Modulation and Coding Classification for Adaptive Power Control in 5G Cognitive Communications

Transcription

1 Modulation and Coding Classification fo Adaptive Powe Contol in 5G Cognitive Communications Anestis Tsakmalis, Symeon Chatzinotas and Bjön Ottesten SnT - secuityandtust.lu, Univesity of Luxemboug {anestis.tsakmalis, symeon.chatzinotas, bjon.ottesten}@uni.lu (Invited Pape) Abstact A key concept suggested fo 5G netwoks is spectum shaing within the context of Cognitive Communications (CC). This efficient spectum usage has been exploed intensively the last yeas. In this pape, a mechanism is poposed to allow a cognitive use, also called Seconday Use (SU), to access the fequency band of a Pimay Use (PU) opeating based on an Adaptive Coding and Modulation (ACM) potocol. The Spectum Sensing (SS) technique used consides Highe Ode Statistical (HOS) featues of the signal and log-likelihood atios (LLRs) of the code syndomes in ode to constantly monito the modulation and coding scheme (MODCOD) of the PU espectively. Once the Modulation and Coding Classification (MCC) is completed, a Powe Contol (PC) scheme is enabled. The SU can attempt to access the fequency band of the PU and incease its tansmitting powe until it causes a change of the PU s tansmission scheme due to intefeence. When the SU detects the change of the PU s MODCOD, then it educes its tansmitting powe to a lowe level so as to egulate the induced intefeence. The poposed blind Adaptive Powe Contol (APC) algoithm conveges without any intefeence channel infomation to the afoementioned intefeence limit and guaantees the pesevation of the PU link thoughput. Keywods Cognitive Communications, Spectum Sensing, Modulation and Coding Classification, Highe Ode Statistics, Log- Likelihood Ratios, Code Syndomes, Adaptive Powe Contol, Adaptive Coding and Modulation I. INTRODUCTION Dynamic Spectum Access (DSA) is a key technique to achieve the coexistence of some sevices in specific fequency bands [1]. Towads that diection, the development of CC has enabled many aspects of DSA [2]. The ealization of the CC begins with the sensing pat. One way of enhancing CC with envionment awaeness is signal detection. This new adio must be able to identify all kinds of signals and a simple appoach can be the ecognition of thei modulation and coding schemes. This SS mechanism concening the modulation and coding detection is temed Modulation and Coding Classification (MCC) and has been ealized by extacting featues of the signal and classifying it based on them. As fa as modulation classification is concened, the featues exploited in this pape ae the signal cumulants of 2nd, 3d, 4th, 6th and 8th ode [3 6], which have distinctive theoetical values among diffeent modulation schemes and even though they demand a geat amount of samples, they ae easy to calculate. These statistical chaacteistics ae fed into a poweful classification tool, the SVM, which has been fequently used in the liteatue [4], [5], [7 11]. Fo the coding identification pat, the most common statistical featues in pevious wok ae the LLRs of the eceived symbol samples [12], [13]. The detection technique in this case involves the compaison of the aveage LLRs of the eo syndomes deived fom the paitycheck elations of each code. Anothe aspect of the DSA concept that has to be examined is the PC stategy unde which the SU is accessing the fequency band of the PU. An inteesting appoach in this topic, suitable fo a CC netwok is the distibuted one. In this case, we focus on a 2 2 cognitive channel, consisting of a PU link and a SU link and without any contol channel between them. Most studies in this field have employed iteative methods such as picing models o one bit contol channel and usually they povide a game theoetic famewok to pove thei convegence to an equilibium [14 16]. In this pape, an integated application is demonstated which concens an SU and focuses on both MCC and PC. The examined scenaio consides a PU link changing its modulation and coding scheme based on an ACM potocol and opeating in its assigned band togethe with an SU link enteing this band. In this wok, it is poposed the cognitive use to apply SS techniques in ode to contol its intefeence to the PU. The intefeence contol is achieved by having the coexisting cognitive SU constantly sensing the tansmission scheme of the PU, which changes dynamically based on the ACM potocol. The tansmitting powe is adapted wheneve it degades the modulation o coding scheme of the PU. The poposed DSA application concens only the SU s side without adding any complexity in the infastuctue o a contol channel between the two links in ode to exchange infomation about the channel o the induced intefeence and the APC mechanism is a simple powe scaling with a vaiable step. The emainde of this pape is stuctued as follows: Section II povides the system and signal model. Section III intoduces the MCC implementation. Section IV analyzes the APC technique. Section V shows the esults obtained by the combination of the above. Finally, Section VI gives the concluding emaks and futue wok in this topic. II. SYSTEM AND SIGNAL MODEL In this pape, the 2 2 cognitive system consists of a PU link and a SU link in the same fequency band as shown in Fig. 1. Futhemoe, the signal fom the PU link tansmitte is eceived

2 by the cognitive use using a seconday omnidiectional antenna only fo sensing and assuming popagation in an AWGN channel. As fa as the intefeence to the PU link is concened, this is caused by the tansmitte pat of the SU link to the eceive of the PU link. Consideing a LOS intefeence link, this may have a sevee effect on the modulation and coding scheme chosen by the PU link. Moeove, the intefeence fom the PU link to the eceive of the SU link is egaded to be negligible. In this scenaio, the fome intefeence is analyzed and it contibutes to the fomulation of the APC poblem. Fig. 1. RPU RSU PU Intefeence Link hi Seconday Use Link hsu The 2 2 cognitive system Pimay Use Link hpu Amplifie Modulato Data Souce Sensing Link hs SU MODCOD Classification Powe Contol In addition, the eceived symbol samples can be witten 1 as: TSU TPU SU [i] = h S s P U [i] + n SU [i] (1) whee h S is the sensing channel gain, s P U [i] is the tansmitted symbol fom the PU and n SU N (0, N SU ) is the Additive White Gaussian Noise (AWGN). On the PU side, the eceived symbol samples can be witten as: P U [i] = h P U s P U [i] + h I s SU [i] + n P U [i] (2) whee h P U is the PU channel gain, h I is the intefeence channel gain, s SU [i] is the tansmitted symbol fom the SU and n P U N (0, N P U ) is the AWGN. It also has to be emaked that the channels used in this pape ae flat and thei gains ae not vaying. Additionally, the tansmitting powes of the PU and the SU ae expessed as: and the SINR of the PU is defined as: P P U = E{s P U s P U } (3) P SU = E{s SU s SU} (4) 1 The SU achieves symbol synchonization in sensing the PU signal. ( hp U 2 ) P P U SINR P U = 10 log h I 2. (5) P SU + N P U Fom a system pespective, an abitay ACM scheme is adopted close to the technical specifications of the potocol set. Accoding to that, the PU cannot vay its tansmitting powe P P U, its tansmitted symbol s[i] can be of QPSK, 16QAM o 64QAM modulation scheme and the coding used in bit level is based on a binay low-density paity-check (LDPC) code of ates 1/2, 2/3, 3/4 o 5/6. Thei combinations povide us with the available MODCOD set: QPSK 1/2, QPSK 3/4, 16QAM 1/2, 16QAM 3/4, 64QAM 2/3, 64QAM 3/4 and 64QAM 5/6. In ode to incease the efficiency of the MCC, these blind identification techniques can opeate given a pedefined candidate set of modulation schemes and code ates. This means that if the cognitive use intends to access the fequency band of the PU link, it needs to have knowledge of the afoementioned MODCOD set and also of the eo coecting code the PU uses. III. MODULATION AND CODING CLASSIFICATION Statistical pocessing of communication signals can povide us with citical featues indicating thei natue. Assuming the signal model descibed in (1), we can obtain the 2nd, 4th, 6th and 8th ode mixed cumulants of the P U complex eceived signal C2,0, C2,1, C4,0, C4,1, C4,2, C6,0, C6,1, C6,2, C6,3, C8,0, C8,1, C8,2, C8,3, C8,4. Cumulants ae best expessed in tems of aw moments. A geneic fomula fo the joint cumulants of seveal andom vaiables X 1,..., X n is C X1,...,X n = ( π 1)!( 1) { } π 1 E X i (6) π B π i π whee π uns though the list of all patitions of 1,..., n, B uns though the list of all blocks of the patition π and π is the numbe of pats in the patition. Consequently, the pthode mixed cumulant C p,q of the complex eceived signal can de deived fom the joint cumulant fomula in (6) as: Cp,q = C,...,,,..., }{{}}{{} (p-q) times (q) times whee is the the complex conjugate signal. Because of the symmety of the consideed signal constellations pth-ode mixed cumulant fo p odd ae equal to zeo and also it can be easily poven that fo p even Cp,q = Cp,p q. The estimates of the pevious statistical chaacteistics ae going to be the featues fed into a patten ecognition stuctue which will decide the modulation scheme the signal belongs to. A poweful and new classification tool that pevious eseaches used is the SVM. Its mathematical foundation is statistical leaning theoy and it has been developed by Vapnik [17]. The SVMs opeate by finding a hypeplane in a high dimensional space which divides the taining samples in two classes. This hypeplane is chosen so that the distance fom it (7)

3 to the neaest data points on each side is maximized. Nonlinea sepaation of data is also possible with some small adaptations and using the kenel tick. An indiect mapping of input featue vectos into a highe dimensional space can be achieved in which they become linealy sepaable. The multi-class classification of a test signal into one of the 3 available modulation schemes of the ACM, the classes, is implemented by combining binay classifies to find to which class it most likely belongs compaed to evey othe one. Following this one-against-one appoach, the most usual stategy fo labeling a test signal is to cast a vote to the esulting class of each binay classifie. Afte epeating the pocess fo evey pai of classes, the test signal is assigned to the class with the maximum numbe of votes. As fa as the LDPC code ate classification is concened, pevious wok [12], [13] has been based on the unique paitycheck matix that each code ate has. A candidate LDPC encode θ has an exclusive paity-check matix H θ Z N θ Nc 2, whee N θ is the numbe of paity check elations of the candidate encode and N c is the length of the poduced by the encode θ codewod, which fo the examined LDPC code ates is always equal to Given a codewod c θ Z Nc 1 2 fom encode θ, in a noiseless envionment the following H θ c θ = 0 (8) holds ove the Galois field GF(2) if and only if θ = θ. Due to noise in the codewod though, some eos occu in (8) even when choosing the coect encode θ. These eos ae called code syndomes e k and fo a candidate encode θ in vecto fom they ae defined as e θ = H θ c θ (9) whee e θ Z N θ 1 2 and evey line epesents a paity-check elation. In ode to use the code syndomes e θ in code ate identification, a soft decision metic was intoduced in [18] and exploited by late eseaches. This featue is the aveage LLR of the code syndomes and it is consideed as a eliability estimate of the syndomes. To compute this, one needs to calculate the LLR of each bit of the codewod c θ, which afte some pocesses in the log-likelihood domain is obtained as LLR(c[m] SU [n]) = LLR( SU [n] c[m]) (10) whee c[m] is the consideed bit and SU [n] is the coesponding eceived symbol sample. This is the esult of the log-likelihood soft decision demodulation. Subsequently, if e k θ is the syndome deived fom the k th paity check elation of the candidate encode θ e k θ = c[k 1] c[k 2 ]... c[k Nk ] (11) whee N k is the numbe of codewod bits taking pat in the XOR opeations of the paity check elation, then the LLR of e k θ is given by ( Nk ) LLR(e k θ ) = 2tanh 1 tanh (LLR(c[k q ])/2). (12) q=1 Finally, the aveage syndome LLR is calculated as Γ θ = N θ k=1 LLR(e k θ ). (13) N θ Once, the aveage syndome LLRs of all the candidate encodes ae calculated, the estimated encode can be identified as ˆθ = ag max Γ θ (14) θ Θ whee Θ is the set of the LDPC encode candidates. IV. ADAPTIVE POWER CONTROL The pupose of the MCC pocedue is to act as a feedback to a closed-loop PC algoithm, which will instuct the SU how to egulate its tansmitting powe and thus the induced intefeence to the PU. Based on this PC scheme, the cognitive use does not need to exchange infomation with the PU and obtain any diect knowledge of the induced intefeence. A blind method fo mitigating the intefeence is the SU to adapt its powe with adjustable steps and monito the eaction of the PU. Simila PC schemes exist in liteatue [14] with poven convegence to the optimum solution. In this pape, a compaable algoithm is poposed consideing an AWGN intefeence channel. In this cognitive scenaio, the SU tansmitting powe P SU must convege to an unidentified theshold P max ove which it causes the PU to lowe its MODCOD. The suggested iteative APC algoithm, pesented in Algo. 1, is a heuistic method fo solving this PC poblem with a minimal numbe of P max violations. Initially, a desciption of its paametes must be given. M ODCOD(n) is the sensed tansmission scheme of the time instant n, P min is the minimum powe the SU can tansmit, N P LV is the numbe of P max violations fom the beginning of time, N max is the maximum acceptable numbe of the P max violations, (n) is the adjustable tansmitting powe step and T p is the peiod afte a P max violation duing which the P SU is set to a powe level below P max. Algoithm 1 Adaptive Powe Contol algoithm Sense M ODCOD(0) Tansmit P SU = P min Sense M ODCOD(1) if MODCOD(1) MODCOD(0) then Do not tansmit at all else Incease P SU by step (1) end if epeat Sense M ODCOD(n) if MODCOD(n) MODCOD(n 1) then Set P SU to pevious level and epose fo time T p else Incease P SU by step (n) end if until N P LV N max o P SU conveges

4 Accoding to this APC method, the SU stats tansmitting the minimum P SU and then gadually boosts it until a P max violation occus with inceasing step (n), which depends on its pevious value (n 1). Afte evey P max violation, the SU sets P SU to the pecedent level not alteing the modulation scheme of the PU, eposes fo a peiod of time T p and afte that stats inceasing it again. The taget of the algoithm is the moe P max violations happen the moe cautious the SU should become to incease P SU. This is achieved by detemining T p as an ascending function of N P LV and (n) as a descending function of N P LV. Eventually, P SU conveges to a value below P max without beaching this powe limit many times. V. RESULTS In this section, the pefomance of the MCC method and the pogess of the P SU and thoughputs vs time ae pesented. Initially, it must be noted that the eceived PU signal though the sensing link is of lowe SNR level than the one in the eceive of the PU link. Additionally, the pefomance of the MCC method is tested in the SNR ange of [ 11, 14]. Also, the numbe of symbol samples consideed to be sensed in the simulations is N s = which fo QPSK, 16QAM and 64QAM constellation schemes coesponds to 2, 4 and bit fames espectively. Moeove, the taining and testing pocedues wee pefomed using numbe of the signals N tain = and N test = 1000 fom each modulation scheme. The metic used to measue the detection pefomance of the MCC method fo a class j is the pobability of coect classification (P cc ), which is defined as: P cc = N cc N test (15) whee N cc is the numbe of coectly classified signals of class j. Pcc Fig. 2. QPSK 1/2 QPSK 3/4 16QAM 1/2 16QAM 3/4 64QAM 2/3 64QAM 3/4 64QAM 5/ SNR (db) P cc vs SNR fo Ns = symbol samples In Fig. 2, the P cc of the simulations is shown. Initially, an obvious emak is that the highe the SNR of the test signal, the highe the P cc. Futhemoe, one can notice that the lowe the ode of the constellation o the code ate to be classified, the easie it is to ecognize it. Also, the P cc cuves ae vey steep, mostly due to the pefomance of the code ate classifie. One moe conclusion which has to be noted is that fo P cc = 1 in all classes, the minimum equied SNR is 14dB. Following, the pogess of the P SU and thoughputs vs time ae pesented based on the APC algoithm descibed in the pevious section. The examined scenaio consides a cognitive SU, that ecognizes pefectly the tansmission scheme of an ACM PU link. The sensing of the PU signal is implemented with an omnidiectional seconday antenna of low gain. In Fig. 3, the P SU vs time diagam can be seen, whee the initial P SU and the unknown theshold P max ae consideed to be 15dBm and 25dBm espectively and the tansmitting powe update happens evey 100ms. Seconday Use Tansmitting Powe (dbm) Time (ms) Fig. 3. P SU vs time The main pinciples of the APC algoithm can be obseved in the P SU diagam. At the beginning, P SU inceases aggessively, until a P max violation occus. Afte each violation, it can be seen that the SU ests to a non violating value of P SU fo a peiod popotional to the total numbe of violations. Also, the moe violations the SU pefoms, the moe eluctant it becomes to incease its powe and finally it conveges to the acceptable P max = 25dBm. Using a paticula set of paametes, only 3 times the SU exceeds the unknown powe limit and it equies 45 powe adjustments to achieve that. Anothe aspect of the APC algoithm is pesented in Fig. 4. Hee, the thoughput of the SU, the PU and the total one can be viewed in time. They ae depending on the instant value of P SU and what has to be maked is the distinct thoughput dops of the PU and in total wheneve a P max violation occus and the convegence of the last one to a maximum value. This poves that a consideable total thoughput gain is achieved using Algo. 1 while peseving the PU thoughput level.

5 Thoughput (Mbps) Total Thoughput PU Thoughput SU Thoughput Time (ms) Fig. 4. The SU, PU and Total Thoughputs VI. CONCLUSIONS In this pape, an integated solution fo intefeence management in a CC context is poposed using a poweful MCC technique as feedback fo a closed-loop PC algoithm. The MCC technique exploits HOS featues and aveage code syndome LLRs of the PU signal in ode to detect even in low SNR level when the PU tansmission scheme changes and thus adjust P SU to innocuous values. The poposed APC method pefoms a powe scaling with flexible steps, so that the induced to the PU intefeence is mitigated. Though simulations, it is shown that the pefomance of the suggested system is excellent with contollable chaacteistics which affect convegence speed and numbe of P max violations. Tackling a numbe of ideal assumptions of the descibed system model can guide ou futue wok. The fist assumption, concening the code ate ecognition stage, is that the cognitive eceive is able to identify the beginning of a PU fame. Solutions on fame synchonization exist in the liteatue and they can be used to complete this wok. Futhemoe, ealistic channel modelling can be taken into account, like fading. Anothe pactical concept which can lead to futue wok is the intoduction of channel pediction in ode to use it in a PC algoithm. Also, futhe studies can focus on the behaviou of the PC scheme, like its convegence consideing a lage numbe of SUs o P cc < 1. Finally, othe PC algoithms can be examined which ae based on picing and tested using a vaiety of utility and pice functions. VII. ACKNOWLEDGEMENT This wok was suppoted by the National Reseach Fund, Luxemboug unde the CORE poject SeMIGod: SpEctum Management and Intefeence mitigation in cognitive adio satellite netwoks. REFERENCES [1] Q. Zhao and B. Sadle, A Suvey of Dynamic Spectum Access, IEEE Signal Pocessing Magazine, pp , [2] J. Mitola, Cognitive adio an integated agent achitectue fo softwae defined adio, Ph.D. dissetation, KTH Royal Institute of Technology Stockholm, Stockholm, Sweden, [3] O. Dobe, A. Abdi, Y. Ba-Ness, and W. Su, Cyclostationaity-based modulation classification of linea digital modulations in flat fading channels, Wieless Pesonal Communications, pp , [4] M. Petova, P. Mahonen, and A. Osuna, Multi-class classification of analog and digital signals in cognitive adios using Suppot Vecto Machines, 7th Intenational Symposium on Wieless Communication Systems (ISWCS), pp , [5] R. Kannan and S. Ravi, Second-ode Statistical Appoach fo Digital Modulation Scheme Classification in Cognitive Radio using Suppot Vecto Machine and K-Neaest Neighbou Classifie, Jounal of Compute Science, pp , [6] B. Ramkuma, Automatic Modulation Classification fo Cognitive Radios Using Cyclic Featue Detection, IEEE Cicuits and Systems Magazine, pp , [7] D. Liu and J. Liu, A Novel Signal Recognition Algoithm Based on SVM in Cognitive Netwoks, 12th IEEE Intenational Confeence on Communication Technology (ICCT), pp , [8] K. M. Thilina, K. W. Choi, N. Saquib, and E. Hossain, Patten classification techniques fo coopeative spectum sensing in cognitive adio netwoks: SVM and W-KNN appoaches, IEEE Global Communications Confeence (GLOBECOM), pp , [9] Z. Dandan and Z. Xuping, SVM-Based Spectum Sensing in Cognitive Radio, 7th Intenational Confeence on Wieless Communications, Netwoking and Mobile Computing (WiCOM), [10] H. Yang, X. Xie, and R. Wang, SOM-GA-SVM Detection Based Spectum Sensing in Cognitive Radio, 8th Intenational Confeence on Wieless Communications, Netwoking and Mobile Computing (WiCOM), [11] H. Liu, D. Yu, and X. Kong, A New Appoach to Impove Signal Classification in Low SNR Envionment in Spectum Sensing, 3d Intenational Confeence on Cognitive Radio Oiented Wieless Netwoks and Communications, [12] T. Xia and H. Wu, Novel Blind Identification of LDPC Codes Using Aveage LLR of Syndome a Posteioi Pobability, IEEE Tansactions on Signal Pocessing, pp , [13] R. Moosavi and E. Lasson, A Fast Scheme fo Blind Identification of Channel Codes, IEEE Global Telecommunications Confeence (GLOBECOM), pp. 1 5, [14] J. D. Hedtne and E. K. P. Chong, Analysis of a Class of Distibuted Asynchonous Powe Contol Algoithms fo Cellula Wieless Systems, IEEE Jounal on Selected Aeas on Communications, pp , [15] C. U. Saayda, N. B. Mandayam, and D. J. Goodman, Efficient Powe Contol via Picing in Wieless Data Netwoks, IEEE Tansactions on Communications, pp , [16] T. Alpcan, T. Basa, R. Sikant, and E. Altman, CDMA Uplink Powe Contol as a Noncoopeative Game, Poceedings of the 40th IEEE Confeence on Decision and Contol, 2001, pp , [17] V. N. Vapnik, The Natue of Statistical Leaning Theoy. Spinge, [18] J. Hagenaue, E. Offe, and L. Papke, Iteative decoding of binay block and convolutional codes, IEEE Tansactions on Infomation Theoy, pp , 1996.