Laurent Hardouin, Carlos Andrey Maia, Bertrand Cottenceau, and Mehdi Lhommeau

538 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 2, FEBRUARY 2010 [2] A. Ilchmann and E. P. Ryan, High-gain control without identification: A survey, GAMM Mitt., vol. 31, no. 1, pp. 115 125, 2008. [3] A. Ilchmann, E. P. Ryan, and C. J. Sangwin, Tracking with prescribed transient behavior, ESAIM: Control, Optim. Calculus Var., vol. 7, pp. 471 493, 2002. [4] A. Ilchmann and H. Schuster, PI-funnel control for two mass systems, IEEE Trans. Autom. Control, vol. 54, no. 4, pp. 918 923, Apr. 2009. [5] I. M. Y. Mareels, A simple self-tuning controller for stably invertible systems, Syst. Control Lett., vol. 4, no. 1, pp. 5 16, 1984. [6] B. Mårtensson, The order of any stabilizing regulator is sufficient a priori information for adaptive stabilization, Syst. Control Lett., vol. 6, no. 2, pp. 87 91, 1985. [7] D. E. Miller and E. J. Davison, An adaptive controller which provides an arbitrarily good transient and steady-state response, IEEE Trans. Autom. Control, vol. 36, no. 1, pp. 68 81, Jan. 1991. [8] A. S. Morse, Recent problems in parameter adaptive control, in Outils et Mod eles Mathématiques Pour L Automatique, L Analyse de Syst emes et le Traitment du Signal, I. D. Landau, Ed. Paris, France: Éditions du Centre National de la Recherche Scientifique (CNRS), 1983, vol. 3, pp. 733 740. [9] C. Sparrow, The Lorenz equations: Bifurcations, chaos, and strange attractors, in Number 41 in Applied Mathematical Sciences. New York: Springer- Verlag, 1982. [10] W. Walter, Ordinary Differential Equations. New York: Springer- Verlag, 1998. [11] J. C. Willems and C. I. Byrnes, A. Bensoussan and J. L. Lions, Eds., Global adaptive stabilization in the absence of information on the sign of the high frequency gain, in Proc. 6th INRIA Conf., Nice, France, 1984, pp. 49 57. Observer Design for Linear Systems Laurent Hardouin, Carlos Andrey Maia, Bertrand Cottenceau, and Mehdi Lhommeau Abstract This technical note deals with the state estimation for maxplus linear systems. This estimation is carried out following the ideas of the observer method for classical linear systems. The system matrices are assumed to be known, and the observation of the input and of the output is used to compute the estimated state. The observer design is based on the residuation theory which is suitable to deal with linear mapping inversion in idempotent semiring. Index Terms Dioid, discrete event dynamics systems, idempotent semirings, max-plus algebra, observer, residuation theory, state estimation, timed event graphs (TEGs). I. INTRODUCTION Many discrete event dynamic systems, such as transportation networks [12], [21], communication networks, manufacturing assembly lines [3], are subject to synchronization phenomena. Timed event Manuscript received April 09, 2008; revised January 15, 2009. First published January 12, 2010; current version published February 10, 2010. This work was supported by CAPES (Brazil) and COFECUB (France). Recommended by Associate Editor S. Haar. L. Hardouin, B. Cottenceau, and M. Lhommeau are with the Laboratoire d Ingénierie des Systèmes Automatisés, Université d Angers, Angers 49100, France (e-mail: laurent.hardouin@univ-angers.fr; bertrand.cottenceau@univ-angers.fr; mehdi.lhommeau@univ-angers.fr) C.A. Maia is with the Departamento de Engenharia Elétrica, Universidade Federal de Minas Gerais,, Belo Horizonte 31270-010, MG, Brazil (e-mail: maia@cpdee.ufmg.br). Color versions of one or more of the figures in this technical note are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TAC.2009.2037477 graphs (TEGs) are a subclass of timed Petri nets and are suitable tools to model these systems. A timed event graph is a timed Petri net of which all places have exactly one upstream transition and one downstream transition. Its description can be transformed into a (max; +) or a (min; +) linear model and vice versa [1], [5]. This property has advantaged the emergence of a specific control theory for these systems, and several control strategies have been proposed, e.g., optimal open loop control [4], [16], [19], [20], and optimal feedback control in order to solve the model matching problem [6], [14], [18], [19] and also [22]. This technical note focuses on observer design for (max; +) linear systems. The observer aims at estimating the state for a given plant by using input and output measurements. The state trajectories correspond to the transition firings of the corresponding timed event graph, their estimation is worthy of interest because it provides insight into internal properties of the system. For example these state estimations are sufficient to reconstruct the marking of the graph, as it is done in [10] for Petri nets without temporization. The state estimation has many potential applications, such as fault detection, diagnosis, and state feedback control. The (max; +) algebra is a particular idempotent semiring, therefore Section II reviews some algebraic tools concerning these algebraic structures. Some results about the residuation theory and its applications over semiring are also given. Section III recalls the description of timed event graphs in a semiring of formal series. Section IV presents and develops the proposed observer. It is designed by analogy with the classical Luenberger [17] observer for linear systems. It is done under the assumption that the system behavior is (max; +)-linear. This assumption means the model represents the fastest system behavior, in other words it implies that the system is unable to be accelerated, and consequently the disturbances can only reduce the system performances i.e., they can only delay the events occurrence. They can be seen as machine breakdown in a manufacturing system, or delay due to an unexpected crowd of people in a transport network. In the opposite, the disturbances which increase system performances, i.e., which anticipate the events occurrence, could give an upper estimation of the state, in this sense the results obtained are not equivalent to the observer for the classical linear systems. Consequently, it is assumed that the model and the initial state correspond to the fastest behavior (e.g., ideal behavior of the manufacturing system without extra delays or ideal behavior of the transport network without traffic holdup and with the maximal speed) and that disturbances only delay the occurrence of events. Under these assumptions a sufficient condition allowing to ensure equality between the state and the estimated state is given in proposition 4 in spite of possible disturbances, and proposition 3 yields some weaker sufficient conditions allowing to ensure equality between the asymptotic slopes of the state and the one of the estimated state, that means the error between both is always bounded. We invite the reader to consult the following link http://www.istia.univ-angers.fr/ ~hardouin/observer.html to discover a dynamic illustration of the observer behavior. II. ALGEBRAIC SETTING An idempotent semiring S is an algebraic structure with two internal operations denoted by 8 and. The operation 8 is associative, commutative and idempotent, that is, a 8 a = a. The operation is associative (but not necessarily commutative) and distributive on the left and on the right with respect to 8. The neutral elements of 8 and are represented by " and e respectively, and " is an absorbing element for the law (8a 2 S; " a = a " = "). As in classical algebra, the operator will be often omitted in the equations, moreover, a i = a i01 a and 0 a = e. In this algebraic structure, a partial order 0018-9286/$26.00 2010 IEEE

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 2, FEBRUARY 2010 539 relation is defined by a b, a = a 8 b, b = a ^ b (where a ^ b is the greatest lower bound of a and b), therefore an idempotent semiring S is a partially ordered set (see [1], [12] for an exhaustive introduction). An idempotent semiring S is said to be complete if it is closed for infinite 8-sums and if distributes over infinite 8-sums. In particular > = x2s x is the greatest element of S (> is called the top element of S). Example 1 ( max ): Set max = [f01; +1g endowed with the max operator as sum and the classical sum + as product is a complete idempotent semiring, usually denoted max, of which " = 01 and e =0. Theorem 1 (see [1], th. 4.75): The implicit inequality x ax 8 b as well as the equation x = ax 8 b defined over S, admit x = a 3 b as the least solution, where a 3 = i2 a i (Kleene star operator). Properties 1: The Kleene star operator satisfies the following well known properties (see [9] for proofs, and [13] for more general results): a 3 =(a 3 ) 3 ; a 3 a 3 = a 3 (1) (a 8 b) 3 = a 3 (ba 3 ) 3 =(a 3 b) 3 a 3 ; b(ab) 3 =(ba) 3 b: (2) Thereafter, the operator a + = i2 a i = aa 3 = a 3 a is also considered, it satisfies the following properties: a + =(a + ) + ; a 3 = e 8 a + (3) (a 3 ) + =(a + ) 3 = a 3 ; a + a 3 : (4) Definition 1 (Residual and Residuated Mapping): An order preserving mapping f : D 0! E, where D and E are partially ordered sets, is a residuated mapping if for all y 2Ethere exists a greatest solution for the inequality f (x) y (hereafter denoted f ] (y)). Obviously, if equality f (x) =y is solvable, f ] (y) yields the greatest solution. The mapping f ] is called the residual of f and f ] (y) is the optimal solution of the inequality. Theorem 2 (see [1], [2]): Let f : (D; ) 0!(C; ) be an order preserving mapping. The following statements are equivalent (i) f is residuated. (ii) there exists an unique order preserving mapping f ] : C 0! D such that f f ] C and f ] f D. Example 2: Mappings 3 a : x 7! a x and 9 a : x 7! x a defined over an idempotent semiring S are both residuated ([1, p. 181]). Their residuals are order preserving mappings denoted respectively by 3 a(x) ] =a nx and 9 a(x) ] =x =a. This means that a nb (respectively, b =a) is the greatest solution of the inequality a x b (respectively, x a b). Definition 2 (Restricted Mapping): Let f : D!Cbe a mapping and BD. We will denote by f jb : B!Cthe mapping defined by f jb = f jb where jb : B!D;x7! x is the canonical injection. Identically, let ECbe a set such that f E. Mapping Ej f : D! E is defined by f = je Ej f, where je : E!C;x7! x. Definition 3 (Closure Mapping): A closure mapping is an order preserving mapping f : D!Ddefined on an ordered set D such that f D and f f = f. Proposition 1 (see [6]): Let f : D! D be a closure mapping. Then, fjf is a residuated mapping whose residual is the canonical injection j f. Example 3: Mapping K : S 0! S;x 7! x 3 is a closure mapping (indeed a a 3 and a 3 =(a 3 ) 3 see (1)). Then ( KjK) is residuated and its residual is ( KjK) ] = j K. In other words, x = a 3 is the greatest solution of inequality x 3 a if a 2 K, that is x a 3, x 3 a 3. Example 4: Mapping P : S 0! S;x 7! x + is a closure mapping (indeed a a + and a + =(a + ) + see (3)). Then ( P jp ) is residuated and its residual is ( P jp ) ] = j P. In other words, x = a + is the greatest solution of inequality x + a if a 2 P, that is x a +, x + a +. Remark 1: According to (4), (a 3 ) + = a 3, therefore K P. Properties 2: Some useful results involving these residuals are presented below (see [1] for proofs and more complete results) a na =(a na) 3 a =a =(a =a) 3 (5) a(a n(ax)) = ax ((xa) =a)a = xa (6) b na nx =(ab) nx x =a =b = x =(ba) (7) a 3 n(a 3 x)=a 3 x (a 3 x) =a 3 = a 3 x (8) (a nx) ^ (a ny) =a n(x ^ y) (x =a) ^ (y =a) =(x ^ y) =a: (9) The set of n2n matrices with entries in S is an idempotent semiring. The sum, the product and the residuation of matrices are defined after the sum, the product and the residuation of scalars in S, i.e., (A B) ik = j=1...n (a ij b jk ) (10) (A 8 B) ij = a ij 8 b ij (11) (A nb) ij = = k=1...n k=1...n (a ki nb kj ); (B =A) ij (b ik =a jk ): (12) The identity matrix of S n2n is the matrix with entries equal to e on the diagonal and to " elsewhere. This identity matrix will also be denoted e, and the matrix with all its entries equal to " will also be denoted ". Definition 4 (Reducible and Irreducible Matrices): Let A be a n2n matrix with entries in a semiring S. Matrix A is said reducible, if and only if for some permutation matrix P, the matrix P T AP is block upper triangular. If matrix A is not reducible, it is said to be irreducible. III. TEG DESCRIPTION IN IDEMPOTENT SEMIRING Timed event graphs constitute a subclass of timed Petri nets i.e., those whose places have one and only one upstream and downstream transition. A timed event graph (TEG) description can be transformed into a (max; +) or a (min; +) linear model and vice versa. To obtain an algebraic model in max, a dater function is associated to each transition. For transition labelled x i, x i (k) represents the date of the kth firing (see [1], [12]). A trajectory of a TEG transition is then a firing date sequence of this transition. This collection of dates can be represented by a formal series x() = k2 x i (k) k where x i(k) 2 max and is a backward shift operator 1 in the event domain (formally x(k) =x(k 0 1)). The set of formal series in is denoted by max [[]] and constitutes a complete idempotent semiring. For instance, considering the TEG in Fig. 1, daters x 1, x 2 and x 3 are related as follows over max: x 3 (k) =1x 1 (k)82x 2 (k)82x 3 (k01). Their respective -transforms, expressed over max[[]], are then related as x 3 () =1x 1 () 8 2x 2 () 8 2x 3 (): In this technical note TEGs are modelled in this setting, by the following model: x = Ax 8 Bu 8 Rw y = Cx (13) where u 2 ( max [[]]) p, y 2 ( max [[]]) m and x 2 ( max [[]]) n are respectively the controllable input, output and state vector, i.e., each of their entries is a trajectory which represents the collection of firing 1 Operator plays a role similar to operator z in the Z transform for the conventional linear systems theory.

540 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 2, FEBRUARY 2010 Fig. 1. Timed event graph, u controllable and w uncontrollable inputs. dates of the corresponding transition. Matrices A 2 ( max[[]]) n2n, B 2 ( max [[]]) n2p, C 2 ( max [[]]) m2n represent the links between each transition, and then describe the structure of the graph. Vector w 2 ( max [[]]) l represents uncontrollable inputs (i.e., disturbances 2 ). Each entry of w corresponds to a transition which disables the firing of internal transition of the graph, and then decreases the performance of the system. This vector is bound to the graph through matrix R 2 ( max [[]]) n2l. Afterwards, each input transition u i (respectively w i) is assumed to be connected to one and only one internal transition x j, this means that each column of matrix B (respectively, R) has one entry equal to e and the others equal to " and at most one entry equal to e on each row. Furthermore, each output transition y i is assumed to be linked to one and only one internal transition x j, i.e., each row of matrix C has one entry equal to e and the others equal to " and at most one entry equal to e on each column. These requirements are satisfied without loss of generality, since it is sufficient to add extra input and output transition. Note that if R is equal to the identity matrix, w can represent initial state of the system x(0) by considering w = x(0) 0 8111(see [1, p. 245], for a discussion about compatible initial conditions). By considering theorem 1, this system can be rewritten as x = A 3 Bu 8 A 3 Rw (14) y = CA 3 Bu 8 CA 3 Rw (15) where (CA 3 B) 2 ( max [[]]) m2p (respectively (CA 3 R) 2 ( max [[]]) m2l ) is the input/output (respectively, disturbance/output) transfer matrix. Matrix (CA 3 B) represents the earliest behavior of the system, therefore it must be underlined that the uncontrollable inputs vector w (initial conditions or disturbances) is only able to delay the transition firings, i.e., according to the order relation of the semiring, to increase the vectors x and y. If the TEG is strongly connected, i.e., there exists at least one path between transitions x i;xj 8i; j, then matrix A is irreducible. If A is reducible, according to definition 4, there exists a permutation matrix such that A = A 11 A 12... A 1k " A 22 111 A 2k........ " " 111 A kk (16) where k is the number of strongly connected components of the TEG, and each matrix A ii is an irreducible matrix associated to the component i. Matrices A ij (with i 6= j) represent the links between these strongly connected components. 2 In manufacturing setting, w may represent machine breakdowns or failures in component supply. Fig. 2. Periodic series s =(e 8 1 8 3 ) 8 (5 8 6 )(3 ). Consequently, for the TEG depicted Fig. 1, the following matrices 4 1 6 " " are obtained: A = 2 2 " ; B = e ", " " 3 " e e " " C = e " " ; R = " e ", and leads to the following " " e " " e A 3 matrix: A 3 = (4) 3 1(4) 3 6(4) 3 2(4) 3 e 8 2 8 4 2 8 6 3 8 9 4 (4) 3 6 2 (4) 3 : " " (3) 3 According to assumptions about matrices C, B, and R, the matrices (CA 3 B) and (CA 3 R) are composed of some entries of matrix A 3. Each entry is a periodic series [1] in the max [[]] semiring. A periodic series s is usually represented by s = p8qr 3, where p (respectively q) is a polynomial depicting the transient (respectively, the periodic) behavior, and r = is a monomial depicting the periodicity allowing to define the asymptotic slope of the series as 1(s) = = (see Fig. 2). Sum, product, and residuation of periodic series are well defined (see [9]), and algorithms and software toolboxes are available in order to handle periodic series and compute transfer relations (see [7]). Below, only the rules between monomials and properties concerning asymptotic slope are recalled t 8 t =max(t ;t ) ; t 8 t =t ; t t =(t +t ) ; (t ) =(t )=(t ) n(t )=(t 0 t ) ; (s 8 s )=min( (s); (s )) (17) (s s ) = min( (s); (s )) (18) (s ^ s ) = max( (s); (s )) (19) if (s) (s ) then (s ns)= (s); else s ns= (20) ": Let us recall that if matrix A is irreducible then all the entries of matrix A 3 have the same asymptotic slope, which will be denoted 1(A). If A is a reducible matrix assumed to be in its block upper triangular representation, then matrix A 3 is block upper triangular and matrices (A 3 ) ii are such that (A 3 ) ii = A 3 ii for each i 2 [1;k]. Therefore, since A ii is irreducible, all the entries of matrix (A 3 ) ii have the same asymptotic slope 1((A 3 ) ii ). Furthermore, entries of each matrix (A 3 ) ij with i <jare such that their asymptotic slope is lower than or equal to min(1((a 3 ) ii );1((A 3 ) jj )).

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 2, FEBRUARY 2010 541 and by recalling that (LCA 3 ) 3 LCA 3 =(LCA 3 ) +, this equation may be written as follows: ^x = A 3 (LCA 3 ) 3 Bu 8 A 3 (LCA 3 ) + Bu 8 A 3 (LCA 3 ) + Rw: Equation (4) yields (LCA 3 ) 3 (LCA 3 ) +, then the observer model may be written as follows: ^x = A 3 (LCA 3 ) 3 Bu 8 A 3 (LCA 3 ) + Rw =(A 8 LC) 3 Bu 8 (A 8 LC) 3 3 LCA Rw: (24) Fig. 3. Observer structure. As said previously the objective considered is to compute the greatest observation matrix L such that the estimated state vector ^x be as close as possible to state x, under the constraint ^x x, formally it can be written IV. MAX-PLUS OBSERVER Fig. 3 depicts the observer structure directly inspired from the classical linear system theory (see [17]). The observer matrix L aims at providing information from the system output into the simulator, in order to take the disturbances w acting on the system into account. The simulator is described by the model 3 (matrices A, B, C) which is assumed to represent the fastest behavior of the real system in a guaranteed way 4, furthermore the simulator is initialized by the canonical initial conditions (i.e., ^x i(k) = "; 8k 0). These assumptions induce that y ^y since disturbances and initial conditions, depicted by w, are only able to increase the system output. By considering the configuration of Fig. 3 and these assumptions, the computation of the optimal observer matrix Lx will be proposed in order to achieve the constraint ^x x. Optimality means that the matrix is obtained thanks to the residuation theory and then it is the greatest one (see definition 1), hence the estimated state ^x is the greatest which achieves the objective. Obviously this optimality is only ensured under the assumptions considered (i.e., ^y y). As in the development proposed in conventional linear systems theory, matrices A, B, C and R are assumed to be known, then the system transfer is given by (14) and (15). According to Fig. 3 the observer equations are given by ^x = A^x 8 Bu 8 L(^y 8 y) = A^x 8 Bu 8 LC ^x 8 LCx ^y = C ^x: (21) By applying Theorem 1 and by considering (14), (21) becomes (A 8 LC) 3 Bu 8 (A 8 LC) 3 3 LCA Rw 3 A Bu 8 3 A Rw 8(u; w) or equivalently (A 8 LC) 3 B A 3 B (25) (A 8 LC) 3 LCA 3 R A 3 R: (26) Lemma 1: The greatest matrix L such that (A 8 LC) 3 B = A 3 B is given by L 1 =(A 3 B) =(CA 3 B): (27) n2m max Proof: First let us note that L = " 2 is a solution, indeed (A 8 "C) 3 B = A 3 B. Consequently, the greatest solution of the inequality (A 8 LC) 3 B A 3 B will satisfy the equality. Furthermore, according to (2), (A 8 LC) 3 B =(A 3 LC) 3 A 3 B. So the objective is given by 3 (A LC) 3 3 A B 3 A B, (A 3 LC) 3 (A 3 B) =(A 3 B) (see example 2), 3 (A LC) 3 3 ((A B) 3 =(A B)) 3 (see eq: (5)), (A 3 LC) (A 3 B) =(A 3 B) (see example 3), L 3 3 A n(a B) 3 =(A B) =C (see example 2), L 3 3 A n(a B) 3 =(CA B) (see eq: (7)), L 3 (A B) 3 =(CA B)=L 1 (see eq: (8)): ^x =(A8LC) 3 Bu 8 (A 8 LC) 3 LCA 3 Bu 8(A 8 LC) 3 3 LCA Rw: (22) By applying (2) the following equality is obtained: by replacing in (22): (A 8 LC) 3 = A 3 (LCA 3 ) 3 (23) ^x = A 3 (LCA 3 ) 3 Bu 8 A 3 (LCA 3 ) 3 LCA 3 Bu 8A 3 (LCA 3 ) 3 LCA 3 Rw 3 Disturbances are uncontrollable and a priori unknown, then the simulator does not take them into account. 4 Unlike in the conventional linear system theory, this assumption means that the fastest behavior of the system is assumed to be known and that the disturbances can only delay its behavior. Lemma 2: The greatest matrix L that satisfies (A 8 LC) 3 LCA 3 R A 3 R is given by Proof: See L 2 =(A 3 R) =(CA 3 R): (28) (A 8 LC) 3 3 LCA R 3 A R, A 3 (LCA 3 ) 3 LCA 3 R A 3 R (see eq: (23));, 3 (LCA ) 3 3 LCA R 3 3 3 A n(a R)=A R (see example 2 and eq: (8); with x = R);, 3 (LCA ) 3 3 3 3 LCA A R =(LCA ) + 3 A R 3 A R (see eq: (1) and a + denition);, 3 (LCA ) + 3 (A R) 3 3 =(A R)=((A R) 3 =(A R)) 3 (see eq: (5))

542 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 2, FEBRUARY 2010 according to remark 1 the right member is in P, then by applying the result presented in example 4, this inequality may be written as follows: 3 LCA 3 3 (A R)=(A R), L 3 3 3 3 3 3 (A R)=(A R)=(CA )=(A R)=(CA A R) (see example 2 and eq: (8)), L (A 3 R)=(CA 3 R)=L 2 (see eq: (1)): Proposition 2: L x = L 1 ^ L 2 is the greatest observer matrix such that ^x = A^x 8 Bu 8 L(^y 8 y) x = Ax 8 Bu 8 Rw 8(u; w): Proof: Lemma 1 implies L L 1 and lemma 2 implies L L 2, then L L 1 ^ L 2 = L x. Corollary 1: The matrix L x ensures the equality between estimated output ^y and measured output y, i.e., C(A 8 L x C) 3 B = CA 3 B (29) C(A 8 L xc) 3 L xca 3 R = CA 3 R: (30) Proof: Let ~ L = e=c be a particular observer matrix. Definition 1 yields ~ LC e then (A 8 ~ LC) 3 = A 3. This equality implies (A 8 ~ LC) 3 B = A 3 B, therefore according to lemma 1 ~ L L 1, since L 1 is the greatest solution. That implies also that L 1 is solution of (29). Equality (A 8 LC) ~ 3 3 = A and inequality LC ~ e yield (A 8 LC) ~ 3 LCA ~ 3 R = A 3 ~LCA 3 R A 3 R then according to lemma 2 L ~ L 2 since L 2 is the greatest solution. That implies also that 3 L ~ and 3 L 2 are such that C(A 8 LC) ~ 3 LCA ~ 3 R C(A 8 L 2 C) L 2 CA R 3 CA R. The assumption about matrix C (see Section III) yields CC T = e and L ~ = e=c = C T, therefore C(A 8 ~LC) 3 LCA ~ 3 R = CA 3 ~LCA 3 R = (C L ~ 8 CA L ~ 8 111)CA 3 R C LCA ~ 3 3 3 R = CC T CA R = CA R. Therefore, since L ~ L 2,we have C(A8 LC) ~ 3 LCA ~ 3 3 3 3 R = C(A8L2C) L 2CA R = CA R and both L ~ and L 2 yield equality (30). To conclude L ~ L 1 ^ L 2 = L x, hence, L x L 1 yields the equality (29) and L x L 2 yields (30). Therefore equality ^y = y is ensured. Remark 2: By considering matrix B = (B R), (12) and (9), matrix L x may be written as: 3 L x =(A B 3 =(CA B). According to the residuation theory (see definition 1), L x yields x =^x if possible. Nevertheless, two questions arise, first is it possible to ensure equality between the asymptotic slope of each state vector entries? Secondly is it possible to ensure equality between these vectors? Below, sufficient conditions allowing to answer positively are given. Proposition 3: Let k be the number of strongly connected components of the TEG considered. If matrix C 2 max[[]] k2n is defined as in Section III and such that each strongly connected component is linked to one and only one output then 1(x i)=1(^x i)8i 2 [1;n]. Proof: First, assuming that matrix A is irreducible (i.e., k = 1), then all entries of matrix A 3 have the same asymptotic slope 1(A 3 ). As said in Section III entries of matrices B, R, and C are equal to " or e, therefore, according to matrices operation definitions [see (10) to (12) and rules (17) to (20)], all the have the entries of matrices 3 A B, 3 A R, 3 CA B, 3 CA R and L x same asymptotic slope which is equal to 3 1(A ). Consequently, by considering (24), 1(((A 8 3 3 L x C) B) ij ) = 1((A B) ij ) and 1(((A 8 3 3 3 L xc) L xca R) ij) = 1((A R) ij), which leads to 1(x i )=1(^x i ) 8i 2 [1;n]. Now the reducible case is considered. To increase the readability, matrices B and R are assumed to be equal to e and the proof is given for a graph with two strongly connected components. The extension for a higher dimension may be obtained in an analogous way. As said in Section III, matrix A 3 is block upper diagonal: A 3 = (A3 ) 11 (A 3 ) 12 " (A 3 ) 22 all the entries of the square matrix 3 (A ) ii have the same asymptotic slope 3 1((A ) ii) and all the entries of matrix 3 (A ) 12 have the same asymptotic slope, 3 3 3 1((A ) 12 ) = min(1((a ) 11 );1((A ) 22 )). Assumption about matrix C 2 max [[]] 22n, i.e., one and only one entry is linked to each strongly connected component, yields the following block upper diagonal matrix: CA 3 = (CA3 ) 11 (CA 3 ) 12 " (CA 3 ) 22 where ((CA 3 ) 11 (CA 3 ) 12 ) is one row of matrix ((A 3 ) 11 (A 3 ) 12 ) and ( " (CA 3 ) 22 ) is one row of matrix ( " (A 3 ) 22 ), hence 1((CA 3 ) ij ) = 1((A 3 ) ij ). Matrix L x is also block upper diagonal 3 L x = A =C = L x11 L x12 " L x22 where ( L x11 " ) T is one column of matrix ((A 3 ) 11 " ) T and ( L x12 L x22 ) T is one column of matrix ((A 3 ) 12 (A 3 ) 22 ) T, hence 1(L xij) = 1((A 3 ) ij). Therefore L xca 3 is block upper diagonal L x CA 3 = Lx11(CA3 ) 11 L x11(ca 3 ) 12 8 L x12(ca 3 ) 22 " L x22 (CA 3 ) 22 3 3 = (L xca ) 11 (L x CA 3 ) 12 (31) " (L xca ) 22 and by considering rules (17) and (20), the sub matrices are such that 1((L 3 3 x CA 3 3 3 ) ij )= 1((A ) ij ). By recalling that (A 8 3 L x C) = A (L x CA ), we obtain 1(((A 8 3 3 L x C) ) ij ) = 1((A ) ij ) and 1(((A 8 3 3 3 L xc) L xca ) ij) = 1((A ) ij), which leads to 1(x i )=1(^x i ) 8i 2 [1;n]. Proposition 4: If matrix 3 A B is in 9CA B, matrix L x is such that ^x = x. Proof: First, let us recall that 3 A B 3 2 9CA B,9z s:t: A B = zca 3 B, 3 ((A B)=(CA 3 B))(CA 3 B) =A 3 B: If 9z s:t: A 3 B = zca 3 B then by recalling that B =(B L xca 3 B =((A 3 B)=(CA 3 B))CA 3 B 3 =((zca B)=(CA 3 B))CA 3 3 B = zca B = A 3 B (see eq: (6) R), this equality can be written (L x CA 3 B L x CA 3 R)=(A 3 B A 3 R): Therefore (A 8 L x C) 3 L x CA 3 R = A 3 (L x CA 3 ) 3 L x CA 3 R = A 3 (L x CA 3 ) + R = A 3 (L x CA 3 R 8(L x CA 3 ) 2 R 8 (L x CA 3 ) 3 R 8 111) (see (23) and a + definition). Since L xca 3 R = A 3 R, the following equality is satified (L x CA 3 ) 2 R = L x CA 3 A 3 R = L x CA 3 R = A 3 R and more generally (L x CA 3 ) i R = A 3 R, therefore L x ensures equality (A8L xc) 3 L xca 3 R = A 3 (L xca 3 ) + R = A 3 R. On the other hand lemma 1 yields the equality (A 8 L x C) 3 B = A 3 B, which concludes the proof. Remark 3: This sufficient condition gives an interesting test to know if the number of sensors is sufficient and if they are well localized to

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 55, NO. 2, FEBRUARY 2010 543 allow an exact estimation. Obviously, this condition is fulfilled if matrix C is equal to the identity. Below, the synthesis of the observer matrices L x for the TEG of Fig. 1 is given L x = (4) 3 6(4) 3 2 (4) 3 6 2 (4) 3 " (3) 3 : Assumptions of proposition 3 being fulfilled, it can easily be checked, by using toolbox Minmaxgd (see [7]), that 1(x i)= 1(^x i) 8i 2 [1;n] and that Cx = C ^x 8(u; w) according to corollary 1. V. CONCLUSION This technical note 5 has proposed a methodology to design an observer for (max; +) linear systems. The observer matrix is obtained thanks to the residuation theory and is optimal in the sense that it is the greatest which achieves the objective. It allows to compute a state estimation lower than or equal to the real state and ensures that the estimated output is equal to the system output. As a perspective, this state estimation may be used in state feedback control strategies as proposed in [6], [19], and an application to fault detection for manufacturing systems may be envisaged. Furthermore, in order to deal with uncertain systems an extension can be envisaged by considering interval analysis as it is done in [11], [15] and more recently in [8]. REFERENCES [1] F. Baccelli, G. Cohen, G. J. Olsder, and J. P. Quadrat, Synchronization and Linearity: An Algebra for Discrete Event Systems. New York: Wiley, 1992. [2] T. S. Blyth and M. F. Janowitz, Residuation Theory. Oxford, U.K.: Pergamon Press, 1972. [3] G. Cohen, D. Dubois, J. P. Quadrat, and M. Viot, A linear system theoretic view of discrete event processes and its use for performance evaluation in manufacturing, IEEE Trans. Autom. Control, vol. AC-30, no. 3, pp. 210 220, Mar. 1985. [4] G. Cohen, S. Gaubert, and J. Quadrat, Max-plus algebra and system theory: Where we are and where to go now, in Proc. Elsevier Annu. Rev. Control, 1999, vol. 23, pp. 207 219. [5] G. Cohen, P. Moller, J. P. Quadrat, and M. Viot, Linear system theory for discrete-event systems, in Proc. 23rd IEEE Conf. Decision Control, Las Vegas, NV, 1984, pp. 539 544. [6] B. Cottenceau, L. Hardouin, J.-L. Boimond, and J.-L. Ferrier, Model reference control for timed event graphs in dioids, Automatica, vol. 37, pp. 1451 1458, 2001. [7] B. Cottenceau, L. Hardouin, M. Lhommeau, and J.-L. Boimond, Data processing tool for calculation in dioid, in Proc. Wodes 00 Workshop Discrete Event Syst., Ghent, Belgique, Aug. 2000 [Online]. Available: http://www.istia.univ-angers.fr/~hardouin/outils.html [8] M. DiLoreto, S. Gaubert, R. Katz, and J.-J. Loiseau, Duality Between Invariant Spaces for Max-Plus Linear Discrete Event Systems Jan. 2009 [Online]. Available: http://fr.arxiv.org/abs/0901.2915 [9] S. Gaubert, Théorie Des Systèmes Linéaires Dans Les Dioïdes, Ph.D. dissertation, Ecole des Mines de Paris, Paris, France, 1992. [10] A. Giua and C. Seatzu, Observability of place/transition nets, IEEE Trans. Autom. Control, vol. 47, no. 9, pp. 1424 1437, Sep. 2002. [11] L. Hardouin, B. Cottenceau, M. Lhommeau, and E. Le Corronc, Interval systems over idempotent semiring, Linear Algebra Appl., vol. 431, no. 5 7, pp. 855 862, Aug. 2009. [12] B. Heidergott, G.-J. Olsder, and J. van der Woude, Max Plus at Work: Modeling and Analysis of Synchronized Systems: A Course on Max- Plus Algebra and Its Applications. Princeton, NJ: Princeton Univ. Press, 2006. [13] D. Krob, Complete systems of B-rational identities, Theoret. Comput. Sci., vol. 89, no. 2, pp. 207 343, 1991. [14] M. Lhommeau, L. Hardouin, and B. Cottenceau, Optimal control for (max; +)-linear systems in the presence of disturbances, Positive Syst.: Theory Appl., POSTA, Springer LNCIS 294, pp. 47 54, 2003. 5 The authors are grateful to V. Reverdy for her valuable linguistic help [15] M. Lhommeau, L. Hardouin, B. Cottenceau, and L. Jaulin, Interval analysis and dioid: Application to robust controller design for timed event graphs, Automatica, vol. 40, pp. 1923 1930, 2004. [16] M. Lhommeau, L. Hardouin, J. L. Ferrier, and I. Ouerghi, Interval analysis in dioid: Application to robust open loop control for timed event graphs, in Proc. 44th IEEE CDC-ECC 05, Sevilla, Spain, 2005, pp. 7744 7749. [17] D. Luenberger, An introduction to observers, IEEE Trans. Autom. Control, vol. AC-16, no. 6, pp. 596 602, Dec. 1971. [18] C. A. Maia, L. Hardouin, R. Santos-Mendes, and B. Cottenceau, Optimal closed-loop control of timed event graphs in dioid, IEEE Trans. Autom. Control, vol. 48, no. 12, pp. 2284 2287, Dec. 2003. [19] C. A. Maia, L. Hardouin, R. Santos-Mendes, and B. Cottenceau, On the model reference control for max-plus linear systems, in Proc. 44th IEEE CDC-ECC 05, Sevilla, Spain, 2005, pp. 7799 7803 [Online]. Available: http://dx.doi.org/10.1109/cdc.2005.1583422 [20] E. Menguy, J.-L. Boimond, L. Hardouin, and J.-L. Ferrier, Just in time control of timed event graphs: update of reference input, presence of uncontrollable input, IEEE Trans. Autom. Control, vol. 45, no. 11, pp. 2155 2159, Nov. 2000. [21] J.-P. Quadrat, N. Farhi, and M. Goursat, Derivation of the fundamental traffic diagram for two circular roads and a crossing using minplus algebra and petri net modeling, in Proc. CDC 05, Sevilla, Spain, 2005. [22] Y. Shang and M. K. Sain, On zero semimodules of systems over semirings with applications to queueing systems, in Proc. Amer. Control Conf., 2005, pp. 225 230. HOSM Observer for a Class of Non-Minimum Phase Causal Nonlinear MIMO Systems Yuri B. Shtessel, Simon Baev, Christopher Edwards, and Sarah Spurgeon Abstract A higher order sliding mode observer is proposed for asymptotic identification of the full state vector and the vector of unknown inputs for MIMO nonlinear causal systems with unstable internal dynamics. The problem is addressed via consistent application of exact higher order sliding mode differentiators in conjunction with the method of stable system center. A numerical example illustrates the performance of the proposed algorithm. Index Terms Causal, higher order sliding mode (HOSM), nonlinear, nonminimum phase (NMP), observer. I. INTRODUCTION Simultaneous observation of the full state and unknown inputs in causal nonminimum-phase (with unstable zero dynamics) nonlinear systems is a challenging, real-life control problem [1], for example state estimation and disturbance reconstruction for flexible launch vehicles. Manuscript received June 24, 2008; revised February 19, 2009 and July 27, 2009. First published January 08, 2010; current version published February 10, 2010. Recommended by Associate Editor A. Ferrara. Y. B. Shtessel is with the Department of Electrical and Computer Engineering, The University of Alabama, Huntsville, AL 35899 USA (e-mail: shtessel@ece.uah.edu). S. Baev is with the School of Computing and Mathematics, Georgia South- Western State University, Americus, GA 31709-4379 USA (e-mail: baevs@eng. uah.edu). C. Edwards is with the Department of Engineering, University of Leicester, Leicester LE1 7RH, U.K. (e-mail: chris.edwards@le.ac.uk). S. Spurgeon is with the Department of Electrical Engineering, University of Kent, Canterbury, Kent CT2 7NZ, U.K. (e-mail: s.k.spurgeon@kent.ac.uk). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TAC.2009.2037478 0018-9286/$26.00 2010 IEEE