Recognition of Two Types of Air-Targets with KNN-SVM Method

Transcription

1 0 4th Internatonal Conferene on Sgnal Proessng Systems (ICSPS 0) IPCSIT vol. 58 (0) (0) IACSIT Press, Sngapore DOI: /IPCSIT.0.V58.9 Reognton of Two Types of Ar-Targets wth KNN-SVM Method Peng Yuxng + and Lao Dongpng Shool of Eletron Sene and Engneerng Natonal Unversty of Defense Tehnology Changsha, Chna Abstrat. The dfferenes n spetrum strutures of the helopters and fxed-wng arraft are analyzed n ths paper, when Doppler spetrum measurement ablty annot be met. A knd of KNN (k nearest)-svm (support vetor mahne)-based ar subjets reognton method s proposed. KNN s used to sreen the to-be-traned spetrum samples of the helopters to ensure the possblty of obtanng the optma dstnton whh reflets the haratersts of the spetrum strutures of the two researh subjets. Aumulaton method s used to solve the problem of the reognton of the helopter spetrum samples that are smlar to the fxed-wng arraft ones and the orrespondng aumulated strateges are presented. And the results of the experment by raw data ndate that the method employed s effetve. Keywords: target reogton, doppler sperum measurement ablty, K nearest, support vetor mahne, aumulated.. Introduton For the reason that the helopters are qute dfferent from the arrafts wth fxed wngs n the ntmdatory effets and destrutve results, to detet and aurately dentfy the two types of arraft targets s neessary. The reognton of these two types of ar targets for radar s manly based on the haratersts of the ehoes omng bak from the man gyro-wngs. Due to the modulaton of the rotaton of the gyro-wngs, the modulated ehoes shows as sn funton n tme doman and ts nstantaneous frequeny spetrum wll be broadened. As a result, the frequeny spetrum of ehoes s dfferent n shapes between the helopter and the arrafts wth fxed wngs on whh we an make the reognton. Support vetor mahne (SVM) s a new mahne learnng method proposed by Vapnk and hs researh team AT & T Bell Labs [], []. It bulds on the prnple of the mnmum strutural rsk and has the advantages n learnng ablty and generalzaton performane whh are effetve to solve the small sample sze, hgh dmenson, nonlnear problems, loal mnma and other pratal problems. Thus SVM provdes a new potental approah for radar target reognton [3-5]. Ths artle wll apply the method to the ar targets reognton and s qute sgnfant n prate. However, when applyng SVM to the ar target reognton dretly some problems ome forth. To extrat the feature that the gyro-wngs ehoes broaden n frequeny spetrum, we need reeve ehoes whh are as large as possble. That s to say longer beam resdent nterval and hgher pulse repetton frequeny (PRF) are requred for the radar [6], [7]. However, n prate, t s hard to meet the requrements above, resultng we may not get the broadenng spetrum and other nformaton n observatons. In the ase that the radar an t ath the spetrum broadenng, the helopter spetrum demonstrated the same fgure wth the fxed-wngs arraft. And then a tranng set nludng the samples of ths type of helopters frequeny spetrum onsstent, t wll be mpossble to obtan the optmal separatng surfae whh an reflets the spetral haratersts n the two types of ar targets [3-5]. Ths paper proposed a KNN-SVM-based ar target reognton algorthm. Frst, the spetrums of helopter tranng sample sets were sreened usng KNN algorthm to weed out the helopter spetrum samples whh s + Correspondng author. E-mal address: pengyuxng003@tom.om 3

2 smlar wth the fxed-wngs samples. And then the fxed-wngs spetrum samples sample and the left helopter tranng spetrum samples wll be the tranng wth SVM to ensure the possblty of obtanng the optmal separatng surfae whh an reflets the spetral haratersts n the two types of ar targets. Seondly, n order to mprove the reognton rate of the helopter spetrum samples whh s smlar wth the fxed-wngs samples, we proposed a method for proessng through the aumulaton and the orrespondng aumulaton strateges. Expermental results show that, ompared wth the dret use of SVM tranng, the proposed method mproved the reognton rate to a ertan extent and the aumulaton strategy ensure the stablty of the reognton results.. Analyss of Ar Tagets Charaterst When eletromagnet waves rradate rotaton blades, as long as the radar beam rradaton tme s long enough, the radar reever an detet a seres of eho pulses. When eletromagnet waves rradate rotaton blades perpendularty, the returns are strongest, thus the so-alled ehoes "flash" forms. The flash s gven by sn funton, the wdth of the man lobe the flash pulse duraton) s as follows: kλ Δ T = () 4π frot ( l r) Where λ s the radar wavelength, f rot s the rotaton veloty, l s the length of blades, r s the shaft radus. When the blade number N s odd, k =, when the blade number N s even, k =. In order to obtan samplng opportuntes n the flash pulse duraton and detet the targets effetvely and relably, the radar pulse repetton frequeny wll be lmted: PRF () 0.443ΔT To avod sample loss, usually we need sample twe at least wthn a flash pulse whh s possble to meet the for large PRF. but f PRF s small, the flash pulse sn t sampled so and the Doppler spetrum obtaned an t show the target haratersts. Therefore, n reognton on helopters and fxed-wngs arrafts, we should hoose the approprate PRF n the ondtons of determned wavelength of the radar. Flash pulse nterval s: T = (3) p knfrot Obvously, the longer beam resdent nterval help we get more flash pulses and detet the gyro-wngs sgnals. The typal value of rotaton speed f rot generally orresponds to 5 ~5.8Hz and the maxmum possble flash pulse nterval s 50ms, thus the radar requres a wde beam wdth and a show rotaton veloty of antennas to get the beam resdent nterval longer than 50ms, and at least one hane to detet flash pulses. If the deteton apablty need to be mproved further, the longer beam resdent nterval s requred. The reognton of these two types of ar targets for radar s manly based on the haratersts of the ehoes omng bak from the man gyro-wngs. Whether we an get the fgures of gyro-wngs ehoes s deded by the ablty of Doppler spetrum measurements for radar. That s to say the radar requres long beam resdent nterval and hgh PRF. In the ase that the ablty of Doppler measurement s met, under deal ondtons the Doppler spetrum of helopters and fxed wngs arrafts are shown n Fgure and Fgure. a a b f0 fs f0 fs Fg.: Ideal struture of helopter spetrum Fg.: Ideal struture of fxed-wn spetrum 4

3 From Fg. and Fg., f s s the sample frequeny, a s the sngle frequeny loated at f 0 whh s generated by the body wth a radal veloty V, and the Doppler spetrum of helopters and fxed wng arraft both ontan ths frequeny omponents. C s the nose omponent generated by the radar reever s unformly dstrbuted. B s the wde Doppler spetrum generated by gyro-wngs ehoes due to modulaton of gyro-wngs to radar sgnals and ths modulaton s unque for helopters. Clearly, whether there exst b s the man dfferene between the Doppler spetrum of helopters and fxed wngs arrafts. In the ase that the apablty of Doppler measurement s not met, we an t ath the nformaton of modulaton of gyro-wngs due to the nomplete observaton. As a result part of the Doppler spetrums of helopters and fxed-wngs arrafts are exatly the same. 3. KNN-SVM Algorthm SVM lassfer an solve the optmal lassfer desgn problem [8], [9], provdng a potental new approah to solve pratal problem of radar target reognton [0]. However, SVM appled dretly to the ar target reognton may be some problems. Doppler measurement apablty n the ondtons are not met, some helopters and fxed-wng arraft Doppler struture dental struture. At ths pont the spetrum wll ontan data suh helopter tranng samples usng SVM for tranng, t s dffult to get he optmal lassfaton surfae of the ar targets. In many ases, tryng to use a omplex model of lmted tranng samples wll have a dereased ablty to promote [9]. In fat, not havng rotor modulaton spetrum affets the dreton of the optmal lassfaton surfaes and the speed of the lassfer tranng. Therefore, f ths part of the tranng data pror to be deleted, t s possble to get optmal lassfaton refletng the spetrum of two types of ar targets, and t an mprove the tranng speed of the algorthm. Based on ths dea, we propose a KNN-SVM algorthm, whh s based on the known spetrum of helopter and fxed-wng spetrum template. The use of KNN tranng an flter the spetrum of helopters, and the k nearest neghbor spetrums of eah to-be-traned sample wll to be sought. Aordng to ertan deson rules, the spetrum samples of the helopters wll be removed beause of they are smlar to the spetrum samples fxed-wng helopter, then tran agan by usng SVM algorthm. Durng the lassfaton, n order to mprove the reognton rate, we proposed ways to deal wth through the aumulaton. In ths seton, at frst, we analyzed the prnples and methods to delete helopter spetrum samples by usng KNN, and then analyzed and gave an aumulaton strategy. Fnally, the major steps of KNN-SVM algorthm are gven. KNN s the promoton of -NN, the lassfaton hosen nearest neghbors, to dede whh type the majorty of the neghbors belongs to. Suppose there are N samples. N knd and N knd,, N knd, The dstngush funton s be defned as: The rule follows: f ( ) max g x k j =, j ( ),,, g x = k = (4) x ω. So, the use of KNN tranng an treat data set to flter the helopters spetrum samples, and delete the fxed-wng spetrum samples aordng to (4). Spetrum through deleted helopter spetrum samples whh are smlar to the fxed-wng s,t ensures the possblty of the optmal lassfaton surfae. However, n ths ase, most of the helopters spetrum samples wll be deded to fxed-wng. In some spef applatons, the radar system does not need to gve the result of target reognton every tme for eah sub-system, but to gve a lttle stablty n the result of reognton, and aumulated strateges an aheve. For example, wth the use of /n aumulaton strategy, eah observaton s gven an dentfyng ntermedate result durng proessng, and eah reognton deson as n a helopter n n ntermedate results, and the reognton result s the helopter, or s fxed-wng. The stablty of the reognton results an be ensured through ths aumulaton strategy. 5

4 The problem s the determnaton of the number of aumulaton. In fat, wthout onsderng the frequeny of repetton of the premse, to get at least a sngle flash pulse deteton opportuntes requres more than 50ms, whh s the beam on the target dwell tme. For example, a sngle beam dwell tme s0ms, then the target observaton needs at least 5 tmes, and we an get larger than the beam on the target dwell tme 50ms, the number an be taken by n = 50 /0 = 5. Of ourse, ths observaton an not guarantee that there must be rotor modulated nformaton on the exstene, beause the observatons may not be ontnuous, suh as some mehanal sannng radar observatons on the same target has almost sngle frame delay. However, the statstal average from a long vew, ths dvde s qute a reasonable way. KNN-SVM algorthm s gven dretly below the man steps: () Aordng to the sze of the amount of tranng pror estmate of value k. () Seletng Euldean dstane to measure the dstane between the samples, that s, the dstane between to-be-traned helopter spetrum samples and the template spetrum samples s defned as: (5) d ( ), = ( ) ( ) = D x y x j y j (3) Make sure the samples of the tranng helopter spetrum as x, from the template sample set to fnd out the k nearest neghbors. (4) Aordng to equaton (5) for deson, delete the sample whh s deded as fxed-wng spetrum. (5) Repeat the above four steps untl all x are dealt wth, to get the remanng to-be-traned samples of helopter spetrum. (6) Wth SVM, tranng the remanng to-be-traned samples of helopter spetrum and the to-be-traned fxed-wng spetrum. (7) Aordng to radar observatons of a sngle beam dwell tme to determne the number n of aumulaton. (8) Usng the traned SVM on the test set for testng, wth the use of aumulaton strategy/n. If one s deded as helopter n eah n ntermedate result, the reognton result s helopter; otherwse, the reognton result s fxed-wng. 4. Smulaton and Analyss Make experment based on the above algorthm, expermental software envronment: Matlab 7./WIN XP; Hardware Envronment: P Ⅳ. 66 G/5 MB. The experment data s from measured Doppler data set, whh s measured by mllmeter-wave radar. The sample types nlude both helopters and fxed wng arraft. The sample dmenson s 64 (not nludng the lassfaton of property). There are 094 total samples, n whh 547 samples n eah lass. There are 400 tranng samples, 00 samples n eah lass. There are 694 test samples, 347 samples n eah lass. There are 0 KNN templates, and 0 templates per lass. In SVM tranng, C s taken to 000. Wth the use of radal bass funton and the parameter s taken to 0.5. The KNN parameter k s taken to 0. Aordng to radar observatons of a sngle beam dwell tme to determne the number of aumulaton n = 6. Table shows the expermental results of dret SVM and KNN-SVM method. Compared wth the reognton results wth dretly SVM method, KNN-SVM method mproves on the fxed wng of the reognton rate of the sample spetrum (97.4%) wthout usng the ase of the aumulaton strateges, but the sample reognton rate of helopter are not always effetve (3.05%). Through the aumulaton strateges, not redung the reognton rate of fxed-wng spetrum sample of the premse (87.7%), t greatly mproves the reognton rate of the helopter spetrum sample (89.47%). At the same tme, as the helopter KNN spetrum of tranng samples removed, redung the number of samples used for tranng to mprove the tranng speed. 5. Conlusons 6

5 Ths paper presents a KNN-SVM-based ar target reognton algorthms. At frst, flter the to-be-traned samples set of helopter spetrum wth KNN algorthm, and delete the helopter spetrum samples of a smlar spetrum wth fxed-wng. It an ensure the possblty of optmal lassfaton whh reflets the haratersts of these two types of ar target spetrum. Then, n order to mprove the reognton rate of helopter spetrum whh s smlar to fxed-wng spetrum, the use of methods to deal wth through the aumulaton s proposed. And the orrespondng aumulaton orrespondng are taken. Measured data results show that the method s effetve. 6. Referees [] Cortes C, Vapnk V.N.Support vetor networks[j]. Mahne Learnng, 995,0: 44~5 [] Vapnk V.N, The Nature of Statstal Learnng Theory[M], New York: Sprnger, 995 [3] Zhang L, Zhou Weda, Jao Lheng, Radar Target Reognton Based on Support Vetor Mahne[C], Proeedngs of the 000 Internatonal Conferene on Sgnal Proessng Proeedngs, 000, 3: [4] L Yng, Ren Yong, Shan Xumng, Radar HRRP Classfaton wth Support Vetor Mahnes[C], Proeedngs of the 00 Internatonal Conferene on Info-teh and Info-net, 00, : 8- [5] Kreβel U.Parwse lassfaton and support vetor mahnes[a].in: Shölkopf B, Burges C.J.C, Smola A.J, Edts, Advanes n Kernel Methods: Support Vetor Learnng[C].Cambrdge, MA: MIT Press, 999: 55~68 [6] Joahms T.Makng large-sale support vetor mahne learnng pratal[a].in: Shölkopf B, Burges C. J. C, Smola A.J, Edts, Advanes n Kernel Methods: Support Vetor Learnng[C]. Cambrdge, MA: MIT Press, 998: 69~84 [7] Platt J.C.Fastng tranng of support vetor mahnes usng sequental mnmal optmzaton[a].in: Shölkopf B, Burges C.J.C, Smola A.J, Edts, Advanes n Kernel Methods: Support Vetor Learnng[C].Cambrdge, MA: MIT Press, 998: 85~08 [8] Burges C. J. C, A Tutoral on Support Vetor Mahnes for Pattern Reognton[M], Boston: Kluwer Aadem Publshers, 998 [9] Roobaert D.DretSVM: a fast and smple support vetor mahne perepton[c]. Proeedngs of the 000 IEEE Sgnal Proessng Soety Workshop on Neural Networks, Sydey, Australa, 000, : 356~365 [0] Dng Alng, Lu Fang, L Yng.Pre-extratng support vetor by adaptve projetve algorthm[c].proeedngs of the 6th Internatonal Conferene on Sgnal Proessng, 00(): ~4 [] Syed N.A, Lu Huan, Sung K.K.Inremental learnng wth support vetor mahnes[c].proeedngs of Workshop on Support Vetor Mahnes at the Internatonal Jont Conferene on Artfal Intellgene, Sweden: Stokholm, 999: 7~76 [] Wang Xaodan, Wang Jqn.Support vetor mahne for HRRP lassfaton[c]. Proeedngs of the 7th Internatonal Symposum on Sgnal Proessng and Its Applatons, 003, : 337~340 [3] Bralovsky V.L, Barzlay O, Shahave R.On global, loal, mxed and neghborhood kernels for support vetor mahnes[j].pattern Reognton letters, 999, 0: 83~90 [4] Yang M. H, Ahuja N.A geometr approah to tran support vetor mahnes[c]. Proeedngs of the 000 IEEE Computer Soety Conferene on Computer Vson and Pattern Reognton, 000, : 430~437 [5] Zhang L, Zhou Weda, Jao Lheng.Pre-extratng support vetors for support vetor mahne[c].proeedngs of the 5th Internatonal Conferene on Sgnal Proessng, 000(3): 43~435 Method Dret SVM KNN-SV M Table I: Dret SVM method and KNN-SVM method n the measured results on the data set Samples of the Reognton The Aumulaton atual number Tranng Reognton rate rate of average strateges of tranng tme(seond) of helopter fxed-wng reognton ( /n ) helopter arraft rate / % 5.78% 57.89% none 79.8% 73.9% 76.5% / % 87.7% 88.59% none 3.05% 97.4% 60.3% 7