Professional Documents
Culture Documents
68
ICNSER 2022, May 27, 28, 2022, Nanjing, China Jihui Xu et al.
69
Aerial Target Threat Estimation in Unmanned Aerial Vehicle Reconnaissance Based on Neural Network ICNSER 2022, May 27, 28, 2022, Nanjing, China
1.0H ≥ 5000
TH = 0.5 + 0.0001H − 5000 ≤ H < 5000
(8)
0.1H < −5000
Figure 2: Trajectory of the air combat data
4) Distance threat index TR :
When calculating the distance threat index of target aircraft, it
is necessary to consider both the enemy airborne radar detection
distance and the air-to-air missile range. Set Rr F and RrT as the
maximum detection distance of the airborne radar of the UAV and
the target aircraft respectively. RmF and RmT are the maximum
range of the air-to-air missile of the UAV and the target aircraft
respectively. Reconnaissance UAVs do not normally carry missiles,
so the launch range of the UAV can be viewed as 0. Then, there are
2 situations to be discussed:
a) RrT , Rr F , RmT > RmF ,
0R ≥ RrT
0.4 − 0.4 • RR−R
R ≤ R < RrT
rF
rT −R r F r F
R−RmT
TR = 1 − 0.6 • Rr F −RmT RmT ≤ R < Rr F
(9) Figure 3: Target threat estimation model
R−RmF
0.5 + 0.5 • RmT −RmF RmF ≤ R < RmT
0.5 + 0.3 • RmF −R R < R
RmF mF
70
ICNSER 2022, May 27, 28, 2022, Nanjing, China Jihui Xu et al.
3 SIMULATION EXPERIMENTS
3.1 Experiment Settings
The estimation model is used to simulate air combat target threat
estimation, and 336 groups of sample data are divided into training
samples and test samples. Among them, the first 236 groups of
data are used as training samples, and the remaining 100 groups of
data are used as test samples. In order to avoid the prediction error
caused by the wide range of values of various parameters, all data
are normalized.
According to the input and output dimensions of the model, it
can be determined that the number of input neurons of the network Figure 4: Test results of the hidden layer neurons.
is 6, and the number of output neurons is 1. In addition, in order
to obtain a predictor with good accuracy, it is necessary to select
the number of hidden neurons of the network. For this reason, 100 is relatively good. Therefore, when training the ELM predictor, the
groups of test samples are divided into two parts, which are set as number of hidden neurons is set as L = 18.
test sample 1 and test sample 2, and the number of samples is 50.
The test sample 1 is used to conduct test experiments to determine 3.3 Accuracy Analysis of Estimation Model
the parameter L; the test sample 2 is used to conduct performance
After the network structure is determined, test sample 2 is used to
test on the finally obtained predictor.
test the estimation accuracy of the BP neural network and the ELM
The simulation experiment is conducted on a PC and the oper-
network, respectively. Among them, the parameters of BP network
ating environment is: Intel(R) Core(TM) i7-4590 3.3GHz processor
are set as follows, the target training error is 0.001, the maximum
with 8GB of memory. Win7 32-bit operating system with a running
number of iterations is 2000, the number of hidden neurons is set
platform of MATLAB 2010a. In order to make the experiment more
as 28 by the test experiment. The estimation results are shown in
convincing, the following simulation results are the average value
Fig. 5.
of 30 calculations.
It can be seen from Fig. 5 that the prediction error of the ELM
network is smaller than that of the BP network, and the prediction
3.2 Setting the Number of Hidden Neurons accuracy of the ELM network is significantly improved and the
The number of hidden neurons of the ELM neural network has a threat estimation can be performed more accurately.
great influence on the prediction accuracy. If the number is too
small, the ELM network cannot learn well and the prediction error 3.4 Complexity Analysis of Estimation Model
is large. If the number is too large, the training time of the network The execution time is calculated to compare the computational
increases, and the phenomenon of over-fitting easily occurs [24].In complexity of different algorithms. The results are shown in Fig. 6.
practical applications, the optimal number of hidden neurons is It can be seen from Fig. 6 that under a certain times of estimation,
usually determined by test experiment [25]. According to the Kol- the execution time of each algorithm is ranked by the ELM network
mogorov Theorem [26], for a single hidden layer neural network, <threat index method <BP neural network. This is because the two
if the number of input neurons is n, the number of hidden neurons neural network algorithms can realize the calculation of "situation
should be set as at least 2n + 1. In order to take into account the parameters-threat value" after training. Compared with the calcula-
network performance and training costs, the test experiment is tion of "situation parameters-threat index-threat value", it simplifies
used to find the optimal number of hidden neurons of the ELM the calculation process. Moreover, the training of the ELM network
network in the interval of [13, 30]. can be completed once without iteration, so the execution time is
In the test experiment, the network initial input layer weight w the shortest. While more iterative training is needed to correct the
and the hidden layer threshold b are set randomly. The activation network weights in the BP neural network, so the execution time
function д(x) is set by default, i.e.д(x) = 1+e1 −x . Then, the training is longer.
samples are used to train the network with different number of
hidden neurons, and the test sample 1 is predicted and the predicted
mean square error (MSE) is output. The computational formula of
4 CONCLUSIONS
MSE is: A new threat estimation method for UAV reconnaissance based on
N1
the ELM neural network is proposed aiming at the shortcomings
1 Õ 2 of the existing methods. The ELM neural network is proposed to
MSE = t j − t¯j (12)
N 1 i=1 use as the predictor to perform target threat estimation, and the
simulation results show that the algorithm has high estimation
where N 1 is the number of test samples, t j and t¯j are the actual accuracy and low computational complexity.
output and calculation output of the i − th sample respectively. The The shortcomings of this paper are that the training samples of
test results are shown in Fig. 4. the model are not rich enough, and only the threat estimation of
From Fig. 4, it can be seen that in this interval, when the number one UAV vs one target aircraft is studied. At present, UAV swarm
of hidden neurons is 18, the prediction accuracy of the ELM network operation is the mainstream of UAV reconnaissance field, in the
71
Aerial Target Threat Estimation in Unmanned Aerial Vehicle Reconnaissance Based on Neural Network ICNSER 2022, May 27, 28, 2022, Nanjing, China
future, the threat estimation for swarm detection of multiple targets 1102–1109, 2003.
[10] Qu, C. W. and He, Y. “A method of threat estimation using multiple attribute
by UAVs needs to be further studied. decision making.” Proceedings of the Sixth International Conference on Signal
Processing Proceedings, 1091–1095, 2002.
REFERENCES [11] Liang, Q. and Cheng, X. “Knowledge-based ubiquitous and persistent sensor
networks for threat estimation.” IEEE Transactions on Aerospace and Electronic
[1] Roux, J. N. and Vuuren J. H. V. “Threat evaluation and weapon assignment
Systems, 44(3), 1060-1069, 2008.
decision support: a review of the state of the art”. Journal of the Operations
[12] Wang, J. and Jiang, C. S. “Target threat estimation based on LSRBF neural network
Research Society of South Africa, 23(2), 151–187, 2017.
for air combat.” Electronics Optics & Control, 14(4), 43-48, 2007.
[2] Hinman, M. L. “Situation estimation and impact estimation activities in infor-
[13] Qiu, L. B., Liu, Z.L. and Liu, M. “A threat estimation algorithm by using the
mation fusion.” Proc. of the SPIE Sensor Fusion: Architectures, Algorithms, and
neural network techniques.” Journal of Air Force Engineering University: Natural
Applications IV, 351–361, 2020.
Science Edition, 3(6), 25-28, 2002.
[3] Huma, N. and Asif, M. “An optimal dynamic threat evaluation and weapon
[14] Wang, X. H., Qin, Z. and Liu, Y., et al. “RBF neural network for threat sequencing.”
scheduling technique.” Knowledge-Based Systems, 23(3), 337–342, 2020.
Journal of System Simulation, 16(7), 1576-1579, 2004.
[4] Baxter, L., Simpson, R., Prebola, J., et al. “Spiral 1 space threat estimation testbed
[15] Guo, H., Xu, H. J. and Liu, L. “Target threat estimation of air combat based
development.” ITEA Journal, 30(3), 345-353, 2019.
on support vector machines for regression”. Journal of Beijing University of
[5] Zhang, H, X., Kang B, Y. and Li, Y., et al. “Target threat estimation based on
Aeronautics and Astronautics, 36(1), 123-126, 2010.
interval data fusion.” Journal of Computational Information Systems, 6(8), 2609-
[16] Huang, G. B., Zhu, Q. Y. and Siew, C. K. “Extreme learning machine: Theory and
2616, 2012.
applications.” Neurocomputing, 70, 489-501, 2006.
[6] Azimirad, E. and Haddadnia, J. “Target threat estimation using fuzzy sets theory.”
[17] Zuo, J. L., Yang, R. N. and Zhang, Y., et al. “Reconstruction and evaluation of close
International Journal of Advances in Intelligent Informatics, Vol 1, No 2, 57-74,
air combat decision making process based on fuzzy clustering.” Acta Aeronautica
2015.
et Astro nautica Sinica, 36(5), 1650-1660, 2015.
[7] Chen, D, F., Feng, Y. and Liu, Y, X. “Threat estimation for air defense operations
[18] Huang, G. B., Zhu, Q. Y. and Siew, C. K. “Extreme learning machine: a new
based on intuitionistic fuzzy logic.” Procedia Engineering, 29, 3302-3306, 2012.
learning scheme of feedford neural networks.” Proceedings of International Joint
[8] Ma, S. D., Zhang H. Z. and Yang, G. Q. “Target threat level estimation based
Conference on Neural Networks, 25-29 July, Budapest, Hungary, 2004.
on cloud model under fuzzy and uncertain conditions in air ombat simulation.”
[19] Luo, X., Chang, X. H. and Ban, X.-J. “Regression and classification using extreme
Aerospace Science and Technology, 67, 49-53, 2017.
learning machine based on L1-norm and L2-norm.” Neurocomputing, 174,179-186,
[9] Okello, N. and Thoms, G. “Threat estimation using Bayesian net-works.” Pro-
2016.
ceedings of the Sixth International Conference on Information Fusion. IEEE,
72
ICNSER 2022, May 27, 28, 2022, Nanjing, China Jihui Xu et al.
[20] Gu, J. J. and Liu,W.H. “WVR air combat situation estimation model based on [24] Guo, R. F., Huang, G. B. and Lin, Q.-P., et al. “Error minimized extreme learn-
weapon engagement zone and kill probability.” Systems Engineering and Elec- ing machine with growth of hidden neurons and incremental learning.” IEEE
tronics, 37(6), 1306-1312, 2015. Transactions on Neural Networks, 20(8), 1352-1357, 2009.
[21] Choi, J.W., Joo, J.W. and L, D. “Situation/threat estimation fusion system.” Pro- [25] Gao, D.W., Wang, P. and Cai, Z.C. “Optimization of hidden neurons and training
ceedings of the Fifth International Conference on Information Fusion. IEEE, times in artificial neural network.” Journal of Harbin Institute of Technology,
1374–1380, 2002. 35(2), 207-209, 2003.
[22] Gross,G., Nagi, R. and Sambhoos, K. “Continuous preservation of situational [26] Kolmogorov, A. N. “On the representation of continuous functions of several
awareness through incremental/stochastic graphical methods.” Proceedings of variables by superposition of continuous functions of one variable and addition.”
the 14th International Conference on Information Fusion. IEEE, 1030–1037, 2011. Doklady Akademiinauk USSR, 114(5), 953–956, 1957.
[23] Wang, W. and Zhang, Y. Z. “Method of obtaining eigenvector for a fuzzy AHP.”
Control and Decision, 21(2), 184-187, 2006.
73