google Patents

GRU-based air target combat intention prediction system and method Download PDF

Info

Publication number
CN115130357A
CN115130357A CN202111402187.4A CN202111402187A CN115130357A CN 115130357 A CN115130357 A CN 115130357A CN 202111402187 A CN202111402187 A CN 202111402187A CN 115130357 A CN115130357 A CN 115130357A
Authority
CN
China
Prior art keywords
intention
characteristic
prediction
air target
combat
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111402187.4A
Other languages
Chinese (zh)
Inventor
滕飞
宋亚飞
王刚
王坚
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Air Force Engineering University of PLA
Original Assignee
Air Force Engineering University of PLA
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Air Force Engineering University of PLA filed Critical Air Force Engineering University of PLA
Priority to CN202111402187.4A priority Critical patent/CN115130357A/en
Publication of CN115130357A publication Critical patent/CN115130357A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/25—Design optimisation, verification or simulation using particle-based methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention belongs to the field of combat intention prediction, and particularly relates to an aerial target combat intention prediction system and method based on GRUs. Comprises a characteristic prediction module and an intention identification module, and identifies the historical air target combat intention into a characteristic set V m Input to a feature prediction module which outputs a set of predicted features W m Will predict the feature set W m Identification characteristic set V for fighting intention with historical aerial target m Inputting the time sequence characteristic data into the intention identification module, calculating the probability of each intention type, and outputting the intention type label with the maximum probability to obtainAnd identifying the result of the target fighting intention in the air. The advantages are that: constructing an air combat intention characteristic set by a layering method, coding to generate numerical time sequence characteristics, and encapsulating domain expert knowledge experience into a label; the air combat characteristics are deeply learned by using the BiGRU network, and the characteristic weight is adaptively distributed by using an attention mechanism, so that the identification accuracy of the air target combat intention is improved.

Description

GRU-based air target combat intention prediction system and method
Technical Field
The invention belongs to the field of combat intention prediction, and particularly relates to an air target combat intention prediction system and method based on GRU.
Background
With the development of military technology and aviation technology, informatization gradually becomes the core content of modern battlefields, and the future war is also certainly the informatization war. Meanwhile, due to the fact that the battlefield information amount is increased greatly due to continuous development and application of high technology, it is difficult to timely and effectively identify enemy intentions from multi-source battlefield data only by means of field expert experience. Therefore, the intelligent method is urgently needed to break through the disadvantages of the manual mode.
In order to meet the requirements of a combat decision system, the conventional enemy target combat intention identification research mainly comprises methods such as an evidence theory, template matching, an expert system, a Bayesian network and a neural network. According to different fighting backgrounds, a certain effect is achieved in identification of the target fighting intention of the enemy, but the time sequence characteristic learning and knowledge representation are insufficient.
Firstly, the target fighting intention is realized through a series of tactical actions, so that the dynamic property of the target and the battlefield environment can present characteristics which change continuously along with time, and the enemy target has certain concealment and deception when carrying out fighting actions, so that the method for judging the enemy target fighting intention by using the characteristic information at a single moment is not scientific enough.
Secondly, the above method requires explicit organization, abstraction and description of experience knowledge of military experts, and knowledge representation and engineering implementation are difficult. Aiming at the defects of the methods, some of the existing methods propose a tactical intention intelligent identification model based on a Long Short-Term Memory (LSTM) network, the input characteristics of the model are continuous 12-frame time sequence characteristics, the shortcoming that the target combat intention of an enemy is judged at a single moment can be effectively overcome, and the model organizes, abstracts and describes the experience knowledge of military experts through an implicit method, so that the difficulty of knowledge representation and engineering realization is low. However, it infers the current information only by using the historical time information, and cannot effectively use the future time information. In addition, the methods proposed above all recognize the air target fighting intention, and do not have the effect of predicting the enemy target fighting intention in advance.
Because the target combat intention in the actual air combat is realized by a series of tactical actions, the target state presents time sequence and dynamic change characteristics. The traditional combat intention identification method only relies on single moment to carry out reasoning, is not scientific and effective enough, and has no effect of predicting enemy intentions in advance.
Disclosure of Invention
The invention aims to provide a GRU-based air target combat intention prediction system and method, which can effectively improve the accuracy and effectiveness of enemy intention prediction.
The technical scheme of the invention is as follows: an air target combat intention prediction system based on GRU comprises a characteristic prediction module and an intention identification module, wherein the characteristic set V is used for identifying historical air target combat intention m Input to a feature prediction module which outputs a set of predicted features W m Will predict the feature set W m Identification characteristic set V for fighting intention with historical aerial target m And inputting the composition time sequence characteristic data into an intention identification module, calculating the probability of each intention type, and outputting a maximum probability intention type label to obtain an identification result of the air target combat intention.
The characteristic prediction module comprises an air target fighting intention recognition characteristic prediction input layer, an air target fighting intention recognition characteristic prediction hidden layer and an air target fighting intention recognition characteristic prediction output layer.
The air target fighting intention recognition characteristic prediction input layer processes the collected air target characteristic data set into a characteristic vector form, including,
(A) reading an acquired data set and cleaning the data;
(B) encoding non-numerical type aerial target characteristic data;
(C) normalizing the encoded non-numerical data and numerical data;
(D) the data was divided into training and test sets by 8: 2.
(E) And constructing a training sample and a testing sample.
The air target combat intention identification feature prediction hidden layer comprises a gate control cycle unit and a BiGRU layer.
The air target fighting intention identifies the characteristic set output layer, and outputs h of the BiGRU network in the hidden layer t And inputting the prediction characteristic values into a full connection layer in an output layer, and outputting the final prediction characteristic values by using a Linear activation function.
The intention identification module comprises an input layer, a hidden layer and an output layer.
The air target combat intention identification input layer is used for constructing intention identification module sample data and predicting the intention in the period of time by utilizing the characteristic data of the first n moments.
The air target combat intention identification hidden layer comprises a BiGRU network layer and an Attention mechanism layer.
The air target combat intention identification output layer comprises the step of inputting the output Y of the Attention mechanism layer into a Softmax classifier and outputting the output Y to obtain an air target intention prediction result.
A GRU-based air target combat intention prediction method comprises the following steps:
step 1: identifying characteristic set V of historical aerial target combat intention m Inputting the feature prediction module;
step 2: the feature prediction module outputs a set of predicted features W m ；
And step 3: set W of prediction features m Identification characteristic set V for fighting intention with historical aerial target m And inputting the composition time sequence characteristic data into an intention identification module, calculating the probability of each intention type, and outputting a maximum probability intention type label to obtain an identification result of the air target combat intention.
The invention has the beneficial effects that: constructing an air combat intention characteristic set by a layering method, coding to generate numerical time sequence characteristics, and encapsulating domain expert knowledge into a label; the air combat characteristics are deeply learned by using the BiGRU network, and the characteristic weight is adaptively distributed by using an attention mechanism, so that the identification accuracy of the air target combat intention is improved. In order to realize the advance prediction of the target intention, an air combat characteristic prediction module is introduced before the intention is identified, and a mapping relation between the prediction characteristic and the type of the combat intention is established. Simulation experiments show that the provided model can predict the fighting intention of the enemy air target in advance by one sampling point on the basis of 89.7% intention recognition accuracy, and has significant meaning in the aspect of improving the intention recognition instantaneity.
Drawings
FIG. 1 is a diagram of an intent representation and reasoning process;
FIG. 2 illustrates the operational intent encoding and pattern parsing;
FIG. 3 is the relative geometric position of air war;
FIG. 4 is an aerial target engagement intention feature set;
FIG. 5 is a model overall structure;
FIG. 6 is a GRU structure;
FIG. 7 is a BiGRU structure;
FIG. 8 is an Attention mechanism model.
Detailed Description
The invention is described in further detail below with reference to the figures and the embodiments.
The invention provides a GRU-based air target combat intention prediction system which comprises a characteristic prediction module and an intention identification module.
The intention identification module introduces a Bidirectional (Bidirectional) propagation mechanism, an Attention (Attention) mechanism and a Particle Swarm Optimization (PSO) on the basis of a Gated cyclic Unit (GRU) to construct an intelligent intention identification model. Compared with the LSTM, the GRU has similar performance, lower structural complexity and shorter time for identification; compared with GRU, the BiGRU can comprehensively make judgment by utilizing not only historical time information but also future time information; the PSO can find the optimal parameters of the BiGRU network; the Attention mechanism layer can further highlight key information influencing the intention and improve the accuracy of intention identification.
The feature prediction module analyzes various collected features by adopting a BiGRU network, predicts to obtain future aerial target features, inputs the future aerial target features into the intention identification module, and establishes a mapping relation between the future aerial target features and the enemy target fighting intention types. Experiments show that the method can predict the air target fighting intention of the enemy in advance by one sampling point with the accuracy of 89.7 percent.
The identification of the air target fighting intention in the invention refers to a process of acquiring real-time data through a plurality of sensors from a dynamic and complex battlefield environment, analyzing the real-time dynamic data, and reasoning the air target fighting intention of an enemy by combining prior knowledge, expert experience of related fighting fields and the like, wherein the process is shown in figure 1.
Airborne target intent recognition is a pattern recognition problem that may be described as a mapping of intent recognition features to airborne target engagement intent types. Wherein, the vector V (t) For the real-time characteristic information of the air war at the time t, the vector P is equal to (P) 1 ，p 2 …p n ) Because of the complexity, high resistance and deceptiveness of the actual air combat environmental conditions, and certain deceptiveness and one-sidedness of the detected air combat real-time characteristic information in a single time, the method for deducing the combat intention of the enemy plane from the air combat characteristic information at continuous moments has higher accuracy and scientificity than that of deducing the combat intention from the air combat characteristic information at the single time. V m Is t 1 To t m A time sequence feature set formed by feature sets at m continuous moments so as to determine a battle intention space set P to a time sequence feature set V m The mapping function P of (a), is as follows:
Figure BDA0003367543570000051
the accurate identification of the air target combat intention needs to be combined with professional military knowledge and combat experience of experts in the air combat field, and the air combat key information is extracted, compared, analyzed and associatedComplex thinking activities such as reasoning and the like are realized, and a single explicit mathematical formula is difficult to establish the idea recognition feature set V m Mapping relation to the battle intention space set P.
The method trains a BiGRU-Attention network structure by using an aerial target fighting intention identification characteristic set, and implicitly establishes a mapping relation between the characteristic set and the fighting intention.
Airborne target combat intention space set description
The target operational intention space set aims at different operational forms, different enemy entities and different imagination backgrounds, and the corresponding target operational intention sets are different. It is therefore desirable to define a spatial set of warfare intentions for an enemy target based on the corresponding context of the battle, the attributes of the enemy target, and the possible performance of the battle mission. A target intention space set established according to the potential threat of the underwater target is { avoidance, patrol and attack }; aiming at the formation of a single group of marine ships of an enemy, establishing a combat intention space set of { retreat, shield, attack and reconnaissance }; the set of the combat intention space of the aerial targets is defined as { reconnaissance, surveillance, attack, defense }. The invention takes unmanned aerial vehicle close-range combat as a research object, and establishes a combat intention space set of an enemy target, wherein the combat intention space set comprises seven intention types of { impersonation attack, monitoring, electronic interference, defense burst, attack, retreat and reconnaissance }.
After determining the enemy fighting intention space set, how to convert the human cognitive model into a label which can be trained by an intelligent model and correspond to the intention type in the fighting intention space set is the key for applying the intelligent identification model to fighting intention identification. Therefore, the cognitive experience of experts in the field of air combat can be packaged into labels to train the intelligent recognition model. Seven label values of {0, 1,2, 3, 4, 5 and 6} are respectively set for seven intention types in the enemy target fighting intention space set up by the invention, and the corresponding fighting intention type coding and model analysis mechanism are shown in FIG. 2. For example, if the intention prediction result output by the model proposed by the present invention is 5, it can be considered that the fighting intention of the enemy object against the my object is withdrawal. Therefore, the knowledge packaging and model analysis can clearly and easily describe the experience knowledge of people, and the model training is convenient.
Identification characteristic set selection for air target combat intention
The enemy air target combat intention is highly relevant to the enemy combat mission, the degree of threat of the two parties to each other, and the tactical action. And selecting the characteristics closely related to the air target combat intention according to the field expert knowledge and the three aspects.
From the perspective of a combat mission, when an enemy unmanned aerial vehicle executes a certain mission, certain aspects of characteristics of the enemy unmanned aerial vehicle need to meet certain conditions, for example, the enemy unmanned aerial vehicle is divided into high-altitude penetration and low-altitude penetration when the penetration mission is executed, and the corresponding height is 10-11 km; the flying speed of the fighter plane is generally 735-1470km/h when the fighter plane executes a large-speed enemy attack. The signal state of the aerial target radar is also in certain relation with the combat mission, for example, a fighter usually starts an aerial radar and electronic interference during aerial combat; sea-facing radars and air-facing radars are typically turned on when performing a reconnaissance mission.
From the perspective of threat degree, the analysis shows that a plurality of factors influence the threat degree between the targets of the two parties, and the invention mainly considers the speed, the flight acceleration, the distance, the flight height, the course angle and the azimuth angle of the fighters of the two parties of the enemy and the my party for facilitating the experimental data acquisition. As shown in FIG. 3, H 1 、H 2 The flight heights of the enemy and the my are obtained; v 1 、V 2 The speed of the enemy and the my double flight; d is the distance between the two parties; psi is the heading angle;
Figure BDA0003367543570000071
is the azimuth angle.
The air combat capability factor is also an important factor influencing the degree of the target threat, and for the air combat capability of a warplane, a single-machine air combat capability threat function is constructed:
C＝[lnε 1 +ln(ε 2 +1)+ln(∑ε 3 +1)]ε 4 ε 5 ε 6 ε 7 (2)
in the formula, epsilon 1 ～ε 7 Respectively representing the maneuvering performance, onboard weapon performance, onboard detection capability and aircraft operation of the warplaneSeven parameters of performance, survivability of the fighter plane, combat range of the fighter plane and electronic information fighting capacity. The air combat capability threat is the inherent attribute of the fighter, air combat capability factors of various fighters of both the enemy and the my are calculated through a formula in a certain period, and are stored in a database, and the database is updated at any time according to the information mastered by the party.
The realization of the air target combat intention is closely related to the maneuvering action of the warplane. There are two kinds of commonly used maneuver library, which are respectively a "typical tactical maneuver library" based on the typical air combat tactical aircraft maneuver as the design basis and a "basic control maneuver library" based on the air combat basic operation maneuver. The invention carries out intention identification according to time sequence characteristics, carries out target combat intention identification by taking continuous 12 time characteristics as one sample, and the control algorithm of a typical tactical action library has complex solution and difficult determination of action exit and conversion time nodes. As for the traditional 'basic operation action library', the library has only seven maneuvering modes, the combined maneuvering actions are not rich enough, and limit operation modes are adopted, so that the library is not in line with the actual air combat condition. Therefore, the invention adopts an improved basic operation action library which comprises 11 operation actions including left turning, right turning, accelerating forward flight, uniform speed forward flight, decelerating forward flight, climbing, left climbing, right climbing, diving, left diving and right diving.
In summary, the air target engagement intention feature set of the invention is 16-dimensional feature vector of { enemy aircraft flight altitude, i aircraft altitude, i aircraft flight speed, i aircraft acceleration, i aircraft air engagement capability factor, course angle, two-party distance, azimuth angle, air radar state, sea-to-sea radar state, maneuver type, interference state } and can be divided into numerical type and non-numerical type features, as shown in fig. 4.
As shown in FIG. 5, the air target fighting intention prediction system based on GRU provided by the invention comprises a characteristic prediction module and an intention identification module, wherein the characteristic prediction module is constructed based on a BiGRU network, and a historical air target fighting intention identification characteristic set V is used m As an inputObtaining a prediction characteristic set W by utilizing the default activation function Linear output of the full connection layer m . Set W of prediction features m Identification characteristic set V for fighting intention with historical aerial target m And inputting the composition time sequence characteristic data into an intention identification module constructed by a BiGRU-Attention network, calculating the probability of each intention type by using a Softmax function, outputting a maximum probability intention type label, and obtaining an identification result of the air target fighting intention.
The feature prediction module and the intent recognition module are explained below.
The feature prediction module comprises: the device comprises an air target fighting intention identification characteristic prediction input layer, an air target fighting intention identification characteristic prediction hidden layer and an air target fighting intention identification characteristic prediction output layer. The prediction of each characteristic independently is higher than the overall prediction precision. Thus, the input node of the module is (8,1), where 8 represents the step size; 1 represents a feature dimension; the output node output is 1, which indicates that the output characteristic dimension is 1. The components are described in detail below.
The air target combat intention identification characteristic prediction input layer is mainly used for preprocessing an acquired air target characteristic data set, namely processing the data set into a characteristic vector form which can be directly received and processed by the BiGRU layer. The method specifically comprises the following steps:
(A) reading an acquired data set and data cleansing
(B) The non-numerical type air target characteristic data is coded, and the four attributes of the interference state, the interfered state, the air radar state and the sea radar state are coded into 1 and 0, wherein 0 represents a closed state, and 1 represents an open state. And (3) quantizing the maneuvering type by using Millier 9-level quantization theory to obtain numerical data, thereby completing the coding of non-numerical data.
(C) And normalizing the encoded non-numerical data and the numerical data. The data normalization can improve the network convergence speed and precision, prevent model gradient explosion and the like. The embodiment of the invention normalizes 11 kinds of numerical data and 5 kinds of encoded non-numerical data, and 16 kinds of aerial target characteristic dataAnd (4) transforming. For ith dimension characteristic data G i ＝[g i1 ,g i2 ,…g ix ,…,g in ](i ═ 1,2, …, 16); where n is the total number of data. Normalizing the xth original data in the ith dimension feature to an interval [0,1]Result of (b) is h' ix The formula is as follows:
Figure BDA0003367543570000091
in the formula maxG i Is the ith dimension characteristic G i Maximum value of (d); minG i Is the ith dimension characteristic G i Of the measured value (c).
(D) The data was partitioned into training and test sets in 8: 2.
(E) Training and test samples were constructed as follows.
The method of single characteristic sequential prediction is adopted, and the method of predicting the air target combat intention characteristic of the height of the enemy plane is taken as an example. Suppose that the enemy height H at the n +1 th moment is predicted by using the enemy height data at the first n moments n+1 . The function mapping relation is as follows:
H n+1 ＝f(h 1 ,h 2 ,…,h n ) (4)
in the formula, h i And i e (1,2 …, n) is the flight height of the enemy plane at the ith moment. Selection of h 1 ～h n For the first set of input data, the label is h n+1 (ii) a By h 2 ～h n+1 For inputting data, h n+2 Is a label. By analogy, the following training sample input data and training sample labels are formed; the construction method of the test data is consistent with the training sample data.
Figure BDA0003367543570000093
[h n+1 h n+2 …h n+m+1 ] (6)
After the five steps, the collected air target combat intention characteristic set V m Is converted into a hidden layer capable of direct acceptanceAnd processed feature vector form.
Air target combat intention recognition feature prediction hidden layer
Building gated cycle units (GRUs): the Gated Recurrent Unit (GRU) as a variant of Recurrent Neural Network (RNN) also has a similar recursive structure to RNN, and has a "memory" function of processing time series data. Meanwhile, the GRU can effectively relieve the problems of gradient disappearance and gradient explosion which possibly occur in the RNN training process, so that the problem of long-term memory is effectively solved. The long-short memory (LSTM) network is also a variant of RNN, and has almost the same performance as GRU, but GRU is simpler in structure, and can reduce the amount of calculation and improve the training efficiency.
The internal structure of the GRU is shown in fig. 6. The GRU has two inputs, respectively the output state h at the previous moment t-1 And the input sequence value x of the current time t Output as the state h of the current time t . Mainly through a reset gate r t And an update gate z t To update the model state, reset gate r t Controlling the degree of forgetting historical state information to enable the network to lose unimportant information; updating the door z t And controlling the proportion of the state information of the previous moment brought into the current state to help the network to memorize long-term information. The internal calculation formula is as follows:
Figure BDA0003367543570000101
where σ is a sigmoid activation function, which acts to convert an intermediate state to [0,1 ]]Within the range; h is a total of t-1 、h t The output states at the time t-1 and the time t respectively; x is the number of t Is the input sequence value at the time t;
Figure BDA0003367543570000102
is a candidate output state; w r 、W z 、
Figure BDA0003367543570000103
U r 、U z And
Figure BDA0003367543570000104
the weight coefficient matrixes corresponding to the parts are obtained; tan h is a hyperbolic tangent function; as a hadamard product of the matrix.
Establishing a BiGRU layer: the conventional GRU structure generally propagates in a single direction along a sequence transmission direction, the acquired information is history information before the current time, so that future information is ignored, the BiGRU structure is composed of a forward GRU and a backward GRU, and has the function of capturing front and back information features, and the model structure is shown in fig. 7.
In the figure, the GRU is a forward GRU; GRU 2 Is a backward GRU. It can be seen that the output state h of the BiGRU at the time t t Can pass through the forward output state
Figure BDA0003367543570000111
And backward output state
Figure BDA0003367543570000112
Two part to find the forward output state
Figure BDA0003367543570000113
From input x at time t t And the output state of the forward GRU at time t-1
Figure BDA0003367543570000114
Determining a backward output state
Figure BDA0003367543570000117
From input x at time t t Backward GRU output state at time t +1
Figure BDA0003367543570000116
And (6) determining.
Air target combat intention recognition characteristic prediction output layer
Output h of BiGRU network in hidden layer t Inputting the predicted characteristic values into a full connection layer in an output layer, and outputting the final predicted characteristic values by using a Linear activation function。
The intention identification module comprises: the air target combat intention identifies an input layer, a hidden layer and an output layer. Wherein, the hidden layer is composed of a BiGRU layer and an Attention layer. Input node (12,16) and output node (7) of the network, wherein 12 represents a step size; 16 denotes a feature dimension; and 7 represents the total number of intent categories. The components are described in detail below.
An input layer: the method mainly comprises the steps of constructing the intention recognition module sample data. And (3) predicting the intention in the period of time by using the characteristic data of the first n moments, wherein the function mapping relation is as follows:
Q 1 ＝f(v 1 ,v 2 ,…,v 11 ,w 12 ) (8)
in the formula, Q 1 Representing the predicted intention type in the time period of 1-12 moments; v. of i I e (1,2 …,11) represents the historical feature data at the i-th time instant, w 12 Representing the feature data predicted by the feature prediction module at time 12.
Selecting (v) 1 ,v 2 ,…,v 11 ,v 12 ) The label is an intention type q corresponding to a 1-12 time period for a first group of input data 1 (ii) a With (v) 2 ,v 3 ,…,v 12 ,v 13 ) Labeling the second group of input data with an intention type q corresponding to a time period of 2-13 2 . By analogy, the training sample input data and training sample labels shown below are composed. The construction method of the test data is similar to that of the training sample data, and the difference is that the feature v of the last moment of each sample is used t Features w predicted by the feature prediction module t I.e. the input data is (v) i ,v i+1 ,…,v i+10 ,w i+11 ) The label is an intention type q corresponding to the time period of i to i +11 i 。
Figure BDA0003367543570000121
[q 1 q 2 …q m ] (10)
After the sample data is constructed, the intention label can be input into the underlying hidden layer through one-hot encoding processing.
Hidden layer: including a BiGRU network layer and an Attention mechanism layer.
Establishing a BiGRU layer: the conventional GRU structure generally performs unidirectional propagation along a sequence transmission direction, the obtained information is history information before the current time, so that future information is ignored, while the BiGRU structure is composed of forward GRUs and backward GRUs, has the function of capturing characteristics of forward and backward information, and has a model structure as shown in fig. 7.
In the figure, the GRU is a forward GRU; GRU 2 Is a backward GRU. It can be seen that the output state h of the BiGRU at the time t t Can pass through the forward output state
Figure BDA0003367543570000122
And backward output state
Figure BDA0003367543570000123
Two part to find the forward output state
Figure BDA0003367543570000124
From input x at time t t And the output state of the forward GRU at time t-1
Figure BDA0003367543570000125
Determining a backward output state
Figure BDA0003367543570000126
From the input x at time t t Backward GRU output state at time t +1
Figure BDA0003367543570000127
And (6) determining.
The Attention mechanism layer is described in detail below.
The Attention mechanism is similar to the principle of observing objects by the human brain, and focuses on the local content of the objects according to the purpose. In the invention, the Attention mechanism highlights the characteristics with larger occupation ratio to the prediction result by calculating the weight of the characteristic vector output in the BiGRU network at different moments, so that the whole model shows more excellent results. In the identification of the air target combat intention, a neural network focuses on some key features through an Attention mechanism in a training process, and the method is realized by distributing weight coefficients. Firstly, the importance degree of each feature is learned, and then corresponding weight coefficients are distributed according to the importance degree of each feature, for example, when an enemy plane is a defense intention, the features such as the flight height and the heading angle of the enemy plane are distributed with higher weights by an Attention mechanism to deepen the memory of a model.
The structure of the Attention mechanism model is shown in fig. 8.
Firstly, outputting the t-th feature vector h output by the BiGRU network t Inputting the initial state vector S into the Attention mechanism layer t (ii) a Next, the attention weight e is learned by equation (11) t (ii) a Then, the attention weight is subjected to probability by the formula (12) namely a softmax function to obtain a weight probability vector alpha t (ii) a Finally, the equation (13) and h t The multiplication and accumulation sum yields the final state vector Y. The formula is as follows:
e t ＝tanh(w w s t +b w ) (11)
Figure BDA0003367543570000131
Figure BDA0003367543570000132
in the formula: w is a w Is a weight coefficient matrix; b w Is a bias coefficient matrix; u. of w Is a randomly initialized attention matrix.
An output layer: and (3) inputting the output of the Attention mechanism layer into a multi-classification activation function Softmax, outputting an air target combat intention label with the highest probability, and analyzing the label according to the graph 2 to identify the air target combat intention of the enemy. The specific formula is as follows:
y k ＝softmax(wY+b) (14)
in the formula: w represents a weight coefficient matrix with training; b represents a bias matrix corresponding to the band training; y is k Representing the output prediction tag.
A GRU-based air target combat intention prediction method comprises the following steps:
step 1: identifying characteristic set V of historical aerial target combat intention m Inputting the feature prediction module;
step 2: the feature prediction module outputs a set of predicted features W m ；
The characteristic prediction module establishes an air target fighting intention recognition characteristic prediction input layer, an air target fighting intention recognition characteristic prediction hidden layer and an air target fighting intention recognition characteristic prediction output layer. The input node of the module is (8,1), wherein 8 represents the step length; 1 represents a feature dimension; the output node output is 1, which means that the output feature dimension is 1. The components are described in detail below.
The construction process of the air target combat intention identification characteristic prediction input layer is as follows: and preprocessing the acquired aerial target feature data set, namely processing the data set into a feature vector form which can be directly received and processed by a BiGRU layer. The method specifically comprises the following steps:
step 21: reading an acquired data set and data cleansing
Step 22: the non-numerical type air target characteristic data is coded, and four attributes of an interference state, an interfered state, an air radar state and a sea radar state are coded into 1 and 0 forms, wherein 0 represents a closed state, and 1 represents an open state. And (3) quantizing the maneuvering type by using Millier 9-level quantization theory to obtain numerical data, thereby completing the coding of non-numerical data.
Step 23: and normalizing the encoded non-numerical data and the numerical data. The data normalization can improve the network convergence speed and precision, prevent model gradient explosion and the like. The embodiment of the present invention combines 11 kinds of numerical data and 5 kinds of encoded nonnumbersThe value type data is normalized by 16 kinds of air target characteristic data. For ith dimension characteristic data G i ＝[g i1 ,g i2 ,…g ix ,…,g in ](i ═ 1,2, …, 16); where n is the total number of data. Normalizing the xth original data in the ith dimension characteristic to an interval [0,1]Result of (b) is h' ix The formula is as follows:
Figure BDA0003367543570000141
in the formula, maxG i Is the ith dimension characteristic G i The maximum value of (a); minG i Is the ith dimension characteristic G i Is measured.
Step 24: the data was divided into training and test sets by 8: 2.
Step 25: training and test samples were constructed as follows.
The method of single characteristic sequential prediction is adopted, and the method of predicting the air target combat intention characteristic of the height of the enemy plane is taken as an example. Suppose that the enemy height H at the n +1 th moment is predicted by using the enemy height data at the first n moments n+1 . The function mapping relation is as follows:
H n+1 ＝f(h 1 ,h 2 ,…,h n ) (4)
in the formula, h i And i e (1,2 …, n) is the flight height of the enemy plane at the ith moment. Selection of h 1 ～h n For the first set of input data, the label is h n+1 (ii) a By h 2 ～h n+1 To input data, h n+2 Is a label. By analogy, the following training sample input data and training sample labels are formed; the construction method of the test data is consistent with the training sample data.
Figure BDA0003367543570000151
[h n+1 h n+2 …h n+m+1 ] (6)
After the five steps, the collected air isTarget operational intention characteristic set V m It is converted into a form of feature vector that the hidden layer can directly accept and process.
The process of establishing the air target combat intention recognition characteristic prediction hidden layer is as follows:
building gated cycle units (GRUs): the Gated Recurrent Unit (GRU) as a variant of Recurrent Neural Network (RNN) also has a similar recursive structure to RNN, and has a "memory" function of processing time series data. Meanwhile, the GRU can effectively relieve the problems of gradient disappearance and gradient explosion which possibly occur in the RNN training process, so that the problem of long-term memory is effectively solved. The long-short memory (LSTM) network is also a variant of RNN, and has almost the same performance as GRU, but GRU is simpler in structure, and can reduce the amount of calculation and improve the training efficiency.
The internal structure of the GRU is shown in fig. 6. The GRU has two inputs, respectively the output state h at the previous moment t-1 And the input sequence value x of the current time t Output as the state h of the current time t . Mainly through a reset gate r t And an update gate z t To update the model state, reset gate r t Controlling the degree of forgetting historical state information to enable the network to lose unimportant information; updating the door z t And controlling the proportion of the state information at the previous moment brought into the current state to help the network memorize long-term information. The internal calculation formula is as follows:
Figure BDA0003367543570000161
where σ is a sigmoid activation function, which acts to convert an intermediate state to [0,1 ]]Within the range; h is t-1 、h t The output states at the time t-1 and the time t respectively; x is the number of t Is the input sequence value at the time t;
Figure BDA0003367543570000162
is a candidate output state; w r 、W z 、
Figure BDA0003367543570000163
U r 、U z And
Figure BDA0003367543570000164
the weight coefficient matrix corresponding to each part; tan h is a hyperbolic tangent function; as a hadamard product of the matrix.
Establishing a BiGRU layer: the conventional GRU structure generally propagates in a single direction along a sequence transmission direction, the acquired information is history information before the current time, so that future information is ignored, the BiGRU structure is composed of a forward GRU and a backward GRU, and has the function of capturing front and back information features, and the model structure is shown in fig. 7.
In the figure, the GRU is a forward GRU; GRU 2 Is a backward GRU. It can be seen that the output state h of the BiGRU at the time t t Can pass through the forward output state
Figure BDA0003367543570000165
And backward output state
Figure BDA0003367543570000166
Two part to find the forward output state
Figure BDA0003367543570000167
From input x at time t t And the output state of the forward GRU at time t-1
Figure BDA0003367543570000168
Determining a backward output state
Figure BDA00033675435700001611
From input x at time t t Backward GRU output state at time t +1
Figure BDA00033675435700001610
And (6) determining.
The process of establishing the air target combat intention recognition characteristic prediction output layer is as follows: output h of BiGRU network in hidden layer t And inputting the prediction characteristic values into a full connection layer in an output layer, and outputting the final prediction characteristic values by using a Linear activation function.
And step 3: set W of prediction features m Identification characteristic set V for fighting intention with historical aerial target m And inputting the composition time sequence characteristic data into an intention identification module, calculating the probability of each intention type, and outputting a maximum probability intention type label to obtain an identification result of the air target combat intention.
The specific process is as follows:
an input layer, a hidden layer, and an output layer of the intent recognition module are constructed. Wherein, the hidden layer comprises a BiGRU layer and an Attention layer. Input node (12,16) and output node (7) of the network, wherein 12 represents a step size; 16 denotes a feature dimension; and 7 represents the total number of intent categories. The components are described in detail below.
Establishing an input layer: the method mainly comprises the steps of constructing the sample data of the intention identification module. And (3) predicting the intention in the period of time by using the characteristic data of the first n moments, wherein the function mapping relation is as follows:
Q 1 ＝f(v 1 ,v 2 ,…,v 11 ,w 12 ) (8)
in the formula, Q 1 Representing the predicted intention type in the time period of 1-12 moments; v. of i I e (1,2 …,11) represents the historical feature data at time i, w 12 Representing the feature data predicted by the feature prediction module at time 12.
Selecting (v) 1 ,v 2 ,…,v 11 ,v 12 ) The label is an intention type q corresponding to a 1-12 time period for a first group of input data 1 (ii) a With (v) 2 ,v 3 ,…,v 12 ,v 13 ) Labeling the second group of input data with an intention type q corresponding to a time period of 2-13 2 . By analogy, the training sample input data and training sample labels shown below are composed. The construction method of the test data is similar to that of the training sample data, and the difference is that the feature v of the last moment of each sample is used t Features w predicted by the feature prediction module t I.e. the input data is (v) i ,v i+1 ,…,v i+10 ,w i+11 ) The label is an intention type q corresponding to the time period of i to i +11 i 。
Figure BDA0003367543570000171
[q 1 q 2 …q m ] (10)
After the sample data is constructed, the intention label can be input into the underlying hidden layer through one-hot encoding processing.
And establishing a hidden layer, wherein the hidden layer comprises a BiGRU network layer and an Attention mechanism layer.
Establishing a BiGRU layer: the conventional GRU structure generally propagates in a single direction along a sequence transmission direction, the acquired information is history information before the current time, so that future information is ignored, the BiGRU structure is composed of a forward GRU and a backward GRU, and has the function of capturing front and back information features, and the model structure is shown in fig. 7.
In the figure, the GRU is a forward GRU; GRU 2 Is a backward GRU. It can be seen that the output state h of the BiGRU at the time t t Can pass through the forward output state
Figure BDA0003367543570000181
And backward output state
Figure BDA0003367543570000182
Two part to find the forward output state
Figure BDA0003367543570000183
From input x at time t t And the output state of the forward GRU at time t-1
Figure BDA0003367543570000184
Determining, backward outputting state
Figure BDA00033675435700001810
From the input at time tx t Backward GRU output state at time t +1
Figure BDA0003367543570000189
And (6) determining.
The Attention mechanism layer is described in detail below.
The Attention mechanism is similar to the principle of observing objects by the human brain, and focuses on the local content of the objects according to the purpose. In the invention, the Attention mechanism highlights the characteristics with larger occupation ratio to the prediction result by calculating the weight of the characteristic vector output in the BiGRU network at different moments, so that the whole model shows more excellent results. In the identification of the air target combat intention, the neural network focuses on some key features through an Attention mechanism in the training process, and the method is realized by distributing weight coefficients. Firstly, the importance degree of each feature is learned, and then corresponding weight coefficients are distributed according to the importance degree, for example, when an enemy plane is a defense intention, the features such as the flying height and the heading angle of the enemy plane are distributed with higher weights by an Attention mechanism to deepen the model memory.
An Attention mechanism layer is established, and the structure of the Attention mechanism layer is shown in fig. 8.
Firstly, outputting the t-th feature vector h output by the BiGRU network t Inputting the initial state vector S into the Attention mechanism layer t (ii) a Next, the attention weight e is learned by equation (11) t (ii) a Then, the attention weight is subjected to probability by the formula (12) namely a softmax function to obtain a weight probability vector alpha t (ii) a Finally, the equation (13) and h t The multiplication and the accumulation sum result in the final state vector Y. The formula is as follows:
e t ＝tanh(w w s t +b w ) (11)
Figure BDA0003367543570000187
Figure BDA0003367543570000188
in the formula: w is a w Is a weight coefficient matrix; b w Is a bias coefficient matrix; u. of w Is a randomly initialized attention matrix.
And (3) establishing an output layer, inputting the output of the Attention mechanism layer into a multi-classification activation function Softmax, outputting an air target combat intention label with the highest probability, and analyzing the label according to the graph in FIG. 2 to identify the enemy air target combat intention. The specific formula is as follows:
y k ＝softmax(wY+b) (14)
in the formula: w represents a weight coefficient matrix with training; b represents a bias matrix corresponding to the band training; y is k Representing the output prediction tag.

Claims (10)

1. A GRU-based air target combat intention prediction system is characterized in that: comprises a characteristic prediction module and an intention identification module, and identifies the historical air target combat intention into a characteristic set V m Input to a feature prediction module which outputs a set of predicted features W m Will predict the feature set W m Identification characteristic set V for fighting intention with historical aerial target m And inputting the formed time sequence characteristic data into an intention identification module, calculating the probability of each intention type, and outputting a maximum probability intention type label to obtain an air target combat intention identification result.
2. The GRU-based air target engagement intention prediction system of claim 1 wherein: the characteristic prediction module comprises an air target fighting intention recognition characteristic prediction input layer, an air target fighting intention recognition characteristic prediction hidden layer and an air target fighting intention recognition characteristic prediction output layer.
3. The GRU-based air target engagement intention prediction system of claim 2, wherein: the air target fighting intention recognition characteristic prediction input layer processes the collected air target characteristic data set into a characteristic vector form, including,
(A) reading an acquired data set and cleaning the data;
(B) encoding non-numerical type aerial target characteristic data;
(C) normalizing the encoded non-numerical data and numerical data;
(D) the data was divided into training and test sets by 8: 2.
(E) And constructing a training sample and a testing sample.
4. The GRU-based air target engagement intention prediction system of claim 2 wherein: the air target combat intention identification feature prediction hidden layer comprises a gate control cycle unit and a BiGRU layer.
5. The GRU-based air target engagement intention prediction system of claim 2, wherein: the air target fighting intention identifies the characteristic set output layer, and outputs h of the BiGRU network in the hidden layer t And inputting the prediction characteristic values into a full connection layer in an output layer, and outputting the final prediction characteristic values by using a Linear activation function.
6. The GRU-based air target engagement intention prediction system of claim 1 wherein: the intention identification module comprises an input layer, a hidden layer and an output layer.
7. The GRU-based air target engagement intention prediction system of claim 6 wherein: the air target combat intention identification input layer is used for constructing intention identification module sample data and predicting the intention in the period of time by utilizing the characteristic data of the first n moments.
8. The GRU-based air target engagement intention prediction system of claim 6 wherein: the air target combat intention identification hidden layer comprises a BiGRU network layer and an Attention mechanism layer.
9. The GRU-based air target engagement intention prediction system of claim 6 wherein: the air target combat intention identification output layer comprises the step of inputting the output Y of the Attention mechanism layer into a Softmax classifier and outputting the output Y to obtain an air target intention prediction result.
10. A GRU-based air target combat intention prediction method is characterized by comprising the following steps:
step 1: identifying characteristic set V of historical aerial target combat intention m Inputting the feature prediction module;
step 2: the feature prediction module outputs a set of predicted features W m ；
And step 3: set W of prediction features m Identification characteristic set V for fighting intention with historical aerial target m And inputting the composition time sequence characteristic data into an intention identification module, calculating the probability of each intention type, and outputting a maximum probability intention type label to obtain an identification result of the air target combat intention.
CN202111402187.4A 2021-11-22 2021-11-22 GRU-based air target combat intention prediction system and method Pending CN115130357A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111402187.4A CN115130357A (en) 2021-11-22 2021-11-22 GRU-based air target combat intention prediction system and method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111402187.4A CN115130357A (en) 2021-11-22 2021-11-22 GRU-based air target combat intention prediction system and method

Publications (1)

Publication Number Publication Date
CN115130357A true CN115130357A (en) 2022-09-30

Family

ID=83375799

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111402187.4A Pending CN115130357A (en) 2021-11-22 2021-11-22 GRU-based air target combat intention prediction system and method

Country Status (1)

Country Link
CN (1) CN115130357A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116362390A (en) * 2023-03-20 2023-06-30 中国人民解放军军事科学院战略评估咨询中心 Marine volt-ampere prediction method and system based on probabilistic neural network
CN116484954A (en) * 2023-03-20 2023-07-25 中国人民解放军军事科学院战略评估咨询中心 Artificial intelligence offshore voltaic model generation method, computer program and storage medium

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116362390A (en) * 2023-03-20 2023-06-30 中国人民解放军军事科学院战略评估咨询中心 Marine volt-ampere prediction method and system based on probabilistic neural network
CN116484954A (en) * 2023-03-20 2023-07-25 中国人民解放军军事科学院战略评估咨询中心 Artificial intelligence offshore voltaic model generation method, computer program and storage medium
CN116362390B (en) * 2023-03-20 2023-09-12 中国人民解放军军事科学院战略评估咨询中心 Marine volt-ampere prediction method and system based on probabilistic neural network
CN116484954B (en) * 2023-03-20 2023-11-10 中国人民解放军军事科学院战略评估咨询中心 Artificial intelligent offshore voltaic model generation method and storage medium

Similar Documents

Publication Publication Date Title
US20220197281A1 (en) Intelligent decision-making method and system for unmanned surface vehicle
CN115130357A (en) GRU-based air target combat intention prediction system and method
Xue et al. Panoramic convolutional long short-term memory networks for combat intension recognition of aerial targets
CN114266355A (en) Tactical intention identification method based on BilSTM-Attention
Teng et al. A GRU‐Based Method for Predicting Intention of Aerial Targets
CN112668804B (en) Method for predicting broken track of ground wave radar ship
Teng et al. An air target tactical intention recognition model based on bidirectional GRU with attention mechanism
CN112749761A (en) Enemy combat intention identification method and system based on attention mechanism and recurrent neural network
López-Randulfe et al. Spiking neural network for fourier transform and object detection for automotive radar
CN114818853B (en) Intention recognition method based on bidirectional gating circulating unit and conditional random field
Yang et al. Online hierarchical recognition method for target tactical intention in beyond-visual-range air combat
CN116592883A (en) Navigation decision method based on attention and cyclic PPO
Wang et al. Tactical intention recognition method of air combat target based on BiLSTM network
CN110766140A (en) XGboost and LSTM-based multi-sensor real-time threat level classification method
Wang et al. Learning embedding features based on multisense-scaled attention architecture to improve the predictive performance of air combat intention recognition
CN115661576A (en) Method for identifying airplane group intention under sample imbalance
Wang et al. Passive tracking of underwater acoustic targets based on multi-beam LOFAR and deep learning
Mulgund et al. Olipsa: On-line intelligent processor for situation assessment
Zhang et al. An intention inference method for the space non-cooperative target based on BiGRU-Self Attention
CN115204286A (en) Target tactical intention online identification method based on deep learning in simulation environment
BUZ et al. A novel approach and application of time series to image transformation methods on classification of underwater objects
Yongcui Situation Awareness and Target Recognition of Marine Big Data Battlefield based on Deep Learning
Chen et al. Target intent recognition method based on evidence fusion in TimeSeries networks
Yang et al. Study on the trajectory positioning and prediction framework of unmanned aerial vehicle based on long short-term memory neural network
Zhang et al. Combat Intention Recognition of Air Targets Based on 1DCNN-BiLSTM

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination