Advaces i Egieerig Research, volume 8 d Iteratioal Coferece o Automatio, Mechaical Cotrol ad Computatioal Egieerig (AMCCE 7) Radar emitter recogitio method based o AdaBoost ad decisio tree Tag Xiaojig, a, Che Weigao ad Zhu Weigag Equipmet Academy, Beijig 64, Chia. a tag_675@6.com Keywords: AdaBoost, decisio tree, classifier, radar emitter recogitio Abstract: For the poor real-time, robustess ad low recogitio accuracy of traditioal radar emitter recogitio algorithm i the curret high desity sigal eviromet, this paper studied a kid of radar source recogitio algorithm based o decisio tree ad AdaBoost. Firstly, the iformatio gai ca be used to costruct sigle decisio tree. The usig AdaBoost algorithm to trai the weak classifier, ad get a strog classifier. Fially, get the recogitio results through the strog classifier. Simulatio results show that the recogitio accuracy of proposed method ca reach 9.78% with % parameter error, ad the time cosumptio is lower tha.5s, which has a good recogitio effect.. Itroductio Emitter recogitio ca aalyze the itercepted radar sigals by the idetificatio database, ad provide referece for grasp eemy radar techical parameters, activity rule, threat level ad developmet []. But i curret high desity complex sigal eviromet, the traditioal radar source recogitio methods such as template matchig[], expert kowledge[, 4] ad eural etwork [5-7], have problem of poor recogitio accuracy rate or poor timeliess, this will ievitably lead to the subsequet estimatio of threat situatio of icomplete, iaccurate, ad lack of key iformatio to guide iterferece [8]. To solve this problem, this paper studied a radar emitter recogitio algorithm based o decisio tree ad AdaBoost [9]. Firstly, by chagig the weights of traiig samples, AdaBoost method build a strog classifier from learig multiple weak classifiers to obtai better classificatio results. The due to the decisio tree method of simple i priciple ad strog robustess, it s very suitable for the weak classifier. Therefore, cosiderig to combie the advatage of two algorithms to costruct the ew radar source recogitio method.. Basic priciples of AdaBoost algorithm The Boostig algorithm origiated [] from Probably Approximately Correct learig model proposed by Valiat. It s a good method to improve the accuracy of ay give learig algorithm. The basic priciple of boostig is the classificatio problem, this method use traiig sample to lear weak classifiers, chage the weight ad get a strog classifier. Adaptive Boostig algorithm does ot eed to kow the exact distributio of sample space, oly through comparig the classificatio accuracy for each sample to determie the weight of samples. The superiority of AdaBoost are reflected i the followig aspects: high-precisio; improve the geeralizatio ability by reducig the traiig error; strog flexibility.. The basic priciple of decisio tree algorithm based o iformatio gai Decisio Tree [, ] is oe of the importat algorithms i machie learig ad data miig. It s easy to uderstad ad implemet, has good robustess ad predictive performace, so it s widely used i various fields. Decisio Tree is a predictio model of classificatio tree structure, which represets the mappig attribute relatioship betwee the each object. The decisio Stump is a typical simple decisio tree, it s oly based o a sigle feature to make decisios. Because the tree has oly a split process, the data processig is very rapid, simple, ad good i real time, but the recogitio result is ot good, so the Decisio Stump is very suitable as a weak classifier. Copyright 7, the Authors. Published by Atlatis Press. This is a ope access article uder the CC BY-NC licese (http://creativecommos.org/liceses/by-c/4./). 6
Advaces i Egieerig Research, volume 8 4. The step of radar emitter recogitio based o AdaBoost ad decisio tree Comparig with support vector machie, artificial eural etwork ad other complex learig algorithms, the performace of sigle decisio tree algorithm is weakess. However, a large umber of research results show that the performace of the decisio tree itegratio algorithm is ofte superior to other algorithms ad their itegratio i may practical fields. Cosiderig the curret coditio faced by radar source recogitio, large volume of data, accuracy, real-time, ew ukow sigal recogitio ad other practical requiremets, combiig with the AdaBoost decisio tree method has the advatages of high accuracy, strog geeralizatio ability ad good real-time, this paper studied a radar emitter recogitiod method, which based o decisio tree ad AdaBoost algorithm. The brief procedure as follows: easb ioisc A priori kowledge database Fid the characteristic parameter of small classificatio error i ag o it a m ro f i m th rio g la ee rt e d re y la el g i S Partitio data set Weight distributio of traiig data update Calculate the classificatio error rate Weight of classifier combiatio Get the basic classifier Ukow radar sigal recoaissace received Get the fial classifier Get the fial recogitio result Fig. Radar source idetificatio process based o AdaBoost ad decisio tree ①The iformatio gai of each feature is calculated based o the prior kowledge of database, ad the feature of the maximum iformatio gai is selected to partitio data set. ②The iitial weights of the traiig data are set, the classificatio error rate is calculated, ad updated the weight distributio of traiig data accordig to the classificatio error, get the weight of weak classifier. ③With each iteratio, util the traiig error is zero or weak classifier umber reach the specified value, costruct the strog classifier by calculatig the weights of weak classifiers, at last usig the strog classifier to classifier the received ukow radar sigal. 5. Simulatio experimet aalysis I order to verify the effectiveess of the radar source recogitio algorithm based o decisio tree ad AdaBoost, 9 kids of radar sources are simulated as kow prior kowledge, DOA, PA, PW, RF, PRF, BW are selected as features, each radar source sample feature iformatio as show i table. 7
Advaces i Egieerig Research, volume 8 Table Kow radar emitter sample characteristic iformatio Radar DOA/( ) PA PW/(μs) 48 5 5 5..4 6 6..4 68 7 6 4 56 58 5 6 RF/(MHz) PRF/(Hz) BW/(Hz) Sample umber 58/68/74. 4 7. stagger 8/85/9/95/ 8.6 4 55 75 shake 7 9 9. 9 95 96 8/8/88 5. 95/96 499 5.6 /7/6 998.8 6 78/8/97 99 6.6 4 98 99 56 564. Type Rage Type Rage 4 5 agile 75 85 shake 6.8 6.9 8/ 97 6..5 agile 8 6.6 6.8 49 5..4 7 5 56 6. 6. 8 65 68 6 6. 9 6 8 4 5 6 6. ()Robustess aalysis Set 9 kids of radar sources listed i Table as traiig data, a total of 97 samples. Geerate test data by addig % ad % parameter error with referece to the parameters i Table. The proposed radar source idetificatio method is used to trai the prior sample, geerate the strog classifier, ad the the test data is idetified by the strog classifier. The recogitio process as follows. The RF-PW-DOA distributio of traiig data radar radar radar radar4 radar5 radar6 radar7 radar8 radar9 7 6 DOA/ 5 4.5 5 5.5 PW/µs 5 RF/MHz Fig. Traiig data distributio Fig. The test data distributio with % parameter error Fig.4 The idetificatio results with % parameter error The RF-PW-DOA distributio of traiig data radar radar radar radar4 radar5 radar6 radar7 radar8 radar9 7 6 DOA/ 5 4.5 5 5.5 PW/µs 5 RF/MHz Fig.5 Traiig data distributio Fig.6 The test data distributio Fig.7 The Idetificatio results with % parameter error with % parameter error Total umber of test data is 975, the each umber of 9 kids sample is, 7, 49, 9, 6, 8, 6, 47,. For the test data with % parameter error, whe the umber of weak classifiers is 46, the 9.78% recogitio accuracy ca be achieved, the recogitio results are show i Fig 4. Compare with Fig., 6 ad 4, 7, the parameter error icreased from % to %, the test data distributio is more complicated, at the same time recogitio accuracy is decreased. Eve whe the parameter error is % ad the umber of weak classifiers is set to 6, 8.8% recogitio accuracy ca be achieved. Ideed, the parameter error is geerally less tha 5%, so the proposed radar emitter recogitio method ca achieve a better recogitio accuracy. ()Idetificatio effect aalysis The recogitio results are aalyzed by idetifyig accuracy ad time cosumptio. Traiig ad testig sample data as show i experimet (), the parameter error is set to %, from the 8
Advaces i Egieerig Research, volume 8 begiig of to, icreasig the umber of weak classifiers, time cosumptio ad recogitio accuracy are show as follows. The relatioship betwee weak classifier ad time cosumptio The relatioship betwee weak classifier ad idetificatio accuracy.5 Idetificatio accuracy/% Time cosumptio/s 9.5 8 7 6 5 4 4 5 6 7 8 9 The umber of weak classifier Fig. 8 The time cosumptio of the algorithm varies with the weak classifier 4 5 6 7 8 9 The umber of weak classifier Fig. 9 The accuracy rate of the algorithm is chaged with the weak classifier. From Fig. 8, we ca see that the time cosumptio of the proposed algorithm is approximately liear distributio with icreasig the umber of weak classifiers. But whe the umber of weak classifiers is, the time cosumptio is less tha.5s. As show i Fig. 9, whe weak classifier umbers icreased from to 5, the algorithm of recogitio accuracy icreased rapidly, ca get 9%. Cotiue to icrease the umber of weak classifiers, the icrease of recogitio accuracy is ot obvious, eve has ot icreased but decreased, thus selectig the appropriate umber of weak classifiers has great ifluece with the recogitio accuracy. 6. Coclusios With the rapid developmet of iformatio techology, promoted may ew techologies of radar system, ad created more complex sigal patter, so i the future battlefield, we will face the threat of ukow radar sigal. I order to cope with the challege, from the perspective of itelligece aalysis, aim at the problem of the existig traditioal radar source idetificatio method with poor recogitio accuracy ad robustess, this paper studied a radar emitter idetificatio method based o decisio tree ad AdaBoost. Firstly, costruct the sigle decisio tree based o iformatio gai. The use the AdaBoost algorithm to trai the classifier, costruct the strog classifier. Fially classificatio of test data by the strog classifier, ad get the recogitio results. Through the simulatio aalysis, it s proved that the proposed radar source recogitio algorithm ot oly has good recogitio accuracy ad robustess, but also has good timeliess. However, i the experimetal process, it s foud that the selectio of weak classifiers still ca optimize, which will be the focus of ext step. 7.Refereces [] Yag Zhutia. Research o radar emitter classificatio ad recogitio based o machie learig [D]. Iformatio ad commuicatio egieerig, Harbi Istitute of Techology,. [] Tag J, Qig L I. Fast template matchig algorithm[j].joural of Computer Applicatios,, (6): 558 559. [] Ford B P, Middlebrook V S. Usig a kowledge based system for emitter classificatio ad ambiguity resolutio[c]. Aerospace ad Electroics Coferece, 989. NAECON 989., Proceedigs of the IEEE 989 Natioal, 989. [4] Li Doghai. Radar idetificatio method based o expert system [J]. Shipboard Electroic Coutermeasure,4, 7(5):. [5] Wag Log. Applicatio of eural etworks i radar target recogitio [D]. North Uiversity of Chia, 5. 9
Advaces i Egieerig Research, volume 8 [6] Hu Houli,Wei Wei,Hu Mega. Priciples ad practices of deep learig [J]. Iformatio Techology,5,(): 75 77. [7] Li Yue. Recogitio techology for remote soudig radar emitter [D]. Harbi Istitute of Techology,. [8] Che Chagxiao, He Mighao, Xu Jig, et al. Progress of study o recogitio techology of radar emitter [J]. Joural of Air Force Radar Academy, 4,(): 5. [9] Rätsch G, Ooda T, Müller K R. Soft Margis for AdaBoost[J]. Machie Learig,, 4():87. [] Li hag. Statistical learig method [M]. Tsighua Uiversity press,. [] Wag Lili,Liu Xueju. Applicatio of C4.5 algorithm i aalysis of studets' performace [J].Joural of Hea Istitute of Egieerig: Natural Sciece Editio,4,(4): 69 7.