Aerospace Science and Technology 132 (2023) 108005

Contents lists available at ScienceDirect

Aerospace Science and Technology

journal homepage: www.elsevier.com/locate/aescte

Quick intention identification of an enemy aerial target through

information classification processing
Yinhan Wang a,b , Jiang Wang a,b , Shipeng Fan a,b,∗ , Yuchen Wang a,b
a
School of Aerospace Engineering, Beijing Institute of Technology, Beijing, CO 100081, China
b
Beijing Key Laboratory of UAV Autonomous Control, Beijing Institute of Technology, Beijing, CO 100081, China

a r t i c l e i n f o a b s t r a c t

Article history: Rapidly and accurately identifying the tactical intention of an enemy aerial target is an important issue
Received 30 March 2022 for combat decision making. To this end, a quick intention identification model based on hybrid neural
Received in revised form 2 November 2022 network is established in this paper. With available feature sequential measurements of the enemy
Accepted 7 November 2022
target as inputs of the model, possibilities of different intentions are calculated timely. To increase
Available online 21 November 2022
Communicated by Qinglei Hu
the training efficiency and accuracy of recognition, the measurement information is processed using
different neural network. Maneuvering data with large variations over time are processed using gated
Keywords: recurrent unit (GRU), while other data are processed using back propagation (BP) neural network. Besides,
Intention recognition the fitting cubic sample interpolation is adopted to deal with incomplete information. Monte Carlo
Artificial neural network simulations demonstrate the robustness and accuracy of the established model, and training comparison
Gated recurrent unit with conventional models shows that the proposed method has higher training efficiency and better
Back propagation neural network identification performance.
Information process
© 2022 Elsevier Masson SAS. All rights reserved.

1. Introduction

As an important participant of battlefield, the aircraft has been widely used in reconnaissance, surveillance, attack, feint, electronic
interference, transportation, and other aspects [1–5]. In order to enable commanders to efficiently analyze the battlefield situation and
promote optimal decision-making, it is of great significance to recognize the tactical intention of an enemy aerial target rapidly and
accurately.
However, with the development of modern military technology, the complexity and dynamics of battlefield environment is dramatically
increasing [6], and it becomes more and more difficult for commanders to manually analyze a large number of battlefield data and extract
tactical intention in a short time. Compared with human brain, the processing speed of computer is much higher. Therefore, it is necessary
and possible to build a decision-making aided combat system to improve the efficiency of enemy aerial target intention recognition.
Intention recognition of an aerial target is a significant part of informationization warfare, requiring to obtain possible intention of a
target according to available information. There are few literatures on intention recognition of an aerial target. In [7], a hostile air-targets
intention recognition method based on Bayesian deduction is proposed. The divided lays ideas are integrated with Bayesian deduction
and thus the calculation time is reduced. In [8], a curvature feature of the target trajectory is defined for the intention recognition, and
its implementation model is established. The model is based on fuzzy inference, combining the defined feature and expert knowledge.
In [9], a decision tree model is built to classify tactical intention based on low correlation features. The support vector machine (SVM)
is employed when building the model to reduce data dimension. Unfortunately, considering the tactical intention of an aerial target is
performed through a series of actions and the target could take decoy maneuvers, it may not be credible enough for the above references
to identify the intention based on combat data at a single time.
With the development of artificial intelligence (AI), ANN has been widely used in the area of intention recognition [10–17]. Back
propagation (BP) neural network is a kind of ANN with a simple structure and is generally used in static questions due to is simple
structure and quick calculation speed [18–20]. Long short-term memory (LSTM) is another kind of ANN and is generally used to deal with

* Corresponding author at: School of Aerospace Engineering, Beijing Institute of Technology, Beijing, CO 100081, China.
E-mail addresses: [email protected] (Y. Wang), [email protected] (J. Wang), [email protected] (S. Fan), [email protected] (Y. Wang).

https://doi.org/10.1016/j.ast.2022.108005
1270-9638/© 2022 Elsevier Masson SAS. All rights reserved.
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 1. Diagram of relative position.

dynamic problems [21–26]. Compared with BP neural network, the LSTM put more emphasis on the change of input information over time.
A simplified form of LSTM, called GRU, is proposed in [27]. The training speed of GRU is increased, while the accuracy is identical [28–30].
In [6], an air target combat intention recognition model based on temporal conventional network and bidirectional gated recurrent unit
(BiGRU) is designed. The model is validated in comparison with conventional methods. Ref. [31] proposes a panoramic conventional LSTM
network to improve the recognition ability. However, the above methods don’t consider the deviation between the obtained measurement
information and true data. What’s more, the attributes of this information are different. It may not be precise enough to process all the
input data using a single kind of ANN (e.g., BP, LSTM, and GRU). Besides, only the statistical experiments are carried out in previous
studies. None of above-mentioned papers in this field have given a sample run to show the transition process of recognition model during
the measurement and calculations.
To deal with these limitations and effectively recognize the tactical intention of an aerial target, a novel quick intention recognition
model through information classified processing is proposed in this paper. Maneuvering data with large variation over time are processed
using GRU, while other data are processed using BP neural network. That is to say, the identification model proposed in this paper is
a transverse hybrid neural network instead of a traditional longitudinal hybrid neural network. The information fusion is carried on in
output layer. In other words, the identification model proposed in this paper is a transverse hybrid neural network instead of a traditional
longitudinal hybrid neural network. The information fusion is carried on in output layer. To the best of our knowledge, this is the first
paper in the area of aerial target intention recognition using transverse hybrid neural network. The transition process of a sample run is
also carried on to demonstrate the robustness of the model. Besides, the influence of batch size on training efficiency is explored.
The remainder of this paper is organized as follows. The next section presents the detailed description of the aerial target intention
recognition. Section 3 presents the structure of the intention recognition model, database establishment and calculation process of training.
Performance analysis of the proposed model is presented in Section 4, followed by the conclusions.

2. Description of aerial target intention recognition

We consider a scenario in which an enemy aerial target takes tactical action against a military objective, as shown in Fig. 1. The
velocity, flight-path angle, and azimuth angle are denoted by V , θ and ψ , respectively. The range and relative height between the aerial
target and the military architecture are denoted by R and h, respectively. The line-of-sight (LOS) angle in vertical and horizontal plane are
denoted by q y and q z , respectively.
In this paper, we divide intentions of the enemy aerial target into the following seven types:
(1) Reconnaissance & Strike: an aircraft equipped with a radar flies around the military objective, and fires air-to-ground missile.
(2) Surveillance & Search: an aircraft flies around a certain area and monitors the situation.
(3) Penetration & Direct attack: an aircraft with bomb, directly flying towards the target according to certain guidance law.
(4) Lure & Feint: an aircraft means to deceive the enemy and make the enemy’s judgment wrong.
(5) Electronic interference: an aircraft uses the carried electromagnetic interference equipment to realize electronic suppression or
communication.
(6) Transportation: an aerial vehicle transports goods on the battlefield.
(7) Others: other intentions.
We assume that the military objective is equipped with combat information sensors (e.g., a Doppler radar). Thus, the combat informa-
tion of the enemy aerial target can be measured timely. We also assume that the measurement M t , t ∈ T p are mutually independent and
can be acquired at discrete time instances t T p , where T p is a fixed measurement period. Furthermore, M t are assumed to contaminated
by zero-mean white Gaussian noise. Thus, the measurement can be expressed as:

M (t ) = [ V , θ, ψ, R , h, q z , RC S , f e , f g ] T + ν (t ) (1)
where RC S ∈ [0.1, 3] represents the radar cross section (RCS); f e ∈ {0, 1} represents the electronic-jamming status; f g ∈ {0, 1} represents
the air-to-ground radar status; and ν (t ) represents the Gaussian noise:

ν (t ) ∼ N ([0]9×1 , Q ) (2)
2 2 2 2 2 2 2
Q =diag (σ V, σθ , σψ , σ σ σ
R, h, qz , σ RC S , 0, 0) (3)
where σ represents the variances of corresponding Gaussian noise.

2
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 2. Diagram of data flow during engagement.

The diagram of data flow during engagement is shown in Fig. 2. The sensors on the enemy aerial target measure the information
of military architecture. Then the information is inputted into a onboard control system, in which the control command is generated
according to current tactical intention. Different tactical intentions lead to different combat operations. The state of enemy aerial target
varies with combat operations and is measured by combat sensors of the military architecture. These measurement information are time-
varying, bringing great uncertainty to the intention recognition [32]. After processing the obtained information, the intention is recognized
with certain method and the defense operation is carried out.
From the above analysis, we can find that there is an implicit mapping relationship between the action of aerial target and its intention,
i.e.,

γ (t ) = f ( M (t )) (4)
where γ (t ) denotes the tactical intention at time t, whose mathematical definition will be presented in subsection 3.2 and f represent
the mapping relationship.
The relationship is difficult to obtain through direct calculation. Moreover, the actual air confrontation environment is complex, and
the enemy target might take inducement operation. It is not credible enough to identify the tactical intention of the enemy aerial target
with characteristic information at a single moment [6]. Thus, the recognition basis should be the characteristic information at multiple
continuous moments. The identification basis is extended from data at a single moment in Eq. (1) to data from t 1 to t K :

I (t 1 , t K ) = [ M (t 1 ), M (t 2 ), ..., M (t K )] (5)

Then Eq. (4) can be expressed as:

γ (t ) = f ( I (t 1 , t K )) = f ( M (t ))|(t1 ,t K ) (6)
Thus, the key to solve this problem turns to find the mapping relationship between a period of available combat information of the
enemy aerial target and its operational intention.

3. Intention recognition model

In this section, we firstly present the structure of intention recognition model, in which information data with different attributes are
processed with different neural networks. Then, we present the database establishment, followed by the training process of the model.

Notation 1. For simplicity, in the remainder of this paper, we denote sigmoid, tanh, ReLU and sof tmax functions in italics, as follows:

1
sigmoid(ζ ) = (7)
1 + e −ζ
e ζ − e −ζ
tanh(ζ ) = −ζ (8)
e + eζ

ζ, ζ ≥ 0
ReLU (ζ ) = (9)
0, ζ < 0
e ζi
sof tmax(ζi ) =  ζ (10)
je
j

3.1. Structure of model

Remark 1. The GRU performs better than BP neural network in dealing with time-related issues. But the calculation formulas of the
former are more complex, and its training is more difficult than the latter. To increase the calculation speed and accuracy of intention
recognition, we use these two neural networks to process data with different attributes. Maneuvering data with large variations over time
are processed using a GRU neural network, while other data are processed using a BP neural network.

3
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 3. Diagram of a basic GRU neuron.

Fig. 4. Diagram of a basic BP neuron.

The diagram of a basic GRU neuron is shown in Fig. 3.

A basic GRU neuron consists of four parts: a reset gate (rt ), an update gate (zt ), a candidate state (
ht ) and an output state (ht ).
Calculation equations of them are as follows:

rt = sigmoid( w hr ht −1 + w xr xt ) (11)
zt = sigmoid( w hz ht −1 + w xz xt ) (12)

ht = tanh( w hh rt × ht −1 + w x
h xt ) (13)
ht = (1 − zt ) × ht −1 + zt × ht (14)
where w is the weight matrix and × is the element-wise multiplication. Note that in a GRU neural network, the input data xt is equal to
the output state (ht ) of the previous layer.
The diagram of a basic BP neuron with three inputs is shown in Fig. 4, whose calculation function is:

y = f A ( w T x + b) (15)
where b is the bias vector and f A is the activation function.
The structure of recognition model is shown in Fig. 5, where x(i , j ) is the jth dimension of input data at time i and o(i , j ) is the
possibility of jth intention at time i.
Inputs of the model are available information of the enemy aerial target in Eq. (5). The input step l is set in the input layer based on
the period of input information κ and the preset input step of intention recognition model K :

l = min( K , κ ) (16)

I 1 , I 2 , ..., I κ , κ < K
x1 , x2 , ..., xl = (17)
I κ − K +1 , I κ − K +2 , ..., I κ , κ ≥ K
Because of the difference between attributes of measurements, the range of original characteristics is different, and features with a
small range might contribute much less to the intention recognition than those with a large range [31]. To deal with this problem and
increase the convergence efficiency of model, the input data should be normalized into dimensionless values with the same range. In this
paper, the maximum-minimum standardized method is adopted as follows:

ζ  (i ) − I (i )min
ζ (i ) = (18)
I (i )max − I (i )min
where ζ (i ) is the data after processed; ζ  (i ) is the initial data; I (i )min and I (i )max are the minimum and maximum values of the ith input
data, respectively.
Meanwhile, the missing input data is fixed in the input layer using cubic spline interpolation. A sample of data patching is shown in
Fig. 6.

4
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 5. Structure of the model.

Fig. 6. Sliding Window.

Hidden layers are composed of two parts: a GRU part and a BP part. The GRU part is composed of based GRU neurons shown in Fig. 3,
used to process maneuvering data with large variation over time, i.e.,

xG (t ) = [ψ(t ), R (t ), q z (t )] T (19)

The BP part is composed of basic BP neurons shown in Fig. 4, used to process data with little variation over time, i.e.,

x B (t ) = [ V (t ), θ(t ), h(t ), RC S (t ), f e (t ), f g (t )] T (20)

Remark 2. Note that we use ( R , h, q z ) instead of ( R , q y , q z ) to represent the position of enemy aerial target, and the information of enemy
aerial target in the vertical plane (h and θ ) and the velocity of it are processed using BP neural network. This is because that the aerial
target is generally in cruise at a certain height when taking tactical operation, in which case its height h, flight-path angle θ , and velocity
V only have a small change and other maneuvering attribute variables (ψ , R, and q z ) change more with time. Thus it is reasonable to
see them approximated as constants over a period of time. As for RC S, f e , and f g , they are inherent attributes of the target, and hardly
change during a specific mission.

The activation function of input layer is generally set as tanh, sigmoid or ReLU , which processes the input data into [−1, 1], [0, 1]
and [0, ∞], respectively. In this paper, we choose tanh as the activation function of input layer to increase the nonlinear fitting ability of
model, and choose ReLU as the activation function of BP layers to increase the training efficiency.
The information fusion is carried on in the output layer:

xo ( y ) = w G o G (t ) + w B o B (t ) + bo (21)

where xo (t ) is the input of output layer; w is the weight matrix; o G (t ) is the output of GRU part; o B (t ) is the output of BP part; and bo
is the bias of output layer.
To transform human cognitive knowledge into statistical data, the tactical intentions of an enemy aerial target should be encapsulated
into labels. The tactical intention encoding and decoding relationship are shown in Fig. 7.
The possibilities of different intentions are calculated through sof tmax function:

o(t ) = sof tmax( y o (t )) = sof tmax( w o xo (t ) + bo ) (22)

5
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 7. Labels of intentions.

Fig. 8. Experimental device.

3.2. Database establishment

Considering there is no public database of enemy aerial targets, we build experimental datasets based on the aircraft control device
shown in Fig. 8.
We carry out a large number of simulations by randomly setting the initial simulation values using the mathematical simulation
software embedded in the aircraft control device. The enemy’s tactical actions are given by experts in the tactical combat field. The
inherent information of the aircraft, such as RCS, is given by using the diversified aircraft models built into the software. Then we use
sliding windows to extract information from experiments, as shown in Fig. 9. The inputs of the samples at a specific time are available
information about the enemy aerial target shown in Eq. (1). The labels are designated intention labels shown in Fig. 7, which are created
through the one-hot method.
The samples are randomly divided into two databases: a training database and a testing database. The former is used to train the
intention recognition model, and the latter is used to test the generalization ability of the trained model.

Remark 3. Note that the data used for training the model are accurate information, while the data used for testing are contaminated by
Gaussian noise. The purpose is that training with accurate data could enable the model to better find the mapping relationship between
the combat information and tactical intentions, and the contaminated data could test the robustness of trained model in a more realistic
environment.

3.3. Training process

The training process of a neural network is to adjust its trainable parameters (e.g., different types of weights in BP and GRU neural
networks) to reduce value of loss function. The initial values of these parameters are set randomly using Gaussian distribution. The model
established in this paper is divided into two parts. Both parts are trained based on back propagation. The loss function of model is cross
entropy:

6
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 9. Database establishment.


p
H (t ) = − y i log(o(t , i )) (23)
i =1
where H (t ) is the value of loss function at time t and p is the number of batch size. The value of s has great important impact on the
training efficiency. If the batch size is too small, the training process of the model may fluctuate greatly and become hard to converge. If
the batch size is too big, then the calculation speed for each iteration will decrease and thus increase the training time.
The activation function of BP neural network layers is ReLU , whose derivation is either 0 or 1. Thus, we present back propagation of
the output layer and GRU part respectively in the following content.

3.3.1. Back propagation of output layer

Notation 2. For simplification, the subscript t representing a specific time is omitted in this part.

For a weight matrix element w o (i , j ) in Eq. (22), the gradient is:

∂H
g o , w (i , j ) = (24)
∂ w (i , j )
Considering the label of a sample is one-hot, Eq. (24) can be expressed as
∂H ∂ H ∂ o(k) ∂ y (i )
g o , w (i , j ) = = (25)
∂ w (i , j ) ∂ o(k) ∂ y (i ) ∂ w (i , j )
where o(k) represents the output of true intention’s label.
Substituting Eq. (23) into (25) yields
⎧
⎪ 1
⎨ o(i )(o(i ) − 1)x( j ) = (o(i ) − 1)x( j ), i = k
g o , w (i , j ) = o ( i) (26)
⎪ 1
⎩ o(k)o(i )x( j ) = o(i )x( j ), i = k
o(k)

3.3.2. Back propagation of GRU part

Define the error gradient of a neuron’s output in the lth layer at time t as follows:
∂H
δtl = (27)
∂ hlt
7
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

where hlt is the GRU neuron’s output.

The error gradient for a certain weight matrix (taking the weight matrix w hz in Eq. (12) as an example) could be expressed as:

∂ H ∂ hlt
gt = (28)
∂ hlt ∂ w hz
where hlt is the GRU neuron’s output.
The error gradient of GRU neurons is calculated in two ways, including back propagation along time and back propagation along layers.
According to Eqs. (11) to (14), the back propagation along time and layers can be expressed as:
⎧ ⎛ ⎞⎫T
∂ H ∂ hlt ⎨ ∂ hl ∂ zl ∂ I l ∂ ht ∂ h̃lt ⎝ ∂ I h̃,t
l
∂ Il l ⎬
t t z,t h̃,t ∂ rt ⎠
δtl −1
= l l = − l l l
+ (1 − zt ) + + δtl
∂ ht ∂ ht −1 ⎩ ∂ zt ∂ I z,t ∂ ht −1 l l ∂I l
∂ h̃t h̃,t ∂ h l
t −1 ∂ r l
t ∂ h l
t −1
⎭
(29)
   T  T
= (ĥlt − hlt −1 ) ∗ zlt ∗ (1 − zlt ) w hz + (1 − zlt ) + zlt ∗ (1 − ĥlt −1 ∗ ĥlt −1 ) ∗ w hĥ rtl + w hĥ ∗ hlt −1 ∗ rtl ∗ (1 − rtl ) w hr δtl
⎧ ⎛ ⎞⎫T
∂ H ∂ hlt ⎨ ∂ hl ∂ zl ∂ I l ∂ h ∂ h̃ l ∂ Il ∂ Il ∂ r l ⎬
δtl−1 = l l−1 = − lt l t t ⎝ t ⎠
z,t t h̃,t h̃,t
+ + δtl
∂ ht ∂ ht ⎩ ∂ zt ∂ I z,t ∂ hl−1 ∂ h̃l ∂ I l ∂ hlt−1 ∂ rtl ∂ hlt−1 ⎭
t t h̃,t
⎧ ⎛ ⎞⎫T
⎨ ∂ hl ∂ zl ∂ I l ∂ h ∂ h̃ l ∂I l
∂ I ∂ rl ⎬
l
(30)
z ,t t t ⎝ h̃ ,t h̃ ,t t⎠
= − lt l t + + δtl
⎩ ∂ zt ∂ I z,t ∂ xlt ∂ h̃lt ∂ I h̃l ,t ∂ xlt ∂ rtl ∂ xlt ⎭
   T T
= (ĥlt − hlt −1 ) ∗ zlt ∗ (1 − zlt ) w xz + zlt ∗ (1 − ĥlt −1 ∗ ĥlt −1 ) ∗ w xĥ + w hĥ ∗ hlt −1 ∗ rtl ∗ (1 − rtl ) w xr δtl

where ∗ represents multiply by element.

Substituting Eqs. (11) to (14) into Eq. (28), the inner partial derivative term ∂ hlt /∂ w hz can be expressed as

∂ hlt ∂ hl ∂ zt ∂ I z,t
= t = (hˆt − ht −1 ) ∗ zt ∗ (1 − zt )htT−1 (31)
∂ w hz ∂ zt ∂ I z,t ∂ w hz
where I z,t is the input of update gate
Substituting Eqs. (29) to (31) into Eq. (28) yields the error gradient:

∂ H (t ) ∂ hlt
gt = = δtl ∗ (hˆt − ht −1 ) ∗ zt ∗ (1 − zt )htT−1 (32)
∂ hlt ∂ w hz

3.3.3. Adam algorithm

After obtaining the error gradient, the parameters of the recognition model are adjusted with Adam algorithm [33]:
η
c i +1 = c i −  mi (33)
v i +

mi β1mi −1 + (1 − β1 ) gt
mi = = (34)
1 − β1i 1 − β1i
vi β1mi −1 + (1 − β2 ) gt2
v i = = (35)
1 − β2i 1 − β2i

where c is the trainable parameter; i is the training iteration; η is the learning rate; v i is the bias-corrected second raw moment estimate;
is the smoothing term in case of v i = 0; mi is the bias-corrected first moment estimate; mi is the biased first moment estimate; v i is
the biased second raw moment estimate; β1 , β2 ∈ [0, 1) are exponential rates of moment estimates; β1i , β2i are i power of β1 and β2 ,
respectively.

4. Performance analysis

In this section, the effectiveness of the enemy aerial target intention recognition model is verified by numerical simulations. We firstly
present the training iteration process of the model, followed by a sample run. Then, the influence of measurement noise, data missing,
and batch size on the performance of the established model are demonstrated. Finally, the training comparison between a GRU-based
model, a BP-based model and the proposed model is carried out. The simulations are carried on based on TensorFlow-1.13.0. The graphic
card is GeForce RTX 3090 24G. The type of central processing unit is Intel(R) Xeon(R) Gold 6248R CPU, and the SAM is 256GB.

Remark 4. Note that for a fair comparison, the training data of a BP network also includes the sliding window to include the historical
measurement data.

8
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 10. Training iteration process. (For interpretation of the colors in the figure(s), the reader is referred to the web version of this article.)

Fig. 11. Sample run.

4.1. Training result

The experimental samples are extracted using the database establishing method presented in Section 3.2. The total number of samples
is 14,400 including 12,000 samples used for training the model and 2,400 samples used for testing the generalization ability. The number
of samples for each intention except the label ‘others’ is 2000. The measurement rate of the radar is f m = 20 H z. The number of hidden
layers is 3, with each hidden layer containing 128 basic GRU or BP neurons. The number of training iterations is 1,000. The batch size
for each training iteration is 512. The preset set step of the established model is K = 12. The learning rate is 0.002, with a decaying
rate setting as 0.95/100 iterations. The Adam algorithm parameters are properly chosen as = 10−8 , η1 = 0.9 and η2 = 0.999. To prevent
overfitting of the trained model, dropout regularization is adopted in this paper, with the dropout rate set to 50%.
The accuracy and loss of the proposed model during the training process is presented in Fig. 10. As can be seen from the figure, the
convergence reaches around 800 iterations. The accuracy and loss after trained are 98.74% and 0.0488, respectively.
Then we use the testing database to test the performance of model in a more realistic environment. The inputs of testing samples
are contaminated by Gaussian noise, as shown in Eqs. (1) to (3). The measurement noises are σ V2 = (5 m/s)2 , σθ2 = σψ2 = (1 mrad)2 ,
σ R2 = σh2 = (5 m)2 , σq2z = (1 mrad2 )2 , and σ RC
2
S = (0.1 m) , respectively. The accuracy and loss under the noise condition are 95.66% and
2

0.52548, respectively. The results validate the accuracy and robustness of the model, and demonstrate the application value in auxiliary
decision-making.
To intuitively show the performance of the proposed model, a sample run is carried on in this subsection before turning to statistical
Monte Carlo simulations. Fig. 11 presents the posterior probability of each intention being true during the engagement. The inputs of the
samples used for the sample run are contaminated by the noise mentioned before. Note that after approximately 0.4 s the Reconnaissance
& Strike has been identified as the tactical intention of the enemy aerial target, and the recognition result remains stable during the
following simulation, which proves the rapidness and stability of the established model.

4.2. Monte Carlo simulations

In this section, Monte Carlo simulations are carried on to explore the influence of training batch size, measurement noise level, and
missing data on the performance of model.
The influence of batch size on the training result is presented in Fig. 12, where the successful training rate is the number of cases in
which the accuracy of trained model is higher than 95% divided by the total number of simulations. For each batch size condition, we
perform 20 simulations and calculate the successful training rate respectively. To deeply analyze the influence of batch size, the training
iteration is increased to 2000.
As can be seen from Fig. 12, with the increase of training iteration the successful training rate gets higher. It also becomes easier for
the model to be successfully trained with the increase of batch size. Note that at the beginning, the successful training rate increases

9
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 12. Influence of batch size.

Fig. 13. Performance with regard to noise.

quickly. This is because that the value of loss function is the mean of samples in current iteration, as shown in Eq. (23). When the batch
size is not large enough, the gradient may not be the optimal training direction for the whole question and thus the training efficiency is
low. With the increase of batch size, the samples in a specific train iteration could better represent the whole sample space and thus the
training efficiency gets higher. When the batch size reaches approximately 512, the successful training rate increases much slower. The
reason is that the batch size is large enough, and the training direction is close to the optimal descending direction. Thus the increase of
batch size doesn’t contribute much to the improvement of training efficiency.
We then explore the influence of measurement noise on the performance of the model. The noise level is defined as:

M (t ) = [ V , θ, ψ, R , h, q z , RC S , f e , f g ] T + ν (t ) L N (36)
where ν (t ) is presented in Eqs. (2) and (3); and L N is the noise level.
The simulation result is presented in Fig. 13. It can be found that with the increase of measurement noise, the intention recognition
accuracy gets lower and the loss gets higher. A extra experiment is carried out to explore the influence of a certain variable. We take
the measurement noise of RCS as an example. The result is shown in Fig. 14. As can be seen from the figure, the identification accuracy
decreases slightly with the increase of the variance of RCS. The phenomenon is similar to Fig. 13. We can also find that the decreasing rate
in Fig. 14 is lower than that in Fig. 13. This is because, in Fig. 14, only the measurement variance of RCS is the test variable, while other
measurement noise does not change. Note that the above simulation results, while not practical in the sense of Gaussian noise, can be
improved using other scenarios while real data are not available. Meanwhile, this phenomenon proves the practicability of the proposed
model when the noise level is unknown.
The accuracy and loss with regard to the number of missing steps are presented in Fig. 15. The trend of the curve is similar to Fig. 13,
i.e., with the increase of data missing, the performance of model deteriorates. Note that although the identification effect decreases, the
results could still be used as a reference for decision-making to a certain extent, and Fig. 15 also validates the data patching method
shown in Fig. 6.

4.3. Comparison

The successful training rate comparison between a GRU-based model, a BP-based model, and the proposed model is carried on in this
subsection to demonstrate the advantages of the proposed model. The simulation results are shown in Fig. 16. As can be seen from the
figure, the training efficiency of the proposed model is significantly better than the other two models. This is because the proposed model
is built through information classification processing, combining the respective advantages of GRU and BP neural networks, as shown in
Section 3.1. We can also find that the BP-based model performs poorly in the current problem, proving it is difficult for a BP neural
network to solve a time-related problem.

10
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 14. Performance with regard to noise of RCS.

Fig. 15. Performance with regard to miss data.

Fig. 16. Comparison between the proposed model with conventional models.

Note that although the training efficiency is much lower than the proposed model, the GRU-based model could also be successfully
trained in some cases, especially when the batch size and iteration number are high enough. This phenomenon can be explained by the
fact that as a widely used recurrent neural network, GRU has good generalization ability in dealing with time-related problems through
the complex internal structure shown in Fig. 3. However, the training result of the GRU-based model is sensitive to initial conditions
and training method, and its training efficiency is much lower than the proposed model. This is because that the error gradient of the
GRU-based model needs a long back propagation to reach the input layer, as shown in Section 3.3. This can better extract the information

11
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

Fig. 16. (continued)

12
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

from the input data with large variation over time. But as for data that changes little with time, such back propagation might decrease
the training efficiency. In this paper, we use BP neural network to process these data, and set ReLU , whose derivative is 0 or 1, as the
activation of BP neurons. In this way, the error gradient could be quickly back propagated to the input layer without complex calculation
and processing, so as to increase the training efficiency of the proposed model.

5. Conclusion

To enable combat commanders to efficiently analyze the battlefield situation and promote optimal decision-making, a quick intention
recognition model for an incoming aerial target based on ANN is established through information classification processing in this paper.
Inputs of the model are available information of the enemy aircraft, and outputs are possibilities of different intentions. To increase
the training efficiency and accuracy, the inputting information is processed using different neural network. Maneuvering data with large
variations over time are processed using GRU, while other data are processed using BP neural network. The training result validates the
rapidness and accuracy of the established model. Meanwhile, the influence of batch size on the training efficiency is explored. We find
that the successful training rate gets higher with the increase of batch size. Besides, the performance of the established model under the
effect of measurement noise and missing data are probed. Simulation results prove the robustness of the model.
The comparison of training efficiency between a GRU-based model, a BP-based model, and the proposed model are carried out. We
find that the training efficiency of the proposed model is highest because of information classification processing method proposed in this
paper, which could combine the respective advantages of GRU and BP neural networks. We also find that the GRU-based model could
be successfully trained in some cases, especially when the batch size and iteration number are high enough, but its training efficiency is
much lower than the proposed model. This is because the error gradient in the proposed model could be quickly back propagated to the
input layer, so as to increase the training speed.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:
Shipeng Fan reports financial support was provided by the Airforce Advance Research Project of China (No. 3030209).

Data availability

No data was used for the research described in the article.

Acknowledgement

This study was supported by the Airforce Advance Research Project of China (No. 3030209).

References

[1] G. Goh, S. Agarwala, G. Goh, V. Dikshit, S. Sing, W. Yeong, Additive manufacturing in unmanned aerial vehicles (uavs): challenges and potential, Aerosp. Sci. Technol. 63
(2017) 140–151.
[2] J.W. Woo, J.-Y. An, M.G. Cho, C.-J. Kim, Integration of path planning, trajectory generation and trajectory tracking control for aircraft mission autonomy, Aerosp. Sci.
Technol. 118 (2021) 107014.
[3] P.D. Bravo-Mosquera, A.M. Abdalla, H.D. Cerón-Muñoz, F.M. Catalano, Integration assessment of conceptual design and intake aerodynamics of a non-conventional air-to-
ground fighter aircraft, Aerosp. Sci. Technol. 86 (2019) 497–519.
[4] D. Xu, Z. Hui, Y. Liu, G. Chen, Morphing control of a new bionic morphing uav with deep reinforcement learning, Aerosp. Sci. Technol. 92 (2019) 232–243.
[5] L. Wang, H. Yin, Y. Guo, T. Yue, X. Jia, Closed-loop motion characteristic requirements of receiver aircraft for probe and drogue aerial refueling, Aerosp. Sci. Technol. 93
(2019) 105293.
[6] F. Teng, Y. Song, X. Guo, Attention-tcn-bigru: an air target combat intention recognition model, Mathematics 9 (19) (2021).
[7] L. Tang, L. Fuxian, Z. Feng, G. Dong, Hierarchical recognition method of hostile air- targets in sea battlefields based on bayesian deduction, Fire Control Command Control
43 (2020) 44–48.
[8] J. Xu, L. Zhang, D. Han, Air target intention recognition based on fuzzy inference, Command Inf. Syst. Technol. 11 (2020) 86–93.
[9] M. Guanglei, Z. Runnan, W. Biao, Z. Mingzhe, W. Yu, L. Xiao, Target tactical intention recognition in multiaircraft cooperative air combat, Int. J. Aerosp. Eng. 2021 (2021).
[10] W. Zhou, P. Yao, J. Zhang, X. Wang, S. Wei, Combat intention recognition for aerial targets based on deep neural network, Acta Aeronaut. Astronaut. Sin. 39 (011) (2018)
195–203.
[11] F. Koochaki, L. Najafizadeh, A data-driven framework for intention prediction via eye movement with applications to assistive systems, IEEE Trans. Neural Syst. Rehabil.
Eng. (2021).
[12] Z. Fang, A.M. López, Intention recognition of pedestrians and cyclists by 2d pose estimation, IEEE Trans. Intell. Transp. Syst. 21 (11) (2019) 4773–4783.
[13] A. Zyner, S. Worrall, E. Nebot, A recurrent neural network solution for predicting driver intention at unsignalized intersections, IEEE Robot. Autom. Lett. 3 (3) (2018)
1759–1764.
[14] M. Huang, Y. Xu, L. Qian, W. Shi, Y. Zhang, W. Bao, N. Wang, X. Liu, X. Xiang, A bridge neural network-based optical-sar image joint intelligent interpretation framework,
Space Sci. Technol. 2021 (2021).
[15] J. Chen, W. Li, S. Li, H. Chen, X. Zhao, J. Peng, Y. Chen, H. Deng, Two-stage solar flare forecasting based on convolutional neural networks, Space Sci. Technol. 2022 (2022).
[16] Z. Ye, Q. Zhou, Performance evaluation indicators of space dynamic networks under broadcast mechanism, Space Sci. Technol. 2021 (2021).
[17] B. Wang, S. Li, J. Mu, X. Hao, W. Zhu, J. Hu, Research advancements in key technologies for space-based situational awareness, Space Sci. Technol. 2022 (2022).
[18] Q. Meng, M. Huang, Y. Xu, N. Liu, X. Xiang, Decentralized distributed deep learning with low-bandwidth consumption for smart constellations, Space Sci. Technol. 2021
(2021).
[19] M. Mirsadeghi, M. Shalchian, S.R. Kheradpisheh, T. Masquelier, Stidi-bp: spike time displacement based error backpropagation in multilayer spiking neural networks,
Neurocomputing 427 (2021) 131–140.
[20] J. Kim, D. Kwon, S.Y. Woo, W.-M. Kang, S. Lee, S. Oh, C.-H. Kim, J.-H. Bae, B.-G. Park, J.-H. Lee, Hardware-based spiking neural network architecture using simplified
backpropagation algorithm and homeostasis functionality, Neurocomputing 428 (2021) 153–165.
[21] Y. Sun, Z. Xie, Y. Chen, Q. Hu, Accurate solar wind speed prediction with multimodality information, Space Sci. Technol. 2022 (2022).
[22] W. Li, X. Gao, H. Liu, Efficient prediction of transonic flutter boundaries for varying Mach number and angle of attack via lstm network, Aerosp. Sci. Technol. 110 (2021)
106451.

13
Y. Wang, J. Wang, S. Fan et al. Aerospace Science and Technology 132 (2023) 108005

[23] C. Che, H. Wang, Q. Fu, X. Ni, Combining multiple deep learning algorithms for prognostic and health management of aircraft, Aerosp. Sci. Technol. 94 (2019) 105423.
[24] K. Li, J. Kou, W. Zhang, Unsteady aerodynamic reduced-order modeling based on machine learning across multiple airfoils, Aerosp. Sci. Technol. 119 (2021) 107173.
[25] S. Allison, H. Bai, B. Jayaraman, Wind estimation using quadcopter motion: a machine learning approach, Aerosp. Sci. Technol. 98 (2020) 105699.
[26] T. Wu, H. Wang, Y. Yu, Y. Liu, J. Wu, Hierarchical fault-tolerant control for over-actuated hypersonic reentry vehicles, Aerosp. Sci. Technol. 119 (2021) 107134.
[27] K. Cho, B. van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, Y. Bengio, Learning phrase representations using RNN encoder–decoder for statistical
machine translation, in: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), Association for Computational Linguistics,
Doha, Qatar, Oct. 2014, pp. 1724–1734 [Online]. Available: https://aclanthology.org/D14-1179.
[28] Y. Wang, S. Fan, G. Wu, J. Wang, S. He, Fast guidance law identification approach for incoming missile based on GRU network, Acta Aeronaut. Astronaut. Sin. (2021).
[29] Y. Wang, S. Fan, J. Wang, G. Wu, Quick identification of guidance law for an incoming missile using multiple-model mechanism, Chin. J. Aeronaut. (2021) [Online].
Available: https://doi.org/10.1016/j.cja.2021.10.032.
[30] L. Tao, T. Zhang, D. Peng, J. Hao, Y. Jia, C. Lu, Y. Ding, L. Ma, Long-term degradation prediction and assessment with heteroscedasticity telemetry data based on gru-garch
and md hybrid method: an application for satellite, Aerosp. Sci. Technol. 115 (2021) 106826.
[31] J. Xue, J. Zhu, J. Xiao, S. Tong, L. Huang, Panoramic convolutional long short-term memory networks for combat intension recognition of aerial targets, IEEE Access 8
(2020) 183 312–183 323.
[32] Z. Sun, H. Piao, Z. Yang, Y. Zhao, G. Zhan, D. Zhou, G. Meng, H. Chen, X. Chen, B. Qu, et al., Multi-agent hierarchical policy gradient for air combat tactics emergence via
self-play, Eng. Appl. Artif. Intell. 98 (2021) 104112.
[33] D.P. Kingma, J. Ba, Adam: a method for stochastic optimization, arXiv preprint, arXiv:1412.6980, 2014.

14