Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 9, No. 3, June 2019, pp. 1732 1741 ISSN: 2088-8708, DOI: 10.11591/ijece.v9i3.pp1732-1741 r 1732 Blind separation of complex-v alued satellite-AIS data f or marine sur v eillance: a spatial quadratic time-fr equency domain appr oach Omar Cherrak 1 , Hicham Ghennioui 2 , Nad ` ege Thirion Mor eau 3 , El Hossein Abarkan 4 1 LREA, Institut sup ´ erieur du G ´ enie Appliqu ´ e, Maroc 1,2,4 LSSC, Uni v ersit ´ e Sidi Mohamed Ben Abdellah, F acult ´ e des Sciences et T echniques, Maroc 3 Aix-Marseille Uni v ersit ´ e, CNRS, France 3 Uni v ersit ´ e de T oulon, France Article Inf o Article history: Recei v ed Jun 30, 2018 Re vised No v 21, 2018 Accepted Des 15, 2018 K eyw ords: Blind source separation Joint zero-(block) diagonalization Marine surv eillance Matrix decompositions Satellite-automatic identification system Spatial generalized mean ambiguity function Spatial time-frequenc y based approach ABSTRA CT In this paper , the problem of the blind separation of c omple x-v alued Satellite-AIS data for marine surv eillance is addressed. Due to the specific properties of the sources un- der consideration: the y are c yclo-stationary signals with tw o close c yclic frequencies, we opt for spatial quadratic time-frequenc y domain methods. The use of an additional di v ersity , the time delay , is aimed at making it possible to undo the mixing of signals at the multi-sensor recei v er . The suggested method in v olv es three main stages. First, the spatial generalized mean Ambiguity function of the observ ations acros s the array is constructed. Second, in the Ambiguity plane, Delay-Doppler re gions of high mag- nitude are determined and Delay-Doppler points of peak y v alues are selected. Third, the mixing matrix is estimated from these Delay-Doppler re gions using our proposed non-unitary joint zero-(block) diagonalization algorithms as to perform separation. Copyright c 2019 Institute of Advanced Engineering and Science . All rights r eserved. Corresponding A uthor: Omar Cherrak, LREA, Institut sup ´ erieur du G ´ enie Appliqu ´ e, 279 Bd Bir Anzarane, Casablanca, Maroc. omar .cherrak@ig a.ac.ma 1. INTR ODUCTION This paper concerns the spatial automatic identification system (S-AIS) dedicated to marine surv eil- lance by satellite. It co v ers a lar ger area than the t errestrial automatic identification system [1], [2]. The idea of switching to satellite monitoring w as introduced because of the f ast and dynamic de v elopment of the ma- rine traf fic [3–5]. It w as an emer genc y to adopt a method that operates a global monitoring with reliability , ef ficienc y and rob ustness. Ho we v er , this generalization to space in v olv es se v eral phenomena. Among these phenomena, we found: (a) The speed of the satellite mo v ement generates the Doppler ef fect which creates frequenc y of fsets at the S-AIS signals [6], (b) The propag ation delay of the signals and their attenuation due to the satellite altitude [7], (c) When a wide area is co v ered by the satellite, it certainly includes se v eral traditional AIS cells. In f act, the ti me propag ation of signals transmitted from v essels to the satellite v ary according to the position J ournal homepage: http://iaescor e .com/journals/inde x.php/IJECE Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1733 of the ships and the maximum co v erage area of t he satellite antenna. Due to these tw o problems, it mainly af fects the or g anizational mechanism of S-AIS signals. It results a collision data, as illustrated in the Figure 1, issued by v essels located in dif ferent AIS cells b ut recei v ed at the antenna of the same satellite [8], [9]. F or this reason, we present ne w approaches to address this problem where the Doppler ef fect and the propag ation delay are also tak en into consideration. Figure 1. Collision problem: The AIS signals from tw o dif ferent SO-TDMA cells recei v ed to the satellite antenna at the same time. In f act, to solv e the collision problem, fe w w orks ha v e focused on blind s eparation of sources ( BSS ) methods [10], [11]. In [11], Zhou et al. present a multi-user recei v er equipped with an array of antennas embedded in Lo w Orbit Eart h (LEO) satellite. The principle of this recei v er is to e xploit spatial multiple xing in a non-stationary asynchronous conte xt. Indeed, the authors consider the equation belo w: X = HG ( S ) + N ; (1) where is the Schur -Hadamard operator , X = [ x 1 ; x 2 ; : : : ; x P N ] 2 C M P N , x n = x ( nT s ) , 1 n P N , is the observ ation matrix, H = [ h 1 ; : : : ; h d ] 2 C M d is the matrix of antenna response, G = diag f g 1 ; g 2 ; : : : ; g d g 2 R d d contains the po wer of the sources, S = [ s H 1 ; s H 2 ; : : : ; s H d ] H 2 C d P N is the matrix of sources and = 0 B B B @ 1 ' 1 1 : : : ' P N 1 1 1 ' 1 2 : : : ' P N 1 2 . . . . . . . . . . . . 1 ' 1 d : : : ' P N 1 d 1 C C C A , where ' k = e j 2 f k T s contains the Doppler frequencies of the sources. The principle of this method is based on joint diagonalization (JD) of matrices in order to reconstruct the S-AIS sourc es from separation matrix estimation [12]. Ho we v er , because of the v ery specific properties of the S-AIS signals (comple x and c yclo-stationary with tw o close c yclic frequencies), we opt for spatial quadratic time-frequenc y domain meth- ods. Our aim is reshaping the collision problem into BSS problem more simpler than (1). W e will sho w ho w another type of decomposition matrix named joint zero-diagonalization (JZD) of matrices set resulting from spatial quadratic time-frequenc y distrib utions allo ws the restitution of S-AIS sources. 2. TRANSMISSION SCHEME 2.1. AIS Frame The AIS frame is a length of 256 bits and occupies one minute. It is di vided into 2250 time s lots where one slot equals 26.67 ms [13]. Its structure as illustrated in Figure 2 contains a trai ning sequence (TS) consisting zero and on e which tak es 24 bits. The start flag (SF) and the end flag (EF) for information tak es 8 bits. A Frame Check Sequence (FCS) (or 16 bits Cyclic Redundanc y Code (CRC)) is added to the data information (168 bits) in which a zero is inserted after e v ery v e cont inuous one. The binary sequence f a k g 0 k K of the AIS frame tak es the v alues f 1 ; +1 g since the NRZI encoding is used. Moreo v er , the modulation specified Blind separ ation of comple x-valued satellite-AIS data ... (Omar Cherr ak) Evaluation Warning : The document was created with Spire.PDF for Python.
1734 r ISSN: 2088-8708 by S-AIS standard is Gaussian Minimum Shift K e ying (GMSK) [14]. The encoded message is modulated and transmitted at 9600 bps on 161.975 MHz and 162.025 MHz frequencies carrier . Figure 2. AIS Frame. 2.2. GMSK modulation The resulting sequence after the bit st uf fing and NRZI coding procedure is modulated with GMSK which is a frequenc y-shift k e ying modulation producing constant-en v elope and continuous-phase. Hence, the signal can be written as s g ( t ) = P + 1 k =0 a k g ( t k T s ) , where a k are the transmitted symboles, T s is the symbol period and g ( t ) = q 2 log 2 B exp 2 log 2 ( B t ) 2 represents the shaping Gaussian filter where B is the band- width of the Gaussian filter . The GMSK m odulation is described by the bandwidth-time (BT) product where S-AIS uses BT = 0 : 4 and T s = 1 9600 s ). Making the signal on one of the frequencies carrier f c , produces a signal of spectral characteristic which is adapted to the band-pass channel transmission. The GMSK signal is, thus, e xpressed as : s ( t ) = <f e j (2 f c t + ( t )) g = I ( t ) cos(2 f c t ) Q ( t ) sin(2 f c t ) ; where <fg is the real part of a comple x number , ( t ) = 2 h P + 1 k =0 a k g ( t k T s ) is the instantaneous phase of s g ( t ) where, in the AIS system, the modulation inde x is theoretically equal to h = 0 : 5 [15], I ( t ) (resp. Q ( t ) ) modulates the frequenc y carrier in phase (resp. in phase quadrature). All steps of the GMSK modulati on can be presented in the Figure 3. sin ( t ) Gaussian filter Inte grator cos ( t ) a k s g ( t ) ( t ) sin(2 f c t ) cos(2 f c t ) I ( t ) Q ( t ) s ( t ) Figure 3. GMSK modulator scheme. 3. PR OBLEM ST A TEMENT : COLLISION & BSS IN INST ANT ANEOUS CONTEXT 3.1. Mathematical model of collision pr oblem The collision problem can be simply e xpressed as follo ws: x ( t ) = J X j =1 h j s j ( t j ) e i 2 f j t + n ( t ) ; (2) where x ( t ) is the recei v ed signal by the satellite, s j is the transmitted signal by the j th v essel, h j , j and f j are respecti v ely the coef ficients of the channel, the delay and the Doppler shift corresponding to the j th v essel with J is the number of v essels and n ( t ) is an additi v e stationary white Gaussian noise, mutually uncorrelated, independent from the s j , with the v ariance 2 n . 3.2. Reshape the collision pr oblem into BSS pr oblem W e sho w , here, that (2) can be written in BSS nomenclature in which the delay and the Doppler shift caused by the satellite speed are considered. Ho we v er , before an y reformulation, we notice that the mixing Int J Elec & Comp Eng, V ol. 9, No. 3, June 2019 : 1732 1741 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1735 matrix considered for S-AIS application is an inst antaneous mixture due to the absence of obstacles in the ocean. Thus, we set J = n , the collision problem can be easily modeled in a BSS problem as follo ws: x ( t ) = Hs ( t ) + n ( t ) ; (3) where H is a ( m n ) mixing matrix, s ( t ) = [ s 1 ( t ) ; s 2 ( t ) ; : : : ; s n ( t )] T is a ( n 1) sources v ector with s j ( t ) = s j ( t j ) exp f i 2 f j t g ; 8 j = 1 ; : : : ; n and x ( t ) = [ x 1 ( t ) ; x 2 ( t ) ; : : : ; x m ( t )] T ; n ( t ) = [ n 1 ( t ) ; n 2 ( t ) ; : : : ; n m ( t )] T are respecti v ely the ( m 1) observ ations and noises v ectors. The superscript ( : ) T denotes the transpose operator . Our de v elopments are based on the follo wing assumptions: Assumption A : The noises n j ( t ) for all j = 1 ; : : : ; m are stationary , white, zero-mean, mutually uncorrelated random signals and independent from the sources with v ariance 2 n . Assumption B : F or each s j of the n sources, there Delay-Doppler points of only one source is present in the Ambiguity plane. Assumption C : The number of sensors m and the number of sources n are both kno wn and m n to deal with an o v er -determined model (the under -determined case is outside of the scope in this paper). 4. PRINCIPLE OF THE PR OPOSED METHODS B ASED ON THE SP A TIAL GENERALIZED MEAN AMBIGUITY FUNCTION W e sho w , here, ho w the algorithms proposed in [16], [17] adresses the problem of the separation of instantaneous mixtures of S-AIS data. The principle of the proposed methods are based on three m ain steps: first, the SGMAF of the observ ations across the array is constructed. Second, in the Ambiguity plane, Delay- Doppler re gions of high magnitude are determined and Delay-Doppler points of peak y v alues are selected. Third, the mixing matrix is estimated from these Delay-Doppler re gions so as to perform separation and to undo the mixing of signals at the multi-sensor recei v er . 4.1. The Spatial Generalized Mean Ambiguity Function W ith re g ard to BSS , it has been sho wn that spatial time-frequenc y distrib utions are an ef fecti v e tool when signature of the sources dif fer in certain points of the time-frequenc y plan [18]. Ho we v er , in the c yclo- stationary sources case, the Delay-Doppler frequenc y domain seems to be a more natural field for the re- estimation of sources than the time-frequenc y domain. As mentioned in [19], the approaches based on infor - mation deri v ed from spatial Ambiguity function ( SAF ) or on SGMAF should be used. In f act, for an y v ectorial comple x signal z ( t ) , the SGMAF is e xpressed as [20–22]: A z ( ; ) = Z 1 1 r z ( t; ) e j 2 t dt = E fh z ; s ; z ig ; (4) where ( s ; z ) is the operator of elementary Delay-Doppler translations of z defined by ( s ; z )( t ) = z ( t ) e j 2 ( t ) and r z ( t; ) = R z  t + 2 ; t 2  = E z t + 2 z H t 2  , where R z ( t; ) stands for the correlat ion matrix of z ( t ) , E f : g stands for the mathematical e xpectation operator and superscript ( : ) H denotes the conjug ate transpose operator . A z ( ; ) characterizes the a v erage correlation of all pairs se pa- rated by in time and by in frequenc y [21], [22]. Notice that the diagonal terms of the matrix A z ( ; ) are called auto-terms, while the other ones are called cross-terms. 4.2. Selection of peak y Delay-Doppler points Under the linear data model in (3), the SGMAF of observ ations across the array at a gi v en Delay- Doppler point is a ( m m ) matrix admits the follo wing decomposition: A x ( ; ) = H A s ( ; ) H H + A n ( ; ) ; = H A s ( ; ) H H + R n ( ) ; (5) where A s ( ; ) represents the ( n n ) SGMAF of sources defined similarly to A z ( ; ) in (4) and R n ( ) = 2 n ( ) I m with ( ) = R 1 1 e j 2 t dt and I m is the m m identity matrix. It is kno wn that the matrix A s ( ; ) for an y and has no special structure. Ho we v er , there are some Delay-Doppler points where this matrix has a specific algebraic structure : (a) Diagonal, for points where each of them corresponds to a single auto-source term for all source signals, Blind separ ation of comple x-valued satellite-AIS data ... (Omar Cherr ak) Evaluation Warning : The document was created with Spire.PDF for Python.
1736 r ISSN: 2088-8708 (b) Zero-diagonal for points where each of them correspond t o all tw o by tw o cross-source term (this struc- ture is e xploited because the signature of the sources dif fer in cert ain points of the Delay-Doppler plan on the zero-diagonal part (as sho wn in section 5.). Our aim is to tak e adv antage of these properties of the A x ( t; ) in (5) since the element of this is no more (zero) diagonal matrices due to the mixing ef fect in order to estimate the separation matrix B (the pseudo-in v erse of matrix H ) and restore the unkno wn sources. 4.3. Construction of M (set of Delay-Doppler matrices of the obser v ations acr oss the array at the cho- sen Delay-Doppler points) W e use the detector suggested in [23] (denoted C Ins ) for the instantaneous mixture consi dered without pre-whitening of the observ ations. The idea is to find “useful” Delay-Doppler points which consists in k eeping Delay-Doppler points with a suf ficient ener gy , then using the rank-one property to detect single cross-source terms (we don’ t mak e an y assumptions on the kno wledge of c yclic frequencies) in the follo wing w ay: 8 > < > : k A x ( t; ) k > 1 ; max A x ( t; ) k A x ( t; ) k 1 > 2 ; (6) where 1 , 2 are (suf ficiently) small positif v alues and max f : g is the lar gest eigen v alue of a matrix. 4.4. Non-unitary joint zer o-(block) diagonalization algorithms ( NU JZ ( B ) D ) The matrices belonging to the set M (whose size is denoted by N m ( N m 2 N )) all admit a particular structure since the y can be decomposed into H A s ( ; ) H H with A s ( ; ) a zero-diagonal matrix with only one non null term on its zero-diagonal. One possible w ay to reco v er the mixing matrix B is to directly joint ze ro- diagonalize the matrix set M . It has to be noticed that the reco v ered sources (after multiplying the observ ations v ector by the estimated matrix B ) are obtained up to a permutation (among the classical indetermination of the BSS ). Hence, tw o BSS methods can be deri v ed. The first called JZD CG DD algorithm based on conjug ate gradient approach [16]. The second JZD LM DD algorithm based on Le v enbre g-Marquardt scheme [17]. T o tackle that problem, we propose here, to consider the follo wing cost function [16], [17], C Z B D ( B ) = P N m i =1 k Bdiag ( n ) f BM i B H gk 2 F ; where the matrix operator Bdiag ( n ) f : g is defined as follo ws: Bdiag ( n ) f M g = 0 B B @ M 11 0 12 : : : 0 1 r 0 21 M 22 . . . 0 2 r . . . . . . . . . . . . 0 r 1 0 r 2 : : : M r r 1 C C A ; where M is a N N ( N = n ( L + L 0 ) where L is the order of the FIR filter and L 0 is the number of delays considered when the con v olutif mixture is considered) square matrix whose block components M ij for all i; j = 1 ; : : : ; r are n i n j matrices (and n 1 + : : : + n r = N ) denoting by n = ( n 1 ; n 2 ; : : : ; n r ) . Note that when L = 0 , L 0 = 1 we find the instantaneous model since A x are no more matrices b ut scalars. Thus, it leads to the minimization of the follo wing cost function: C Z D ( B ) = N m X i =1 k Diag f BM i B H gk 2 F ; (7) where M i = A x i is the i th of the N m matrices belonging to M . W e suggest to use conjug ate gradient and Le v enber g-Marquardt algorithms [16], [17] to minimize the cost function gi v en by Equation .(7) in order to estimate the matrix B 2 C n m . It means that B is re-estimated at each iteration m (denoted B ( m ) or b ( m ) when the v ector b ( m ) = vec B ( m ) is considered). The matrix B (or the v ector b ) is updated according to the follo wing adaptation rule for all m = 1 ; 2 ; : : : Int J Elec & Comp Eng, V ol. 9, No. 3, June 2019 : 1732 1741 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1737 Conjugate gradient appr oach b ( m +1) = b ( m ) ( m ) d ( m ) B ; d ( m +1) B = g ( m +1) + ( m ) d ( m ) B ; (8) where is a positi v e small f actor called the step-size, d B is the direction of search, is an e xact line search and g = vec ( r a C Z D ( B )) is the v ectorization of the comple x gradient matrix G = r a C Z D ( B ) = 2 P N m i =1 [ Diag f BM i B H g BM H i + Diag f BM i B H g H BM i i (see the proof pro vided in [16] ho w the optimal step-size opt , r a C Z D ( B ) and are calculated at each iteration). Le v enber g-Mar quardt appr oach b ( m ) = b ( m 1) h H ( m 1) e + I m 2 i 1 g ( m 1) ; (9) where [ : ] 1 denotes the in v erse of a matrix, is positi v e a small damping f actor , I m 2 is the m 2 m 2 identity matrix, H e =   H e B ; B = A 00 + A T 11 2 H e B ; B = A 01 + A T 01 2 H e B ; B = A 10 + A T 10 2 H e B ; B = H e B ; B T ! is the Hessian matrix of C Z D ( B ) com- posed of four comple x matrices with: A 00 = M T i B T I N T T Bof f ( B M i I N ) + M i B T I N T T Bof f B M T i I N + M i OBdiag ( n ) f BM i B H g + M T i OBdiag ( n ) f BM H i B H g = A 11 ; (10) A 10 = K T N ;M I N M i B H T T Bof f ( B M i I N ) + K T N ;M I N M H i B H T T Bof f B M T i I N = A 01 ; (11) where the operator denotes the Kroneck er product, K N ;M is a square commutation ma trix of size N M N M and T Bof f = I N 2 T Diag , is the N 2 N 2 “transformation” matrix, with T Diag = diag f vec ( BDiag f 1 N g ) g , 1 N is the N N matrix whose components are all ones, diag f a g is a square diagonal matrix whose diagonal elements are the elements of the v ector a , I N 2 = Diag f 1 N 2 g is the N 2 N 2 identity matrix, and Diag f A g is the square diagonal matrix with the same diagonal elements as A . 4.5. Summary of the pr oposed methods The proposed methods JZD CG DD and JZD LM DD combine the NU JZD algorithms which are JZD CG and JZD LM together with the detector C Ins . Its principles are summarized belo w: Data: Consider the N m matrices of set M : f A x 1 ; A x 2 ; : : : ; A x N m g , stopping criterion , step-size (for conjug ate gradient), max. number of iterations M max Result: Estimation of joint zero diagonalizer B initialize: B (0) ; (0) ; m = 0 ; D (0) (for conjug ate gradient); Conjugate gradient r epeat if m mo d M 0 = 0 then restart else Calculate ( m ) opt Compute g ( m ) Compute B ( m +1) Compute ( m ) PR Compute d ( m +1) B m = m + 1 ; end until (( k B ( m +1) B ( m ) k 2 F ) ou ( m M max )) ; Le v enber g-Mar quardt r epeat Calculate g ( m ) Calculate the diagonal of H e Calculate b ( m +1) Calculate the error e ( m ) = 1 N m C Z D ( B ( m +1) ) m = m + 1 ; if e ( m ) e ( m 1) then ( m ) = ( m 1) 10 , e ( m ) = e ( m 1) else ( m ) = 10 ( m 1) end until (( k B ( m +1) B ( m ) k 2 F ) ou ( m M max )) ; 5. COMPUTER SIMULA TIONS Computer simulations are performed to illustrate the good beha vior of the suggested methods and to compare them with the same kind of e xisting approach denoted by JZD Chab riel DD proposed in [24] with the Blind separ ation of comple x-valued satellite-AIS data ... (Omar Cherr ak) Evaluation Warning : The document was created with Spire.PDF for Python.
1738 r ISSN: 2088-8708 Delay-Doppler point C Ins detector . W e consider m = 3 mixtures of n = 2 frames of 256 bits correspond to tw o v essels with dif ferent characteristics. The frames are generated according to the S-AIS recommendation as mentioned in the Figure 2 (see also [11], [10]). These frames are encoded with NRZI and modulated in GMSK with a bandwidth-bit-time product parameter BT = 0 : 4 . The transmission bit rate is = 9600 bps and the order g aussian filter is OF = 21 . The frequenc y carrier of the first source (resp. the second source) is 161.975 MHz (resp. 162.025 MHz), taking into account a delay of 10 ms and a Doppler shift of 4 kHz (resp. a delay of 0 ms and the Doppler shift of 0 Hz). These sources correspond to 1400 time samples which are mix ed according to a mixture matrix H whose components stands for: H = 0 @ 1 : 1974 1 : 3646 0 : 8623 1 : 6107 0 : 1568 0 : 9674 1 A : (12) The real part and the imaginary part of their SGMAF is gi v en on the left and on the right of t he Figure 4 respecti v ely . Then, the SGMAF of the observ ations x is then calculated by (5) and finally the 100 resulting SGMAF are a v eraged. W e ha v e chosen 1 = 0 : 07 and 2 = 0 : 08 for the detector C Ins in order to construct the set M to be joint zero-diagonalized. The signal-to-noise ratio SNR is defined by SNR = 10 log ( 1 2 N ) of mean 0 and v ariance 2 n . The selected Delay-Doppler points using the proposed detector are represented in the Figure 5 for SNR = 10 dB and 100 dB. The SGMAF of the source 1 Cyclic frequency (Doppler) [Hz] 100 200 300 400 500 600 700 −6.481 −4.8608 −3.2405 −1.6203 0 1.6203 3.2405 4.8608 x 10 8 The SGMAF between the source 1 and 2 100 200 300 400 500 600 700 −6.481 −4.8608 −3.2405 −1.6203 0 1.6203 3.2405 4.8608 x 10 8 The SGMAF between the source 2 and 1 Time in samples 100 200 300 400 500 600 700 −6.481 −4.8608 −3.2405 −1.6203 0 1.6203 3.2405 4.8608 x 10 8 The SGMAF of the source 2 100 200 300 400 500 600 700 −6.481 −4.8608 −3.2405 −1.6203 0 1.6203 3.2405 4.8608 x 10 8 The SGMAF of the source 1 Cyclic frequency (Doppler) [Hz] 100 200 300 400 500 600 700 −6.481 −4.8608 −3.2405 −1.6203 0 1.6203 3.2405 4.8608 x 10 8 The SGMAF between the source 1 and 2 100 200 300 400 500 600 700 −6.481 −4.8608 −3.2405 −1.6203 0 1.6203 3.2405 4.8608 x 10 8 The SGMAF between the source 2 and 1 Time in samples 100 200 300 400 500 600 700 −6.481 −4.8608 −3.2405 −1.6203 0 1.6203 3.2405 4.8608 x 10 8 The SGMAF of the source 2 100 200 300 400 500 600 700 −6.481 −4.8608 −3.2405 −1.6203 0 1.6203 3.2405 4.8608 x 10 8 Figure 4. Left : The SGMAF real part of the S-AIS sources. Right : The SGMAF imaginary part of the S-AIS sources. T o measure the quality of the estimation, the ensuing error inde x is used [25] : I ( T ) = 1 n ( n 1) 2 4 n X i =1 0 @ n X j =1 k T i;j k 2 F max ` k T i;` k 2 F 1 1 A + n X j =1 0 @ n X i =1 k T i;j k 2 F max ` k T `;j k 2 F 1 1 A 3 5 ; (13) where ( T ) i;j for all i; j 2 1 ; : : : ; n is the ( i; j ) -th element of T = ^ BH . The separation is perfect when the error inde x I ( ) is close to 0 in a linear scale ( 1 in a log arithmic scale). All the displayed results ha v e been a v eraged o v er 30 Monte-Carlo trials. W e plot, in the Figure 6, the e v olution of the error inde x v ersus the SNR in order to emphasize the influence of this in the quality of the estimation. All algorithms are initialized using the same initialization suggested in [24]. Int J Elec & Comp Eng, V ol. 9, No. 3, June 2019 : 1732 1741 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1739 First, we can deduce from the Figure 4 t hat the di v ersity in the Delay-Doppler re gions is obtained on the zero-diagonal part which supports the use of zero diagonalization algorithms. Then, our analysis are e xamined on the Figure 6 according to noiseless and noisy conte xts . F or the noiseless conte xt (when SNR =100 dB), the JZD CG DD and JZD LM DD reach approximately -64 dB and -60 comparing with JZD Chab ri e l DD method which reaches ' -20 dB. From this comparison, we ha v e check ed the v alidity of the good beha vior of JZD CG DD and JZD LM DD compared to the JZD Chab riel DD approach. Moreo v er , we observ e that the JZD LM DD based on the computation of e xact Hessian matrices is more ef ficient than the JZD CG DD approach. Ev en in a dif ficult (noisy) conte xt (for e xample SNR =15 dB), we note that the best results are generally obtained using the JZD LM DD (-36 dB) then JZD CG DD (-33 dB) especially the JZD LM DD algorithm based on the computation of e xact Hessian matrices. It may be concluded that the approaches e xploiting the Delay-Doppler di v ersity of S-AIS signals seem rather promising. Due to its rob ustness to the noise, it seems to be able to solv e the problem of BSS (i.e the collision problem) in a marine surv eillance conte xt. 400 500 600 700 800 900 1000 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Time in samples Reduced Frequency 400 500 600 700 800 900 1000 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Time in samples Reduced frequency Figure 5. Delay-Doppler points selected with the detector C Ins . left : SNR =100 dB. right : SNR =10 dB. 15 20 30 40 60 80 100 SNR [dB] -70 -60 -50 -40 -30 -20 -10 Error Index I(T) [dB] JZD LM DD JZD CG DD ZDC Chabriel DD Figure 6. Comparison of the dif ferent methods: e v olution of the error inde x I ( T ) in dB v ersus SNR . 6. CONCLUSION In this paper , we ha v e sho wn that the blind source separation based on SGMAF can be performed. W e ha v e considered comple x-v alued S-AIS data for marine surv eillance which can be recei v ed at the same time- slot in where the collision of these data is caused. In addition, it is presented that the collision problem can be reshaped into BSS problem. Moreo v er , it is sho wn that proposed BSS methods are established thanks to an automatic single cross-term selection procedure combined with tw o NU JZD algorithms denoted Conjug ate Gradient and Le v enber g-Marquardt which are based on the minimization of a least-mean-square criterion. Finally , we deduced that the JZD LM DD and JZD CG DD of fer the best perform ances e v en in noisy conte xts. As perspecti v e, a question needing analysis is t o study more realistic and comple x cases in which the number of S-AIS messages recei v ed at the antenna embedded in the satellite w ould be much higher and mixing models could also be considered. Blind separ ation of comple x-valued satellite-AIS data ... (Omar Cherr ak) Evaluation Warning : The document was created with Spire.PDF for Python.
1740 r ISSN: 2088-8708 REFERENCES [1] B. T etreault, “Use of the automatic identification system (ais ) for maritime domain a w areness (mda), in OCEANS, 2005. Pr oceedings of MTS/IEEE , v ol. 2, Sep 2005, pp. 1590–1594. [2] International Associ ation of Maritime Aids to Na vig ation and Lighthouse Authorities (IALA), IALA guidelines on the univer sal automatic identification system (AIS) . IALA, 2002. [3] G. Høye, B. Narheim, T . Eriksen, and B. J. Meland, Euclid JP 9.16, Space-Based AIS Reception f or Ship Identification . FFI, 2004. [4] T . W ahl and G. Høye, “Ne w possible roles of small satellites in maritime surv eillance, Acta Astr onautica , v ol. 56, no. 1–2, pp. 273–277, 2005. [5] T . Eriksen, G. Høye, B. Narheim, and B. Meland, “Maritime traf fic monitoring using a space-based ais recei v er , Acta Astr onautica , v ol. 58, no. 10, pp. 537–549, May 2006. [6] P . Burzigotti, A. Ginesi, and G. Cola v olpe, Adv anced recei v er design for satellite-based ais signal de- tection, in Advanced satellite multimedia systems confer ence and the 11th signal pr ocessing for space communications workshop , Sep 2010, pp. 1–8. [7] International T elecommunications Union Recommendation ITU-R M.2169, Impr o ved Satellite Det ection of AIS , Dec 2009. [8] J. T unale y , An analysis of ais signal collisions, T ech. Rep., Aug 2005. [9] M. Y ang, Y . Zou, and L. F ang, “Collision and detection performance with three o v erlap signal collisions in space-based ais reception, in T rust, Security and Privacy in Computing and Communications (T rustCom), 2012 IEEE 11th International Confer ence on , Jun 2012, pp. 1641–1648. [10] R. Pre v ost, M. Coulon, D. Bonacci, J. LeMait re, J.-P . Millerioux, and J.-Y . T ourneret, “Inte rference miti- g ation and error correction method for ais signals recei v ed by satellite, in Signal Pr ocessing Confer ence (EUSIPCO), 2012 Pr oceedings of the 20th Eur opean , 2012, pp. 46–50. [11] M. Zhou, A.-J. v an der V een, and R. V an Le uk e n , “Multi-user leo-satellite recei v er for rob ust space de- tection of ais messages, in Acoustics, Speec h and Signal Pr ocessing (ICASSP), 2012 IEEE International Confer ence on , Mar 2012, pp. 2529–2532. [12] A.-J. v an der V een and A. P aulraj, An analytical constant modulus algorithm, Signal Pr ocessing , IEEE T r ansactions on , v ol. 44, no. 5, pp. 1136–1155, May 1996. [13] L. Chang, “Study of ais communication protocol in vts, in Signal Pr ocessing Systems (ICSPS), 2010 2nd International Confer ence on , v ol. 1, Jul 2010, pp. 168–171. [14] A. Linz and A. Hendrickson, “Ef ficient implementation of an i-q gmsk modulator , IEEE T r ans. Cir cuits Syst. II, Analo g Digit. Signal Pr ocess. , v ol. 43, no. 1, Jan 1996. [15] R. Pre v ost, M. Coulon, D. Bonacci, J. LeMai tre, J.-P . Millerioux, and J.-Y . T ourneret, “Joint phase- reco v ery and demodulation-decoding of ais signals recei v ed by satellite, in Acoustics, Speec h and Signal Pr ocessing (ICASSP), 2013 IEEE International Confer ence on , May 2013, pp. 4913–4917. [16] O. Cherrak, H. Ghennioui, N. Thirion-Mor eau, E.-H. Abarkan, and E. Moreau, “Non-unitary joint zero- block diagonalization of matrices using a conjug ate gradient algorithm, in Pr oc. EUSIPCO , Sep 2015. [17] O. Cherrak, H. Ghennioui, N. Thirion-Moreau, and E.-H. Abarkan, “Nouv el algorithme de z ´ ero-bloc diag- onalisation conjointe par approche de le v enber g-marquardt, in 25 ` eme Colloque GRETSI , L yon, France, Sep 2015. [18] A. Belouchrani and M. Amin, “Blind source separation based on time-frequenc y signal representations, Signal Pr ocessing , IEEE T r ansactions on , v ol. 46, no. 11, pp. 2888–2897, No v 1998. [19] M. G. Amin, “Blind time-frequenc y analysis for source discrimination in multisensor array processing, V illano v a Uni v ersity , Department of Electrical and Computer Engineering, T echnical Report, Oct 1998. [20] F . Hla w atsch and F . Auger , Eds., T emps-fr ´ equence et tr aitement statistique . Herm ` es Sciences Publica- tions, 2005. [21] W . K ozek, “On the transfe r function calculus for underspread ltv channels , IEEE T r ans. Signal Pr oc. , v ol. 45, no. 1, Jan 1997. [22] ——, “Matched we yl-heisenber g e xpansions of nonstationary en vironments, Ph.D. dissertation, V ienna Uni v ersity of T echnology , 1997. [23] E. F adaili, N. Moreau, and E. Moreau, “Nonorthogonal joint diagonalization/zero diagonalization for source separation based on tim e-frequenc y distrib utions, IEEE T r ans. Signal Pr oc. , v ol. 55, no. 5, pp. 1673–1687, May 2007. [24] G. Chabriel, J. Barrere, N. Thirion-Moreau, and E. Moreau, Algebraic joint zero-diagonalization and Int J Elec & Comp Eng, V ol. 9, No. 3, June 2019 : 1732 1741 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 1741 blind sources separation, IEEE T r ans. Signal Pr oc. , v ol. 56, no. 3, pp. 980–989, Mar 2008. [25] E. Moreau, A generalization of joint-diagonalization criteria for source separation, IEEE T r ans. Signal Pr oc. , v ol. 49, no. 3, pp. 530–541, Mar 2001. BIOGRAPHIES OF A UTHORS Omar Cherrak w as born in Fez Morocco. He recei v ed the Bachelor de gree in 2009 in the field of Electronics T elecommunications and Computer Sciences and the Master de gree of Microelec- tronic, T elecommunications and Computer Industry Systems in 2011, both from the Uni v ersit ´ e Sidi Mohamed Ben Abdella h (USMB A), F acult ´ e des Sci ences et T echniques (FST), Fez Morocco. He obtained his Ph.D. de gree on March 2016 in the are a of ”Signal, T elecommunications, Image and Radar” from Uni v ersit ´ e de T oulon and this thesis w as carried out in cotutelle with USMB A. His main research interests are blind source separation, telecommunications, joint matrix decomposi- tions, maritime surv eillance s ystem, time-frequenc y representation, smart grid and DoA estimation. Hicham Ghennioui obtained the Ma ˆ ıtrise de gree in T elecommunications from the F aculty of Sci- ences and T echnologies (FST), Fez, Morocco, in 2002. He got the D.E.S.A. de gree in Computer and T elecommunications from the F aculty of Sciences, Rabat, Morocco, in 2004 and the Ph.D de gree in engineering sciences in 2008, from Mohamed V Agdal Uni v ersi ty , Morocco, and the Uni v ersity of T oulon, France, respecti v ely . From May 2008 to December 2009, he w as a Research & De v el- opment Engineer at Amesys Bull, and form January 2010 to May 2011, he w as a Signal/Image Researcher at Moroccan foundation for Adv anced Science, Inno v ation and Research (MASCIR), Rabat, Morocco. Since 2011, he is an Assistant Professor at the Electrical Engineering Department of the F aculty of Sciences and T echniques, Sidi Mohamed Ben Abdellah Uni v ersity , Fez, Morocco. His main research interests are signal/image processing including blind sources separation, decon- v olution, deblurring, time-frequenc y representations and cogniti v e radio. Nad ` ege Thirion-Mor eau w as born in Montb ´ eliard France. S he recei v ed the DEA de gree in 1992 and the Ph.D. de gree in 1995, both in the field of signal processing and from the Ecole Nationale Sup ´ erieure de Ph ysique (ENSPG) Institut National Polytechnique de Grenoble (IN PG), France. From 1996 to 1998, she w as an Assistant Professor at the Ecole Sup ´ erieure des Proc ´ ed ´ es Elec- troniques et Optiques (ESPEO), Orl ´ eans, France. Since 1998, she has been with the Institut des Sciences de l’Ing ´ enieur de T oulon et du V ar (ISITV), La V alette, France, as an Assistant Professor , in the Department of T elecommunications. Her main research interests are in deterministic and statistical signal processing including array processing, blind sources separation/equalization, high- order statistics, nonstationary signals, time-frequenc y representations, and decision/ classificat ion. El Hossain Abarkan w as born in 1953, Nador , Morocco. He graduated the bachelor de gree in ph ysics from Mohammed V Uni v ersity , Rabat, Morocco, in 1978. He obtained the DEA, the Doc- torate and the Ph.D de grees from Uni v ersity of Languedoc, Montpellier , France, in 1979, 1981 and 1987 respecti v ely . He got the Ph.D de gree from the Uni v ersity of Sidi Mohamed Ben Abdellah, Fez, Morocco, in 1992. Cur rently , he is a Professor in Sidi Mohamed Ben Abdellah Uni v ersity . From 1996 to 2014, he w as the founder and the director of the Laboratory of Si gnals, Systems and Com- ponents, Sidi Mohamed Ben Abdel lah Uni v ersity . His main r esearch interests are in electronics, Modeling, characterization and CAD in inte grated circuits. Blind separ ation of comple x-valued satellite-AIS data ... (Omar Cherr ak) Evaluation Warning : The document was created with Spire.PDF for Python.