Near-Optmum Blnd ecson Feedback Equalzaton for ATSC gtal Televson Recevers Hyoung-Nam Km, Sung Ik Park, Seung Won Km, and Jae Moung Km Ths paper presents a near-optmum blnd decson feedback equalzer (FE) for the recevers of Advanced Televson Systems Commttee (ATSC) dgtal televson. By adoptng a modfed trells decoder (MT) wth a trace- back depth of for the decson devce n the FE, we obtan a hardware-effcent, blnd FE approachng the performance of an optmum FE whch has no error propagaton. In the MT, the absolute dstance s used rather than the squared Eucldean dstance for the computaton of the branch metrcs. Ths results n a reducton of the computatonal complexty over the orgnal trells decodng scheme. Compared to the conventonal slcer, the MT shows an outstandng performance mprovement n decson error probablty and s comparable to the orgnal trells decoder usng the Eucldean dstance. Reducng error propagaton by use of the MT n the FE leads to the mprovement of convergence performance n terms of convergence speed and resdual error. Smulaton results show that the proposed blnd FE performs much better than the blnd FE wth the slcer, and the dfference s promnent at the trells decoder followng the blnd FE. Keywords: TV, VSB, ATSC, equalzer, blnd FE. Manuscrpt receved Apr., 23; revsed Jan. 3, 24. Ths work was supported n part by Electroncs and Telecommuncaton Research Insttute (ETRI). Hyoung-Nam Km (phone: +82 2394, emal: hnkm@pusan.ac.kr) s wth the epartment of Electroncs Engneerng, Pusan Natonal Unversty, Busan, Korea. Sung Ik Park (emal: ps76@etr.re.kr) and Seung Won Km (emal: swkmm@etr.re.kr) are wth gtal Broadcastng Research vson, ETRI, aejeon, Korea. Jae Moung Km (emal: jaekm@nha.ac.kr) s wth the Graduate School of Informaton Technology and Telecommuncatons, Inha Unversty, Incheon, Korea. I. Introducton ecson feedback equalzers (FEs) are commonly used n dgtal communcaton systems to suppress ntersymbol nterference [], [2]. Advanced Televson Systems Commttee (ATSC) dgtal televson (TV) recevers have also used FEs to equalze the 8-vestgal sdeband (VSB) sgnal, whch s the transmsson standard of ATSC terrestral TV [3]. In a conventonal FE, the data passed nto the feedback secton s the slcer output and, hence, no longer contans any nose, thus ncreasng the accuracy of the nterference cancellaton. However, ths advantage s meanngful only when the slcer output s correct. When there s no tranng sgnal and the channel eye s closed, the FE may have a problem convergng ts tap coeffcents snce the probablty that the slcer mght decde ncorrectly ncreases. Such decson errors result n error propagaton throughout the feedback loop [4]. In a FE for ATSC TV recevers, the error propagaton phenomenon may serously affect the convergence performance because a blnd equalzaton or decson-drected equalzaton usng slcer outputs has to be carred out n most receved symbols. When the tranng sequence exsts, the FE uses the tranng sequence for the feedback flter nput. However, when there s no tranng sequence durng the data segments, the slcer output s generally fed nto the feedback secton. As the symbol error rate (SER) can be as hgh as.2 and the tranng sequence s very short n the terrestral TV recever, error propagaton s unavodable durng the recepton of data symbols []. Ths results n deteroraton of the performance n terms of convergence speed and resdual meansquare error (MSE) because the advantage of the FE s no longer vald. In [6], the error propagaton phenomenon exstng n the ETRI Journal, Volume 26, Number 2, Aprl 24 Hyoung-Nam Km et al.
FE of ATSC TV recevers and the performance upperlmts of the FE were analyzed and presented by comparng error-propagaton and no-error-propagaton cases. For one approach to the performance lmt, the authors consdered a blnd FE adoptng a trells decoder (T) wth a trace-back depth (TB) of as a decson devce. The TB has been set to to remove symbol delay caused by the T because the low-order tap coeffcents of the feedback flter have a large mpact on the equalzer performance. Whle adopton of the T n the FE decreases the decsonerror probablty, t results n an ncrease of computatonal complexty due to the twelve Ts requred for trells code denterleavng [7]. In ths paper, to reduce the computatonal complexty wthout performance loss, we modfy the conventonal T by usng the absolute dstance nstead of the squared Eucldean dstance for computng branch metrcs. The modfed T provdes a near-optmum FE to mnmze the error propagaton as a near-best decson devce. Compared to the conventonal slcer, the MT lowers decson-error probablty and thus reduces error propagaton n the FE. The reducton of error propagaton enhances the performance of the FE, approachng the performance level of an optmum FE. Ths paper s organzed as follows. In secton II, we ntroduce the FE commonly used for ATSC TV recevers and address the error propagaton effects n a FE. As an effectve method to overcome the error propagaton, n secton III we present a blnd FE adoptng the MT wth a trace-back depth of for the decson devce. The performance enhancement of the proposed method s presented through smulaton results n secton IV. Fnally, secton V concludes ths paper. II. FE for ATSC TV Recevers The 8-VSB sgnal s transmtted n frames, as shown n Fg.. Each data frame s composed of two data felds, each contanng 33 segments, of whch the frst segment s the feld sync segment, followed by 32 data segments. Each data segment s composed of 832 symbols, of whch the frst four symbols are the segment sync symbols (, -, -, ) and the remanng 828 symbols are Reed-Solomon-encoded, nterleaved, and trells-encoded symbols drawn from the 8- level pulse ampltude modulaton constellaton (±, ±3, ±, ±7) [3]. The feld sync segment s used for the tranng sequence of the equalzer. urng the feld sync segment, the FE operates wthout error propagaton because the known sgnal s fed nto the feedback flter. However, snce the feld sync segment arrves only once every feld, correspondng to 24 ms, the overall rate of convergence of the equalzer can be qute slow f the adaptaton 33 segments (=24.2 ms) 33 segments (=24.2 ms) 4 828 symbols Segment sync Feld sync # ata+ FEC Feld sync #2 ata + FEC segment (= 77.3 µs) Fg.. An 8-VSB data frame. of the equalzer s carred out only durng the feld sync segment. In addton, such an adaptaton polcy s not effcent for tme-varyng channels. ecson-drected () adaptaton at the data segments does not work well. Snce the SER of the symbols obtaned from the equalzer output sgnal s about.2 at the threshold of vsblty (TOV) [], and the FE has 2 more taps, one error feedback can affect 2 more outputs. To deal wth tme-varyng channels and compensate for the shortness of the tranng sequence, blnd equalzaton algorthms have been ntroduced for the FE durng the data segments [], [8]. The FE can be adapted by usng the least-mean-square (LMS) algorthm or the recursve-least-squares (RLS) algorthm n the tranng mode, and by usng one of the blnd algorthms such as the constant modulus algorthm, the stopand-go (SAG) algorthm, or the SAG dual-mode constant modus algorthm for the data segments [8]. Let x [ be the equalzer nput at tme k, then the output of the FE at tme k, y [, s gven by Nb N a = y[ = b [ x[ k ] c [ yˆ[ k j], () j= where b[ ( =,, L, Nb ) are the forward equalzer taps at tme k, c j[ ( j =, L, N a ) are the feedback taps at tme k, and y ˆ[ s the slcer output, whch s the constellaton pont closest to y [. The LMS update algorthm for the feedforward and feedback flter taps are gven by j 2 Hyoung-Nam Km et al. ETRI Journal, Volume 26, Number 2, Aprl 24
b [ k + ] = b [ µ e c [ k + ] = c [ + µ e j j [ x[ k ], [ yˆ[ k j], (2) FE can mprove by more than 2 db f we mnmze error propagaton. where µ s the step sze and e [ = y[ yˆ[ s the error. In the blnd mode, usng the SAG algorthm, the flter tap coeffcents are updated va b[ k + ] = b[ µ f [ e[ x[ k ], c [ k + ] = c [ + µ f [ e [ yˆ[ k j]. j The SAG flag f [ s defned as f [ = f f j sgn{ e sgn{ e [ } = sgn{ e [ } [ } sgn{ e [ }, where sgn{ } s a sgnum functon defned by sgn{ x} = + when when when and e S [ s the Sato error gven by e S Here, γ s a constant defned by = > S S x < x x, (3) (4) [ = y[ γ sgn{ y[ }. () 2 E[ a[ ] γ =, (6) E[ a[ ] where a [ s the transmtted symbol. Note that the slcer output n (), (2), and (3) s replaced wth the tranng symbol durng the tranng mode. In many crcumstances of TV recevers, the FE operates under the addtve whte Gaussan nose condton of sgnal-tonose rato (SNR) values between 2 db and 3 db. When we consder mult-path fadng channels, the SNR condton worsens at the equalzer output because the FE does not completely compensate for the mult-path channel effect. When the SNR s lower than about 8 db, the eyes of the 8- VSB sgnal are closed and thus error propagaton occurs durng the data segments. Ths results n deteroraton of the convergence performance of the blnd FE [6]. To overcome ths problem, t s necessary to analyze the error propagaton effect and reduce t. In [8], a selectve feedback scheme was proposed to reduce the performance degradaton of the blnd FE caused by error propagaton. Though the performance of the blnd FE can be mproved by ntroducng the selectve feedback scheme, the degree of mprovement s not large and error propagaton stll exsts. The analyss presented n [6] shows that the SER performance of the blnd III. Blnd FE wth a Modfed Trells ecoder The slcer output s usually determned by searchng for the symbol closest to the equalzer output, y [, n a predetermned transmt symbol constellaton. Ths knd of decson devce has the lowest computatonal complexty but may result n error propagaton when the eyes are closed. To mprove the correct-decson probablty of the slcer, the Vterb decoder may be a good canddate as a decson devce to replace the slcer [9]. It s well known that a trace-back depth of δ K results n a neglgble degradaton n the performance relatve to the optmum Vterb algorthm [], where K s the constrant length. Snce K s 3 n the ATSC TV system, the TB must be no less than. Unfortunately, the use of a T such as the Vterb decoder wth a TB of may not be effectve for the FE of ATSC TV recevers due to the long delay ) caused by the TB and the trells code denterleaver [7]. To adopt the T as a decson devce n the blnd FE, the delay caused by the T has to be mnmzed because the low-order tap coeffcents of the feedback flter have a large mpact on the equalzer performance. Accordng to codng theory, the trells-coded 8-VSB has better performance than the un-coded 4-VSB at a TB settng of about [6]. In the case of the FE, however, trells codng s valuable only f the output of the T has better performance than the output of the slcer. Fgure 2 shows that the T wth a TB of has better SER performance than the slcer output by more than db at an SER of about.3. The more mportant fact Symbol error rate - -3 Performance of orgnal and modfed trells decoder Slcer output Orgnal T output (TB=) Modfed T output (TB=) - 2 4 6 8 2 22 24 26 SNR (db) Fg. 2. Symbol error rates of the slcer, the orgnal trells decoder, and the modfed trells decoder. ) The delay becomes ( δ ) 2, where δ s the TB. For example, the TB of produces a 68 symbol delay. ETRI Journal, Volume 26, Number 2, Aprl 24 Hyoung-Nam Km et al. 3
s that the T wth a TB of does not cause any delay n usng the output of the T for the blnd FE. If we acheve an output SNR of the FE of only 7 db, the T wth a TB of produces an SER of about.3. Provded that ths SER value s gven, the FE can avod deteroraton by error propagaton and approach a state of no-error propagaton. However, n the case of the slcer, when the FE acheves an output sgnal-to-nose rato of 7 db, the SER becomes., and thus results n error propagaton, as shown n Fg. 2. Ths error propagaton makes the convergence speed slow down and the resdual MSE ncrease. χ 2 χ Mapper z 2 z 2z z R z z -7 - -3-3 7 Fg. 3. Trells-coded modulaton scheme for the ATSC 8-VSB system: trells encoder, state dagram. m m Current state Next state (m,m ) (m,m ) (-7) ( ) 2 (-3) ( ) 3 2 3 (-) ( 7 ) χ = (-3) ( ) (-7) ( ) (-) ( 7 ) (-) ( 3 ) (-) ( 3 ) χ = Adopton of the trells decoder n the FE decreases the decson-error probablty but results n an ncrease of the computatonal complexty due to the twelve Ts requred for trells code de-nterleavng [7]. To reduce the computatonal complexty wthout performance loss, we modfy the computaton method of the branch metrc. The modfcaton s explaned based on Fg. 3, where the trells-coded modulaton scheme for the ATSC 8-VSB system s llustrated. Fgure 3 presents the 8-VSB trells encoder, and Fg. 3 shows the state dagram derved from the trells encoder. In the orgnal T, the branch metrc s obtaned by usng the squared R Eucldean dstance gven by 2 BM[ y[, ] = ( y[ ), (7) where, ( =,, 2,3) are the branch output symbols of the trells dagram, as shown n Fg. 3. The sum of ths branch metrc and the path metrc of the current state determnes the path metrc of the next state. After comparng each path metrc of four states, the state wth the mnmum path metrc becomes the next state n the trells state transton []. On the other hand, n the modfed T the branch metrc s computed usng the absolute dstance gven by BM[ y[, ] = y[, (8) and the other decodng schemes are equal to the orgnal T. Ths modfcaton reduces the computatonal complexty by about / wth the substtuton of the twelve absolute operatons for the twelve multplers. In spte of ths reducton, shown n Fg. 2, the MT has almost the same SER performance as the orgnal T and thus provdes a near-optmum FE to mnmze error propagaton as a near-best decson devce. Computaton of the branch metrc 2 3 8.7 or.7 6.7 or.3 4.7 or 3.3 2.7 or.3.7.3 3.3 2.7 6.6 or.4 4.6 or 3.4 2.6 or.4.6 or 7.4.4 3.4 2.6.6 9. or. 7. or.. or 2. 3. or 4. Selecton of the mnmum branch metrc 2 3.. 2. 3..7 2..7.4. 3.3 2.6 3.3.6 3.3 3.4 6.7 3.9. 2.. 3.6 4.6 4.8.8.4.2 or 2.8 3.2 or 4.8.2 or 6.8.8 or 8.8 2.8 3.2.2.8.2.2.8 6. 4.8 6.2.4.8 or 6.2.2 or 8.2 2.2 or.2 4.2 or 2.2.8.2 2.2 4.2 Fg. 4. Example of trells decodng wth the modfed trells decoder. To easly compare the performance and operatonal prncple between the conventonal slcer and the MT, we present a smple example n Fg. 4. The transmtted symbols and the nput of the conventonal slcer and the MT are assumed to be as follows:.8.8.2.2 7.8 8...6 4 Hyoung-Nam Km et al. ETRI Journal, Volume 26, Number 2, Aprl 24
Transmtted symbols: (.,.,., 3.,.) Slcer/MT nput symbols: (.7,.4, 2.,.8,.2) The ntal state s and the path metrc of ths state s.. The path metrcs of the other states are all nfnte. After the decodng process, shown n Fg. 4, the outputs of the conventonal slcer and the MT are obtaned as follows: Slcer output: (.,., 3.,.,.) MT output: (.,.,., 3.,.) Whle all the MT output symbols are the same as the transmtted symbols, three of the slcer output symbols do not match the transmtted symbols. Ths result verfes the superorty of the MT over the conventonal slcer. Fgure s the proposed blnd FE adoptng the MT wth a TB of. When we use the LMS algorthm n the tranng mode, the output SNR of the equalzer at the end of the tranng mode fals to reach 7 db. In ths case, to rase the SNR, the use of blnd algorthms n the data segments s requred. On the other hand, wth fast algorthms, such as the RLS, we can obtan an SNR of hgher than 7 db, but only n the tranng mode. In such a case, we can use equalzaton for the FE nstead of blnd algorthms. x [ B(z) + Modfed trells + y [ yˆ [ decoder wth a TB of C(z) Fg.. FE wth the modfed trells decoder. e [ Fnally, note that the blnd FE wth the trells decoder does not cause any problems n the correlaton of the nose sequence at the FE output addressed n [] because we do not assume any no-error propagaton. Therefore, the analyss results gven n [] suggests that ths FE wll converge toward the mnmummean-square-error taps derved usng the error propagaton model and actually reduce the nose correlaton at the FE output; hence, the loss through the trells decoder s reduced. IV. Smulaton Results We performed extensve computer smulatons to verfy the equalzaton performance of the proposed blnd FE for ATSC TV recevers. The transmtted data symbols were generated n frames, as shown n Fg.. The symbols of the frst segment were the same as those of the feld sync segment specfed by the ATSC standard [3]. The other segments were unknown data segments comprsed of 8-level trells-encoded a [ Up samplng s[n] by M g [ n]* 2 g [ n]cosω n Pulse shapng Basebandequvalent VSB channel g [n] h R [n] w[n] Matched flterng & VSB demod. own samplng by /M Fg. 6. Smulaton block dagram of the generaton of the equalzer nputs based on the baseband-equvalent VSB channel model. symbols from 2-bt data ( x, x2), shown n Fg. 3, whch were generated from unformly dstrbuted random numbers. These symbols correspond to a [, n Fg. 6, whch s the smulaton block dagram of the generaton of the nput symbols of the FE based on the baseband-equvalent VSB channel model presented n [2]. The receved SNR was obtaned based on the sgnals shown n Fg. 6 and s defned as SNR = E[ 2 E[ h [ n] [ n] [ n R g a ] ], 2 g[ n] 2 g[ n]cosω n w[ n] ] where s a convoluton operator and n s a sample tme ndex over-sampled by an nteger value, M, that s, T n = M (( k ) M + ), M. Here k s a symbol tme ndex and T s a symbol perod of /.76 MHz (.93 µs) specfed by the ATSC transmsson standard [3]. In (9), h R [n] s an mpulse response of the baseband-equvalent VSB channel model, g [n] s the rased cosne flter wth the roll-off factor of.76, g[n] s the square-root rased cosne flter correspondng to g [n], a [n] s a transmt symbol sequence, w[n] s a whte Gaussan nose process, and ω = 2π.38 MHz (Refer to [2] for a detaled descrpton). The channel profle used for ths smulaton was Ensemble havng fve echoes wth ampltudes, delays, and phases as specfed by the Advanced Televson Technology Center (ATTC) [3], whch s shown n Table. In our smulatons based on the VSB channel model, we Table. Mult-path profle elay (µs) Ampltude (db) Phase (degree).8 +. +.8 +.7 +8. 2 2 8 4 9 2 8 9 x[ (9) ETRI Journal, Volume 26, Number 2, Aprl 24 Hyoung-Nam Km et al.
consdered VSB modulaton and passband-related effects such as phase nformaton and a carrer frequency under Korean TV CH, for whch the center frequency s 479 MHz. The mpulse response and the ampltude spectrum of the basebandequvalent VSB channel correspondng to Ensemble are shown n Fgs. 7 and 8. A FE wth 4 feedforward and 26 feedback taps was adapted usng the LMS or RLS algorthm n the feld sync segment and blnd equalzaton algorthms n the data segments. The step szes of µ n the feld sync segment and the data 4 segments were 2. and 2. (. for the algorthm under SNR values of less than 2 db), respectvely. The forgettng factor n the RLS was.. Ampltude.6.8.4.2 -.2 -.4 -.6 2 Symbol tme (T=.93 µs) Fg. 7. Impulse response of the equvalent VSB channel correspondng to Ensemble. Ampltude.8.6.4.2.8.6.4.2 - - Frequency (MHz) Fg. 8. Ampltude spectrum of the baseband channel mpulse response.. Convergence Performance The convergence performance was checked by the MSE of the equalzer output, whch was computed as follows: 2 MSE[ = E[ y[ a[ ]. () The statstcal results come from the average of ndependent ensembles. For ease of performance vewng, we also used the tme average of symbols. A. LMS n the Tranng Mode Fgures 9 and show the MSE learnng curves of the conventonal FE wth the conventonal slcer (denotng the conventonal FE), the MT (denotng the proposed FE), and no-error propagaton (denotng the optmum FE) at sgnal-to-nose ratos of 6 to 23 db. The LMS algorthm was appled n the tranng mode, and the SAG and algorthms were used durng the data segments. The result shows that error propagaton caused by wrong decsons degrades the convergence performance n terms of both the convergence speed and resdual error. At an SNR of 6 db, as shown n Fg. 9, the output SNR at the end of the tranng mode 2) was about db, where the symbol error rate of the modfed trells decoder wth a TB of was almost the same as that of the slcer and was greater than.4, as can be seen n Fg. 2. Thus, at the startng pont of the blnd adaptaton, error propagaton s nevtable even though we use the MT wth a TB of nstead of the slcer for a decson devce. However, blnd adaptaton usng the SAG algorthm makes the output SNR mprove and thus the performance s better than when the algorthm s used. As the blnd adaptaton usng the SAG algorthm proceeds, the output SNR ncreases and thus the effect of adoptng the MT s promnent compared wth the slcer because the dfference between the SER of the MT and that of the slcer grows large. The effect of error propagaton on the equalzaton performance becomes clearer n Fg. 9. Though error propagaton exsts at the start of blnd adaptaton, the SAG algorthm adopted n the proposed FE enhances the output SNR of the FE and, thus, makes the error propagaton decrease. On the other hand, when we use the algorthm n the conventonal FE, error propagaton caused by slcng error serously affects the convergence performance. The SAG of the conventonal FE s comparable to the of the proposed FE because blnd adaptaton of the SAG algorthm n the conventonal FE compensates for the slcng error, and the MT decreases the decson error probablty n the proposed FE. In concluson, as the SNR ncreases from the low ratos shown n Fg. 9 to the hgh ratos seen n Fg., the proposed FE approaches the optmum FE. We can see that blnd equalzaton s preferable to equalzaton at low SNRs (less than about 8 db) whle equalzaton s superor to blnd 2) Ths corresponds to an 832 symbol tme. Actually, the number of the tranng symbols s 82. In addton, the number may be shortened to 728 f we do not use the reserved symbols. Ths number s not crtcal for analyzng performance trend of the FE. 6 Hyoung-Nam Km et al. ETRI Journal, Volume 26, Number 2, Aprl 24
- MSE learnng curve (6 db) 2 3 4 6 x 4 - MSE learnng curve (8 db) 2 3 4 6 x 4 (c) 2 3 4 6 x 4 Fg. 9. Mean-squared error convergence of the FE (LMS wth µ=.2 n the feld sync segments; SAG or wth µ=.2 n the data segments) at an SNR of 6 db, 7 db, (c) 8 db, and (d) 9 db. - - MSE learnng curve (7 db) MSE learnng curve (9 db) 2 3 4 6 x 4 (d) - MSE learnng curve (2 db) 2 3 4 6 x 4 MSE learnng curve (22 db) - 2 2 3 4 6 x 4 (c) 2 3 4 6 x 4 Fg.. Mean-squared error convergence of the FE (LMS wth µ=.2 n the feld sync segments; SAG or wth µ=.2 n the data segments) at an SNR of 2 db, 2 db, (c) 22 db, and (d) 23 db. - MSE learnng curve (2 db) MSE learnng curve (23 db) - 2 2 3 4 6 x 4 (d) ETRI Journal, Volume 26, Number 2, Aprl 24 Hyoung-Nam Km et al. 7
MSE learnng curve (6 db) MSE learnng curve (7 db) - - - - - - 2 4 6 8 MSE learnng curve (8 db) 2 4 6 8 MSE learnng curve (9 db) - - - - - - 2 4 6 8 (c) 2 4 6 8 (d) Fg.. Mean-squared error convergence of the FE (RLS wth the forgettng factor of. n the feld sync segments; SAG or wth µ=.2 n the data segments) at an SNR of 6 db, 7 db, (c) 8 db, and (d) 9 db. MSE learnng curve (2 db) MSE learnng curve (2 db) - - - Mean Square Error (db) - - - 2 4 6 8 MSE learnng curve (22 db) 2 4 6 8 MSE learnng curve (23 db) - - - Optmal FE - - - 2 4 6 8 (c) 2 4 6 8 (d) Fg. 2. Mean-squared error convergence of the FE (RLS wth the forgettng factor of. n the feld sync segments; SAG or wth µ=.2 n the data segments) at an SNR of 2 db, 2 db, (c) 22 db, and (d) 23 db. 8 Hyoung-Nam Km et al. ETRI Journal, Volume 26, Number 2, Aprl 24
equalzaton at hgh SNRs (more than about 9 db). B. RLS n the Tranng Mode To rase the output SNR at the end of the tranng mode, we used the RLS algorthm, and the results are shown n Fgs. and 2. Under an SNR of 6 db, as show n Fg., the output SNR at the end of the tranng mode was about 3.7 db, where the SER of the proposed FE was.8, as was shown n Fg. 2. Whle ths SER value was about.2, seen n Fg. 2, and was smaller than that of the conventonal FE, t stll caused error propagaton. Accordngly, the performance of the proposed FE was close to that of the conventonal FE. The performance of the proposed FE approached the optmum FE at an SNR of 7 db, whch can be seen n Fg.. In ths case, the output SNR at the tranng mode was about 4.7 db correspondng to an SER of.4. When the SNR was greater than 7 db, the performance of the proposed FE was almost the same as that of the optmum FE. From these results, we found that f the output SNR of the FE wth the MT at the end of the tranng mode was greater than.6 db, correspondng to an SER of.2, error propagaton dd not affect the convergence performance. However, the conventonal FE suffered from error propagaton even when the output SNR was about 9 db, as seen n Fg. (d), because the SER of the slcer output was about., as was shown n Fg. 2. Note that the optmum soluton of LMS-type algorthms s dfferent from that of LS-type algorthms. As llustrated n Fgs. and 2, dscontnutes between the tranng mode and the blnd mode appeared because we used LS-type algorthms such as the RLS n the tranng mode, and LMS-type algorthms such as and SAG n the blnd mode. The dscontnutes become more obvous when ether error propagaton, caused by slcng error n the conventonal FE, or a low output SNR exsts. We also found that adaptaton and blnd adaptaton had almost the same performance n the proposed FE. In the case of the conventonal FE, however, blnd adaptaton showed a lttle better performance than adaptaton n terms of resdual MSE at low SNRs, whch wll be clearly shown n the SER plots presented n the next secton. C. Symbol Error Rate Performance We carred out smulatons to obtan the SER curves of the blnd FE to compare the resdual error performance. The number of smulated segments ncludng one feld sync segment was 3, and thus the number of data segments was 3, whch corresponded to 249,6 symbols. The SER was computed by countng the number of symbol errors exstng n the last 8, symbols after the tap coeffcents converged. In ATSC TV recevers, we are nterested n the performance Symbol error rate of equalzer Symbol error rate of equalzer Equalzer performance n ATTC- channel - -3 6 7 8 9 2 2 22 23 24 2 SNR (db) Equalzer performance n ATTC- channel - Conventonal FE wh SAG Conventonal FE wh -3 6 7 8 9 2 2 22 23 24 2 SNR (db) Fg. 3. Symbol error rate performance of the FE: LMS, RLS. at SNR values of not more than 2 db because the trells decoder followng the equalzer can correct most symbol errors at SNR values greater than 2 db. Fgure 3 shows the SER curves usng the LMS and RLS algorthms n the tranng mode. In both cases, the optmum FE has a better performance by about 3 db than the conventonal FE at an SER of.2, correspondng to the threshold of vsblty (TOV). On the other hand, the proposed FE was about 2 db better than the conventonal FE at the TOV value. The performance of the proposed FE was smlar to that of the optmum FE at most SNR values but degraded to the level of the conventonal FE at very low SNR values when ether the LMS or the RLS algorthms n the tranng mode was used. We found that when error propagaton exsts, blnd adaptaton was preferable to adaptaton n the FE for ATSC TV recevers. We have to note that the performance of the FE affects the SER performance of the trells decoder 3) followng the FE 3) Ths trells decoder s dfferent from the T used as a decson devce n the FE and generally has a TB of about to produce the maxmum SER performance n the trells-coded 8-VSB sgnal. ETRI Journal, Volume 26, Number 2, Aprl 24 Hyoung-Nam Km et al. 9
Symbol error rate of TCM decoder Symbol error rate of TCM decoder - -3 - - -3 - TCM decoder performance n ATTC- channel 6 7 8 9 2 2 22 SNR (db) TCM decoder performance n ATTC- channel 6 7 8 9 2 2 22 SNR (db) Fg. 4. Symbol error rate performance of the trells decoder followng the FE: LMS, RLS. specfed n the ATSC terrestral TV standard [7]. Fgure 4 shows the SER curves of the trells decoder recevng the output of the FE and then decodng wth a TB of. Whle the FE output SER performance of the proposed FE s smlar to that of the optmum FE, the SER performance of the trells decoder between them dffers by db. In the conventonal FE, blnd adaptaton of the SAG algorthm has better performance than adaptaton at SNR values of less than 22 db. When the SNR was greater than 2 db, the SER was smaller than the TOV at the trells decoder, and thus, the comparson s meanngless. V. Conclusons The error propagaton phenomenon s unavodable n the FE for ATSC TV recevers because the tranng sequence s very short and the symbol error rate of the equalzer output can be as hgh as.2. To reduce the error propagaton, we proposed a hardware-effcent blnd FE by ncorporatng a modfed trells decoder for a decson devce. As the MT uses the absolute dstance nstead of the Eucldean dstance, computatonal complexty s reduced by / wthout performance loss, compared to the orgnal T. In the proposed FE, error propagaton was apparently reduced, resultng n a faster convergence speed and mproved SER performance by more than 2 db compared wth the conventonal FE. Acknowledgment The authors would lke to thank the anonymous revewers whose comments mproved ths paper. References [] S.U.H. Quresh, Adaptve Equalzaton, Proc. IEEE, vol. 73, no. 9, Sept. 98, pp. 349-387. [2] N. Al-hahr and J.M. Coff, MMSE ecson-feedback Equalzers: Fnte-Length Results, IEEE Trans. Inform. Theory, vol. 4, no. 4, July 99, pp. 96-97. [3] ATSC, ATSC gtal Televson Standard, Apr. 2, oc. A/3A. [4] A. Fertner, Improvement of Bt-Error-Rate n ecson Feedback Equalzer by Preventng ecson-error Propagaton, IEEE Trans. Sgnal Processng, vol. 46, no. 7, July 998, pp. 87277. [] M. Ghosh, Blnd ecson Feedback Equalzaton for Terrestral Televson Recevers, Proc. IEEE, vol. 86, no., Oct. 998, pp. 278. [6] H.-N. Km, S.I. Park, and S.W. Km, Performance Analyss of Error Propagaton Effects n the FE for ATSC TV Recevers, IEEE Trans. Broadcastng, vol. 49, no. 3, Sept. 23, pp. 2497. [7] ATSC, Gude to the Use of ATSC gtal Televson Standard, Oct. 99, oc. A/4. [8] H.-N. Km, Y.-T. Lee, and S.W. Km, Blnd ecson Feedback Equalzaton for VSB-Based TV Recevers, IEEE Trans. Consumer Electroncs, vol. 48, no. 3, Aug. 22, pp. 629. [9] J.J. Ncolas and J.S. Lm, Equalzaton and Interference Rejecton for the Terrestral Broadcast of gtal HTV, Proc. ICASSP-93, vol. 4, Mnnesota, USA, Apr. 993, pp. 76-79. [] J.G. Proaks, gtal Communcatons, 4th ed., McGraw-Hll, New York, 2, p. 48. [] E. Bgler, B. vsalar, P.J. McLane, and M.K. Smon, Introducton to Trells-Coded Modulaton wth Applcatons, Macmllan, New York, 99. [2] H.-N. Km, Y.-T. Lee, and S.W. Km, Mathematcal Modelng of VSB-Based gtal Televson Systems, ETRI J., vol. 2, no., Feb. 23, pp. 9. [3] Comm. Research Center, gtal Televson Test Results Phase I, CRC Report No. CRC-RP-, Ottawa, Nov. 2, p. 6. Hyoung-Nam Km et al. ETRI Journal, Volume 26, Number 2, Aprl 24
Hyoung-Nam Km receved the BS, MS, and Ph degrees n electronc and electrcal engneerng from Pohang Unversty of Scence and Technology (POSTECH), Pohang, Korea, n 993, 99, and 2, respectvely. From May 2 to February 23, he was wth Electroncs and Telecommuncatons Research Insttute (ETRI), aejeon, Korea, developng advanced transmsson and recepton technology for terrestral dgtal televson. Snce March 23, he has been an Assstant Professor n the epartment of Electroncs Engneerng at Pusan Natonal Unversty, Busan, Korea. Hs research nterests are n the areas of dgtal sgnal processng, adaptve IIR flterng, and radar sgnal processng, n partcular, sgnal processng for dgtal televson, dgtal communcatons, and multmeda systems. r. Km s a member of IEEE and KICS. Sung Ik Park receved the BSEE from Hanyang Unversty, Seoul, Korea, n 2 and MSEE from POSTECH, Pohang, Korea, n 22. Snce 22, he has been wth the Broadcastng System Research Group, ETRI, where he s a member of Research Staff. Hs research nterests are n the area of error correcton codes and dgtal communcatons, n partcular, sgnal processng for dgtal televson. Seung Won Km receved the BSEE and MSEE from Sung Kyun Kwan Unversty, Korea, n February 986 and 988. After graduaton, he served n the Korean Army as a reserved offcer from August 988 to February 989. Snce June 989 he has been employed at ETRI, Korea. He receved hs Ph from Unversty of Florda, USA, n May 999. He s currently a leader of TV system research team at ETRI. Hs man research nterests are n the areas of dgtal communcaton systems, dgtal sgnal processng and TV transmsson systems. Jae Moung Km receved the BS degree from Hanyang Unversty, Korea n 974, the MSEE degree from Unversty of Southern Calforna, USA n 98, and the Ph degree from Yonse Unversty, Korea n 987. He was a Vce Presdent of Rado & Broadcastng Technology Laboratory and rector of Satellte Communcaton System epartment at ETRI from September 982 to March 23. Snce Aprl of 23, he has been a Professor n the Graduate School of Informaton Technology and Telecommuncatons, Inha Unversty. He s a board member of drectors of Korean Insttute of Communcaton Scence (KICS), a Vce Presdent of Korea Socety of Broadcast Engneers (KOSBE) and a senor member of IEEE. Hs research background s telecommuncaton systems modelng and performance analyss of broadband wreless access systems, moble communcatons, satellte communcatons and broadcastng transmsson technologes. ETRI Journal, Volume 26, Number 2, Aprl 24 Hyoung-Nam Km et al.