High Accuracy Signal Recognition Algorithm Based on Machine Learning for Heterogeneous Cognitive Wireless Networks

Transcription

1 Journal of Communcatons Vol., o. 3, March 7 Hgh Accuracy Sgnal Recognton Algorthm Based on Machne Learnng for Heterogeneous Cogntve Wreless etworks Jan Lu, Jbn Wang, and San Umar Abdullah School of Computer and Communcaton Engneerng, USTB, Bejng 83, Chna Emal: lujan@ustb.edu.cn; wjbjekyll@63.com; sanumar.a.ng@eee.org Abstract Heterogeneous Wreless etworks (HWs), ncludng several dfferent wreless technologes, are recent solutons that provde seamless communcaton for moble users. However, wth the development of varous wreless networks, the spectrum detecton of cogntve networks and termnals becomes more complcated, whch decreases the detecton performance. The accuracy and effcency of the spectrum detecton wll be reduced due to ntegratng varous wreless networks wth dfferent characterstcs nto a dverse overlay system. In ths paper, we desgn a hgh accuracy recognton algorthm for Cogntve Rado (CR) sgnal based on machne learnng n HWs, whch can recognze the receved sgnal type through extractng the features. Ths algorthm can recognze the sgnal types blndly wth low complexty, and prevent the nfluence of hostle termnals. Smulaton results ndcate that the algorthm we proposed can acheve hgh recognton accuracy under ether Addtve Whte Gaussan ose (AWG) channel or Raylegh fadng channel. Index Terms Sgnal recognton, cogntve rado, machne learnng, SVM, heterogeneous wreless networks I. ITRODUCTIO In recent years, wreless communcaton technologes are developng so rapdly that there has been a spate of nterest n the ntegraton of varous wreless networks whch wll provde moble termnals wth the seamless communcaton servce. Heterogeneous wreless networks (HWs) are consdered as a promsng way to costeffectvely enhance coverage and capacty of the network [], especally n LTE network []. Cogntve Rado (CR) s provded as a vrtual technology to ncrease the effcency of spectrum effectvely [3], [4]. CR termnals, whch use the dle lcensed spectrum for communcaton, can't cause nterference between lcenses users and adjacent communcaton users [5]. There are two major challenges for cogntve networks and termnals, () The cogntve network and termnals sense the wreless envronment and communcate through usng the vacant spectrum. Thus, cogntve termnals requre effcent spectrum detecton technques. Manuscrpt receved January 8, 7; revsed March 9, 7. Ths work was supported by atonal Major Project under Grant o. 5ZX33-. Correspondng author emal: wjbjekyll@63.com. do:.7/jcm () There wll be more cogntve termnals to access the dle spectrum. But there are always some hostle termnals whch volate or gnore the communcaton protocol. How to solve the "hostle termnal", whch wll nfluence the normal cogntve termnals to search and access dle spectrum, s a very mportant ssue [6], [7]. Faced wth these challenges, a relable and hgheffcency spectrum detecton method s necessary for the cogntve networks and termnals. A lot of works have already been done n spectrum detecton and a lot of achevements have been realzed [8]-[]. The spectrum detecton problems become more complcated due to the varous network deployed n HWs. The transmsson rate and sgnal convergence range sgnfcantly affect the network performance []. As a result, many spectrum detecton methods are unavalable or the detecton results are not deal n HWs. Cogntve sgnal recognton s a useful technque to solve ths detecton problem n cogntve rado under HWs. It can prevent the nfluence of hostle termnals and then mprove the spectrum's utlzaton [3], [4]. The dentfcaton of the cogntve sgnal does not need any pror nformaton of receved sgnals, and the cogntve sgnal recognton can be composed of two parts: decson theoretc approach and feature-based approach [5], [6]. In [7], the authors descrbe a dscrete lkelhood-rato test based on rapd-estmaton approach, amng to dentfy the modulaton solutons blndly for unnterrupted date demodulaton n real tme. Feature-based algorthms use sgnal features such as sgnal statstcs, wavelet transform, sgnal constellaton to dstngush the varous modulaton types and constellatons [8]. In [9], the proposed algorthm s characterzed by hgher order statstcal moments of wavelet transform; the classfer s a mult-layer neural network wth reslent backpropagaton learnng algorthm. In [], the feature parameters of receved sgnal are extracted usng HHT and sngular values of matrx whch s composed of nstantaneous parameters are used as characterstc vector and nputted to the Generalzed Regresson eural etwork (GR) to recognze the modulaton of the sgnal. In ths paper, we descrbe a hgh accuracy cogntve sgnal recognton algorthm based on machne learnng whch can dstngush dfferent rado sgnal types n HWs, and the pror nformaton s not requred n the 7 Journal of Communcatons 73

2 Journal of Communcatons Vol., o. 3, March 7 II. SYSTEM MODEL A. Heterogeneous Wreless etworks A wdeband wreless system n HWs s represented n Fg.. Cogntve termnals detect the whole broadband frequency spectrum and fnd the spectrum holes contnuously. When spectrum holes appeared agan, cogntve termnals can detect t mmedately and make t accessble to communcaton. Macro BS M4 M3 CT CT Femto BS Fg.. A wdeband wreless system n HWs As shown n Fg., the transmt sgnal vectors of femto-bs and macro-bs are denoted as Sf and Sm, and the transmt sgnal vectors of cogntve termnal (CT ) and moble (M) are denoted as Sc, Smob. The Hmc s the channel matrx for macro-bs and cogntve termnal (CT), Hfc s the channel matrx for femto-bs and CT, Hcc s the channel matrx for CT and CT, and Hmobc s the channel matrx for macro-ms moble M and CT respectvely. The receved sgnal vectors at CT can be represented as, rc H mc Sm H fc S f H cc Sc H mobc Smob n 7 Journal of Communcatons Femto BS fc mc Sm bc H mo M H cc M CT CT Macro BS Fg.. Cogntve rado s n actve As can be seen n Fg., assumng there s a workng termnal whch would volate or gnore the communcaton rules or allocated strateges and access spectrum randomly, t could obtan nformaton from all the transmtted sgnals as sent by all the transmtters. So the spectrum holes may always be occuped by the hostle termnal. As a result, the termnal may result n waste of resources n the avalable and dle spectrum, and even destroy all the HWs. We can solve the problem through the followng three steps, () Recognze the sgnal types, whch are occupyng the wdeband frequency. () Analyze the recognzed sgnal types, and dentfy the hostle sgnals. () Avod hostle sgnals wth some effectve strategy. In ths paper, we manly research how to recognze the sgnal types. We descrbe a low-complexty algorthm to recognze sgnal types n HWs, after that, cogntve termnals can formulate a strategy to prevent the nterference of hostle termnals. y(t) A(t) x(t) (t) (3) where y (t) s the receved sgnal, A(t) changes along wth the channel, x(t) represents uncharted source sgnals receved by the cogntve termnals, and (t) s AWG wth varance. III. PROPOSED SIGAL RECOGITIO ALGORITHM () The receved sgnal vectors at macro-ms moble (M) can be wrtten as, rm H mc Sm H fm S f n Sf H B. Channel Model The wreless channel model s composed of AWG channel and Raylegh fadng channel. The average value of nose n the AWG channel s, whch adds to the transmtted sgnal after Raylegh fadng. The receved sgnal at the termnals s wrtten as, M M Femto BS where, n (), (), n represents the channel nose. H algorthm. There are two steps n the proposed algorthm: frst, we extract the features by usng the Daubeches5 wavelet transform and Fractonal Fourer Transform (FRFT), and then we use a classfer based on machne learnng to determne the sgnal types. Through the algorthm, we can get the modulaton nformaton of the receved sgnals. The smulaton results ndcate that the algorthm can acheve good performance whether n AWG channel or Raylegh fadng channel at acceptable SR range. The paper s organzed as follows: Secton ntroduces system model brefly. Sgnal recognton algorthm desgn s provded n Secton 3, whch analyzes the process of feature extracton consderng Daubeches5 wavelet transform and FRFT, and also focuses on the classfcaton method based on machne learnng. In Secton 4, smulaton results and performance analyss are presented. Fnally, conclusons are drawn n Secton 5. () 74 Sgnal recognton has become a key topc for electronc survellance n cogntve rado, and t plays a vtal role n mltary and commercal fled [3]. In ths paper, we proposed a hgh relablty and low complexty recognton algorthm usng feature-based approach under HWs. Fg. 3 shows the process of sgnal recognton,

3 Journal of Communcatons Vol., o. 3, March 7 whch conssts of two blocks: feature extracton and classfer. In the feature extracton block, we use Daubeches5 wavelet transform and FRFT to obtan the sgnal feature. After that, we choose a classfer based on machne learnng to recognze the sgnal type. Then, the recognzed sgnal type can be used for further analyss. Receved sgnal Wavelet transform of db5 Feature extracton block Fractonal fourer transform Fg. 3. The process of sgnal recognton A. Feature Extracton Classfer Recognzed sgnal types ) The necessary features The key step of the sgnal recognton s to extract the features for dentfcaton. The features of most sgnals n tme doman and frequency doman are not complex; hence they are appled n sgnal dentfcaton wdely. In ths paper, we choose fve features that are as follows, () Maxmum of the spectral power densty of the normalzed centered nstantaneous ampltude max max max FFT a cn (4) cn / s a a m (5) s the number of samples, where s acn s the normalzed center nstantaneous ampltude, m a s the average of the ampltude. It s often used to dfferentate the ampltude modulaton sgnal. () Standard devaton of the normalzed centered nstantaneous ampltude aa a s s a a (6) aa cn cn s s () Standard devaton of the absolute nonlnear centered nstantaneous phase ap L L c an a c t an at ap (7) where a t s the threshold used to remove weak sgnals, L s the value of the nonlnear component, c s the number of non-weak sgnal values. It can be used to dstngush the dfferent phase modulaton sgnal. (v) Standard devaton of the non-lnear centered drect phase dp dp L L c an a c t an at (8) (v) Standard devaton of the absolute non-lnear nstantaneous frequency af f where f f (9) af c an ( ) a c t an ( ) at m f f f f m R R R s s s R s s the symbol rate, s s f () f s the nstantaneous frequency. It can be used to dstngush dfferent frequency modulaton sgnals ) Wavelet transform Wavelet transform can acheve better results for analyss of non-statonary sgnals. The wavelet bass s defned as t b (a,b) (t) a () a We use Daubeches5 as wavelet bass n ths paper. The four flters assocated wth Daubeches5 wavelet s shown n Fg. 4. The effect of wavelet transform s to extract the abrupt feature of the receved sgnal. Fg. 4. The four flters assocated wth Daubeches5 wavelet 3) Fractonal fourer transform The FRFT offers a way to solve the problem of addtonal degree of freedom as parameter s used n the partcular sgnal processng []. For any real, the -angle FRFT of a functon f can be denoted as, Fs f x Bmm s; m x () m mm (3) f x B x m m m m R,,,... B f f x x dx m (4) The process of feature extracton can be seen n Fg. 5. The feature obtaned by wavelet transform s reflected by FRFT, whch can show the tme-frequency characterstc and the mutaton characterstc of the orgnal sgnal better. 7 Journal of Communcatons 75

4 Journal of Communcatons Vol., o. 3, March 7 Receved sgnal x(t) Classfer Daubeches5 wavelet transform Fg. 5. Feature extracton process X(ω) B. Machne Learnng Mechansm FRFT Extract features f p(μ) ) Support vector machne Support Vector Machne (SVM) s one of the most common machne learnng, t s also the most practcal part of statstcal theory and t s based on structural rsk mnmzaton prncple. Compared to the neural network classfer, t has many advantages n solvng small sample, non-lnear and hgh-dmensonal pattern recognton, and t has stronger generalzaton ablty. The purpose of support vector machne s to fnd the optmal classfcaton hyperplane for bnary classfcaton problem. Frstly, the nput space s transformed nto a hgh dmensonal space by nonlnear transformaton, and then the optmal lnear hyperplane s obtaned n the new space. Suppose there are tranng data as follows, ( x, y ),( x, y ),,( x, y ), x R n, y, (5) n n where x s the nput pattern set, y s + when x belongs to the frst sort, otherwse y s -.These data can be completely separated from the hyperplane ( w x) b. The problem can be transformed nto solvng optmzaton. mn w wb l C s. t. y w x b,,, l (6) where w s the coeffcent vector, b s the threshold for classfcaton, C s the penalty factor, s the relaxaton varable. We use the Lagrangan multpler to solve the above quadratc programmng problem and can obtan the decson functon as follows, f x sgn y K x, x b l (7) where K x, x represents the kernel functon. The role of the kernel functon s to map the nonlnear sample data to the hghdmensonal space so as to acheve lnear classfcaton n the hgh-dmensonal space. The common kernel functons nclude polynomal kernel functon K x, x x x d, the Radal Bass Functon s the Lagrangan multpler, (RBF) K x, x exp x x /, the sgmod kernel functon K x, x tanh k x x. ) Classfcaton usng SVM Support Vector Machne (SVM) s proposed for bnary classfcaton problems, whle dgtal modulaton sgnal recognton s a typcal mult-class classfcaton problem. Therefore, the above method should be mproved so that support vector machne can be appled to multclassfcaton problem. In ths paper, we adopt the followng method to solve the mult-classfcaton problem,. One Vs All We consder one category of samples as a category and the remanng categores of samples as another category. The problem s stll a bnary classfcaton of the problem.. One Vs One We select two samples randomly to construct a t bnary / classfcaton classfer from samples, and functons are constructed totally. The testng sample s classfed by each classfer, and the category of the testng sample s decded by the method of votng. Although ths method avods the problem of data set nclnaton, but the number of classfers ncreases, resultng n slow speed of tranng and decson. 3. Drected Acyclc Graph (DAG-SVM) In order to solve the problem of msclassfcaton and rejectng classfcaton n the method of One Vs One, Drected Acyclc Graph (DAG) s proposed by Plat, exported from Decson drect acyclc graph (DDAG). The dfference between One Vs One and DAG-SVM s the organzatonal structure of the classfer showed n Fg. 6. vs 5 vs 5 vs 4 3 vs 5 vs 4 vs 3 4 vs 5 3 vs 4 vs 3 vs Fg. 6. The structure of DAG-SVM In the proposed algorthm, we choose the RBF as the kernel functon of SVM classfer. Compare to the polynomal kernel functon, RBF has smaller number of parameters, whch affects the complexty of the tranng model drectly. The classfcaton process s showed n Fg. 7. Extracted features Recognton results Fg. 7. The classfcaton scheme Tranng sample nput Feature selecton Dscrmnant functon Improved dscrmnant functon 7 Journal of Communcatons 76

5 Journal of Communcatons Vol., o. 3, March 7 IV. SIMULATIO RESULTS AD PERFORMACE AALYSIS In ths secton, we perform extensve smulatons and analyss to prove the effcacy of the proposed algorthm. Four conventonal sgnal types (,, and ) are adopted for the recognton algorthm n MATLAB. The smulaton results and performance analyss of the proposed algorthm can be dvded nto two parts: sngle type recognton performance n AWG channel, sngle type recognton performance n Raylegh fadng channel. We set the followng parameters: sgnal symbol rate (Kb/s) and frequency of the carrer wave ( KHz). 4 frames of data are generated n smulaton, the Daubeches5 wavelet transforms and the FRFT are computed, applyng statstcal samplng ponts to extract the set of features. samples are used to tran the SVM model and the remanng 3 samples are used for testng. A. Recognton Performance n AWG Channel Fg. 8 ()-(5) shows the characters of max, aa, ap, dp and af n AWG channel. It can be seen from the characterstc curve, the curve nterval of the character max, aa of, s large, and the curve nterval of the character ap, dp, af of, s large. It s easy to dstngush the sgnal type dependng on the characters. In Fg. 9, the fold lnes ndcate the sngle type recognton performance of the proposed algorthm n AWG channel. The smulaton shows that the correct recognton probablty of n AWG channel s almost % when the SRs vary from db to 4dB. When SR=6dB, the recognton probabltes of the three sgnal types (,, ) are hgher than 96%. Gama m Sgma aa Sgma ap () () (3) Sgma dp dp Sgma af (4) (4) (5) (5) Fg. 8. The characters of sgnals n AWG channel prob of correct classfcaton Fg. 9. The correct recognton probablty n AWG channel B. Recognton Performance n Raylegh Fadng Channel Fg. ()-(5) shows the characters of max, aa, ap, dp and af n Raylegh fadng channel. From the fgure, we can fnd that the curve nterval of the character aa, ap of, s large, and the curve nterval of the character dp, af of, s large. It s easy to dstngush the sgnal type dependng on the characters. Sgma aa aa aa Gama m m Sgma ap () () () () (3) (3) 7 Journal of Communcatons 77

6 prob of correct classfcaton Journal of Communcatons Vol., o. 3, March 7 Sgma dp dp Sgma af af (4) (4) (5) (5) Fg.. The characters of sgnals n Raylegh fadng channel In Fg., the channel s Raylegh fadng channel. From the smulaton results, wth the ncrease of SR, the probablty of correct classfcaton s on the rse. When SR=8dB, the recognton probabltes of the four sgnal types (,,, ) are hgher than 85%. The recognton performance n Raylegh fadng channel s a bt worse than that n AWG, but s stll satsfyng Fg.. The correct recognton probablty n Raylegh fadng channel V. COCLUSIOS In the paper, we propose a hgh relable sgnal recognton algorthm of cogntve rado n heterogeneous wreless networks. Feature-based approach s used n ths algorthm, whch conssts of Daubeches5 wavelet transform and FRFT, and can be appled to complex envronment of HWs. Ths algorthm can be realzed wth low complexty and dstngush dfferent sgnal types. Accordng to the sgnal type, HWs can prevent the ``hostle termnals'' nfluence. Here we smulate wth four sgnal types (,, and ). Based on the smulaton results, we can conclude that the proposed algorthm shows superor performance n terms of low complexty, hgh recognton rate and effcency under the AWG and Raylegh fadng channels respectvely, as compared to tradtonal schemes. What s more, the recognton performance n AWG channel s a bt better than that n Raylegh fadng channel. REFERECES [] A. Gorcn and H. Arslan, Sgnal dentfcaton for adaptve spectrum hyperspace access n wreless communcatons systems, IEEE Wreless Communcatons, vol. 5, no., pp , 4. [] A. Damnjanovc, et al., A survey on 3GPP heterogeneous networks, IEEE Wreless Communcatons, vol. 8, no. 3, pp. -,. [3] T. K. Km, H. M. Km, M. G. Song, and G. H. Im, Improved spectrum-sharng protocol for cogntve rado networks wth multuser cooperaton, IEEE Personal Communcatons, 5. [4] A. Hess, F. Malandrno,. J. Kamnsk, T. K. Wjaya, and L. A. DaSlva, Cogntve rado algorthms coexstng n a network: Performance and parameter senstvty, IEEE Transactons on Cogntve Communcatons and etworkng, vol., no. 4, pp , 6. [5] H. E. Eglmez and A. Ortega, Wavelet-based compressed spectrum sensng for cogntve rado wreless networks, n Proc. IEEE Internatonal Conference on Acoustcs, Speech and Sgnal Processng, 5, pp [6] G. Sklvants, A. Gannon, S.. Batalama, and D. A. Pados, Addressng next-generaton wreless challenges wth commercal software-defned rado platforms, IEEE Communcatons Magazne, vol. 54, no., pp , 6. [7] Z. Guan, et al., Dstrbuted resource management for cogntve ad hoc networks wth cooperatve relays, IEEE/ACM Transactons on etworkng, vol. 4, no. 3, pp , 6. [8] J. So and W. Sung, Group-Based multbt cooperatve spectrum sensng for cogntve rado networks, IEEE Transactons on Vehcular Technology, vol. 65, no., pp , 6. [9] Z. Da, J. Lu, C. Wang, and K. Long, An adaptve cooperaton communcaton strategy for enhanced opportunstc spectrum access n cogntve rados, IEEE Communcatons Letters, vol. 6, no., pp. 4-43,. [] S. Lv and J. Lu, A novel sgnal separaton algorthm based on compressed sensng for wdeband spectrum sensng n cogntve rado networks, Internatonal Journal of Communcaton Systems, vol., no. 7, pp , 3. [] Z. Guan, et al., On the effect of cooperatve relayng on the performance of vdeo streamng applcatons n cogntve rado networks, n Proc. IEEE Internatonal Conference on Communcatons,. [] W. H. Yang, Y. C. Wang, Y. C. Tseng, and B. S. Ln, Energy-effcent network selecton wth moblty pattern awareness n an ntegrated WMAX and WF network, Internatonal Journal of Communcaton Systems, vol. 3, no., pp. 3-3,. [3] D. Lu, C. L, J. Lu, and K. Long, A novel sgnal separaton algorthm for wdeband spectrum sensng n cogntve networks, n Proc. IEEE Global Communcatons Conference,, pp. -6. [4] Z. Guan, T. Meloda, and G. Scutar, To transmt or not to transmt? Dstrbuted queueng games n 7 Journal of Communcatons 78

7 Journal of Communcatons Vol., o. 3, March 7 nfrastructureless wreless networks, IEEE/ACM Transactons on etworkng, vol. 4, no., pp , 6. [5]. Madhavan, A. P. Vnod, A. S. Madhukumar, and A. K. Krshna, Spectrum sensng and modulaton classfcaton for cogntve rados usng cumulants based on fractonal lower order statstcs, AEU-Internatonal Journal of Electroncs and Communcatons, vol. 67, pp , 3. [6] P. Panagotou, A. Anastastasoupoulos, and A. Polydoros, Lkelhood rato tests for modulaton classfcaton, n Proc. IEEE Mltary Communcatons Conference,, pp [7] J. L. Xu, W. Su, and M. Zhou, Software-defned rado equpped wth rapd modulaton recognton, IEEE Transactons on Vehcular Technology, vol. 59, no. 4, pp ,. [8] B. Ramkumar, Automatc modulaton classfcaton for cogntve rados usng cyclc feature detecton, IEEE Crcuts and Systems Magazne, vol. 9, no., pp. 7-45, 9. [9] K. Hassan, et al., Automatc modulaton recognton usng wavelet transform and neural network, n Proc. IEEE 9th Internatonal Conference on Intellgent Transport Systems Telecommuncatons, 9. [] Y. Hou and H. Tan, An automatc modulaton recognton algorthm based on HHT and SVD, n Proc. IEEE 3rd Internatonal Congress on Image and Sgnal Processng,. [] J. Lu and Q. Luo, A novel modulaton classfcaton algorthm based on Daubeches5 wavelet and fractonal fourer transform n cogntve rado, n Proc. IEEE 4th Internatonal Conference on Communcaton Technology,, pp. 5-. Jan Lu receved hs B.S. degree n Automatc Control Theory and Applcatons from Shandong Unversty, Chna, n, and the Ph.D. degree n School of Informaton Scence and Engneerng from Shandong Unversty n 8. He s currently an assocate professor of Unversty of Scence and Technology (USTB), Bejng, Chna. Hs research nterests nclude cogntve rado networks, moble mesh networks, and LTE-A. He s an IEEE member snce 9. Jbn Wang receved hs Bachelor degree n Communcaton Engneerng from Unversty of Scence and Technology Bejng (USTB), Bejng, Chna, n 3. He s currently studyng for a master's degree at Unversty of Scence and Technology Bejng (USTB), Bejng, Chna. Hs research nterests nclude cogntve rado networks and modulaton dentfcaton. San Umar Abdullah receved hs B.Eng. degree n Electrcal Engneerng from Ahmadu Bello Unversty Zara, gera n 8 and an M.S. degree n Moble and Satellte Communcatons from the Centre for Communcatons and Systems Research (CCSR) of Unversty of Surrey n 9. He s currently pursung a Ph.D. degree wth the School of Communcatons and Computer Engneerng, Unversty of Scence and Technology (USTB), Bejng, Chna. Hs research nterests nclude resource allocaton and nterference management n heterogeneous networks, cogntve rado, and use of stochastc geometry n moble wreless networks. 7 Journal of Communcatons 79