Indonesian Journal of Electrical Engineering and Computer Science V ol. 3, No . 2, A ugust 2016, pp . 410 419 DOI: 10.11591/ijeecs .v3.i2.pp410-419 410 Using A Fuzzy Number Err or Correction Appr oac h to Impr o ve Algorithms in Blind Identification Elmostafa Atify * , Cherki Daoui , and Ahmed Boumezzough Labor ator y of Inf or mation Processing and Decision Suppor t F aculty of Sciences and T echniques P .B 523 Bni-Mellal * e-mail: at.elmost@gmail.com Abstract As par t of a detailed study on b lind identification of Gaussian channels , the main pur pose w as to propose an algor ithm based on cum ulants and fuzzy n umber approach in v olv ed throughout the whole process of identification. Our objectiv e w as to compare the ne w design of the algor ithm to the old one using the higher order cum ulants , namely Alg1, Algat and the Giannakis algor ithm. W e w ere ab le to demonstr ate that the proposed method -fuzzy n umber error correction- increases the perf or mance of the algor ithm b y calculating the r atio of squared errors of ALGaT and AlgatF . The method can be applied to an y algor ithm f or more impro v ement and effinciency . K e yw or ds: Blind identification, Fuzzy n umber , FIR channel,Gaussien channel, cum ulant. Cop yright c 2016 Institute of Ad v anced Engineering and Science 1. Intr oduction process of identification has no w become cr ucial and is prominent in se v er al fields , in- cluding astroph ysics , geology , data tr ansmission, r adio comm unication, mobile r adio . Thanks to Y .Sato that the issue of identification in its v ar ious aspects w as r aised. It has contib ueted to the resolution of man y prob lems [1]. The tr ansmission of inf or mat ion through a ph ysical medium ma y undergo se v er al ph ysical alter ations or modifications , aff ecting the nature and e v en the direction of the initial inf or mation. The y are essentially ph ysical phenomena whose impact is quite consider ab le on the authenticity of the message induced b y the inf or mation. Major e xamples include absor ption, refr actio n, re- flection or diffusion. These cases of impact gener ate a signal distor tion through atten uation and interf erence betw een symbols (IES). Moreo v er , in digit al comm unication, such phenomena alter- ing the amplitude and the phase can be modeled b y a m ulti-path tr ansmission channel inf ected with a white noise . Such a model uses digital Finite Impulse Response filter (FIR) inf ected b y a Gaussian white noise [2, 3, 4, 5, 6, 7]. Identification methods allo w us to dete r mine the channel impulse response of FIR. T ac k- ing into account on related liter ature , one can identify more methods and identifications that can be classified into three categor ies according to resolution methods [8] [9]. This in v olv es the use of o v ersiz ed linear algebr aic systems , e xplicit solutions and solutions using cum ulants which are easy to implement throughly . Ho w e v er ,f or the first tw o methods of resolution there is a clear ineffi- ciency since minimizing functions presents man y neighbor ing local mini ma whose computational comple xity is big. The third method of resolution using a higher order cum ulant is also unreliab le and less efficient. It is f or this reason that w e think action should be tak en to impro v e the reliability of the identification through the use of fuzzy n umber method in the implementation of higher order cum ulants used in the case of non-Gaussian frequency distr ib utions . The fuzzy n umber error cor- rection method allo ws us to eliminate e xtreme v alues that ma y aff ect the calculation of targeted v alues . Our goal is to optimiz e efficiency the processes of b lind identification and the sensib le use of higher order cum ulants whose v alue f or a Gaussian distr ib ution is z er o . Moreo v er , their use in the case of identification of the noisy chann el b y a Gaussian white noise is v er y frequent. Our Receiv ed Apr il 9, 2016; Re vised J uly 12, 2016; Accepted J uly 25, 2016 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 411 study will f ocus on the b lind identification of linear non-Gaussian process adjusted a v er age (MA) [2, 3, 4, 6, 7]. T o achie v e this goal,w e will present a typology of order cum ulants , algor ithm using three and f our cum ulants to impro v e the perf or mance of standard algor ithms . Ne xt, w e will implement the assumptions related to the channel model (MA) to identify it with its useful relationships . W e will mo v e on to the presentation of tw o estimation methods based on cum ulants including Alg1 algor ithm [1 0] et al, the algor ithm C (q, k) Giannakis [3]. Finally , w e will pro vide a detailed presentation of our algor ithm, which will be f ollo w ed b y a sim ulation to compare and assess the eff ectiv eness of diff erent algor ithms presented in this w or k. 2. Model and Fundamental realtionships 2.1. Hypothesis and model W e alw a ys design a model of a Sin gle Input and Single Output (SISO) channel with a m ultipath phenomenon,Fig.1, b y using a linear digital FIR. The equation of the finite diff erences model f or the FIR mo ving a v er age channel (Mobile A v er age: x(k)  y(k)  z(k) + additive noi se Canal( ) n(k) B Figure 1. FIR Model channel MA), is represented b y the f ollo wing [10], [11], with out noiseless: Y ( k ) = q X j =0 b ( j ) :X ( k j ) ; b (0) = 1 ( outnoisel ess ; ) (1) and with noise: Z ( k ) = Y ( k ) + n ( k ) (2) where X ( k ) is a non-Gaussian e xcitation inaccessib le to independent components an d identically distr ib uted (iid) with z ero mean, of v ar iance 2 x ,with at least one non-z ero m > 2 order and chec king E [ X 2 m ( k )] < 1 . n ( k ) is a white Gaussian noise , independent of the input X ( k ) and unkno wn po w er spectr al den- sity . B = [ b (0) ; b (1) ; ::; b ( q )] represent the impulse response of FIR channel. b ( i ) are constant f or a stationar y time-in v ar iant channel. q is the order of the channel, assumed to be kno wn [12]. 2.2. Fundamental Relationships Related liter ature sho ws that there are man y impor tant algor ithms based on higher order cum ulant. In this par ag r aph, w e present the fundamental relationships using the cum ulants in the case of a stationar y time-in v ar iant Gaussian noisy channel. 2.2.1. Moment and Cum ulant In this section, w e presen t some definitions of higher order sta tistics , moments and cu- m ulants . Let x(k), where 1 k N, is a real discrete stationar y process with N length, so its moment of order m is giv en b y [3] [4] [10] [5] M m;x ( t 1 ; t 2 ; :::; t m 1 ) = E f x ( k ) x ( k + t 1 ) x ( k + t 2 ) :::x ( k + t m 1 ) g (3) Fuzzy Number F or Blind cum ulants Identification (Elmostaf a Atify) Evaluation Warning : The document was created with Spire.PDF for Python.
412 ISSN: 2502-4752 Where E f : g represents the mathematical e xpectation. The cum ulant of order n of a non-Gaussian stationar y process is giv en b y: C m;x ( t 1 ; t 2 ; :::; t m 1 ) = M m;x ( t 1 ; t 2 ; :::; t m 1 ) M m;G ( t 1 ; t 2 ; :::; t m 1 ) (4) This relationship sho ws the impor tance of cum ulants estimators relativ ely to the time when it comes noise Gaussian in nature . 2.2.2. Higher or der cum ulants The most used moments in pr actice are moment s of order m lo w er or equal to 5. In this section w e giv e the e xpression cum ulant based moments . The giv en e xpressions are simplified in the case of the samples adjusted to a z ero mean (centered System C.S ). The cum ulant of order m = 1 is giv en b y: C 1 ;x = M 1 ;x = E f x ( k ) g : (5) is equal to 0 f or a z ero-mean sample: centered sample . The e xpression (5) is equal to 0 f or a z ero-mean sample: centered sample ( C.S ). The cum ulant of order m = 2 is giv en b y: C 2 ;x ( t 1 ) = M 2 ;x ( t 1 ) ( M 1 ;x ) 2 (6) In the case of a system ( C.S ) e xpression (6) becomes: C 2 ;x ( t 1 ) = M 2 ;x ( t 1 ) (7) The cum ulant of order m = 3 is wr itten as: C 3 ;x ( t 1 ; t 2 ) = M 3 ;x ( t 1 ; t 2 ) M 1 ;x ( M 2 ;x ( t 1 )+ M 2 ;x ( t 1 t 2 )) + 2( M 1 ;x ) 2 (8) F or a system ( C.S ) e xpression (8) becomes: C 3 ;x ( t 1 ; t 2 ) = M 3 ;x ( t 1 ; t 2 ) (9) F or a system ( C.S ), the cum ulant of order m = 3 is wr itten as: C 4 ;x ( t 1 ; t 2 ; t 3 ) = M 4 ;x ( t 1 ; t 2 ; t 3 ) M 2 ;x ( t 1 ) M 2 ;x ( t 3 t 2 ) M 2 ;x ( t 2 ) M 2 ;x ( t 3 t 1 ) M 2 ;x ( t 3 ) M 2 ;x ( t 2 t 1 ) (10) 2.2.3. Brilling er and Rosenb latt Equation The common point of all con v entional methods of identifying adjusted a v er age (MA) mod- els is the use of Br illinger and Rosenb latt f or m ula [2] which, under the abo v e assumptions is: C m;Z ( 1 ; :::; m 1 ) = C m;Y ( 1 ; ::: m 1 ) = m;x P q i =0 b ( i ) b ( i + 1 ) :::b ( i + m 1 ) (11) F or m = 2 , the autocorrelation is: C 2 ;Z ( ) = C 2 ;Y ( ) + C 2 ;N ( ) (12) where C 2 ;N ( ) is the autocorrelation of the noise sk e wing results and C 2 ;Y ( ) is the autocorrela- tion of the non-noisy signal e xpressed b y: C 2 ;Y ( ) = 2 ;x q X i =0 b ( i ) b ( i + ) ; ( 2 ;x = 2 x ) (13) IJEECS V ol. 3, No . 2, A ugust 2016 : 410 419 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 413 According to (11), one can easily demonstr ate that the order cumm ulants m et n , with ( m > n ), meet the f ollo wing relationship: P q i =0 b ( i ) C m;Y ( i + 1 ; ::i + n 1 ; n ; :::; m 1 ) = " m;n P q i =0 b ( i ) h Q m 1 j = n b ( i + j ) i C n;Y ( i + 1 ; ::i + n 1 ) (14) Where " m;n = m;x n;x . This gener al equation estab lishes se v er al basic algor ithms and will also be the basis of our proposed algor ithm. 2.3. Based unique or der cumm ulants algorithms The algor ithms based only on higher order cumm ulants are interesting when the pro- cessed signal is contaminated b y an additiv e Gaussian noise . Indeed cumm ulants of higher or equal to three orders of a Gaussian distr ib ution is z ero . 2.3.1. Algorithm Based on 4th Or der Cum ulant using equations 2q +1: Alg1 F rom equation (11) The matr ix f or m of the algor ithm is giv en b y Alg1 [11] 0 B B B B B B B B B B B @ 0 0 C 4 ;y ( q ; q ; 0) . . . . . . . . . 0 C 4 ;y ( q ; q ; 0) C 4 ;y ( q ; q ; q ) . . . . . . 0 . . . C 4 ;y ( q ; q ; q ) 0 0 1 C C C C C C C C C C C A 0 B B B B B B B B B B @ 1 b 2 ( q ) . . . b 3 ( i ) b 2 ( q ) . . . b 3 ( q ) b 2 ( q ) 1 C C C C C C C C C C A = 0 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 A (15) in a more compact f or m, the system of equations (15) can be wr itten as f ollo ws: M b q = d (16) with M , and h q are defined in the equation system (15). The solution in the sense of least squares , LS , of the system of equation (16) is giv en b y: b h ( q ) = ( M T M ) 1 M T d (17) this solution giv es us an estimate of the quotient of par ameters b 3 ( i ) and b 3 ( q ) , b y: h q ( i ) = \ b 3 ( i ) b 3 ( q ) ; i = 1 ; :::; q : (18) So , to estimate the par ameters b b ( i ) , i = 1 ; :::; q w e proceed as f ollo ws: The par ameters b ( i ) f or i = 1 ; :::; q 1 are estimated from estimates of b h q ( i ) v alues using the f ollo wing equation: b b ( i ) = sig n h b h q ( i )( b h q ( q )) 2 i f abs ( b h q ( i ))( b h q ( q )) 2 g 1 = 3 (19) a v ec sig n ( x ) = 8 < : 1 ; if x > 0; 0 ; if x = 0; 1 ; if x < 0 : and abs ( x ) = j x j indicates the absolute v alue of x. Fuzzy Number F or Blind cum ulants Identification (Elmostaf a Atify) Evaluation Warning : The document was created with Spire.PDF for Python.
414 ISSN: 2502-4752 The par ameter b b ( q ) is estimated as f ollo ws: b b ( q ) = 1 2 sig n h b h q ( q ) i 8 < : abs ( b h q ( q )) +   1 b h q (1) ! 1 = 2 9 = ; (20) 2.3.2. Algorithme ’C(q,k)’ of Giannakis F rom (11), Giannakis sho w ed that the coefficients (FIR) can be e xpressed b y the f ollo wing f or m ula: b ( ) = C m;Y ( q ; ; 0 ; :::; 0) C m;Y ( q ; 0 ; :::; 0) (21) with = 0,...,q and the cum ulant of order m of e xcitation is: m;x = C 2 m;Y ( q ; 0 ; :::; 0) C m;Y ( q ; q ; :::; 0) (22) F or m = 3, w e ha v e: b ( ) = C 3 ;Y ( q ; ) C 3 ;Y ( q ; 0) et 3 ;x = C 2 3 ;Y ( q ; 0) C 3 ;Y ( q ;q ) 2.4. Pr oposed Algorithm In this section the impulse response B = [ b (0) ; b (1) ; :::; b ( q )] is proposed to estimate a q order RIF channel using an algor ithm that combines cum ulants of order 3 and 4, as a pre viously proposed h ypothesis . It also e xplains the method that impro v es the proposed algor ithm. 2.4.1. General equation Equation (14) is tr ansf or med into an equation which links m and n such that m = n + 1 as f ollo wing: P q i =0 b ( i ) C m;Y ( i + 1 ; ::i + n 1 ; n ) = " m;n P q i =0 b ( i ) b ( i + n ) C n;Y ( i + 1 ; ::i + n 1 ) (23) 2.4.2. Appr oac h combining 3 and 4 cum ulants or der Especially m = 4 et n = 3 , Equation (23) becomes: P q i =0 b ( i ) C 4 ;Y ( i + 1 ; i + 2 ; 3 ) = " 4 ; 3 P q i =0 b ( i ) b ( i + 3 ) C 3 ;Y ( i + 1 ; i + 2 ) (24) W e tak e 1 = 2 = q et 3 = , the equation (24) becomes: q X i =0 b ( i ) C 4 ;Y ( i + q ; i + q ; ) = " 4 ; 3 q X i =0 b ( i ) b ( i + ) C 3 ;Y ( i + q ; i + q ) (25) giv en that C 4 ;Y ( 1 ; 2 ; 3 ) = C 3 ;Y ( 1 ; 2 ) = 0, si i > q ; the equation (25) becomes: b (0) C 4 ;Y ( q ; q ; ) = " 4 ; 3 b (0) b ( ) C 3 ;Y ( q ; q ) (26) W e deduce: b ( ) = C 4 ;Y ( q ; q ; ) " 4 ; 3 C 3 ;Y ( q ; q ) (27) with " 4 ; 3 = 4 ;x 3 ;x (28) IJEECS V ol. 3, No . 2, A ugust 2016 : 410 419 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 415 According to equation (22), w e deduce: " 4 ; 3 = C 2 4 ;Y ( q ; 0 ; 0) C 4 ;Y ( q ; q ; 0) C 3 ;Y ( q ; q ) C 2 3 ;Y ( q ; 0) (29) then b ( ) = C 4 ;Y ( q ; q ; 0) C 2 4 ;Y ( q ; 0 ; 0) C 2 3 ;Y ( q ; 0) C 3 ;Y ( q ; q ) C 4 ;Y ( q ; q ; ) C 3 ;Y ( q ; q ) (30) 2.4.3. AlgatF The reduction of n umer ical calculations and the perf or mance of the used statistical es- timator can be a source of some div ergence of v alues compared to the tr ue v alue . T o minimiz e these error diff erences sign w e will also proposes a selectiv e choice of estimated v alues of im- pulse responses from the pre vious algor ithms in the f ollo wing f or mat: Since each calculated v alue is accompanied b y an error , it is theref ore considered as a fuzzy n umber [13] defined b y an inter v al in the set R b y the f ollo wing figure ,Fig.2: x x x x x Figure 2. Fuzzy n umber representation Fig.3, represents fuzzy v alues obtained b y iter ativ e sim ulation. Fuzzy v alues ma y be intersecting or not. W e remo v ed fuzzy e xtreme v alues ha ving a z ero intersection with the other fuzzy v alues . Indeed, these fuzzy v alues are f ar from the tr ue v alue . Note that the n umber of fuzzy x x x x x Figure 3. Representation of a fuzzy n umber of estimated ser ies v alues , remaining after remo v al of the end m ust be g reater than at least half of the iter ations . AlgatF is the method of selection applied on ALGaT giv en that the fuzzy v ar iab le is se- lected b y: B = q X i =0 b ( i ) (31) where 2x B is the siz e fuzzy inter v al. The sum is f ed to remo v e the div ergence due to the undesired occurrence of the min us sign in one of the component of the impulse response . 3. Sim ulation In this sim ulation, w e tak e 100 iter ations and each time a ne w sample is tak en b y a noisy Gaussian noise with z ero mean. The diff erent algor ithms pro vide estimates f or the same samples in siz es 400, 800 and 1200 respectiv ely . T o compare the samples using the mean square error defined as f ollo ws: E QM = 1 q + 1 q X i =0 ( b ( i ) h ( i )) 2 (32) Fuzzy Number F or Blind cum ulants Identification (Elmostaf a Atify) Evaluation Warning : The document was created with Spire.PDF for Python.
416 ISSN: 2502-4752 Consider the channel, non-minim um phase (there is a z ero of the tr ansf er function outside the unit circle), figure (4) belo w , ha ving the impulse response H = [1 1 ; 083 0 ; 95 0 ; 95] . The 0 2 −1 1 −1.5 −0.5 0.5 1.5 0 −1 1 −1.2 −0.8 −0.6 −0.4 −0.2 0.2 0.4 0.6 0.8 1.2 Axe réel Axe des imaginaires Pôles Zéros Zéros et pôles de transmission Figure 4. The z eros and poles f or channel par ag r aphs belo w summar iz e sim ulation on channel 1 f or the v ar ious algor ithms presented abo v e in the case where the noise signal to noise r atio SNR equal to 10 dB and in case SNR equal to 20 dB . With S N R = 10 Log 10   2 y 2 br uit ! (33) where 2 i is the standard de viation of the statistical distr ib ution ( i ). Giv en that h (1) = 1. The least precise v alue of h ( i ) comes three significant digits . Our choice of the error on B is also to 3 significant fig ures in the f ollo wing is tak en into sim ulation B = 0 ; 03 . 3.1. Case SNR is 10 dB The f ollo wing tab le summar iz es the results obtained f or the proposed channel, the f our algor ithms namely Alg1, Alg of Gianakis , AlgaT and AlgatF . The AlgaT corrected b y the proposed selection method in case S N R = 10 dB . The descr iptiv e data tab le (1), allo ws us to see a clear impro v ement of EQM. Indeed, f or a sample siz e of 400, w e note that the proposed method ensures amelior ation, EQM b y a f actor of 2. In addition, the 800 sample reaches a f act or of about 5.1, more than doub le . This f actor will increase and reach about 20 in the case of the siz e of the sample 1200 f or the same method. This increase ensures thereb y a minimizing of the EQM v ersus other algor ithm. The impro v ement associated with, according to the v ar iab le fuzzifier , the method of the prob lem of the sign is remar kab le and is also the e xample of the estimated h 4 AlgatF b y the sample siz e to 800. This error sign is corrected b y AlgatF . Thus the cr iter ion of choice is cr ucial and e v en decisiv e in impro ving the div ergence of the calculation. Note at this le v el the e xample of the estimated h 4 b y Alga T at AlgatF f or 1200. According to the sample siz e Figure 5, w e note that the cur v e coincides perf ectly with the AlgatF ideal channel cur v e . IJEECS V ol. 3, No . 2, A ugust 2016 : 410 419 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 417 10 −1 10 0 2.0x10 −1 4.0x10 −1 6.0x10 −1 8.0x10 −1 0 −20 20 −30 −10 10 Normalized Frequency (x %pi rad/sample) Magnitude (dB) Magnitude 10 −1 10 0 2.0x10 −1 4.0x10 −1 6.0x10 −1 8.0x10 −1 0 −200 200 −100 100 −150 −50 50 150 Normalized Frequency (x %pi rad/sample) Phase (Degrees) Ideal chanel ALG 1 Chanel Gianakis Chanel ALGaT Chanel ALGaTF Chanel Phase Figure 5. N = 1200 and SNR = 10 dB Magnitude and phase representation 10 −1 10 0 2.0x10 −1 4.0x10 −1 6.0x10 −1 8.0x10 −1 0 −20 −30 −10 10 −25 −15 −5 5 Normalized Frequency (x %pi rad/sample) Magnitude (dB) Magnitude 10 −1 10 0 2.0x10 −1 4.0x10 −1 6.0x10 −1 8.0x10 −1 0 −200 200 −100 100 −150 −50 50 150 Normalized Frequency (x %pi rad/sample) Phase (Degrees) Ideal chanel ALG 1 Chanel Gianakis Chanel ALGaT Chanel ALGaTF Chanel Phase Figure 6. N = 800 and SNR = 20 dB Fuzzy Number F or Blind cum ulants Identification (Elmostaf a Atify) Evaluation Warning : The document was created with Spire.PDF for Python.
418 ISSN: 2502-4752 T ab le 1. Estimed v alues of h i f or SNR = 10 dB and N = 400,800,1200 Algor ithms/Sample siz e h 1 h 2 h 3 h 4 EQM Ideal channel 1 -1,083 -0,95 0,95 0 Alg1/400 1 -0,346 -0,253 0,553 0.488 AlgGianakis/400 1 2,119 -1,724 1,984 1,544 AlgaT/400 1 -0,240 -0,494 0,220 0,539 AlgatF/400 1 -0,672 -0,592 0,699 0,268 Alg1/800 1 -0,610 -0,434 0,734 0,328 AlgGianakis/800 1 -0,555 0,155 -0,672 0,909 AlgaT/800 1 -0,227 -0,342 -0,226 0,706 AlgatF/800 1 -1,317 -0,833 1,116 0,139 Alg1/1200 1 -0,904 -0,458 0,917 0,235 AlgGianakis/1200 1 1,150 -2,407 2,539 1,388 AlgaT/1200 1 -0,615 -1,566 3,554 1,215 AlgatF/1200 1 -1.121 -0.874 0,847 0,060 3.2. Case SNR is 20 dB The obtained results are summar iz ed in T ab le 2, f or channel 1, f or the f our ALG1 algo- r ithms Alg of Gianakis , ALGaT and ALGatF the ALGa T corrected b y the selection method pro- posed in case S N R = 20 dB . T ab le 2. Estimed v alues of h i f or SNR = 20 dB and N = 400,800,1200 Algor ithms/Sample siz e h 1 h 2 h 3 h 4 EQM Ideal channel 1 -1,083 -0,95 0,95 0 Alg1/400 1 -0,417 -0,324 0,562 0,444 AlgGianakis/400 1 1,0773 -1,846 0,467 1,068 AlgaT/400 1 -1,340 -0,970 1,734 0,369 AlgatF/400 1 -0,986 -0,966 1,005 0,051 Alg1/800 1 -0,724 -0,562 0,845 0,242 AlgGianakis/800 1 -0,206 -0.080 0,045 0,685 AlgaT/800 1 -0,751 -0,720 1,063 0,188 AlgatF/800 1 -1,085 -1,040 0,871 0,053 Alg1/1200 1 -0,874 -0,487 0,961 0,227 AlgGianakis/1200 1 -0,055 -0,233 -0,253 0,777 AlgaT/1200 1 -2,572 -1,058 2,962 1,121 AlgatF/1200 1 -1,284 -0,935 1,177 0,136 According to the descr iptiv e data T ab le 2, w e can also see a big impro v ement in the EQM f or SNR = 20 dB . Indeed, f or 400 the siz e sample , one notes that the proposed method ensures the impro v ement of the EQM of a f actor 7 : 3 . Fur ther more , the sample reaches 800 orders of a f actor 3 : 5 , more than doub le . This f actor will increase and reach the 8 : 2 in the case of the sample siz e 1200. This increase also ensures minimization of the EQM v ersus other algor ithm. The impro v ement asso- ciated with the so called prob lem of the method ensures good con v ergence to the tr ue v alues of the impulse response . Figure 6 sho w that the cur v es of AlgatF coincides perf ectly with the cur v e of the ideal channel. IJEECS V ol. 3, No . 2, A ugust 2016 : 410 419 Evaluation Warning : The document was created with Spire.PDF for Python.
IJEECS ISSN: 2502-4752 419 4. Conc lusion Se v er al b lind identification algor ithms based on higher order cum ulants are usually used. Among them three e xamples Alg1, Alg of Gi annakis and AlgaT , w ere selected . W e applied the method based on the concept of fuzzy n umber on the latter one to obtain the corrected algor ithm AlgatF . In sim ulation , w e considered a non-minim um phase channel and the estimated impulse response of 100 iter ations f or SNR of about 10 dB and 20 dB f or the v ar ious algor ithms . W e w ere ab le to demonstr ate that the proposed method increases the perf or mance of the algor ithm b y calculating the r atio of squared errors of ALGaT and AlgatF . The method can be applied to an y algor ithm f or more impro v ement and efficiency . F or future research, w e intend to test the eff ect of the method on a small n umber of iter ations so as to minimiz e the e x ecution time of the algor ithms . Ref erences [1] Y .Sato , “A method of self-reco v er ing equalizatio n f or m ultile v el amplitude-modulation sys- tems , IEEE tr ansaction, comm , v ol. 23, no . 6, pp . pp 679 682, J une 1975. [2] D . Br illinger and L. Rosenb latt, Computation and inter pretation of kth order spectr a . Spectr al Analysis of Times Signals , Ne w Y or k : Wile y , 1967. [3] G. Giannakis and A. Sw ami, “Higher order statistics , Else vier Science Pub l , 1997. [4] M. I.BADI, E.A TIFY and S .SAFI, “Blind identification of tr ansmission channel with the method of higher-order cumm ulants , Inter national Jour nal of Adv ances in Science and T echnology , v ol. 6, no . 3, 2013. [5] E. K.Abid-Mer iam and F .Loubaton, “Predection erreur method f or second-ordre b lind identifi- cation, IEEE T r ansaction,Signal, Processing , v ol. 45, no . 3, pp . pp 694 705, 1997 - March. [6] S . Safi and A. Zeroual, “Blind par ametr ic identification of linear stochastic non gaussian fir systems using higher order cum ulants , Inter national Jour nal of Systems Sciences T a ylor F r ancis , Signal Processing , v ol. 44, no . 15, pp . pp:855–867, 2004. [7] X. D . Zhang and Y . S . Zhang, “Fir system identification using higher order statistics alone , IEEE T r ansactions , Signal Processing , v ol. 42, no . 12, pp . pp:2854 2858, 1994. [8] D . Dembl, “Identification du modle ar ma lineaires l’aide de statistiques d’ordres ele vs . appli- cation l’egalisation a v eugle , Ph.D . disser tation, J uillet-1995. [9] G.F a vier , “Identification de modles par amtr iques ar ,ma et ar ma a v ec des statistiques d’ordres supr ieur et analyse des prf or mances , GRETSI , pp . pp:137 140, Septembre-1993. [10] I. Badi, M. Boutalline , S . Safi, and B . Bouikhalene , “Blind identification and equalization of channel based on higher-order cumm ulants: Application of mc-cdma systems , in Multimedia Computing and Systems (ICMCS), 2014 Inter national Conf erence on , Apr il 2014, pp . 800– 807. [11] A. S .Safi, “Ma system identification using higher ordre cum ulants applications to modelling solar r adiation, Jour nal of Statistical Computation and Sim ulation , v ol. 72, no . 7, pp . pp 533 548, 2002. [12] A. S .Alsh ebeili and F . Cetin, “Cum ulant based identification approaches f or minim um phase fir system, IEEE T r ansaction, Signal Processing , v ol. 41, no . 4, pp . pp 1576 1588, 1993 - Apr . [13] A. N. Gani and S . N. M. Assar udeen, “An algor ithmic approach of solving fuzzy linear system using f our ier motzkin elimination method, Adv ances in Fuzzy Sets and Systems , v ol. 10, no . 2, pp . pp:95 109, 2011. Fuzzy Number F or Blind cum ulants Identification (Elmostaf a Atify) Evaluation Warning : The document was created with Spire.PDF for Python.