robust pole placement, Robotyka
[ Pobierz całość w formacie PDF ]
43rd IEEE Conference on Decision and Control December 14-17, 2004 Atlantis, Paradise Island, Bahamas TuB12.3 Robust pole placement by static output feedback Jérôme Bosche, Olivier Bachelier and Driss Mehdi Abstract —This paper tackles the problem of pole placement by static output feedback. The considered systems are LTI and subject to both polytopic and norm-bounded uncertainties i.e. the uncertain closed-loop state matrix can be written A o = A + BK o C + J EL -regions and of the polytopic and norm-bounded uncertainties in order to state the precise problem. It is solved in the third part from the nominal point of view and in the fourth one from the robust point of view. An illustration is proposed in the fifth section before to conclude. EEMI . Matrices A , B , C , J and L belong K o is the matrix associated with the static output feedback. It is aimed to compute K o so as to place the spectrum of A o in an EEMI - region, denoted by D U , while maximizing the acceptable 2- norm of E ensuring the D U -stabity of A o . Nevertheless, the pole placement by static output feedback presents two major difficulties : - the strict pole placement by static output feedback is usually possible only if Kimura’s condition holds. - the non strict pole placement methods through Lyapunov based approaches leads to non convex problems such as BMI (Bilinear Matrix Inequality) rather than LMI (Linear Matrix Inequality). This paper proposes a heuristic technique to compute K o .It aims at circumventing the difficulties mentionned herebefore. This technique is based on genetic algorithms (not detailed here) and on the resolution of LMI . E is unknown. the Hermitian expression M + M .The Kronecker product is denoted by H (M ) ||M || 2 is the 2-norm of matrix M induced by the Euclidean vector-norm, i.e. the maximal singular value of M .I n ⊗ . is the identity matrix is a null matrix of suitable dimension. Matrix inequalities are considered in the sense of Löewner i.e. “ O ≺ 0 ”(“ 0 ”) means negative (semi-)definite and ”) positive (semi-)definite. HPD stands for Hermitian Positive Definite. Small letters are used for scalar numbers and vectors while capital letters denote matrices or sets. In the Hermitian matrices, notation ( 0 ”(“ 0 • )’ prevents us I. INTRODUCTION Pole placement arouses the interest of many authors working on linear state representations [3], [6], [12]. Indeed, transient performances are strongly influenced by the location of the closed-loop state matrix eigenvalues. It can then be desirable that all those eigenvalues belong to aregion from repeating the symmetric blocks. At last, i denotes the imaginary unit. D of the complex plane. This property, known II. P ROBLEM S TATEMENT In this part are introduced all the preliminary useful concepts. First, the uncertain state matrix is presented. The- reafter, the description of the considered clustering regions is specified. Once these preliminaries established, a nominal D U -stabilization condition is recalled before to formulate our precise purpose. -stability, is an essential concept in this article. The regions considered in this work are referred to as D EEMI - regions [2] ( E xtended E llipsoidal M atrix I nequality) D U . We then talk about D U -stability. This class of regions enables to handle unions of several possible disjoint and non symmetric subregions. State-space LTI models are considered. They are affected by polytopic and additive norm-bounded uncertainties. The objective of this work is to compute some output feedback matrix K o so as to guarantee the A. The uncertain matrix Consider a complex uncertain matrix A ∈ C n×n defined by : A = A + E with E = J EL . (1) D U -stability of the closed-loop state matrix, while maximizing the size of the uncertainty domain. In other words, it is aimed at robustly Matrices A and E are both uncertain. E is an additve uncertainty in which E ∈ C q×r is unknown. E is assumed to belong to B(ρ) , the ball of (q × r) complex matrices D U -stabilizing the closed-loop system by a static output feedback. checking ||E|| 2 ≤ ρ . Define matrix M by : AJ L O A(θ) J(θ) L(θ) O . LMI ( L inear M atrix I nequality) M = = (2) D U -stabilization exist [4] and can involve P arameter- D ependent L yapunov M atrix ( PDLM ) [10], [17]. The paper is organized as follows : after this introduction, Matrix M is assumed to belong to a polytope of matrices : M = ˘ M = M(θ) ∈ C (n+r)×(n+q) | This work was not supported by any organization J. Bosche, O. Bachelier and D. Mehdi are with the Laboratoire d’Automatique et d’Informatique Industrielle - Ecole Supérieure d’Ingénieurs de Poitiers, Bâtiment de Mécanique, 40 Avenue du Recteur Pineau, 86022 POITIERS C EDEX ,F e Jerome.Bosche@esip.univ-poitiers.fr X N ) M(θ)= (θ i M i ); θ ∈ Θ (3) i=1 where Θ is the set of all barycentric coordinates : 0-7803-8682-5/04/$20.00 ©2004 IEEE 869 Authorized licensed use limited to: Biblioteka Glowna i OINT. Downloaded on May 20, 2009 at 05:48 from IEEE Xplore. Restrictions apply. the second section recalls formulations of to polytopes of matrices whereas Notations : We denote by M , the transpose conjugate of M ,by of order n , ” as and are generically denoted by conditions for robust ⎧ ⎨ ⎫ ⎬ only if there exists a set P of k HPD matrices P k ∈ C n×n , ⎡ ⎣ ⎤ ⎦ ∈{ k =1, ..., k , such that : U (A, P k )= II θ . θ N N II d ⊗ A U(P k ) Θ= θ = IR + } N | θ i =1 . (4) II dn ≺ (8) ⎩ ⎭ dn II ⊗ A i=1 d Extreme matrices M i , i =1, ..., N are the vertices of M : with : k A i J i L i . M i = (5) U(P k )= (R k ⊗ P k ) . (9) O k=1 Proof : see [2]. .It is well known that this formulation encompasses the case where M linearly depends on some parameter deflections around a nominal value. If no polytopic dependence is assumed (only one vertex for A can read A = A(θ)+J(θ)EL(θ) is only subject to a classical norm-bounded uncertainty [18]. M )then A Remark 1: In the previous theorem, the matrices are consi- dered as complex. Thus, to specify R k as complex enables to consider nonsymmetric subregions which is an advantage in term of variety of regions. The possible complexity of A is however pratically useless in Automatic. Nevertheless, it turns out that the tools and theorems pre- sented in this paper are valid for complex matrices (and thus a fortiori for real matrices). In practice, if A i , J i , L i can be considered as real, it is impossible to impose the reality of E which is unknown. The non consideration of this realness may induce some conservatism. E -regions Definition 1: [2] Let EEMI R be a set of k Hermitian matrices R k defined by : ⎧ ⎨ R k = R k = R k 00 R k 10 R k 10 ∈ C 2d×2d (6) ⎩ D. The precise purpose In this paper, we consider the closed-loop state matrix of a multivariable system that can be written. R k 11 0,R k 11 ∈ C d×d ∀k ∈{1, ..., k} . The set of points D U definedby: < 2 4 ω . ω k 3 5 ∈{ IR + } k A o = A o (θ, E)=A(θ)+B(θ)K o C(θ)+J(θ)EL(θ) . (10) We will adopt formulation (10) afterwards. Formulation (10) allows, for example, to consider fragility analysis. Indeed, as it is impossible in practice to precisely implant a control law, we can suppose that the feedback matrix equals to K o nom +∆ K o where K o nom D U = : z ∈ C |∃ω = | f D U (ω, z) X k (ω k ˆ II d z II d ˜ R k » II d z II d – = ; = ) ≺ 0 (7) k=1 ∆ K o an uncertain one. The uncertain closed-loop state matrix is then is a nominal term and is called an open EEMI -region of degree d . When k =1 , this description reduces to EMI formu- lation ( E llipsoidal M atrix I nequality) proposed in [17]. A o (θ, E)=A(θ)+B(θ)K o nom C(θ)+B(θ)∆ K o C(θ (11) which is a particular writing of (10). The present purpose is to compute an output feedback matrix K o which assigns the eigenvalues of the uncertain closed-loop state matrix -region is possibly nonconnected and can result from the union of possibly complex and disjoint EEMI EMI - -subregions are convex and can be, for instance, shifted and nonsymmetric half planes, classical or hyperbolic sectors, vertical or horizontal strips, discs or insides of ellipses... Here, R k is allowed to be complex, i.e. EMI EMI D U defined as in (7). K o is derived with calculating a bound on ||E|| 2 , denoted by ρ , as large as possible, such that A o in a EEMI -region -subregions can be nonsymmetric (with respect to the real axis). Some necessary and sufficient conditions (NSC) for D U - stability of is guaranteed, for any radius ρ ≤ ρ and for any θ ∈ Θ D U - stability of real or complex matrices exist [2]. Such a NSC is now recalled. . The obtained bound is known as a robust D U -stability bound. This pole placement technique must implicitly involve PDLM in order to take the polytopic structure of the uncertainty into account. D U -stability of a complex matrix A reformulation of the conditions suggested in [2] is now proposed : III. NOMINAL POLE PLACEMENT In this part, its first explained that the strict pole placement by static output feedback is generically difficult if Kimura’s condition is not satisfied. Then, a Theorem 1: [2] Let D U be an EEMI -region as introduced D U -stabilization condition in the nominal case is presented. 870 in definition 1. A matrix A ∈ C n×n is D U -stable if and Authorized licensed use limited to: Biblioteka Glowna i OINT. Downloaded on May 20, 2009 at 05:48 from IEEE Xplore. Restrictions apply. Therefore matrix and thus B. R k 11 An subregions. Such C. Nominal A. Eigenstructure assignment and Kimura’s condition Eigenstructure assignment by output feedback is a major topic in multivariable systems [7], [15], [22] and many results have been presented since the middle of the seventies [8], [23]. These techniques use the degrees of freedom ( dof ) offered by eigenvectors. Nevertheless, the m×p dof offered by the output feedback matrix are not always sufficient to place the all desired spectrum. Several conditions linking m , n and p allowing a complete pole placement by static output feedback were proposed [20], [24]. It seems that if the problem is restricted to real matrices, the condition is mp > n . However, when it is aimed to apply effective methods of eigenstructure assignment, Kimura’s condition [11], presented as generically sufficient to place n poles, also proves generically necessary for methods [5] and [22] to be applied. This is why we retain Kimura’s condition which is now presented. where k is the number of subregions and U (•) is defined in (8) : U (A s ,P k ) ≺ 0 U (A o ,P k ) ≺ 0 (13) A s and A o are state matrices defined by : A s = A + BK s and A o = A + BK o C. (14) K s and K o are then respectively associated with state and output feedback. Theorem 2 proposes a NSC for LMI system (13) to hold. -region defined as in (7). Let also K s be a matrix associated with a static state feedback such that A s D U ,bean EEMI = A + BK s is D U -stable. The LMI system (13) is satisfied if and only if there ,madeupby k HPD matrices P k ∈ exist a set P IR n×n , k ∈{1, .., k} , a nonsingular matrix G ∈ IR m×m as well as H ∈ IR m×p such that the following LMI holds : Consider a multivariable realization, reachable and obser- vable, with n states, m imputs and p outputs, if the condition » II dn II d ⊗ A s O – » – II dn O II d ⊗ A s II d ⊗ B H U = II d ⊗ B U(P k ) + m + p>n (12) j» O – ( II d ⊗ G) ˆ − II dm ˜ ff H II d ⊗ K s − + is satisfied, then it is possible to place any of the spectrum for the closed-loop output state matrix A o = A + BK o C as long as a slight modification of the required spectrum is tolerable. According to this condition, two classes of systems are considered : 1) systems of class II dm j» – ( II d ⊗ H) ˆ II d ⊗ C O ˜ ff O II dm H ≺ 0 (15) with U(P k ) defined in (9). S + : their models check Kimura’s condition ( i.e. m + p> n ). It is then possible to place any spectrum by static output feedback. Several approaches, not detailed in this paper, can be exploited [5], [19], [22]. 2) systems of class An output feedback matrix is then given by : K o = G −1 H. (16) Proof : According to (16), H = GK o . Once developed, inequality (15) becomes : S − : their models do not check Kimura’s condition ( i.e. m + p ≤ n ). Concerning these systems, finding a strict placement law by static output feedback becomes a much more delicate task and we will propose a conservative method of nonstrict placement in an Ψ k + O II d ⊗ S G ≺ 0 (17) (•) − II d ⊗ (G + G ) with 2 4 3 5 EEMI -region (which can X U (A s ,P k ) (R k 10 ⊗ P k B + R k 11 ⊗ A s P k B) S + ). The next part proposes a technique that may enable to compute an output feedback matrix (if Kimura’s condition is satisfi ed or not ). Ψ k = k=1 X (•) (R k 11 ⊗ B P k B) k=1 D U -stabilization This part proposes a method to compute a robust control law by static output feedback for any nominal multivariable linear systems. This method is based on the resolution of tractable and S = K o C − K s . conditions and relies on an idea originally proposed in [1], [16] and then reformulated in [13]. This idea is here extended to the concept of N v = II dn II d ⊗ S (18) respectively the left and the right orthogonal comple- ments of O and D U -stability and consists in finding a state feedback and an output feedback checking the same D U -stability property through the exis- tence of the same set of Lyapunov matrices . It amounts to solving the following U = O and V = II d ⊗ S − II dm . (19) II dm ∀ k ∈{1, .., k} LMI system (13) 871 Authorized licensed use limited to: Biblioteka Glowna i OINT. Downloaded on May 20, 2009 at 05:48 from IEEE Xplore. Restrictions apply. Theorem 2: Let a LTI system be modelled by the triple of matrices ( A; B; C )et k also be used for systems of class k B. Nominal Define matrices N u = II dn LMI First, notice that A o are respectively the uncertain state matrices of the systems controlled by static state and output feedbacks. A sufficient condition for the existence of a solution to system (22) is now proposed : Ψ k + H (U ( II d ⊗ G)V ) ≺ 0 (20) corresponds to condition (15). Then, a simple algebraic manipulation shows that N u Ψ k N u = U (A s ,P k ) N v Ψ k N v = U (A o ,P k ) . (21) It becomes clear that, in virtue of the elimination lemma [21], Theorem 3: The LMI system (22) is satisfied if there and exist N sets P i , each one made up by k HPD matrices P k,i ∈ C n×n , k ∈{1, .., k} , a nonsingular matrix G ∈ IR m×m as well as F ∈ IR d(2n+m)×dn and H ∈ IR m×p LMI system (13) is equivalent to condition (15). ∀ i ∈{1, ..., N} for a given state feedback associated with the matrix K s D U -stabilizing the pair LMI (24) is satisfied . It is the idea suggested in [1] and [16]. Although this constraint is conservative, it allows, by checking the same property of P (A(θ),B(θ)) . 2 4 U(P i ) O II d ⊗ L i O 3 5 D U -stability for A s and A o , to circumvent a H U i = O O O O − O II problem. Indeed, if the state feedback K s which D U -stabilizes the system is not initially calculated, the condition (15) corresponds to a dq II d ⊗ L i O O −γ II dr < : 2 4 3 5 ˆ II d ⊗ A s i = ; .Matrix K s can thus be considered as an initialization of the technique presented in this part, enabling the computation of a control law by static output feedback. This « initialization step » is very important for this technique because among the infinity of matrices K s which BMI + H F O O − II dn II d ⊗ B i II d ⊗ J i O ˜ < : 2 4 O O 3 5 d ⊗ G ˆ − II O O ˜ = ; + H II dm O O II d ⊗ K s O − II dm D U -stabilize the model, some may not lead to a static output feedback checking (13) and (15). In the following section a heuristic resolution enabling to use the D U -stabilization condition of theorem 2 is proposed for the more general case of the robust placement. + H < : 2 4 O O II dm O O 3 5 II d ⊗ H ˆ II d ⊗ C i O O O O ˜ = ; ≺ 0 . (24) IV. ROBUST POLE PLACEMENT The model is now affected by a polytopic and norm- bounded uncertainty and it aims at computing a robust control law by static output feedback. A robust with k U(P i )= (R k ⊗ P k i ) et γ = ρ −2 . (25) k=1 D U -stabilization condition by static output feed- back is presented in the first part whereas in the second part, a heuristic of robust placement is proposed. An output feedback matrix is then given by : K o = G −1 H (26) Proof : For any vector of barycentric coordinates θ ∈ Θ , D U -stabilization The uncertain closed-loop state matrix is the one defined in (10). The objective is to compute the matrix K o which leads to the largest size of let us consider the convex combination N H U = (θ i H U i ) . B(ρ) , i.e. which leads to the i=1 largest bound ρ . To do that, the idea presented in the previous section consists in finding a state feedback and an output feedback checking the same property by the same (or the same ones) set of Lyapunov matrices, is still exploited. This will be checked if there exists a solution to the following Since all θ i are positive, it comes : H U ≺ 0 . (27) In the sequel, we will adopt the following notations : A s (θ)=A s , B(θ)=B , C(θ)=C , J(θ)=J and . Let us define the matrix LMI S = K o C−K s , the condition (27) becomes with (26) : system : H U = U ( A s ,P k (θ)) ≺ 0 U ( A o ,P k (θ)) ≺ 0 2 4 3 5 ∀θ ∈ Θ . (22) U(P(θ)) II d ⊗ S G O O O II d ⊗ L O (•) − II dq ⊗ (G + G ) O O O with − II dq (•) (•) −γ II dr U ( A • ,P k (θ)) = < : 2 4 3 5 ˆ II d ⊗ A s = ; ≺ 0 . (28) ˆ II dn ˜ k X » II dn II d ⊗ A • – + H F O O − II dn II d ⊗ B II d ⊗ J O ˜ II d ⊗ A • (R k ⊗ P k (θ)) ≺ 0 (23) k=1 872 Authorized licensed use limited to: Biblioteka Glowna i OINT. Downloaded on May 20, 2009 at 05:48 from IEEE Xplore. Restrictions apply. A s and such that the Here, it is important to emphasize that matrices A s and A o « share » the same set of Lyapunov matrices BMI A. Robust L(θ)=L Since γ 0 , applying Schur’s lemma to (28) gives : The technique used for the numerical example is that proposed in [14]. Step 2 : A robust control law by static output feedback is computed, if possible , for each matrix K s thanks to the tool presented by theorem 3. In other words, when a matrix K s enables to compute a robust control law by static output feedback, we obtain a matrix K o and its corresponding robustness bound ρ .When K s does not enable to satisfy LMI H U = 2 4 U(P(θ)) II d ⊗ S G O 3 5 (•) − II d ⊗ (G + G ) + H ˘ F ˆ II d ⊗ A s − II dn II d ⊗ B ˜¯ ). Te s t : Then we can define a convergence criterion which, if it is satisfied, apriori picks up the « best » K o (denoted by K o ) in the sense of the criterion, and stops the process. If the convergence criterion is not satisfied, an evolution process based on genetic algorithms (not detailed in this paper) is used to improve the robustness of the control law, iteration after iteration, until the convergence criterion is finally satisfied. (24), the robustness bound is imposed zero ( ρ =0 + » II d ⊗ ρL O – ˆ II d ⊗ ρL O ˜ + F ( II d ⊗ JJ )F ≺ 0 . (29) By virtue of the lemma proposed in [25], it exists a matrix E checking E E ≤ II dr such that : 2 4 U(P(θ)) 3 5 H U = II d ⊗ S G O (•) − II d ⊗ (G + G ) V. NUMERICAL ILLUSTRATION The model considered for this example, inspired from [9], is that of a satellite. The state, input and output matrices of the system are respectively given hereafter : + H ˘ F ˆ II d ⊗ (A s + Jρ EL) − II dn II d ⊗ B ˜¯ ≺ 0 . (30) 2 4 0 1 0 0 −10Γ −10δ 10Γ 10δ 0 3 5 ; B = 2 4 0 0 0 1 3 5 ; Let consider E = ρ E ,itcomes: ⎡ ⎣ U(P(θ)) ⎤ A(Γ,δ)= 0 0 1 II d ⊗ S G O Γ δ −Γ −δ H U = ⎦ » – −2 −5 −7 −2 −2, 5 , 3 , 5 −3, 3 (•) − II d ⊗ (G + G ) C = . + H F II d ⊗ A s II dn II d ⊗ B ≺ 0 . (31) − Then, the considered system belongs to the class S − ,i.e. At this stage, the reasoning is the same as the one used in the proof of theorem 2 but with considering equation (17) with A s = A s , B = B , C = C and P k = P k (θ) .Itthen Kimura’s condition is not satisfied ( m + p<n ). The torque is affected by a polytopic uncertainty such that Γ=Γ 0 +∆ Γ where the nominal term is Γ 0 =0, 35 and .Inthesame way, the viscous damping δ is also affected by such an uncertainty and it is assumed to read δ |∆ Γ |≤0, 03 leads to LMI system (22). = δ 0 +∆ δ with B. Heuristic resolution As we already noticed, the technique presented in this paper requires an « initialization step » which consists in computing a state feedback matrix K s . We can thus speak about « good initializations » and « bad initializations » with respect to the system robustness conferred by K s . Indeed, according to the value of the matrix K s ,the approach leads or not to a more or less robust control law by static output feedback. In other words, each matrix K s can lead to a bound ρ , significant of the system robustness. Nevertheless, for the time being, there is no effective method enabling to determine a « good initialization » with respect to the bound, even less « the best one » . δ 0 =0, 028 and |∆ δ |≤0, 001 . The uncertain open-loop state matrix A can then be written : 2 4 3 5 = 0 1 0 0 −10∆ Γ −10∆ δ 10∆ Γ 10∆ δ A = A(Γ 0 ,δ 0 )+ 0 0 0 1 ∆ Γ ∆ δ −∆ Γ −∆ δ A(Γ 0 ,δ 0 )+ 2 4 0 000 −10 0 10 0 0 000 10−10 3 5 ∆ Γ + 2 4 0000 0 −10 0 10 0000 010−1 3 5 ∆ δ . and the describes polytope with four vertices. D U considered in this example results from the union of three subregions : D R 1 , D R 2 and D R 3 . D R 1 and D R 2 are two discs of radius 1 , centered around −2 ± 2 i whereas D R 3 is a vertical half-plane defined by The resolution method that we propose is as follows : x<−4 . The three subregions are then disjoint and D U is symmetric. The optimization problem to be solved is : max Step 1 : Several state feedbacks K s , each one that D U - P i ,F,G,H ρ ), are computed by using a beforehand specified eigenstructure assignment technique and exploiting the dof on the eigenvalues and eigenvectors. , B(θ) constraint (24). The evolution process based on the genetic algorithms 873 under the LMI Authorized licensed use limited to: Biblioteka Glowna i OINT. Downloaded on May 20, 2009 at 05:48 from IEEE Xplore. Restrictions apply. the uncertain term is such that The clustering region stabilizes the pair ( A(θ) [ Pobierz całość w formacie PDF ] |