Indonesian Journal of Electrical Engineering and Computer Science V ol. 2, No . 2, Ma y 2016, pp . 334 343 DOI: 10.11591/ijeecs .v2.i2.pp334-343 334 Extension of Linear Channels Identification Algorithms to Non Linear Using Selected Or der Cum ulants Mohammed Zidane 1,* , Said Safi 2 , and Mohamed Sabri 1 1 Depar tment of Ph ysics , F aculty of Sciences and T echnology , Sultan Moula y Slima ne Univ ersity , Morocco 2 Depar tment of Mathematics and Inf or matics , P olydisciplinar y F aculty , Sultan Moula y Sliman e Univ ersity , Morocco * corresponding author , e-mail: zidane .ilco@gmail.com Abstract In this paper , w e present an e xtension of linear comm unication channels identification algor ithms to non linear channels using higher order cum ulants (HOC). In the one hand, w e de v elop a theoretical analysis of non linear quadr atic systems using second and third order cum ulants . In the other hand, the relationship linking cum ulants and the coefficients of non linear channels presented in the linear case is e xtended to the gener al case of the non linear quadr atic systems identification. This theoretical de v elopment is used to de v elop three non linear algor ithms based on third and f our th order cum ulants respectiv ely . Numer ical sim- ulation results e xample sho w that the proposed methods ab le to estimate the impulse response par ameters with diff erent precision. K e yw or ds: Higher order cum ulants , Blind identification, Non linear quadr atic systems . Cop yright c 2016 Institute of Ad v anced Engineering and Science 1. Intr oduction Applications of higher order cum ulants theo r y in the b lind identification domain are widely used in v ar ious w or ks [1]-[3], [6, 7]. Se v er al models are identified in the liter ature such as the lin- ear and non linear systems . In the par t of the linear case , w e ha v e impor tant results estab lished that the b lind identification is possib le only from the second order cum ulants (autocorrelation func- tion) of the output stationar y signal, b ut t hese methods is not ab le to identify correctly the channel models e xcited b y non Gaussian signal and aff ected b y Gaussian noise , because the additiv e Gaussian noise will be v anish in the higher order cum ulants domain. The sensitivity of the second order cum ulants to the additiv e Gaussian noise appealed to other b lind identification methods e x- ploiting the cum ulants of order super ieur than tw o [4, 5]. There are se v er al motiv ations behind this interest, first the methods based on HOC are b lind to an y kind of a Gaussian process , whereas autocorrelation function (second order cum ulants) is not. Consequently , cum ulants based meth- ods boost signal to noise r atio (SNR) when signals are corr upted b y Gaussian measurement noise . Second, the HOC methods are useful in identifying non minim um phase systems and in reconstr ucting non minim um phase signals when the signals are non Gaussian. The linear models are not efficient f or representing and modeling all systems , because the major ity of systems are represented b y non linear models [7, 8]. Ho w e v er , wh en linear modeling of the channe l is not adequate , the non linear modeling appeared lik e an alter nativ e efficient solution in most real cases . Moreo v er , quadr atic non linear systems are widely used in v ar ious engineer ing fields such as signal processing, system filter ing, predicting, identification and equalization [9, 10]. In this contr ib ution, firstly w e present a theoretical de v elopment of non linear quadr atic systems using higher order cum ulants . Indeed, w e de v elop the relationship linking third order cum ulants and the coefficients of non linear channels , then the method de v eloped [11] f or linear channels , is e xtended to the gener a l case of the non linear quadr atic systems identification. Sec- ondly , these theoretical analyses are used to de v elop an e xtension of linear algor ithms based on third and f our th order cum ulants proposed b y safi, et al : [5] and Abderr ahim, et al : [12] f or linear Receiv ed J an uar y 21, 2016; Re vised March 27, 2016; Accepted Apr il 11, 2016 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 335 systems to non linear algor it hms f or identification of quadr atic systems . Num er ical sim ulations results are giv en to illustr ate the accur acy of the proposed methods . 2. Non linear comm unication c hannel and h ypotheses W e consider a non linear channel (Fig. 1) descr ibed b y the f ollo wing equation: y ( k ) = q X i =0 h ( i; i ) x 2 ( k i ) + w ( k ) ; (1) where y ( k ) represent the output model is gener ated through a quadr atic non linear model dr iv en b y a stationar y input sequence x ( k ) , h ( i; i ) and q are the par ameters and the order of non linear channel, respectiv ely , w ( k ) is additiv e Gaussian noise . N on  l i ne a r c om m uni c a t i on  c ha nne l ) ( k x ) ( k z ) ( k y ) ( k w Figure 1. Non linear channel model F or this system w e assume that: H1: The order q is kno wn; H2: The input sequence x ( k ) is independent and identically distr ib uted (i.i.d) z ero mean, station- ar y , non Gaussian and with: C k ;x ( 1 ; 2 ; :::; k 1 ) = k ;x ; 1 = 2 = ::: = k 1 = 0 ; 0 otherwise where k ;x denotes the k th order cum ulants of the input signal x ( k ) at or igin H3: The system is supposed causal and bounded, i.e . h ( i; i ) = 0 if i < 0 and i > q with h (0 ; 0) = 1 ; H4: The measurement noise sequence w ( k ) is assumed to be z ero mean, i.i.d, Gaussian, inde- pendent of x ( k ) with unkno wn v ar iance; H5: The system is supposed stab le , i.e . j h ( i; i ) j < 1 . 3. Theoretical de velopment f or non linear comm unication c hannels using HOC In this section, w e firstly f ocus on additional concepts and definitions used throughout this paper . Indeed, w e discuss the identifiability of the quadr atic non linear model in the second and third order cum ulants domain, then based an the relationship proposed b y Stogioglou and McLaughlin [11] in the linear case , w e de v elop an e xtension of this relation in non linear case . 3.1. Identification of quadratic non linear models in the second and thir d or der cum ulants In this subsection, w e use the Leono v Shir y ae v f or m ula [6] and the definition of the cum u- lants to demonstr ate the equations linking the thr id order cum ulants and the diagonal par ameters of non linear systems . Thus , the cum ulants , order r , and the moments of the stochastic signal are link ed b y the f ollo wing relationships of the Leono v and Shir y a y e v f or m ula [6]: C u m [ z 1 ; :::; z r ] = X ( 1) k 1 ( k 1)! E h Y i 2 v 1 z i i :E h Y j 2 v 2 z j i :::E h Y k 2 v p z k i ; (2) where the addition oper ation is o v er all the set of v i , 1 i p r and v i compose a par tition of 1 ; 2 ; :::; r . In Eq. (2) k is the n umber of elements compose a par tition. F or the second order cum ulants , w e are r = 2 and 1 p 2 . The possib le par titions are: (1,2) and (1)(2) thus: C u m [ z 1 ; z 2 ] = E [ z 1 z 2 ] E [ z 1 ] E [ z 2 ] ; (3) Extension of Linear Channels Identification Algor ithms to Non Linear Using Selected Order Cum ulants (M. Zidane) Evaluation Warning : The document was created with Spire.PDF for Python.
336 ISSN: 2502-4752 C 2 ;z ( ) = C um ( z 1 ; z 2 ) = C um h z ( k ) ; z ( k + ) i = E h q X i =0 h ( i; i ) x 2 ( k i ) q X j =0 h ( j ; j ) x 2 ( k + j ) i E h q X i =0 h ( i; i ) x 2 ( k i ) i E h q X j =0 h ( j ; j ) x 2 ( k + j ) i (4) C 2 ;z ( ) = q X i =0 h ( i; i ) h ( i + ; i + ) E h x 2 ( k i ) x 2 ( k i ) i q X i =0 h ( i; i ) h ( i + ; i + ) E h x 2 ( k i ) i E h x 2 ( k i ) i (5) Under the assumption H2 and Eq. (5), the second order cum ulants becomes: C 2 ;z ( ) = ( 4 ;x 2 2 ;x ) q X i =0 h ( i; i ) h ( i + ; i + ) (6) F or the third order cum ulants , w e are r = 3 and 1 p 3 . The possib le par titions are: (1,2,3), (1)(2,3), and (1)(2)(3) so: C um [ z 1 ; z 2 ; z 3 ] = E [ z 1 z 2 z 3 ] E [ z 1 ] E [ z 2 z 3 ] E [ z 2 ] E [ z 1 z 3 ] E [ z 3 ] E [ z 1 z 2 ] + 2 E [ z 1 ] E [ z 2 ] E [ z 3 ] (7) Thus , the third order cum ulants becomes: C 3 ;z ( 1 ; 2 ) = C um h z ( k ) ; z ( k + 1 ) ; z ( k + 2 ) i = E h q X i =0 h ( i; i ) x 2 ( k i ) q X j =0 h ( j ; j ) x 2 ( k + 1 j ) q X l =0 h ( l ; l ) x 2 ( k + 2 l ) i [3] E h q X i =0 h ( i; i ) x 2 ( k i ) i E h q X j =0 h ( j ; j ) x 2 ( k + 1 j ) q X l =0 h ( l ; l ) x 2 ( k + 2 l ) i + [2] E h q X i =0 h ( i; i ) x 2 ( k i ) i E h q X j =0 h ( j ; j ) x 2 ( k + 1 j ) i E h q X l =0 h ( l ; l ) x 2 ( k + 2 l ) i (8) Under the assumption H2 and Eq. (8), the third order cum ulants becomes: C 3 ;z ( 1 ; 2 ) = ( 6 ;x 3 2 ;x 4 ;x + 2 3 2 ;x ) q X i =0 h ( i; i ) h ( i + 1 ; i + 1 ) h ( i + 2 ; i + 2 ) (9) In the gener al case , the relationship betw een n th order cum ulants and t he coefficients of non linear impulse response channel can be wr itten under f or m: C n;z ( 1 ; 2 ; :::; n 1 ) = n;x q X i =0 h ( i; i ) h ( i + 1 ; i + 1 ) :::h ( i + n 1 ; i + n 1 ) ; (10) where n;x = P i p i m i i;x , with p i 2 Z and m i 2 N IJEECS V ol. 2, No . 2, Ma y 2016 : 334 343 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 337 3.2. Relationship linking cum ulants and the coefficients of non linear c hannel In this secti on, w e de v elop a non linear e xtension of the Stogioglou and McLaughlin’ s linear relationship [11]. Pr oof: Let : Z s;n = X i X j h ( i; i ) h ( j ; j ) h s Y k =1 h ( i + k ; i + k ) ih s Y k =1 h ( j + k ; j + k ) ih n 1 Y k = s +1 h ( i + j + k ; i + j + k ) i ; (11) where s is an arbitr ar y integer n umber satisfying: 1 s n 2 . Changing the order of summation in Eq (11), and w e m ultiply this equation b y n;x , w e will obtain: n;x Z s;n = X i h ( i; i ) h s Y k =1 h ( i + k ; i + k ) i n;x X j h ( j ; j ) h s Y k =1 h ( j + k ; j + k ) ih n 1 Y k = s +1 h ( i + j + k ; i + j + k ) i ; (12) where , n;x represents the n th order cum ulants of the e xcitation signal at or igin in non linear case defined b y n;x = P i p i m i i;x . F rom Eqs (10) and (12) w e obtain: n;x Z s;n = q X i =0 h ( i; i ) h s Y k =1 h ( i + k ; i + k ) i C n;y ( 1 ; :::; s ; i + 1 ; :::; i + n s 1 ) (13) In the same w a y , if w e sum on i afterw ards on j in (11), w e will find: n;x Z s;n = X j h ( j ; j ) h s Y k =1 h ( j + k ; j + k ) i n;x X i h ( i; i ) h s Y k =1 h ( i + k ; i + k ) ih n 1 Y k = s +1 h ( i + j + k ; i + j + k ) i (14) The same , from Eqs (10) and (14) w e obtain: n;x Z s;n = q X j =0 h ( j ; j ) h s Y k =1 h ( j + k ; j + k ) i C n;y ( 1 ; :::; s ; j + 1 ; :::; j + n s 1 ) ; (15) F rom Eqs . (13) and (15) w e obtain Stogioglou-McLaughlin relation in non linear case: q X j =0 h ( j ; j ) h s Y k =1 h ( j + k ; j + k ) i C n;y ( 1 ; :::; s ; j + 1 ; :::; j + n s 1 ) = q X i =0 h ( i; i ) h s Y k =1 h ( i + k ; i + k ) i C n;y ( 1 ; :::; s ; i + 1 ; :::; i + n s 1 ) ; (16) where 1 s n 2 . 4. Extension of linear algorithms to non linear algorithms In this section, w e descr ibe an e xtension of linear comm unication channels identification algor ithms to non linear using HOC . Ho w e v er , the linear algor ithms based on third and f our th order cum ulants proposed in the liter ature [5, 12] f or linear case is e xtended to the non linear case f or identification of quadr atic systems . Extension of Linear Channels Identification Algor ithms to Non Linear Using Selected Order Cum ulants (M. Zidane) Evaluation Warning : The document was created with Spire.PDF for Python.
338 ISSN: 2502-4752 4.1. Fir st algorithm: Algcum1 The F our ier tr ansf or m of Eqs . (6) and (9) giv es us the bispectr a and the spectr um respec- tiv ely: S 3 ;z ( ! 1 ; ! 2 ) = ( 6 ;x 3 2 ;x 4 ;x + 2 3 2 ;x ) H ( ! 1 ! 2 ; ! 1 ! 2 ) H ( ! 1 ; ! 1 ) H ( ! 2 ; ! 2 ) (17) S 2 ;z ( ! ) = ( 4 ;x 2 2 ;x ) H ( ! ; ! ) H ( ! ; ! ) (18) If w e suppose that ! = ! 1 + ! 2 , Eq. (18) becomes: S 2 ;z ( ! 1 + ! 2 ) = ( 4 ;x 2 2 ;x ) H ( ! 1 ! 2 ; ! 1 ! 2 ) H ( ! 1 + ! 2 ; ! 1 + ! 2 ) (19) Then, from Eqs . (17) and (19) w e obtain the f ollo wing equation: S 3 ;z ( ! 1 ; ! 2 ) H ( ! 1 + ! 2 ; ! 1 + ! 2 ) = 6 ;x 3 2 ;x 4 ;x + 2 3 2 ;x 4 ;x 2 2 ;x H ( ! 1 ; ! 1 ) H ( ! 2 ; ! 2 ) S 2 ;z ( ! 1 + ! 2 ; ! 1 + ! 2 ) (20) The in v erse F our ier tr ansf or m of Eq. (20) demonstr ates that the third order cum ulants , the auto- correlation function and the impulse response channel par ameters are combined b y the f ollo wing equation: q X i =0 C 3 ;z ( 1 i; 2 i ) h ( i; i ) = 6 ;x 3 2 ;x 4 ;x + 2 3 2 ;x 4 ;x 2 2 ;x q X i =0 h ( i; i ) h ( 2 1 + i; 2 1 + i ) C 2 ;z ( 1 i ) (21) If w e use the autocorrelation function proper ty of the stationar y process such as C 2 ;z ( ) 6 = 0 only f or q q and v anishes else where if w e tak e 1 = q , Eq. (21) tak es the f or me: q X i =0 C 3 ;z ( q i; 2 i ) h ( i; i ) = 6 ;x 3 2 ;x 4 ;x + 2 3 2 ;x 4 ;x 2 2 ;x h (0 ; 0) h ( 2 + q ; 2 + q ) C 2 ;z ( q ) ; (22) else , if w e suppose that 2 = q , Eq. (22) will become: C 3 ;z ( q ; q ) h ( q ; q ) = 6 ;x 3 2 ;x 4 ;x + 2 3 2 ;x 4 ;x 2 2 ;x h (0 ; 0) C 2 ;z ( q ) (23) Using Eqs . (22) and (23) w e obtain the f ollo wing relation: q X i =0 C 3 ;z ( q i; 2 i ) h ( i; i ) = C 3 ;z ( q ; q ) h ( 2 + q ; 2 + q ) (24) The system of Eq. (24) can be wr itten in matr ix f or m as: 0 B B B B B B @ C 3 ;z ( q 1 ; q 1) ::: C 3 ;z ( 2 q ; 2 q ) C 3 ;z ( q 1 ; q ) ::: C 3 ;z ( 2 q ; 2 q + 1) : : : : : : : : : C 3 ;z ( q 1 ; 1) ::: C 3 ;z ( 2 q ; q ) 1 C C C C C C A 0 B B B B B B B B B B B B @ h (1 ; 1) : : : h ( i; i ) : : : h ( q ; q ) 1 C C C C C C C C C C C C A = 0 B B B B B B B B B B @ 0 C 3 ;z ( q ; q + 1) : : : : : C 3 ;z ( q ; 0) 1 C C C C C C C C C C A ; (25) where = C 3 ;y ( q ; q ) . Or in more compact f or m, the Eq. (25) can be wr itten as f ollo ws: M h s = d; (26) IJEECS V ol. 2, No . 2, Ma y 2016 : 334 343 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 339 where M is the matr ix of siz e ( q + 1) ( q ) elements , h s is a column v ector constituted b y the un- kno wn impulse response par ameters h ( i; i ) f or i = 1 ; :::; q and d is a column v ector of siz e ( q + 1) as indicated in the Eq. (26). The least squares solution of the system of Eq. (26), per mits b lindly identification of the par am- eters h ( i; i ) and without an y inf or mation of the input selectiv e channel. Thus , the solution will be wr itten under the f ollo wing f or m: b h s = ( M T M ) 1 M T d (27) 4.2. Second algorithm: Algcum2 Safi, et al : [5] use the Stogioglou-McLaughlin relation f or de v eloping an algor ithm based only on f our th order cum ulants in the linear case , this algor ithm is e xtended to the non linear case . Thus , if w e tak e n = 4 into (16) w e obtain the f ollo wing equation: q P j =0 h ( j ; j ) h ( j + 1 ; j + 1 ) h ( j + 2 ; j + 2 ) C 4 ;y ( 1 ; 2 ; j + 1 ) = q P i =0 h ( i; i ) h ( i + 1 ; i + 1 ) h ( i + 2 ; i + 2 ) C 4 ;y ( 1 ; 2 ; i + 1 ) (28) If 1 = 2 = q et 1 = 2 = 0 , (28) tak e the f or m: h (0 ; 0) h 2 ( q ; q ) C 4 ;y (0 ; 0 ; 1 ) = q X i =0 h 3 ( i; i ) C 4 ;y ( q ; q ; i + 1 ) ; (29) with q 1 q (30) Then, from (29) and (30) w e obtain the f ollo wing system of equations: 0 B B B B B B B B B B B B @ C 4 ;y ( q ; q ; q ) ::: C 4 ;y ( q ; q ; 0) : : : : : : : : : C 4 ;y ( q ; q ; 0) ::: C 4 ;y ( q ; q ; q ) : : : : : : : : : C 4 ;y ( q ; q ; q ) ::: C 4 ;y ( q ; q ; 2 q ) 1 C C C C C C C C C C C C A 0 B B B B B B B B B B B B B @ 1 h 2 ( q ;q ) : : : h 3 ( i;i ) h 2 ( q ;q ) : : : h 3 ( q ;q ) h 2 ( q ;q ) 1 C C C C C C C C C C C C C A = 0 B B B B B B B B B B B B @ C 4 ;y (0 ; 0 ; q ) : : : C 4 ;y (0 ; 0 ; 0) : : : C 4 ;y (0 ; 0 ; q ) 1 C C C C C C C C C C C C A (31) In more compact f or m, the system of (31) can be wr itten in the f ollo wing f or m: M b q = d; (32) where M , b q and d are defined in the system of (32). The least squares solution of the system of (32) is giv en b y: b b q = ( M T M ) 1 M T d (33) This solution giv e us an estimation of the quotient of the par ameters h 3 ( i; i ) and h 2 ( q ; q ) , i.e ., b q ( i; i ) = \ h 3 ( i;i ) h 2 ( q ;q ) , i = 1 ; :::; q . Thus , in order to ob tain an estimati on of the par ameters b h ( i; i ) , i = 1 ; :::; q w e proceed as f ollo ws: The p ar ameters h ( i; i ) f or i = 1 ; :::; q 1 are estimated from the estimated v alues b b q ( i; i ) using the f ollo wing equation: b h ( i; i ) = sig n h b b q ( i; i ) ( b b q ( q ; q )) 2 in abs ( b b q ( i; i )) ( b b q ( q ; q )) 2 o 1 = 3 (34) Extension of Linear Channels Identification Algor ithms to Non Linear Using Selected Order Cum ulants (M. Zidane) Evaluation Warning : The document was created with Spire.PDF for Python.
340 ISSN: 2502-4752 The b h ( q ; q ) par ameters is estimated as f ollo ws: b h ( q ; q ) = 1 2 sig n h b b q ( q ; q ) in abs ( b b q ( q ; q )) + 1 b b q (1 ; 1) 1 = 2 o (35) 4.3. Thir d algorithm: Algcum3 Abderr ahim, et al : [12] use also the relationship (16), based on f our th order cum ulants , f or de v eloping an algor ithm based an f our th order cum ulants in linear case . This algor ithm is e xtended to the gener al case of the non linear quadr atic systems , with 1 = 2 = 0 , 1 = q and 2 = 0 : q X i =1 h 3 ( i; i ) C 4 ;y ( q ; 0 ; i + 1 ) h ( q ; q ) C 4 ;y (0 ; 0 ; 1 ) = C 4 ;y ( q ; 0 ; 1 ) ; (36) where 1 = q ; :::; q : In more compact f or m, the system of Eq. (36) can be wr itten in the f ollo wing f or m: M = A (37) = [ h ( q ; q ) h 3 (1 ; 1) h 3 (2 ; 2) ::::h 3 ( q ; q )] T is a column v ector of siz e ( q + 1) ; A = [0 ::: 0 C 4 y ( q ; 0 ; 0) C 4 y ( q ; 0 ; 1) :::: C 4 y ( q ; 0 ; q )] T is a v ector of siz e (2 q + 1) ; The least squares solution of the system of Eq. (37), will be wr itten under the f ollo wing f or m: b = ( M T M ) 1 M T d (38) The par ameters h ( i; i ) f or i = 1 ; :::; q are estimated from the estimated v alues b ( i; i ) using the f ollo wing equation: b h ( i; i ) = 3 q b ( i + 1 ; i + 1) (39) 5. Sim ulation results and Anal ysis In this section, w e testy the p roposed approach to identify the non linear impulse re- sponse par ameters of the non linear channels . F or this reason, w e select the model (Eq. 40) char acter izing a non linear quadr atic system, with kno wn par ameters and then w e tr y to reco v er these par ameters using proposed algor ithms . y ( k ) = x 2 ( k ) + 0 : 15 x 2 ( k 1) 0 : 35 x 2 ( k 2) + 0 : 90 x 2 ( k 3) z eros: z 1 = 1 : 1439 ; z 2 = 0 : 4969 + 0 : 7348 i ; z 3 = 0 : 4969 0 : 7348 i: (40) The model of the selected channel is a non minim um phase because one of its z eros are outside of the unit circle (Fig. 2). The sim ulation is perf or med with MA TLAB softw are an d f or diff erent signal to noise r atio (SNR) defined b y the f ollo wing relationship: S N R = 10 l og 10 h 2 z ( k ) 2 n ( k ) i (41) T o measure the accur acy of the diagonal par ameter estimation with respect to the real v alues , w e define the Nor maliz ed Mean Square Error (NMSE) f or each r un as: N M S E = q X i =0 h h ( i; i ) b h ( i; i ) h ( i; i ) i 2 (42) IJEECS V ol. 2, No . 2, Ma y 2016 : 334 343 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 341 −1 −0.5 0 0.5 1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 3 Real Part Imaginary Part Figure 2. The z eros of Model In order to test the rob ustness of the proposed algor ithms w e conseder a non minim um phase channel descr ibe b y (Eq. 40). The sim ulation results , using selected algor ithms , of impulse response par ameters estimation are sho wn in the T ab le 1 f or diff erent v alues of SNR a nd f or N=2400. F rom T ab le 1, w e can conclude that f or all v alues of SNR considered, the NMSE v alues of the T ab le 1. Estimated par ameters of the first channel f or diff erent S N R and e xcited b y sample siz es N = 2400 S N R b h ( i; i ) std ALGcum 1 ALGcum 2 ALGcum 3 b h (1 ; 1) std 0 : 1391 0 : 0960 0 : 0770 0 : 4375 0 : 1310 0 : 3395 0 dB b h (2 ; 2) std 0 : 4225 0 : 0940 0 : 5359 0 : 1832 0 : 5057 0 : 1666 b h (3 ; 3) std 0 : 8326 0 : 1120 0 : 8250 0 : 1420 0 : 5371 0 : 2283 N M S E 0 : 0538 0 : 5261 0 : 3766 b h (1 ; 1) std 0 : 1539 0 : 0622 0 : 1016 0 : 3243 0 : 2059 0 : 2406 8 dB b h (2 ; 2) std 0 : 3747 0 : 0527 0 : 3376 0 : 2095 0 : 4090 0 : 0708 b h (3 ; 3) std 0 : 7879 0 : 0757 0 : 8153 0 : 0984 0 : 6404 0 : 1436 N M S E 0 : 0212 0 : 1141 0 : 2507 b h (1 ; 1) std 0 : 1523 0 : 0455 0 : 1366 0 : 2959 0 : 1742 0 : 2425 16 dB b h (2 ; 2) std 0 : 3680 0 : 0500 0 : 2640 0 : 2366 0 : 3718 0 : 0724 b h (3 ; 3) std 0 : 8155 0 : 0589 0 : 8395 0 : 0826 0 : 6524 0 : 1158 N M S E 0 : 0117 0 : 0729 0 : 1056 b h (1 ; 1) std 0 : 1583 0 : 0475 0 : 1282 0 : 2880 0 : 1522 0 : 2609 24 dB b h (2 ; 2) std 0 : 3611 0 : 0547 0 : 2983 0 : 1876 0 : 3662 0 : 1072 b h (3 ; 3) std 0 : 8350 0 : 0609 0 : 8234 0 : 0812 0 : 6565 0 : 1048 N M S E 0 : 0093 0 : 0501 0 : 0755 b h (1 ; 1) std 0 : 1479 0 : 0492 0 : 1664 0 : 2536 0 : 1438 0 : 2580 32 dB b h (2 ; 2) std 0 : 3643 0 : 0451 0 : 2939 0 : 2011 0 : 3722 0 : 1007 b h (3 ; 3) std 0 : 8419 0 : 0643 0 : 8298 0 : 0930 0 : 6723 0 : 0955 N M S E 0 : 0060 0 : 0437 0 : 0697 T r ue par ameters h ( i; i ) h (1 ; 1) = 0 : 150 h (2 ; 2) = 0 : 350 h (3 ; 3) = 0 : 900 first proposed method such as (Al gcum1) are lo w er than the other methods (Algcum2, Algcum3), this is due to the comple xity of the systems of equations f or each algor ithm, non linea r of the par ameters in the (Algcum2, Algcum3) algor ithms . The perf or mance of the (Algcum2) method deg r ade than the (Algcum3) in v er y noise en vironment (SNR=0 dB), b ut it becomes more eff ectiv e than (Algcum3) when the noise v ar iance is relativ ely small, this is due the f act that the higher order cum ulants f or a Gaussian noise are not identically z ero , b ut the y ha v e v alues close to z ero f or higher data length. This is v er y clear in the (Fig. 3). In the par t, of comple xity of these algor ithms the first proposed algor ithm e xploiting ( q + 1) Extension of Linear Channels Identification Algor ithms to Non Linear Using Selected Order Cum ulants (M. Zidane) Evaluation Warning : The document was created with Spire.PDF for Python.
342 ISSN: 2502-4752 0 5 10 15 20 25 30 0 0.1 0.2 0.3 0.4 0.5 Signal to Noise Ratio SNR(dB) Normalized Mean Square Error (NMSE)     NMSE using Algcum1 NMSE using Algcum2 NMSE using Algcum3 Figure 3. NMSE f or each algor ithm and f or diff erent SNR and f or a data length N = 2400 equations , compar ing de second and third proposed methods e xploiting (2 q + 1) f or identify the impulse response par ameters channel. In the Fig. 4 w e ha v e presented the estimation of the magnitude and the phase of the impulse response using the proposed algor ithms , f or data length N = 2400 and v er y noise en vi- ronment SNR =0 dB . F rom the Fig. 4 w e remar k that the magnitude estimation ha v e the same appear ance using tw o first proposed methods b ut using (Algcum3) algor ithm w e ha v e a minor dif- f erence betw een the estimated and tr ue ones . Concer ning the phase estimation, w e ha v e same allure compar ativ ely to the real model using all proposed algor ithms . 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 −600 −400 −200 0 200 Normalized Frequency  ( × π  rad/sample) Phase (degrees) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 −20 −10 0 10 Normalized Frequency  ( × π  rad/sample) Magnitude (dB)     True channel Estimated using Algcum1 Estimated using Algcum2 Estimated using Algcum3 (True, Algcum1, Algcum2) Algcum3 Figure 4. Estimated magnitude and phase of the non linear model channel impulse response when the data input is N=2400 and an SNR=0 dB IJEECS V ol. 2, No . 2, Ma y 2016 : 334 343 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 343 6. Conc lusion In this contr ib ution, w e ha v e considered the prob lem of b lind identification of non linear channel using selected order cum ulants . W e ha v e de v eloped a theoretical analysis f or non linear quadr atic systems , these tools will ser v e us later f or proposing an e xtension of linear algor ithms to non linear . Ho w e v er , w e ha v e de v eloped three approaches based on the three and f our th order cum ulants , respectiv ely , f or b lind identification of diagonal par ameters of quadr atic systems . F rom sim ulation results and compar ison betw een these methods one can see that the first proposed algor ithm (Algcum1) can alw a ys achie v e better perf or mance than other (Algcum1, Algcum2), and is adequate f or estimating diagonal quadr atic systems . The future w or k of this paper is the non linear Broadband Radio Access Netw or k (BRAN) channels identification and equalization especially MC-CDMA systems using the presented meth- ods . Ref erences [1] M. Zidane , S . Safi, M. Sabr i and A. Boumezz ough, “Higher Order Statistics f or Identification of Minim um Phase Channels , W or ld Academ y of Science Engineer ing and T echnology , In- ter national Jour nal of Mathematical, Computational, Ph ysical and Quantum Engi neer ing , v ol 8, No 5, pp . 831-836, (2014). [2] M. Zidane , S . Safi, M. Sabr i and A. Boumezz ough, “Blind Identification Channel Using Higher Order Cum ulants with Application to Equalization f or MC-CDMA System, W or ld Academ y of Science Engineer ing and T e c h nology , Inter national Jour nal of Electr ical, Robotics , Electron- ics and Comm unications Engineer ing , v ol 8, No 2, pp . 369-375, (2014). [3] M. Zidane , S . Safi, M. Sabr i, A. Boumezz ough and M. F r ik el, “Broadband Radio Access Netw or k Channel Identification and Do wnlink MC- CDMA Equalization, Inter national Jour nal of Energy , Inf or mation and Comm unications , v ol. 5, Issue 2, pp .13-34, (2014). [4] S . Safi and A . Zeroual, “Blind non minim um phase channel identification using 3 r d and 4 th order cum ulants , Int. J . Sig. Proces . , v ol. 4, No 1, pp . 158-168, (2008). [5] S . Safi, M. F r ik el, A. Zeroual, and M. M’Saad, “Higher Order Cum ulants f or Identification and Equalization of Multicarr ier Spreading Spectr um Systems , Jour nal of T elecomm unications and Inf or mation T echnology , pp . 74-84, 1/2011. [6] V . P . Leono v and A. N. Shir y ae v , “On a method of calculation of semi-in v ar iants , Theor y of probability and its applications , v ol. 4, No 3, pp . 319-329, (1959). [7] J . Antar i, A. Elkhadimi, D . Mammas , and A. Zeroual, “De v eloped Algor ithm f or Super vising Identification of Non Linear Systems using Higher Order Statistics : Modeling Inter net T r affic , Inter national Jour nal of Future Gener ation Comm unication and Netw or king , v ol. 5, No 4, pp . 17-28, (2012). [8] J . Antar i, S . Cha baab , R. Iqdour , A. Zeroual, S . Safi, “Identificat ion of quadr atic systems using higher order cum ulants and neur al netw or ks: Application to model the dela y of video- pac k ets tr ansmission, Jour nal of Applied Soft Computing (ASOC), Else vier , v ol. 11, No 1, pp . 1-10, (2011). [9] H. Z. T an, T . W . S . Cho w , “Blind identification of quadr atic non linear models using neur al netw or ks with higher order cum ulants , IEEE T r ansactions on Industr ial Electronics , v ol. 47, No 3, pp . 687-696, (2000). [10] H. Z. T an, Z. Y . Mao , “Blind identifiability of quadr atic non linear systems in higher order statistics domain, Inter national Jour nal of Adaptiv e Control and Signal Process , v ol. 12, No 7, pp . 567-577, (1998). [11] A. G. Stogioglou and S . McLaughlin, “MA par ameter estimation and cum ulant enhancement, IEEE T r ansactions on Signal Processing , v ol. 44, No 7, pp . 1704-1718, (1996). [12] K. Abderr ahim, R. B . Abdennour , F . Msahli, M. Ksour i, and G. F a vier , “Identification of non minim um phase finite impulse response systems using the f our th order cum ulants , Prog ress in system and robot analysis and control design, Spr inger , v ol. 243, pp . 41-50, (1999). Extension of Linear Channels Identification Algor ithms to Non Linear Using Selected Order Cum ulants (M. Zidane) Evaluation Warning : The document was created with Spire.PDF for Python.