Underwater Acoustic Multi-target Recognition Algorithm ... - IEEE Xplore
Underwater Acoustic Multi-target Recognition Algorithm Based on Hierarchical Information Fusion Structure Liang YU1,2,Yong-mei CHENG1, Lin SONG1, Zhun-ga LIU1, Ke-zhe Chen3 1. Department of Automatic control, Northwestern Polytechnical University 2. Engineering Practice and Training Center, Northwestern Polytechnical University Xi’an, Shaanxi 710072, China E-mail:[email protected] 3. The twentieth Research Institute of China Electronic Technology Group Corporation Xi’an, Shaanxi 710068, China Abstract— In the complex hydroacoustic countermeasure environment, submarine often release a variety of decoys to disrupt the attack. Because the current electronic technology is able to produce high quality decoys, it is very difficult to correctly recognize the real submarine by the spectral analysis method. During the tracking process, a hierarchical multi-source information fusion structure is designed based on evidence theory for identification of the submarine. In the beginning, the observed acoustic data are divided into two types: target acoustic feature data and motion feature data. The sources of information obtained from different features in the same type are fused respectively at the first fusion level. The two pieces of fusion results are weighted combined at the second level. The sequential feature information (i.e. the multi-temporal fusion results at the second level) in the time domain will be selected according to the current context to enter the fusion process at the third level for the final identification of targets. Simulation results show that the proposed method can effectively identify acoustic targets. Keywords—Acoustic Counter-countermeasuer; Multi-target Recognition; Acoustic feature; Motion feature; Hierarchical Information Fusion Structure
I.
INTRODUCTION
Underwater, in the process of torpedo tracking and attack submarine, submarine generally first found torpedo [1], and will release in the types of soft killing weapons, such as suspended and self-propelled or other decoys, to induce the torpedo deception. These decoys can simulate submarine echo signal and conduct movements according to the set routes, navigational speed, maneuvering mode, making the torpedo to produce the misjudgment so as to miss real targets. Therefore, the underwater acoustic counter-countermeasures technology generated. Underwater acoustic counter-countermeasure technology demands urgently to be able to recognize the decoys and real submarine accurately to realize the effective tracking and precision attack. In response to this problem, most of current studies focused on the identification by acoustic features, from the perspective of time domain feature, frequency-domain characterization, the time-frequency domain features [2-4],
such as light spots, scale, timbre, and so on to identify the target type. A series of methods such as support vector machines combined with genetic algorithm [5], genetic algorithm based fuzzy logic [6], model matching method [7], and adaptive immune feature selection algorithm [8] had been proposed for analysis the spectral data. However, it should be noted that, it is difficult to extract original underwater acoustic target samples, there exist high redundancy rate in the target characteristic extraction process. Furthermore, the development of modern electronic technology enables the decoy could simulate completed almost all of the time domain and frequency domain characteristics of the submarine, Using the time-frequency information only, self-guided system of the torpedo is unable to timely and correctly identify the target and decoy [5]. Therefore, It is difficult to distinguish between the target and decoy accurately just depend on a single post processing signal, which need a higher level, that is, to analyze of multiple acoustic properties of the target for comprehensive sort recognition. And now there is no published literature was researched in this aspect. A similar literature [9] use gray correlation method (GROD), Analytic Hierarchy Process (AHP), the ideal solution (TOPSIS) to choose the most threatening target of UUV. GROD is more effective than other algorithms in this optimum seeking problem, but the method just from two types, cost and benefit type, to get the target probability. However, in the countermeasure environment, models of various properties of true and false target cannot simply come from a large or a small simple measurement; it should be modeled from real degree to the target. To solve these problems, optimization theory combined with acoustic features and motion feature information of targets is introduced during the track, and a hierarchical multitarget recognition algorithm is implemented based on Evidence theory. Acoustic features and motion features are fused respectively at the first level of the fusion. And at the second level of the fusion, weighted integration, which got by sensor characteristic was done on the acoustic features result and the motion result. Third level fusion builds on sequence feature information, which means that the result of the second
level should do an adaptive integration in the time domain. The paper is organized as follows. In Section 2, the general rules of evidence theory are described. In Section 3, hierarchical multi-target recognition framework is given; a new method for automatic determination of discounting factors is presented. A numerical example is given in Section 4. Section 5 concludes this paper. II.
BASICS OF BELIEF FUNCTION THEORY
Belief function theory is also called Dempster–Shafer theory (DST), or evidence theory. Let us consider Θ being a finite discrete set of mutually exclusive and exhaustive hypotheses. Θ is called the frame of discernment of the problem under consideration. The power set of Θ , 2Θ , contains all the subsets of Θ , and R is one power set of the framework 2Θ , if the collection mapping function : R → [0,1], satisfy the following conditions
used to build as the basic confidence function. According to the idea of Evidence Theory, all the attribute values should be normalized respectively, to achieve the quantification of all features. And at the first level of the fusion structure, all acoustic features will be fused and all motion features be fused respectively. At different distance and orientation, the credibility of the motion feature and acoustic feature are changed. So, the fusion results of the first level will be weighted merged at the second level. After that, It can be judge that, whether the difference between largest and second largest probability result of all targets which get by second level smaller than the threshold ε , if the difference bigger that ε , use the second level results as the final results, otherwise, fusion in the time dimension, that is, fusion at the second level and the previous time results will be done to get the final probability results. The algorithm structure is shown in Fig.1.
(1)
Motion features got by tracking
Acoustic features got by spectrum analysis
The subsets A of Θ such that m( A) > 0 are called the focal elements of m(⋅) , and the set of all its focal elements is called the core of m(⋅) .
Built feature belief assignment function
Built feature belief assignment function
Normalizing each attribute
Normalizing each attribute
Fusion between motion feature
Fusion between acoustic feature
⎧⎪m(∅) = 0 ⎨ Θ ⎪⎩∑ {m( A) | A ⊆ 2 } = 1
Assume that there are two different properties evidence B and C on the frame of Θ . Bi and C j is the focal element , and their basic confidence function are m1 and m2 respectively. Mathematically, use the DS combination, m( A) is defined by ⎧ ⎪ 1 ⎪ m( A) = ⎪ ⎪ K ⎨ ⎪ ⎪ ⎪ ⎪ ⎩m(∅) = 0
∑
m1 ( Bi )m2 (C j ), A ≠ ∅, A ⊆ Θ
Bi ∩ C j = A
Where, K = 1−
(2)
Weighted fusion between two types m2 ( t )
∑
m1 ( Ai )m2 ( B j ) .
max( m2 ( t )) − sub(m2 (t ) ) < ε i
i
Y
III.
Fusion of the 2nd level
i
Ai ∩ B j =φ
HIERARCHICAL INFORMATION FUSION ALGORITHM
Hierarchical information fusion algorithm structure for underwater acoustic target recognition is shown in Fig.1. During the targets tracking process, the real time speed, direction, et al, motion information can be got according to the filtering. Radiated noise, target strength, et al, real time acoustic properties information can be got according to the time and frequency domain analysis, and furthermore, bearing and other information of the target can be estimated. The above data that obtained by filtering or spectrum analysis are referred measured value. The real feature value at same time is called as theoretical value; of course, the theoretical model of the feature can be built to verify the correctness of the algorithm. Firstly, the feature theoretical models of the submarine and decoy have been built in section A as below. Then, compute the differences between theoretical value and the measured value of various attributes, the differences are
Fusion of the 1st level
fusion between this and last time
N Fusion of the 3rd level
Final result m (t )
Fig.1. Hierarchical information fusion algorithm structure for underwater acoustic target recognition
A. Target feature models In literature [2-4], many differences, such as target strength, radiated noise, highlights distribution exist between submarine and decoy, they can be chosen as the acoustic features. During the submarine escape process, there exist corresponding escape strategies and elapsed time, angle of chord which that can be helpful to submarine maximum extent dodge or escape homing sector of the torpedo. So they can be
used as the motion feature. In the paper, two acoustic features and two motion features have been given. (1) Target strength (acoustic feature A) The target strength of submarine changing with the sound wave incident angle, a large number of experimental data analysis are found that the target strength of the submarine changing with bearing type distribution and likes a "butterfly"[10], it is shown in figure 2. As the acoustic decoy simulating the target strength, the transducer of it should have wide emission directivity while ensuring the distribution of the target strength like the "butterfly", that means the target strength can change with the angle of incidence of the sound wave. In the paper, it is thought very difficult, and under normal circumstances, the largest target strength that acoustic decoy can simulate is no more than 20 dB, and it is assumed as a constant distribution. The target strength of the submarine and decoy can be calculated by formula (3), (4) respectly.
Fig.3. The radiated noise directivity simulated by submarine and decoy
The radiated noise of the submarine and the decoy is calculated respectively, according to the formula (5) and (6).
SLF 2
⎪⎧25log Vm + 77, Vm ≤ 12 (5) SLF1 = ⎪⎨ ⎪⎪⎩104 + (1.5 ~ 2)(Vm −12), Vm > 12 ⎪⎧SL , π / 6 ≤ α ≤ 5π / 6; -5π / 6 ≤ α ≤ -π / 6 = ⎪⎨ F1 (6) ⎪⎪⎩0, -π / 6＜α＜π / 6;-5π＜α ≤ -π;5π / 6＜α ≤ π;
Where, Vm is the speed (kn) of the target.
Fig.2. The submarine's target strength
TS F 1 = TS0 (16.17 − 2.98cos 2α − 3.083cos 6α ) / 22.233 (3) π
TS F 2 =
1 TS F 1 (α )d α π ∫0
(4)
Where, α is the incident wave angle of the chord, TS 0 is the beam reflections, the value is taken as 25 dB generally, TS F 1 and TS F 2 are the intensity of the target of submarine and decoy. (2) The directivity of the radiated noise (acoustic feature B) When the submarine size is much smaller than the distance between the submarine and the torpedo, it is generally believed that the radiated noise of the submarine is similar isotropic. That is to say that the radiated noise of submarinelaunched directivity is circular external radiation without considering the clutter interference [11]. In fact, there is a distribution of radiated noise influenced by engine noise [1], the influence is ignored in this paper. In establishing the model, the noise radiation directivity of decoy simulated is determined by the directivity of the transmitting transducer. It is assumed that the decoy emission directivity is along the longitudinal axis decoy left and right sides of the horizontal direction of 120 °, and there is no radiation noise in other directions (radiation noise blind area). The submarine radiation noise and radiated noise directivity simulated by decoy is shown in figure 3.
(3) The elapsed time (motion feature C) When submarine found torpedo, submarine escapes torpedo detection sector as far as possible in order to avoid the torpedo homing detection. While the decoy should stay in torpedo sector as much as possible to lure the torpedo [1]. So it can be thought that longer escaping sector goals, the more likely it is decoy. The opposite is more likely it is a submarine. Therefore, the time of target escaped sector is made as a motion feature to determine the properties of a genuine target. It is assumed that the current location of the torpedo is ( xd , yd ) ,the course of the torpedo is ψd , the size of torpedo detection sector is δ ,the current position of the target is ( xt , yt ) , the course of the target is ψt , and the speed is vt .It can be seen in figure 4. y
tmax ( xt , yt ) T
δ
Vt
ψt
C ( xc , yc )
tmin
ψ
D D x ( xd , yd ) Fig.4. The calculating diagram of elapsed time
The border slope of the torpedo sectors is shown as follows
k d = tan(ψd ± δ 2)
(7)
The vertical intercept can be computed as follows
bd = yd − k d ⋅ xd
(8)
The slope of the target course line is expressed as
kt = tan(ψt )
y
α min
(9)
The vertical intercept is got by
bt = yt − kt ⋅ xt
(10)
When the target moving to the sector boundary, the intersection of sector boundary and the course line is shown as follows ⎧⎪ ⎪⎪ xc = bd − bt ⎪⎨ kt − k d ⎪⎪ ⎪⎪⎩ yc = kd ⋅ xc + bd
(11)
( yc − yt ) 2 + ( xc − xt )2 vt
Vt
ψt
α max D ( xd , yd )
θ
x
Fig.5. The target angle of chord
The angle between the connection of the torpedo and target and the x axis direction is shown as
And the elapsed time can be got as follows ts =
α
( xt , yt ) T
(12)
Using the method, the maximum and minimum value of the time to escape from sectors can be calculated. The maximum elapsed time ( tmax )can be got when the target velocity direction is away from the connection to the torpedo and target direction, it is considered as the theoretical escape time of the submarine. Also, the minimum elapsed time ( tmin ) can be got when the target velocity direction is vertical to torpedo recent sector direction, it is considered as the theoretical escape time of the decoy. (4) The target angle of chord (motion feature D) In the process of the fight and the confrontation underwater, according to the classical confrontation strategy [12] ，submarine releases decoy to interfere torpedo, and will go away from the torpedo. As shown in figure 5, under the condition of the direction of the movement is only considered, the shortest direction of the encounter time between targets and the torpedo is most likely to be the decoys. Assume that in the period of time of Δ t , the torpedo moving distance v ⋅ Δt , it is easy to draw that the most likely to be the direction of the decoy is when the target direction opposites to the direction that the line of torpedo to target. Whereas for the submarine. The angle between the line of the torpedo to target and the target velocity direction is defined as the angle of the chord, expressed with α . The larger the angle of chord is, the smaller probability of the submarine is, and the greater probability of the decoy is. On the basis of this, the target is relative to torpedo angle of chord as the second motion feature for modeling.
θ = arctan(
yt − y d ) xt − xd
(13)
The target angle of chord is shown as follows α = θ − ψt
(14)
Where, α , θ , ψt ∈ [−180,180] .And the minimum target angle of chord αmin and maximum angle of the target string αmax can be calculated as 0 ° and 180 ° respectively,and which stands the theoretical angle of chord of the submarine and decoy respectly. B. Fusion of the 1st level Assume that the real target exists during the tracking process, and the submarines and decoy are the object that needs to be distinguished. So, framework of identification Θ = {submarine ( F1 ), decoy ( F2 )} can be defined. When all feature data at the time t of the targets were got, compute the probability of all targets belong to the framework of identification. Time t , the measured value of a probability of a target is ki (t ), k = 1, 2,3...n . Following the tracking results, we can get the direction and position between the torpedo and targets, so, F the theory value ki j (t ) at this direction and position can be got, and the distance measure that the target belongs to the submarine and decoy are d kFi (t ) = ki (t ) − kiFi (t ) , i = 1, 2; k = 1, 2, 3,...n
(15)
Compute the basic belief assignment of all targets: m kt ( Fi ) = (1 − η kt ) exp( − d kFi (t ))
(16)
Where η k is the magnitude of the error of the k ‘s feature. And the uncertain belief assignment can be built as
mtk (Θ) = η kt ⋅ ∏ (1 − mtk ( Fi )) i
(17)
Following the definition of the belief assignment function, all of the features should be normalized as follows:
⎧mtk ( Fi ) = mtk ( Fi ) (mkt (Θ) + ∑ mkt ( Fi )) ⎪ i ⎨ t t t m ( Θ ) = m ( Θ ) ( m ( Θ ) + mkt ( Fi )) ∑ ⎪ k k k i ⎩
(18)
∑
mk ( A)ml ( B)
A∩ B = Fi
(19)
Where, A and B is the core of evidence respectively, and Fi ≠ ∅, Fi ⊆ Θ, K kl = 1− ∑ mk ( A) ml ( B ) . A ∩ B =∅
There are just two kinds of acoustic feature and two kinds of motion feature in the article, so acoustic feature and motion feature can be fused respectively as follows: 1 ⎧ ⎪ m12 ( Fi ) = K ⎨ 12 ⎪ m(∅ ) = 0 ⎩
A∩ B = Fi
1 ⎧ ⎪ m34 ( Fi ) = K 34 ⎨ ⎪ m(∅ ) = 0 ⎩
C ∩ D = Fi
∑
∑
m1 ( A)m2 ( B )
(20) m3 (C )m4 ( D )
Where, A , B , C and D are the core of evidence respectively, m1 , m2 are the belief assignment of the acoustic feature, m3 , m4 are the belief assignment of the motion feature. And,
⎧⎪ K12 = 1− ∑ m1 ( A)m2 ( B) ⎪⎪ A ∩ B =∅ ⎪⎨ . ⎪⎪ K 34 = 1− ∑ m3 (C )m4 ( D) ⎪⎪ C ∩ D =∅ ⎩ C. Fusion of the 2nd level The fusion of the 2nd level is to do the fuse between two feature types. As the detection performance and precision of the sonar is different at different distances and direction. It is difficult to obtain precise values of acoustical properties at far distance, as the distance getting closer, the precision will be better by degrees. Therefore, there exists a confliction between acoustic properties and the final target probability. So at the 2nd fusion level, when acoustic feature and motion feature are fused, weighting factor ω can be set for correcting the fusion process between the two type prosperity
∑
mut ( A)mvt ( B ) +
A∩ B = Fi
∑
So, all probabilities of the two type features can be got. All weights between features when they were fused should base on their reliability. Assume the weights between the two features is K kl , integrate various attributes between each two by Dempster's rule: 1 ⎧ t ⎪ mkl ( Fi ) = K ⎨ kl ⎪ m(∅) = 0 ⎩
muvt ( Fi ) =
[ω (t ) ⋅mut ( Fi )mvt ( B) + (1 − ω )mut ( A)mvt ( Fi )]
(21)
Fi ∩ A =∅ , Fi ∩ B =∅
Where, ω is related to the distance, direction, et al, which can be determined by the self-guided detection performance testing, mut and mvt are the acoustic feature and motion feature fusion result that got at the 1st level. rd
D. Fusion of the 3 level The fusion of the 3rd level is to do the fuse in the time domain. Using the time accumulates to determine the type of the target, the right fusion result will be more clearly got, fusion results will be enhanced, and there will be a smooth probability change of the targets, avoiding the mutation of the probability. The third level fusion base on the results of the 2nd
t t level. If max(mkl ( F1 )) − sub( mkl ( F1 )) < ε , we thought that
the targets were can’t be distinguished at the 2nd level, the result should strengthen to solve the problem, the time dimension fusion should be done: ⎧ mklt ( Fi ) = ∑ mkt −1 ( A)mlt ( B ) ⎪ A ∩ B = Fi (22) ⎨ ⎪⎩ m(∅) = 0 The above formula means that the result of this time and last time has the same weights. Otherwise, the fusion results of the 2nd level will be the final fusion results. The threshold ε is decided by the tracking demand and sensor error performance. IV.
SIMULATION EXPERIMENT
A. Static discrete validation Assume that the torpedo homing sector is 120 °and the detection range is 3000 meters. Submarine along the sector line moves at a constant speed. Random discrete points generated around submarine within the sector are on behalf of decoy. And the speed of random discrete points is the same as the submarine. Submarine with decoy against a phase are randomly generated. 100 discrete Monte Carlo simulations have been performed on the condition of different distance decoy data. By the fusion of the first and second level, the basic belief assignment result of that the target is the submarine is shown in figure 6, it gives out the result of the probability that the targets, decoy and submarine, is a submarine. And below, the probability of the decoy or submarine means the probability that it belongs to the submarine.
1500
1 submarine decoy
0.95
Submarine Decoy
1000
0.9
500
0.85
y/m
Probablity
0
0.8 0.75
-500
-1000
0.7 0.65
-1500
0.6
-2000
0.55
0
100
200
300
400 500 600 Distance/m
700
800
900
-2500
1000
0
500
1000
1500 x/m
2000
2500
3000
Fig.6. The probability changed with the distance
B. Dynamic environment validation The performance of the torpedo is same as section A in this part. According to the classical confrontation strategy [12], the acoustic warfare environment has been established. The motion and acoustic features of the decoy and submarine are built by III (A) in this paper. The torpedo is at the origin of the coordinate position, and the velocity direction of the torpedo is along the horizontal forward. For a real maneuvering submarine, The initial time as the beam orientation of matrix, with a distance of 1500m from the torpedo, the angle between the initial direction of motion and the horizontal is 85°, at a speed of 12kn, and then to circumvent the maximum speed of 20kn, three decoys move at the same speed with an angle to the negative X axis, respectively, are -45 °, -85 °, -135 °, the velocities are same to submarine. The target moving posture is shown in Fig.7.
Fig.7. Movement posture of the targets
Results are shown in Figure 8, 9 and 10 by using the method for target recognition. 0.9 0.8 0.7
Probability
0.6 0.5 0.4 0.3 0.2 0.1 0
Submarine Decoy 0
10
20
30
40
50 Time/t
60
70
80
90
100
Fig.8. Targets probability computed by acoustic features 0.9 0.8 0.7 0.6 Probability
From the result, It can be seen that the probability of the decoy is always less than the submarine, the probability is defined as the target is a submarine. This shows that the algorithm is right in theory, the algorithm can identify the real target correctly. Also, when targets move further, the probability of the targets belongs to submarine declined, for example, the maximum probability is near 1, and the least probability in the figure of the decoy and submarine at 1000m are 0.58 and 0.66, that means that when the probability is declined when the sensor accuracy is lower at a further distance, and it is in line with the actual situation.
0.5 0.4 0.3 0.2 0.1 0
Submarine Decoy 0
10
20
30
40
50 Time/t
60
70
80
90
100
Fig.9. Targets probability computed by acoustic and motion features
considering the motion feature and without time dimension fusion.
0.9
0.7
1
0.6
0.9
0.5 0.8 0.4 Probability
Probability
0.8
0.3 0.2 0.1 0
Submarine Decoy 0
10
20
30
40
50 Time/t
60
70
80
90
0.7
0.6
0.5 100
0.4 Submarine Decoy
Fig.10.Targets probability computed by hierarchical fusion method
Figure 8 is the probability that the target is submarine when regardless the target motion properties and without time dimension fusion. As can be seen, at first, the real targets can be rightly recognized before the simulation time of 70, after that, false target has a greater probability. The reason of that is clear, the closer distance of the two decoys to the torpedo, the more elevated probability is. Conversely, the submarine becomes further, with the distance increased, the farther the submarine moves, the sensor performance is lower. The uncertainty is higher, the probability is less. So, it can be seen that, under the condition without considering the motion feature, the correctness of the target probability will be decreased to some extent with the distance increased. Figure 9 is the probability that the target is a submarine using two types’ features without time dimension fusion. Compared with Fig. 8 ， the correctness of the target probability is being extended, the wrong probability appeared after the time of 80. It can be seen that with the motion feature added, the submarine can be correctly estimated in a certain degree, but it cannot guarantee the result is correct in the whole simulation process. Given the threshold ε = 0.1 , the target probability that the target belongs to submarine computed by the hierarchical fusion method is shown in Fig.10. It can be seen throughout the whole simulation process, the probability of the submarine is the biggest and real target has been correctly calculated. It shows the proposed method can be well applied to underwater acoustic multi-objective optimum seeking. At the time between 40 and 50，due to changes in the orientation of the targets, the acoustic features change and it make the final fusion results changed, with respect to Fig.8 and Fig.9, fluctuated changes style appeared in Fig.10, which is caused by the time dimension fusion. Time dimension fusion can make the probability stabilized and make the affection by one feature change minimized. Compared to the related work us the same probability of each attribute, the result computed by GROD method is shown in figure 11. It can be seen that when the target becomes farther away from the torpedo, that is, after the time 58, the calculation results don’t match the actual situation. So, even if the GROD method can solve the greatest threat problem of the UUV. But in solving the true and false targets optimum seeking issue, the correct rate of the method is far below the method in this paper, even it below the method without
0
10
20
30
40
50 Time/t
60
70
80
90
100
Fig.11. Targets probability computed by GROD method
V.
CONCLUSION
A hierarchical fusion method has been proposed to solve the problem that true and false of underwater acoustic target are difficult to identify, which need combining with acoustic information and motion information for recognition under tracking progress. Firstly, the feature models are built, then, how to build the membership functions of the feature attribute is given, after that, evidence theory for underwater acoustic target recognition algorithm implementation process has been proposed. Three level fusion structures have been done on the same type of features, and then in two types, finally on time dimension respectively. Experiment results show that this method can effectively identify the true and false targets, can significantly improve the probability of correctly identify the target just based on acoustic feature information, and is more effective than the traditional GROD method to deal with the problem. It provides a new idea and method for underwater acoustic multi- target recognition. Next, this method should be tested in more complex countermeasure environments, and also the movement of the torpedo should be considered to verify the reliability of the algorithm. Acknowledgment
This research was financially supported by national natural science foundation of China (61135001), the Xi'an (China) Science and Technology Project in 2014 and 2013 (CXY1350 (2)).
References [1]
[2]
[3]
Jung-Min Pak, Dong-Gi Woo, Bonhwa Ku, et al. Target Search Method for a Torpedo to the Evading Ship Using Fuzzy Inference, ICROS-SICE International Joint Conference, Japan, 2009, pp.5279-5284. Wang Mingzhou, Huang Xiaowen, Hao Chongyang. “Model of an Underwater Target Based on Target Echo Highlight Structure. Journal of System Simulation”, vol. 15, pp. 21-25. 2003,. Wang Mingzhou, Hao Chongyang, Huang Xiaowen. “On Underwater Target Dimension Recognition Based on Bearings Analysis of Signal
[4]
[5]
[6]
[7]
Correlation Feature”. Journal of Northwestern Polytechnical University, vol. 21 , pp.317-320, 2003. Chen Kean, Wang Na, Wu Ying, et al. “Subjective evaluation experiments of timbre attribute based acoustic target identification”. Chinese Sci Bull, vol. 55 , pp. 651-659, 2010. Yang Honghui, Dai Jian, SUN Jincai. “A New Adaptive Immune Feature Selection Algorithm for Underwater Acoustic Target Classification”. Journal of Xi’an Jiao Tong University, vol. 45, pp. 2831, 2011. Chu-Kuei Tu, Chung Yuan, Chung Li, Tseng-Hsien Lin. Applying genetic algorithms on fuzzy logic system for underwater acoustic signal recognition, Underwater Technology, 2000. Proceedings of the 2000 International Symposium on, pp: 405-410. TUCKER S, BROWN G J. “Classification of transient sonar sounds using perceptually motivated features”. IEEE Journal of Ocean Engineering, vol. 30 , pp.588-600, 2005.
[8]
J. Acoust, Soc. Am. “Perception-based automatic classification of impulsive-source active sonar echoes”. Journal of the Acoustical Society of America, vol. 122 , pp. 1502-1517, 2007. [9] Wang Baihe，Huang Jianguo，Zhang Qunfei. “Model and Algorithm for Attacked Target Optimization in Underwater Multi-targets Tracing System”. Torpedo Technology, vol. 15 , pp. 22-25, 2007. [10] Meng Qingyu, Zhang Jingyuan, Song Baowei. Analysis of Operational Effectiveness for Torpedo. Beijing: National Defense Industry Press, 2003. [11] Xiao Juan, Gu Hao, Cao Zhimin. “Modeling and Simulation of Acoustic Decoy”. Computer Simulation, vol. 25 , pp. 25-27, 2008. [12] Ko-Hsin Liang, Kuei-Ming Wang. Using Simulation And Evolutionary Algorithms to Evaluate the Design of Mix Strategies of Decoy and Jammers in Anti-Torpedo Tactics. Proceedings of the 2006 Winter Simulation Conference. Monterey, CA: IEEE. pp. 1299-1306.
I.
INTRODUCTION
Underwater, in the process of torpedo tracking and attack submarine, submarine generally first found torpedo [1], and will release in the types of soft killing weapons, such as suspended and self-propelled or other decoys, to induce the torpedo deception. These decoys can simulate submarine echo signal and conduct movements according to the set routes, navigational speed, maneuvering mode, making the torpedo to produce the misjudgment so as to miss real targets. Therefore, the underwater acoustic counter-countermeasures technology generated. Underwater acoustic counter-countermeasure technology demands urgently to be able to recognize the decoys and real submarine accurately to realize the effective tracking and precision attack. In response to this problem, most of current studies focused on the identification by acoustic features, from the perspective of time domain feature, frequency-domain characterization, the time-frequency domain features [2-4],
such as light spots, scale, timbre, and so on to identify the target type. A series of methods such as support vector machines combined with genetic algorithm [5], genetic algorithm based fuzzy logic [6], model matching method [7], and adaptive immune feature selection algorithm [8] had been proposed for analysis the spectral data. However, it should be noted that, it is difficult to extract original underwater acoustic target samples, there exist high redundancy rate in the target characteristic extraction process. Furthermore, the development of modern electronic technology enables the decoy could simulate completed almost all of the time domain and frequency domain characteristics of the submarine, Using the time-frequency information only, self-guided system of the torpedo is unable to timely and correctly identify the target and decoy [5]. Therefore, It is difficult to distinguish between the target and decoy accurately just depend on a single post processing signal, which need a higher level, that is, to analyze of multiple acoustic properties of the target for comprehensive sort recognition. And now there is no published literature was researched in this aspect. A similar literature [9] use gray correlation method (GROD), Analytic Hierarchy Process (AHP), the ideal solution (TOPSIS) to choose the most threatening target of UUV. GROD is more effective than other algorithms in this optimum seeking problem, but the method just from two types, cost and benefit type, to get the target probability. However, in the countermeasure environment, models of various properties of true and false target cannot simply come from a large or a small simple measurement; it should be modeled from real degree to the target. To solve these problems, optimization theory combined with acoustic features and motion feature information of targets is introduced during the track, and a hierarchical multitarget recognition algorithm is implemented based on Evidence theory. Acoustic features and motion features are fused respectively at the first level of the fusion. And at the second level of the fusion, weighted integration, which got by sensor characteristic was done on the acoustic features result and the motion result. Third level fusion builds on sequence feature information, which means that the result of the second
level should do an adaptive integration in the time domain. The paper is organized as follows. In Section 2, the general rules of evidence theory are described. In Section 3, hierarchical multi-target recognition framework is given; a new method for automatic determination of discounting factors is presented. A numerical example is given in Section 4. Section 5 concludes this paper. II.
BASICS OF BELIEF FUNCTION THEORY
Belief function theory is also called Dempster–Shafer theory (DST), or evidence theory. Let us consider Θ being a finite discrete set of mutually exclusive and exhaustive hypotheses. Θ is called the frame of discernment of the problem under consideration. The power set of Θ , 2Θ , contains all the subsets of Θ , and R is one power set of the framework 2Θ , if the collection mapping function : R → [0,1], satisfy the following conditions
used to build as the basic confidence function. According to the idea of Evidence Theory, all the attribute values should be normalized respectively, to achieve the quantification of all features. And at the first level of the fusion structure, all acoustic features will be fused and all motion features be fused respectively. At different distance and orientation, the credibility of the motion feature and acoustic feature are changed. So, the fusion results of the first level will be weighted merged at the second level. After that, It can be judge that, whether the difference between largest and second largest probability result of all targets which get by second level smaller than the threshold ε , if the difference bigger that ε , use the second level results as the final results, otherwise, fusion in the time dimension, that is, fusion at the second level and the previous time results will be done to get the final probability results. The algorithm structure is shown in Fig.1.
(1)
Motion features got by tracking
Acoustic features got by spectrum analysis
The subsets A of Θ such that m( A) > 0 are called the focal elements of m(⋅) , and the set of all its focal elements is called the core of m(⋅) .
Built feature belief assignment function
Built feature belief assignment function
Normalizing each attribute
Normalizing each attribute
Fusion between motion feature
Fusion between acoustic feature
⎧⎪m(∅) = 0 ⎨ Θ ⎪⎩∑ {m( A) | A ⊆ 2 } = 1
Assume that there are two different properties evidence B and C on the frame of Θ . Bi and C j is the focal element , and their basic confidence function are m1 and m2 respectively. Mathematically, use the DS combination, m( A) is defined by ⎧ ⎪ 1 ⎪ m( A) = ⎪ ⎪ K ⎨ ⎪ ⎪ ⎪ ⎪ ⎩m(∅) = 0
∑
m1 ( Bi )m2 (C j ), A ≠ ∅, A ⊆ Θ
Bi ∩ C j = A
Where, K = 1−
(2)
Weighted fusion between two types m2 ( t )
∑
m1 ( Ai )m2 ( B j ) .
max( m2 ( t )) − sub(m2 (t ) ) < ε i
i
Y
III.
Fusion of the 2nd level
i
Ai ∩ B j =φ
HIERARCHICAL INFORMATION FUSION ALGORITHM
Hierarchical information fusion algorithm structure for underwater acoustic target recognition is shown in Fig.1. During the targets tracking process, the real time speed, direction, et al, motion information can be got according to the filtering. Radiated noise, target strength, et al, real time acoustic properties information can be got according to the time and frequency domain analysis, and furthermore, bearing and other information of the target can be estimated. The above data that obtained by filtering or spectrum analysis are referred measured value. The real feature value at same time is called as theoretical value; of course, the theoretical model of the feature can be built to verify the correctness of the algorithm. Firstly, the feature theoretical models of the submarine and decoy have been built in section A as below. Then, compute the differences between theoretical value and the measured value of various attributes, the differences are
Fusion of the 1st level
fusion between this and last time
N Fusion of the 3rd level
Final result m (t )
Fig.1. Hierarchical information fusion algorithm structure for underwater acoustic target recognition
A. Target feature models In literature [2-4], many differences, such as target strength, radiated noise, highlights distribution exist between submarine and decoy, they can be chosen as the acoustic features. During the submarine escape process, there exist corresponding escape strategies and elapsed time, angle of chord which that can be helpful to submarine maximum extent dodge or escape homing sector of the torpedo. So they can be
used as the motion feature. In the paper, two acoustic features and two motion features have been given. (1) Target strength (acoustic feature A) The target strength of submarine changing with the sound wave incident angle, a large number of experimental data analysis are found that the target strength of the submarine changing with bearing type distribution and likes a "butterfly"[10], it is shown in figure 2. As the acoustic decoy simulating the target strength, the transducer of it should have wide emission directivity while ensuring the distribution of the target strength like the "butterfly", that means the target strength can change with the angle of incidence of the sound wave. In the paper, it is thought very difficult, and under normal circumstances, the largest target strength that acoustic decoy can simulate is no more than 20 dB, and it is assumed as a constant distribution. The target strength of the submarine and decoy can be calculated by formula (3), (4) respectly.
Fig.3. The radiated noise directivity simulated by submarine and decoy
The radiated noise of the submarine and the decoy is calculated respectively, according to the formula (5) and (6).
SLF 2
⎪⎧25log Vm + 77, Vm ≤ 12 (5) SLF1 = ⎪⎨ ⎪⎪⎩104 + (1.5 ~ 2)(Vm −12), Vm > 12 ⎪⎧SL , π / 6 ≤ α ≤ 5π / 6; -5π / 6 ≤ α ≤ -π / 6 = ⎪⎨ F1 (6) ⎪⎪⎩0, -π / 6＜α＜π / 6;-5π＜α ≤ -π;5π / 6＜α ≤ π;
Where, Vm is the speed (kn) of the target.
Fig.2. The submarine's target strength
TS F 1 = TS0 (16.17 − 2.98cos 2α − 3.083cos 6α ) / 22.233 (3) π
TS F 2 =
1 TS F 1 (α )d α π ∫0
(4)
Where, α is the incident wave angle of the chord, TS 0 is the beam reflections, the value is taken as 25 dB generally, TS F 1 and TS F 2 are the intensity of the target of submarine and decoy. (2) The directivity of the radiated noise (acoustic feature B) When the submarine size is much smaller than the distance between the submarine and the torpedo, it is generally believed that the radiated noise of the submarine is similar isotropic. That is to say that the radiated noise of submarinelaunched directivity is circular external radiation without considering the clutter interference [11]. In fact, there is a distribution of radiated noise influenced by engine noise [1], the influence is ignored in this paper. In establishing the model, the noise radiation directivity of decoy simulated is determined by the directivity of the transmitting transducer. It is assumed that the decoy emission directivity is along the longitudinal axis decoy left and right sides of the horizontal direction of 120 °, and there is no radiation noise in other directions (radiation noise blind area). The submarine radiation noise and radiated noise directivity simulated by decoy is shown in figure 3.
(3) The elapsed time (motion feature C) When submarine found torpedo, submarine escapes torpedo detection sector as far as possible in order to avoid the torpedo homing detection. While the decoy should stay in torpedo sector as much as possible to lure the torpedo [1]. So it can be thought that longer escaping sector goals, the more likely it is decoy. The opposite is more likely it is a submarine. Therefore, the time of target escaped sector is made as a motion feature to determine the properties of a genuine target. It is assumed that the current location of the torpedo is ( xd , yd ) ,the course of the torpedo is ψd , the size of torpedo detection sector is δ ,the current position of the target is ( xt , yt ) , the course of the target is ψt , and the speed is vt .It can be seen in figure 4. y
tmax ( xt , yt ) T
δ
Vt
ψt
C ( xc , yc )
tmin
ψ
D D x ( xd , yd ) Fig.4. The calculating diagram of elapsed time
The border slope of the torpedo sectors is shown as follows
k d = tan(ψd ± δ 2)
(7)
The vertical intercept can be computed as follows
bd = yd − k d ⋅ xd
(8)
The slope of the target course line is expressed as
kt = tan(ψt )
y
α min
(9)
The vertical intercept is got by
bt = yt − kt ⋅ xt
(10)
When the target moving to the sector boundary, the intersection of sector boundary and the course line is shown as follows ⎧⎪ ⎪⎪ xc = bd − bt ⎪⎨ kt − k d ⎪⎪ ⎪⎪⎩ yc = kd ⋅ xc + bd
(11)
( yc − yt ) 2 + ( xc − xt )2 vt
Vt
ψt
α max D ( xd , yd )
θ
x
Fig.5. The target angle of chord
The angle between the connection of the torpedo and target and the x axis direction is shown as
And the elapsed time can be got as follows ts =
α
( xt , yt ) T
(12)
Using the method, the maximum and minimum value of the time to escape from sectors can be calculated. The maximum elapsed time ( tmax )can be got when the target velocity direction is away from the connection to the torpedo and target direction, it is considered as the theoretical escape time of the submarine. Also, the minimum elapsed time ( tmin ) can be got when the target velocity direction is vertical to torpedo recent sector direction, it is considered as the theoretical escape time of the decoy. (4) The target angle of chord (motion feature D) In the process of the fight and the confrontation underwater, according to the classical confrontation strategy [12] ，submarine releases decoy to interfere torpedo, and will go away from the torpedo. As shown in figure 5, under the condition of the direction of the movement is only considered, the shortest direction of the encounter time between targets and the torpedo is most likely to be the decoys. Assume that in the period of time of Δ t , the torpedo moving distance v ⋅ Δt , it is easy to draw that the most likely to be the direction of the decoy is when the target direction opposites to the direction that the line of torpedo to target. Whereas for the submarine. The angle between the line of the torpedo to target and the target velocity direction is defined as the angle of the chord, expressed with α . The larger the angle of chord is, the smaller probability of the submarine is, and the greater probability of the decoy is. On the basis of this, the target is relative to torpedo angle of chord as the second motion feature for modeling.
θ = arctan(
yt − y d ) xt − xd
(13)
The target angle of chord is shown as follows α = θ − ψt
(14)
Where, α , θ , ψt ∈ [−180,180] .And the minimum target angle of chord αmin and maximum angle of the target string αmax can be calculated as 0 ° and 180 ° respectively,and which stands the theoretical angle of chord of the submarine and decoy respectly. B. Fusion of the 1st level Assume that the real target exists during the tracking process, and the submarines and decoy are the object that needs to be distinguished. So, framework of identification Θ = {submarine ( F1 ), decoy ( F2 )} can be defined. When all feature data at the time t of the targets were got, compute the probability of all targets belong to the framework of identification. Time t , the measured value of a probability of a target is ki (t ), k = 1, 2,3...n . Following the tracking results, we can get the direction and position between the torpedo and targets, so, F the theory value ki j (t ) at this direction and position can be got, and the distance measure that the target belongs to the submarine and decoy are d kFi (t ) = ki (t ) − kiFi (t ) , i = 1, 2; k = 1, 2, 3,...n
(15)
Compute the basic belief assignment of all targets: m kt ( Fi ) = (1 − η kt ) exp( − d kFi (t ))
(16)
Where η k is the magnitude of the error of the k ‘s feature. And the uncertain belief assignment can be built as
mtk (Θ) = η kt ⋅ ∏ (1 − mtk ( Fi )) i
(17)
Following the definition of the belief assignment function, all of the features should be normalized as follows:
⎧mtk ( Fi ) = mtk ( Fi ) (mkt (Θ) + ∑ mkt ( Fi )) ⎪ i ⎨ t t t m ( Θ ) = m ( Θ ) ( m ( Θ ) + mkt ( Fi )) ∑ ⎪ k k k i ⎩
(18)
∑
mk ( A)ml ( B)
A∩ B = Fi
(19)
Where, A and B is the core of evidence respectively, and Fi ≠ ∅, Fi ⊆ Θ, K kl = 1− ∑ mk ( A) ml ( B ) . A ∩ B =∅
There are just two kinds of acoustic feature and two kinds of motion feature in the article, so acoustic feature and motion feature can be fused respectively as follows: 1 ⎧ ⎪ m12 ( Fi ) = K ⎨ 12 ⎪ m(∅ ) = 0 ⎩
A∩ B = Fi
1 ⎧ ⎪ m34 ( Fi ) = K 34 ⎨ ⎪ m(∅ ) = 0 ⎩
C ∩ D = Fi
∑
∑
m1 ( A)m2 ( B )
(20) m3 (C )m4 ( D )
Where, A , B , C and D are the core of evidence respectively, m1 , m2 are the belief assignment of the acoustic feature, m3 , m4 are the belief assignment of the motion feature. And,
⎧⎪ K12 = 1− ∑ m1 ( A)m2 ( B) ⎪⎪ A ∩ B =∅ ⎪⎨ . ⎪⎪ K 34 = 1− ∑ m3 (C )m4 ( D) ⎪⎪ C ∩ D =∅ ⎩ C. Fusion of the 2nd level The fusion of the 2nd level is to do the fuse between two feature types. As the detection performance and precision of the sonar is different at different distances and direction. It is difficult to obtain precise values of acoustical properties at far distance, as the distance getting closer, the precision will be better by degrees. Therefore, there exists a confliction between acoustic properties and the final target probability. So at the 2nd fusion level, when acoustic feature and motion feature are fused, weighting factor ω can be set for correcting the fusion process between the two type prosperity
∑
mut ( A)mvt ( B ) +
A∩ B = Fi
∑
So, all probabilities of the two type features can be got. All weights between features when they were fused should base on their reliability. Assume the weights between the two features is K kl , integrate various attributes between each two by Dempster's rule: 1 ⎧ t ⎪ mkl ( Fi ) = K ⎨ kl ⎪ m(∅) = 0 ⎩
muvt ( Fi ) =
[ω (t ) ⋅mut ( Fi )mvt ( B) + (1 − ω )mut ( A)mvt ( Fi )]
(21)
Fi ∩ A =∅ , Fi ∩ B =∅
Where, ω is related to the distance, direction, et al, which can be determined by the self-guided detection performance testing, mut and mvt are the acoustic feature and motion feature fusion result that got at the 1st level. rd
D. Fusion of the 3 level The fusion of the 3rd level is to do the fuse in the time domain. Using the time accumulates to determine the type of the target, the right fusion result will be more clearly got, fusion results will be enhanced, and there will be a smooth probability change of the targets, avoiding the mutation of the probability. The third level fusion base on the results of the 2nd
t t level. If max(mkl ( F1 )) − sub( mkl ( F1 )) < ε , we thought that
the targets were can’t be distinguished at the 2nd level, the result should strengthen to solve the problem, the time dimension fusion should be done: ⎧ mklt ( Fi ) = ∑ mkt −1 ( A)mlt ( B ) ⎪ A ∩ B = Fi (22) ⎨ ⎪⎩ m(∅) = 0 The above formula means that the result of this time and last time has the same weights. Otherwise, the fusion results of the 2nd level will be the final fusion results. The threshold ε is decided by the tracking demand and sensor error performance. IV.
SIMULATION EXPERIMENT
A. Static discrete validation Assume that the torpedo homing sector is 120 °and the detection range is 3000 meters. Submarine along the sector line moves at a constant speed. Random discrete points generated around submarine within the sector are on behalf of decoy. And the speed of random discrete points is the same as the submarine. Submarine with decoy against a phase are randomly generated. 100 discrete Monte Carlo simulations have been performed on the condition of different distance decoy data. By the fusion of the first and second level, the basic belief assignment result of that the target is the submarine is shown in figure 6, it gives out the result of the probability that the targets, decoy and submarine, is a submarine. And below, the probability of the decoy or submarine means the probability that it belongs to the submarine.
1500
1 submarine decoy
0.95
Submarine Decoy
1000
0.9
500
0.85
y/m
Probablity
0
0.8 0.75
-500
-1000
0.7 0.65
-1500
0.6
-2000
0.55
0
100
200
300
400 500 600 Distance/m
700
800
900
-2500
1000
0
500
1000
1500 x/m
2000
2500
3000
Fig.6. The probability changed with the distance
B. Dynamic environment validation The performance of the torpedo is same as section A in this part. According to the classical confrontation strategy [12], the acoustic warfare environment has been established. The motion and acoustic features of the decoy and submarine are built by III (A) in this paper. The torpedo is at the origin of the coordinate position, and the velocity direction of the torpedo is along the horizontal forward. For a real maneuvering submarine, The initial time as the beam orientation of matrix, with a distance of 1500m from the torpedo, the angle between the initial direction of motion and the horizontal is 85°, at a speed of 12kn, and then to circumvent the maximum speed of 20kn, three decoys move at the same speed with an angle to the negative X axis, respectively, are -45 °, -85 °, -135 °, the velocities are same to submarine. The target moving posture is shown in Fig.7.
Fig.7. Movement posture of the targets
Results are shown in Figure 8, 9 and 10 by using the method for target recognition. 0.9 0.8 0.7
Probability
0.6 0.5 0.4 0.3 0.2 0.1 0
Submarine Decoy 0
10
20
30
40
50 Time/t
60
70
80
90
100
Fig.8. Targets probability computed by acoustic features 0.9 0.8 0.7 0.6 Probability
From the result, It can be seen that the probability of the decoy is always less than the submarine, the probability is defined as the target is a submarine. This shows that the algorithm is right in theory, the algorithm can identify the real target correctly. Also, when targets move further, the probability of the targets belongs to submarine declined, for example, the maximum probability is near 1, and the least probability in the figure of the decoy and submarine at 1000m are 0.58 and 0.66, that means that when the probability is declined when the sensor accuracy is lower at a further distance, and it is in line with the actual situation.
0.5 0.4 0.3 0.2 0.1 0
Submarine Decoy 0
10
20
30
40
50 Time/t
60
70
80
90
100
Fig.9. Targets probability computed by acoustic and motion features
considering the motion feature and without time dimension fusion.
0.9
0.7
1
0.6
0.9
0.5 0.8 0.4 Probability
Probability
0.8
0.3 0.2 0.1 0
Submarine Decoy 0
10
20
30
40
50 Time/t
60
70
80
90
0.7
0.6
0.5 100
0.4 Submarine Decoy
Fig.10.Targets probability computed by hierarchical fusion method
Figure 8 is the probability that the target is submarine when regardless the target motion properties and without time dimension fusion. As can be seen, at first, the real targets can be rightly recognized before the simulation time of 70, after that, false target has a greater probability. The reason of that is clear, the closer distance of the two decoys to the torpedo, the more elevated probability is. Conversely, the submarine becomes further, with the distance increased, the farther the submarine moves, the sensor performance is lower. The uncertainty is higher, the probability is less. So, it can be seen that, under the condition without considering the motion feature, the correctness of the target probability will be decreased to some extent with the distance increased. Figure 9 is the probability that the target is a submarine using two types’ features without time dimension fusion. Compared with Fig. 8 ， the correctness of the target probability is being extended, the wrong probability appeared after the time of 80. It can be seen that with the motion feature added, the submarine can be correctly estimated in a certain degree, but it cannot guarantee the result is correct in the whole simulation process. Given the threshold ε = 0.1 , the target probability that the target belongs to submarine computed by the hierarchical fusion method is shown in Fig.10. It can be seen throughout the whole simulation process, the probability of the submarine is the biggest and real target has been correctly calculated. It shows the proposed method can be well applied to underwater acoustic multi-objective optimum seeking. At the time between 40 and 50，due to changes in the orientation of the targets, the acoustic features change and it make the final fusion results changed, with respect to Fig.8 and Fig.9, fluctuated changes style appeared in Fig.10, which is caused by the time dimension fusion. Time dimension fusion can make the probability stabilized and make the affection by one feature change minimized. Compared to the related work us the same probability of each attribute, the result computed by GROD method is shown in figure 11. It can be seen that when the target becomes farther away from the torpedo, that is, after the time 58, the calculation results don’t match the actual situation. So, even if the GROD method can solve the greatest threat problem of the UUV. But in solving the true and false targets optimum seeking issue, the correct rate of the method is far below the method in this paper, even it below the method without
0
10
20
30
40
50 Time/t
60
70
80
90
100
Fig.11. Targets probability computed by GROD method
V.
CONCLUSION
A hierarchical fusion method has been proposed to solve the problem that true and false of underwater acoustic target are difficult to identify, which need combining with acoustic information and motion information for recognition under tracking progress. Firstly, the feature models are built, then, how to build the membership functions of the feature attribute is given, after that, evidence theory for underwater acoustic target recognition algorithm implementation process has been proposed. Three level fusion structures have been done on the same type of features, and then in two types, finally on time dimension respectively. Experiment results show that this method can effectively identify the true and false targets, can significantly improve the probability of correctly identify the target just based on acoustic feature information, and is more effective than the traditional GROD method to deal with the problem. It provides a new idea and method for underwater acoustic multi- target recognition. Next, this method should be tested in more complex countermeasure environments, and also the movement of the torpedo should be considered to verify the reliability of the algorithm. Acknowledgment
This research was financially supported by national natural science foundation of China (61135001), the Xi'an (China) Science and Technology Project in 2014 and 2013 (CXY1350 (2)).
References [1]
[2]
[3]
Jung-Min Pak, Dong-Gi Woo, Bonhwa Ku, et al. Target Search Method for a Torpedo to the Evading Ship Using Fuzzy Inference, ICROS-SICE International Joint Conference, Japan, 2009, pp.5279-5284. Wang Mingzhou, Huang Xiaowen, Hao Chongyang. “Model of an Underwater Target Based on Target Echo Highlight Structure. Journal of System Simulation”, vol. 15, pp. 21-25. 2003,. Wang Mingzhou, Hao Chongyang, Huang Xiaowen. “On Underwater Target Dimension Recognition Based on Bearings Analysis of Signal
[4]
[5]
[6]
[7]
Correlation Feature”. Journal of Northwestern Polytechnical University, vol. 21 , pp.317-320, 2003. Chen Kean, Wang Na, Wu Ying, et al. “Subjective evaluation experiments of timbre attribute based acoustic target identification”. Chinese Sci Bull, vol. 55 , pp. 651-659, 2010. Yang Honghui, Dai Jian, SUN Jincai. “A New Adaptive Immune Feature Selection Algorithm for Underwater Acoustic Target Classification”. Journal of Xi’an Jiao Tong University, vol. 45, pp. 2831, 2011. Chu-Kuei Tu, Chung Yuan, Chung Li, Tseng-Hsien Lin. Applying genetic algorithms on fuzzy logic system for underwater acoustic signal recognition, Underwater Technology, 2000. Proceedings of the 2000 International Symposium on, pp: 405-410. TUCKER S, BROWN G J. “Classification of transient sonar sounds using perceptually motivated features”. IEEE Journal of Ocean Engineering, vol. 30 , pp.588-600, 2005.
[8]
J. Acoust, Soc. Am. “Perception-based automatic classification of impulsive-source active sonar echoes”. Journal of the Acoustical Society of America, vol. 122 , pp. 1502-1517, 2007. [9] Wang Baihe，Huang Jianguo，Zhang Qunfei. “Model and Algorithm for Attacked Target Optimization in Underwater Multi-targets Tracing System”. Torpedo Technology, vol. 15 , pp. 22-25, 2007. [10] Meng Qingyu, Zhang Jingyuan, Song Baowei. Analysis of Operational Effectiveness for Torpedo. Beijing: National Defense Industry Press, 2003. [11] Xiao Juan, Gu Hao, Cao Zhimin. “Modeling and Simulation of Acoustic Decoy”. Computer Simulation, vol. 25 , pp. 25-27, 2008. [12] Ko-Hsin Liang, Kuei-Ming Wang. Using Simulation And Evolutionary Algorithms to Evaluate the Design of Mix Strategies of Decoy and Jammers in Anti-Torpedo Tactics. Proceedings of the 2006 Winter Simulation Conference. Monterey, CA: IEEE. pp. 1299-1306.
