A Full Duplex Transceiver with Reduced Hardware Complexity

A Full Duplex T ranscei ver with Red uced Hardware Comple xity Mus tafa Ema ra , Patrick R oss on , Kilia n R oth , David Da sso n ville Next Gener atio n and Stand ard s, Intel Deu tsch land Gmb H, Neu biberg, Germany Emai l: musta fa.em ara, kilian.r oth @intel.co m CEA, L ETI , MIN A TEC Camp us, F-3 805 4 Grenob le, France Emai l: p atric k.ro sson , david.d assonville @cea .fr Abstract —For future wireless communication syste ms, fu ll dup lex is seen as a possible solution to the ever presen t spec trum shortage. The key asp ect to en able In-Band Full Duplex (IBFD ) is sufficient cancellation of the unav oidable Self-In terfe rence (SI). In th is work we ev aluate th e performance of a low complexity IBFD transceiver , includ ing the required analog and digita l interference cancellati on techniqu es. The Radio Frequenc y Self - Interference Canceler (RFS IC) is b ased on the isolation of a circulator in combin ation with a vector modulator regenera ting the in terference signal, to destructively combine it with t he receiv ed si gnal. On th e digital sid e, a Digit al Self-In terf erence Cancellati on (DS IC) algorith m based on non-l inear adaptiv e filterin g is used. With the simp lified an alog front-en d of a Software Defi ned Radio (SDR ) p latform, SI cancellati on of 90 dB is achieved with the p resence of a receiv ed signal. Index T erms —In-Band Full-D uplex, Self In terference Cancel- lation I . I N T R O D U C T I O N Incr easin g the spectr al efficie ncy is cr uci al fo r ce llular system s, esp ecially in the con text of spec tru m scarci ty bel ow 6 G Hz. IBFD transc eivers rep resen t a m ean to ac hieve th is goal , since th e same time-freq uen cy res our ces are u sed for both transm ission an d rec eptio n [1]. M oreove r, IBFD p rovid es a theo ret ical spectr al efficiency gain o f up to two, for a poin t-to -po int link com par ed to a Frequen cy Division Du ple x (FDD ) system, by incr easin g th e a vailable band wid th if cer t ain con ditio ns are met [2]. Co mp ared to a Time Division Dup lex (TDD ) sy stem , the nu mber of available tim e-slo ts for U p L ink (UL) an d Down Link ( DL) is in crease d. It has b een shown in [3] th at signific ant g ains are ob tained at th e sy stem level, even w ith limite d inte rfer ence ca ncell atio n capab ilitie s. E ven fo r a system th at o nly ac hieves 65 dB of Self- Inte rfer ence Can cella tion (SIC), it is po ssible to im prove the average down link an d uplin k th rou gh pu t by 21 % an d 4.9 % respec tively . Th is perfo rm ance im provem en t incr eases to 69 % an d 81 % f or 8 5 d B o f can cellatio n. The m ain chal leng e o f IBFD is the can cella tion of the SI resultin g f rom the tr ansm it sig nal an d leakin g into the r ece iver fro nt-en d. Sin ce the differen ce b etwee n the SI and the recei ved signal co uld easily exce ed 90 dB, several stag es are req uir e d to red uce the SI p ower . Firstly , d ifferent an tenn a conce pts ca n reduce the SI power . High iso lation can be p rovid ed eith er by using a circu lato r with wel l adap ted an ten nas o r d ual- pola rized an ten nas f or tran smissio n an d recep tion [4]. Utilizi ng the form er so lu- tion assum es chan n el recip ro city in poin t to p oint scenar io s, wher eas th e latte r addresses distinctive users with d iffer ent chan nels [5]. Seco ndly , to avoid satu ratio n of the Low N oise A mp lifier (LNA), it is impo rtan t to ad d a RFSI C that atten uates the SI at the r eceiver’ s f ron t-en d. Th is is usu ally ach ieved by a com bin ation of p assive an d/or ac tive tech niq ues [6] [7]. T h e latter ai ms at r egener atin g th e SI sig nal and then d estru cti vely com bin ing it with the actual received sig nal. On the other han d, th e passive tech niq ues ar e gen eral ly bas ed on filterin g (nar row or b ro adb and ) o f th e tran smitte d sig nal bef ore su b- tractin g it fro m th e rem ainin g rec eived sign al. After the RFSIC, the resid ual signal is m ixed to th e analo g baseb and an d th en d igita lized . It is impo rtan t that th e d yna mic rang e of the A nalo g to Dig ital Converter (AD C) cover s th e super po sition o f the SI an d the Sign al o f In terest (So I). Mor eover, th e RFSI C sho uld n ot add n on -coh eren t no ise that cann ot be rem oved th rou gho ut th e SIC pr oce ss. Moving to the DSIC, its o bje ctive is to cance l th e rem ain ing SI d own to the system ’ s n oise floo r . Th eref ore, it is n ecessar y to pro per ly cance l the rem ainin g linea r as well as th e non linea r com pon ents o f th e SI sig nal . It was shown in [8] that a 1 10 dB SI C level was rea ch ed over an 80 MHz ban dwid th at 2 .4 GHz. Nevertheless, this solutio n in clud es a co mp lex and exp ensive 1 6-ta p Rad io Freq uen cy (RF) filter . Th e se tup in [9] con sider ed a relat ive ly small ban dwid th of 625 kHz, but provid ed u p to 100 dB of self inter feren ce can cellatio n, incl ud ing 41 dB o f an ten na iso l ation by usin g separ ate Tx an d Rx ante nn as. The o bject ive o f this work is to r each an acce ptab le SI C level, w ith a less co mp lex solutio n. For this setup , a classi ca l SDR r adio is u sed with on ly 12 bits resolu tion f or bo th th e Digita l to Analog Converter (D A C) and the ADC. I n contr ast to previo us wo rk, the investigatio n was exten ded to cover th e presen ce of a rec eived SoI. The paper is org aniz ed as follows. I n Se ction II, the d esign of the imple men ted RFSIC alo ng with th e algor ithm of the DSIC ar e p resen ted . Ad ditio nal ly , the sy stem setup and the ev alu ation resu lts are disc ussed in Sectio n III . Finall y , con clusio n o f th e work is su mm ariz ed in Se ction I V. RFSIC DSIC T ran smit signal Rx fro nt-en d Tx fro nt-en d Receiver P A LNA SoI BB BB Fig. 1. Full duplex transcei ver system diagra m. I I . F U L L D U P L E X T R A N S C E I V E R D E S I G N A. System Mo del The system m odel investigated th rou gh out this work is shown in Figur e 1. The gen erat ed baseban d signal den ote d by first pr opag ates th ro ugh the transm itter ’ s fro nt- end , resultin g in the analo g sig nal at the Power Amplifi er (P A) outp ut P A . Afte rward s, th e sign al pas ses v ia a dir ectio nal cou pler to th e ci rcula tor f or tran smissio n an d th e coup led signal g oes towards the RFSIC funct ion . The rec eived sign al den oted by , in clud es the SoI alo ng w ith th e SI signal . After the RFSI C process, th e sig nal LNA den otes th e in pu t to the LNA. This signal is then processe d by the rec eiver , resultin g in th e d igital baseb and sig nal SoI BB whic h inco rpo rates the resid ual linear and no nlin ear SI BB , an d the rec eived SoI SoI . T he DSI C b lock is then utilize d to com pute an estimat e of the SI sign al den oted by BB whic h is subtr acted fro m the actual SI signal resultin g in the resi dual signal after th e d igita l cance llatio n sta ge. B. Radio F requen cy Self- Inte rference Can cella tion The RFSIC sp ans three mai n requir em ents; consider ab le SI red uctio n, low noise ad ditio n an d stable sign al track ing . The fir st req uir em ent aim s at avoidin g th e sa turat ion of th e receiver’ s fron t-en d, esp ecially the AD C [10]. Acc ord ingl y , the receiver g ain nee ds to be ad apte d to the smaller r eceived signal ’ s p ower . I n this pap er, the usage o f th e circula tor wa s emp loyed as a so lutio n to th is aspect. Th e secon d imp orta nt requ irem en t addr esses the added no ise by the RFSI C. If the RFSI C intro duce s ad ditive n on -co heren t no ise, its digi ta l subtr actio n will be im po ssible , du e to the ran do m natu re o f the noise. T he last requ irem ent targets th e o scillatio ns ex isting in th e h ard ware circu it. T he last requ irem ent ta rgets a stab le SI sig nal tr ack ing cap abilit y . T he RFSIC has a capac ity to follow the SI ev olu tion in time, however , it shou ld chan ge slowly to allow th e digita l ca ncele r to co nverge. Mor eover, any power su pply har mo nics that ind uce s fast o scillatio ns at th e RF cance ler m ust b e p erfec tly filter ed. Glo bally , the RFSI C must redu ce th e SI in order to use th e r eceiver gain in th e corr ect settin g an d no t d egrad e th e r eceiver’ s no ise fig ur e. W itho ut the RFSIC, the SI received at the LN A’ s in put can be mo deled as LNA P A LNA (1) The term s P A corr espo nd to th e multi- path com- pon ents co min g f rom th e P A and goin g towards the L NA. It con sists o f ci rcu lator leak age, anten na ret urn loss and ove r th e air scattered wave cl ose to the antenn a [ 11]. The seco nd term LNA corr espo nd s to the two way traveling wa ve due to th e retu rn loss of the r eceiver an d the tr ansm itter . It is ob served that LNA P A due to the ado pted tran sceiver’ s ar ch itectu re. However , th e secon d t er m is no t n eglig ible if 100 dB of can cellatio n is consid ered , as it prov ides ro om fo r f urth er en h ance men t when m odel ed. In ord er to redu ce the SI, th e RF can celer thr ou gh this pap er is p ro pose d to be added b etwee n th e ou tpu t of th e P A and th e inp ut o f th e LNA. Add ition ally , a direc tional co up le r is inse rted at the ou tput of the P A to cap ture the tran smitte d signal . This sign al is then delay ed with a fixed delay an d after wards atte nuat ed an d ro tated b y a vecto r mo dula tor . T h e vector mod ula tor’ s o utpu t is then in jected v ia ano ther dire ctive cou pler bef ore th e LNA. Due to th e lim ited d irec tivity of th e cou pler s or its retur n loss, o ther un desir ed signal s are in j ec ted in the RF cance ler . It is imp orta nt to note that direc tive cou pler s in du ce loss on the dir ect path whic h red uce the tran smitted power on the T x sid e, an d d egrad e the n oise fig ur e on the Rx side. T o com bat this drawbac k, high directivity cou pler s are ad vise d. T o avoid th e usage of an ac tive amplifi e r in the RFS IC pat h, the overal l min imu m loss on th e RFSIC must be lower th an the ante nn a sy stem isola tion . W ith an ac tive amp lifier in th e RFSIC wo rkin g in the linear zon e, th e SI sig nal received at the LNA ca n be mo del ed as LNA P A LNA (2) wher e mod els the n oise gen erated wh en th e RFSIC is add ed. T he vecto r mod ula tor is co ntr olle d in o rder to r edu ce the overall r eceived power . I t can b e obse rved that w ith one fixed delay and o ne variable complex pat h, it is imp ossib le to cance l all th e leak age pat hs pr esen t in th e initia l con figu ra tion . Nevertheles s, when th e delay is tun ed accord ing ly and the vector m odu lato r is co rrec tly set, th e stron ger leak ag e pat h s can be mitigat ed. It is worth mentio nin g that if the RFSI C works in th e no nlin ear zone, th e previo us model need s to be upg rad ed, tak ing into ac cou nt the added n onlin ear effect . Finally, the analo g signal LNA is freq uen cy down converted and then A/D converted. Thro ug h the co min g par t, th e DSI C’ s desig n will be presen ted. C. Digital Sel f-In terference Can cella tion The main o bject ive of the dig ital cance llation is to cance l the res idu al SI after the analog cancella tion. Th is incl ude s the non linea r com pon ents, in p articu lar th e rem ain ing no nlin e ar- ities ad ded by th e RFSIC and the ci rcu lator [8]. Thr oug ho ut this wor k, a no nlin ear ad ap tive digita l can- cellatio n so lutio n based on th e wo rk in [12], utilizin g th e tran sversal recur sive least squar es is adop ted. A pre-ad ap tatio n orth og onal izatio n based on th e Chole sky d eco mpo sitio n is carr ied ou t to fu rth er en han ce th e ad aptatio n p ro cess. A prop er m od eling of th e SI signal affects the system design and perf orm ance . Inspired by [8] an d [13], the overall SI baseb and signal can be mod eled using a paral lel H amm erstein mod el BB (3) wher e dep icts th e mem ory d ep th of the mo del and is the non linea rity order . The sy mb ol repr esen t the th ord er chan nel coefficien ts o f the effective SI chan nel an d den otes the n onlin ear b asis funct ion o f th e b aseb and sig nal . The signal afte r th e D SIC dep en ds on the SI ch an nel coefficien ts , th e received sign al of inte rest SoI and the existe nt Add itive White Gaussia n No ise (A WG N) in th e receiver as fo llows SoI SoI Res idual SI (4) The main target is to p rovid e an accu rate, fast and low- com plex estimation of th e SI chann el coefficients in order to regen er ate the SI signal. Since non linea r bas is fun ction s are gen er ated for every inco min g samp le, the b asis fu nctio ns acr oss differen t non l in - earity or ders are hig hly correl ated and have d ifferent vari - ance . T hus, a slow convergence an d a degrad ed can cellatio n perf orm an ce ca n be o bser ved while estim atin g th e SI effec - tive chan nel coefficie nts. Conse quen tly , an o rtho go naliz a tio n of th e basis fun ctio ns befo re the coefficie nts estimat ion is requ ired [13]. Th e covariance mat rix of the b asis funct ions acro ss sufficien tly large n umb er of sam ple s ca n be com pu ted as H (5) wher e is the exp ectatio n op erat ion . The vecto r T repr esen ts the in stan taneo us basis fun ction s o f the -th sam ple . A transf orm atio n of th e bas is fun ction is carr ied o ut via a wh itenin g tr ansf orm atio n matr ix based o n the Ch olesky dec omp ositio n H (6) wher e is a lower tria ngu lar mat rix w ith positive dia gon al entr ies. T he orth og on alized b asis funct ion s are com pu ted as (7) In order to simp lify the nota tion of the signal mo del in (4), i t can be ref or mula ted such that the data vec tor f or the p reviou s samp les are in clud ed as T T T T (8) wher e is the in put com plex d ata vecto r . Ap ply ing the same nota tion to the estimat ed SI chan nel coefficien ts can get T (9) wher e are the SI channel coefficient to be estim ated . From p lug ging (8 ) an d (9) into (4) , the res idu al sign al afte r the DSIC ca n b e r efo rmu lated to H (10 ) The p rop osed al gor ithm in this work is based o n th e Recur sive Least Squar es (RL S) algo rith m com bin ed w ith com plexity red uctio n techniq ues . I nspir ed by the work in [ 1 4], the sum mar y o f the pro po sed e xp onen tially weighte d RLS algo rithm is pr esente d in T ab le I. In the in itializat ion step r esidu al vecto r is set to th e covariance vecto r . Th e co rrel ation m atrix is se t to an eq u alizatio n m atrix , w her e is an iden tity matr ix of dim en sion . Th e par amet er is cho sen based on the Sign al-to -No ise-Ra tio (SNR) as [15]. T he p aram eter is th e forg etting factor an d is ch o sen as . The first ste p repr esen ts the u pd ate of the corr elatio n matr i x for each incom ing sam ple . Or iginal ly , th e u pd ate sh ou ld con sider all th e com ing in pu t data vecto r . Nevertheles s, followin g the station arity assu mptio n of the inp ut data, on ly the fir st com pon ents of the d ata vecto r are su fficient to reco nstru ct the com ple te correlat ion matr ix. Th ose com - pon ents are fu lly cap tured in . The no tatio n stand s for th e fir st rows of th e cor relat ion m atrix . Thu s, by explo iting the transver sal str uctu re of the in pu t data vect or, a redu ctio n in th e com pu tation al com plexity is achieved. An efficient so lutio n should b e utili zed to so lve step 4 , whic h constitu tes the com plexity bottle nec k o f the algo rit hm . T ABLE I E X P O N E N T I A L LY R E C U R S I V E L E A S T M E A N S Q U A R E S A L G O R I T H M Step Computa tion real real Initial ization : while transmit ting - - 1 H 2 BB H 3 4 5 - T otal: : T otal : - - This step res ults in comp utin g th e coefficients upd ate step alon g with th e resid ual vecto r . T he sym bols and stand f or the co m plexity of rea l mu ltiplic ation s an d add ition s of step 4. T he fo cus h as bee n d irecte d thro ug h- out this work towards th e D icho tom ou s Co or dinat e Desce nt (DCD) algo rithm due to its low co m plexity ad vantag e [16]. It has been sh own that th e max imu m nu mber of add ition s requ ired is up per bo und ed b y and the n umb er of m ultip licatio ns is zer o. I I I . S Y S T E M E V A L U A T I O N A. Platform Ove rview The test ben ch presen ted in this sectio n allows the per- for man ce evaluatio n of th e RFSI C and the DSIC on r eal signal with the pr esen ce of an exter nal received sig nal. A block diagram illu strat ing th e des ign ed test bench is shown in Figu re 2 . A PC run nin g Matla b is conn ecte d to a digita l boar d that supp or ts sim ultan eo us Tx /Rx data flows. The dig ital boar d is co nnec ted to a SD R ARRad io b oar d. This RF boar d inclu des an AD 936 1 SDR ch ip fro m Anal og Devices as th e RF transceiver . It supp or ts the converters ( ADC/DA C), the freq uen cy conversion (up /down) as wel l as a driver amplifi er and a LN A. The o utpu t RF sign al is then connec ted to an add ition al P A and th e RFS IC b oar d. As p resen ted ea rlier, w e con sider the cir cula tor so lutio n, wh ere th e an ten na h as o nl y one inpu t/ou tput por t. The an ten na is r eplace d by a static passive emula tor circuit wh ich em ulate s the reflectio n seen on an an ten na p or t. At th e sy stem level, w e co nsid er th at th e ci rcu lator b elon gs to the antenn a sub -sy stem. W e rec all that in IBFD sy stem con sider ation , two option s ar e possible depen din g o n th e sc e- nario ; either o ne sing le-p ort an tenn a with a circu lato r, or a two por t an ten na addr essing two an tenn a p atter ns (on e for Tx, on e for Rx) . T wo se par ate an ten nas falls into th e same sc enar io as a two p ort anten na. T hr oug ho ut th e fo llowing subsect ion, the sp ecific elem en ts o f the test setup are p resen ted. 1) Antenn a em ula tion with cir culato r: The usag e of an anten na em ula tion circuit repr esen ts an alte rnat ive to a r e al anten na. Su ch a cho ice was motivated by th e avoidan ce of the un certa inty ca used b y the time- varyin g environ men t, reflec tion s an d chan ging chan nel pro pag atio n, wh ich wou ld incr ease th e co mp lexity of the sy stem debug ging . The an ten n a emu lation circu it is an RF passive circu it whic h is con stan t over time an d has a kn own ret urn loss. This solutio n is a first ste p to investigate the d ifferen t in terf eren ce can cel la tion stages. T he time varying reflec tions with a real an tenn a w ill be stud ied in a seco n d step . Figu re 3 pr esents th e RF leakag e due to the com bin ation of the cir cula tor and th e ante nna em ulatio n circu it. The ci rcu lator leakag e with anten na em ula tion is app rox imatel y -2 0 dB at the ce nter f req uen cy o f 90 0 MHz and varie s abou t 1 dB ac ross a band wid th o f 20 MHz. Ad ditio nal time varying RF leakag e would origin ate fr om o ther inter feren ce such as p ower su pp l y cou plin g, acti ve cir cuit co uplin g and th e inte rnal co mp on e nts of the SDR chip . 2) R FS IC b oa r d: Th e RF cance ler b oar d co nnec ts th e RF tran sceiver boar d on o ne sid e and th e circu lato r on the o ther side. In side the RFSIC, a sam ple of the Tx sign al is vector- mod ula ted, del ayed an d amp lified , bef ore bein g add ed to the Rx sign al. T he vec tor-mo dula tor is co ntro lled by two anal og voltages of th e dig ital boar d. 3) T ransceive r b oa r d: The RF tran sceiver boar d is b ased on th e ARRA dio bo ard . It emb eds an AD 936 1 SDR chip [ 17]. A P A is ad ded to increa se the transm itted power to 20 dBm . Due to th e w ide- ban d cap abilit y of the SD R Integr ated Circu i t (IC), extern al ban d filter s are n ecessa ry to p revent aliasin g. As the A D93 61 has a g oo d no ise fig ure, no exter nal LNA is requ ired . 4) D igita l Bo ar d: The d igital b oar d contr ols the con figu - ratio n o f th e AD93 61 ; espec ially the Phase Lock Loop (PLL) and th e tran smitte r an d rec eiver gai n. It is able to send the IQ data to the tr ansc eiver b oar d an d syn chro no usly rec eive the IQ dat a of th e received signal f rom the tr ansceiver boar d . This bo ard has bee n d esign ed to supp ort simulta neo usly Fil t er Bank Multi- Carr ier ( FBMC) tran smissio n an d r ecep tion an d has th e ad ditio nal ca pacity to m anag e IBFD . T he d igital boar d is al so ab le to gen er ate two DC voltag es to con trol th e RFSIC. A G rap hic User In terfac e (G UI) ru nn ing on the PC allows to activ ate different SI ca ncele rs for the IBFD m od es (i.e. time- freq uen cy sy nch ron ized sa mple s generat ion an d acqui si- tion) . The dat a has u p to 12 b its r esolu tion for IQ data. In thi s pap er, we will focu s on te sts in the 9 00 MHz freq uen cy b and . The RF bandw idth is set to 56 M Hz an d the samp le rat e is set to 6 1.4 4 MHz. In th is co nfig ur ation , th e ch ip works in FDD mod e but the Tx Lo cal Osc illato r (LO ) an d the Rx L O are set to the sa me freq uen cy . It is also assu med that th e tran sm issi on and rec eptio n ar e synch ro nized in time. B. Results 1) R FS IC perfo rma nce: T his first showcase presen ts the RFSIC perf orm ance . Fig ure 4 shows the Power Spec trum Density (PSD) mea sured by a FSW signal analyzer of the tran smitted signal over mu ltiple stag es ag ain st the relati ve freq uen cy in MH z. It is o bser ved the RFSIC’ s effect and that with th e p rop ose d solu tion , a value o f 3 6 dB of ca ncella tion Digita l boar d CTRL DC voltages DSIC Tx Rx IP fram es Tx Rx Tx Rx ARRad io RFSIC Circu lator and an tenn a emu lation Fig. 2. System setup incorporating passi ve cancell ation, R FSIC and DSIC. Freq uen cy [MHz] Fig. 3. S21 Circulator with antenna emulation. Relative freq uen cy [MHz] P A ou tput SI RFSIC in SI RFSIC out RFSIC no ise flo or Spectrum ana lyse s noise floor Fig. 4. SIC performance of the IBFD transceiv er with RFSIC on ly . was r ealiz ed. T wo oth er sig nals are measu red ( noise floor of the RFSIC in bla ck, n oise floo r o f th e sig nal anal yze r in y ellow) to ind icate the ca nce llatio n lim its im pose d by th e emp loyed har dware com po nen ts. 2) DSIC per forma nce : This analy sis was f ocu sed o n the DSIC stand alon e case , where the RFSI C was sw itched o ff. In Figu re 5, the SI signal over th e differ ent stag es is shown. In Relative fr equ ency [M Hz] P A output SI DSIC in SI DSIC out Noise floor Fig. 5. SIC performance of the IBFD transceiv er with DSIC onl y and an attenua ted transmitted signal. ord er to p revent the satu ratio n of the r eceiver fron t-en d, t he tran smitted sig nal was attenu ated p rop erly as den ote d by th e DSIC in put in the men tio ned figu re. In this test, th e DSIC target is to cance l the lin ear and no nlin ear comp on ents o f th e tran smit sig nal . For the case af ter th e DSIC, it is ob served that n ot al l in terf eren ce can b e ca nce led with the ch ose n con figu ratio n. T wo r eason s f or th is beh avior are po ssible, either this rem ain ing in terf eren ce has a ran dom sour ce, or i t is n ot covered by o ur sign al mod el u sed f or th e cance llatio n. Due to th e imp erf ectio ns of th e use d rec eiver , the no ise flo or (in black ) is no t fla t in the fr equ ency d omai n. 3) Co mb ined perfo rman ce with receive signa l: The fin al perf orm an ce showcase is presen ted in Figu re 6, w here a SoI is rec eived sim ulta neo usly wh ile tr ansm itting a sign al. T o illustrat e the receptio n of th e signal , w e red uce d th e b and w idth of the sig nal tran smitte d fr om the sig nal gen erato r to 10 MHz . The SI is illu strat ed over th e mu ltiple can cellatio n stages . The RFSIC p erfo rman ce follows that of Figu re 4 , wher eas after the D SIC is now dom inated by the So I. This pr oof s Relative fr equ ency [M Hz] P A ou tput RFSIC out DSIC out SoI Noise floor Fig. 6. SIC per formance of the IBFD transcei ver (RFSIC and DS IC) with recei ved signal of interest. T ABLE II C A N C E L L AT I O N P E R F O R M A N C E . Po wer [dBm] Ca ncell ation [dB] Transmit signal afte r P A 20 - SI afte r circula tor -1 21 SI afte r RFSIC -37 36 SI afte r DSIC -71 34 tota l cancel lati on - 91 the sy stem’ s cap abilit y to ca nce l the SI con sider ably . Such cance llatio n perf orm ance en able s the succes s recep tion of the SoI regar dle ss of th e la rge power d ispar ity b etwee n th e tran smitted sign al an d th e SoI. A fin al su mm ary of the achieved cancella tion values is presen ted in T able II , wher e a total o f 91 dB ca ncella tion was realiz ed. I V . C O N C L U S I O N This work pr esen ted an I BFD arch itectu re b ased on a SDR-RF tran sceiver . The sim ple, d igita lly con trolle d RFSI C con sists of a fixed delay and a vecto r m odu lato r . A single anten na is use d f or tr ansm ission and rec eptio n in the inves- tigated setup . Th e p rop ose d so lutio n fo r th e D SIC is b ased on the expon entia lly wei ghte d tr ansver sal RLS-DCD w ith a pr e-o rtho go naliz ation stage of the non linea r b asis co mp o - nen ts. This solutio n was motiv ated by a good cancella tion perf orm an ce alo ng w ith a low-com plexity target fo r the d igi ta l cance ler . Th e system ’ s evaluation illustr ates th e effec ti veness of the pr opo sed ca nce llation solu tion s fo r both the RFSI C an d the DSI C. The sy stem setu p was furth er exten ded to includ e a rec eived SoI to em ulate p ractic al scen ario s of fu ll du plex system s dep loym ent. A C K N O W L E D G M E N T This work was su pp orte d b y the Fren ch p roje ct DU PLE X and the E uro pea n Com missio n in the f ram ework of th e H20 20- ICT -20 14 -2 pr oject Flex5 Gware( Gr ant ag ree men t no . 671 56 3). R E F E R E N C E S [1] F . Bocca rdi et al., “Five disrupti v e technology directi ons for 5G, ” IEEE Commun. Mag. , vol. 52, no. 2, pp. 74–80, Feb. 2014. [2] A. Goldsmith, W ireless Communications . Ne w Y ork, NY , USA: Cambridge Univ ersity Press, 2005. [3] S. Goyal et al., “Improving small cell capacity with comm on-carrier full duplex radios, ” in Int. Conf.on Communic ations (ICC) 2014 , June 2014, pp. 4987–4993. [4] “DUPLO D2.1: Design and measurement report for RF and ant enna solutions for self-interference cance llation, ” 2014. [On line]. A vai lable: http://www .fp7-duplo.eu/images/docs/Deliv erables/D2 1.pdf [5] T . Dinc , A. Cha krabarti and H. Krishnasw amy, “A 60 GHz CMO S dull-duple x transcei ver and link wit h polari zation-based antenna and RF cance llation , ” IEEE J . Solid- State Circuits , vol. 51, no. 5, pp. 1125– 1140, May 2016. [6] A. Sabharwal et al., “In-band full-duplex wireless: cha llenges and opportuniti es, ” IEEE J. Sel. Areas Commun. , vol. 32, no. 9, pp. 1637– 1652, Sept. 2014. [7] P . Rosson, D. Dassonvil le, X. Popon, and S. Mayrargue, “S DR based test bench to ev aluate analog cancellat ion techniques for i n-band full-duple x transcei ver , ” Fifth Internat ional W orkshop on Cloud T ech- nologie s and Energy Efficiency in Mobile Communication Netw orks (CLEEN) , 2017. [8] D. Bharadia , E. McMilin and Sj. Katti, “Full duplex radios, ” in Pr oceedings of the ACM SIGCOMM 2013 , Hong Kong, China, Aug. 2013, pp. 375–386. [Online]. A v ailable: http://do i.acm.org/10.1145/2486001.2486033 [9] M. Duarte, C. Dick and A. Sabharwal, “Experiment-dri v en characte ri- zation of full-duplex wireless systems, ” IEEE T rans. W ireless Commun. , vol. 11, no. 12, pp. 4296–4307, Dec. 2012. [10] K. E. Kolodzi ej, J. G. McMichael and B. T . Perry, “Multi t ap RF can- celler for in-band full-duplex wireless communications, ” IEEE T rans. W ir eless Commun. , vol. 15, no. 6, pp. 4321–4334, June 2016. [11] X. Wu, Y . Shen and Y . T ang, “The po wer delay profile of the s ingle- antenna full-duplex self-inte rference channe l in indoor e n vironments at 2.6 GHz, ” IEEE Antennas and W ir eless Propa gation Lett. , vol. 13, pp. 1561–1564, 2014. [12] M. E mara, K. Roth, L. G. Baltar , M. Faerbe r , and J. Noss ek , “Nonlinear digital self-interference ca ncellat ion with reduced comp lexit y for full duplex systems, ” in WSA 2017; 21th International ITG W orkshop on Smart Antennas , March 2017, pp. 1–6. [13] D. Korpi et al., “Digital self-interference cance llat ion under nonideal RF components: adv anced algorithms and measured performance , ” in Int. W orkshop on Signal Processing Advances in W irel ess Communi cations (SP A WC) 2015 , Stockholm, Sweden, June 2015, pp. 286–290. [14] Y . V . Zakharo v , G. P . White and J. Liu, “Low-comple xity R L S algo- rithms using dichotomous coordinate descent iterati ons, ” IE EE T rans. Signal Proce ss. , vol. 56, no. 7, pp. 3150–3161, July 2008. [15] S. Haykin, Adaptiv e F ilter Theory (3rd Ed.) . Upper Saddle River , NJ, USA: Prentice-Hall , Inc., 1996. [16] Y . Zakharov and F . Albu, “Coordinat e descent itera tion s in fast affine projecti on al gorithm, ” IE EE Signal Pr ocessing Lett. , vol. 12, no. 5, pp. 353–356, May 2005. [17] AD9361 data sheet , ANALOG DEVICES, av aila ble at http://www .analog.com/media/en/technical-documentat ion/dat a- sheets/AD9361.pdf.