Observability and Constructability

Transcription

1 Capitolo. INTRODUCTION 5. Observability and Constructability Observability problem: compute the initial state x(t ) using the information associated to the input and output functions u(t) and y(t) of the given dynamic system in the time interval t [t,t ]. Definition. The initial state x(t ) of a dynamic system is compatible with the input and output functions u( ) and y( ) in the time interval [t,t ] if relation y(τ) = η(τ,t,x(t ),u( )) holds for τ [t,t ]. Let: E (t,t,u( ),y( )) denote the set of all the initial states x(t ) compatible with the input and output functions u( ) and y( )in the time interval [t,t ]. Constructability problem: compute the final state x(t ) using the information associated to the input and output functions u(t) and y(t) of the given dynamic system in the time interval t [t,t ]. Definition. A final state x(t ) of a dynamic system is compatible with the input and output functions u( ) and y( ) in the time interval [t,t ] if it exists an initial state x(t ) E (t,t,u( ),y( )) such x(t ) = ψ(t,t,x(t ),u( )). Let: E + (t,t,u( ),y( )) denote the set of all the final states x(t ) compatible with the input and output functions u( ) and y( )in the time interval [t,t ]. For time-invariant systems the two sets E and E + are function only of the amplitude t t of the time interval and therefore in this case one can set t = and use the following simplified notation: E (t,u( ),y( )), E + (t,u( ),y( ))

2 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5. Discrete-time linear systems Let us consider the following discrete-time linear system S = (A, B, C, D): { x(k +) = Ax(k)+Bu(k) Observability: y(k) = Cx(k)+Du(k) Definition. The initial state x() is called unobservable in k steps if it belongs to the set E (k) = E (k,, ), that is if it is compatible with the zero input and zero output functions, u(τ) = and y(τ) =, in the time interval τ [, k ]: = y(τ) = CA τ x(), for τ [, k ] that is if:. = y() y(). y(k ) = C CẠ. CA k } {{ } O (k) x() = O (k)x() where O (k) denotes the observability matrix in k steps. ThesetE (k)ofalltheinitialstatesx()whichareunobservable in k steps is a vectorial space which is equal to the kernel of matrix O (k): E (k) = ker[o (k)] TheunobservablesubspacesE (k)satisfythefollowingchainofinclusions (n is the dimension of the state space): E () E ()... E (n) = E (n+) =... ThesmallestunobservablesubspaceE (n)isobtained,atmost,innsteps.

3 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.3 The set E of all the initial states x() unobservable (that is unobservable in any number of steps) is a vectorial state space which is equal to the set E (n) of all the state space unobservable in n steps: E = E (n) = ker[o (n)] The subspace E is called unobservable subspace of the system and can be determined as follows: C E = ker O O = CẠ. CA n where O = O (n) denotes the observability matrix of the system. Definition. A system S is called observable (or completely observable) if the subspace E is equal to the zero state. Property. A system S is observable if and only if one of these relations holds: kero = {} rango O = n The initial state x() E (k,u( ),y( )) is compatible with the input and output functions u( ) and y( ) in the time interval [,k[ if the following relation is satisfied for τ [,k ]: that is if y(τ) = CA τ x()+ τ i= CA τ i Bu(i)+Du(τ) y l (τ) = CA τ x() where y l (τ) = y(τ) τ i= CAτ i Bu(i) Du(τ) is the free output evolution ofthesystem. Functiony l (τ)can be easilycomputed if theinputandoutputfunctionsu(τ)andy(τ)areknownforτ [,k ].

4 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.4 For time-invariant linear system the following relation holds: E (k,u( ),y( )) = E (k,,y l ( )) Note: the set E (k,u( ),y( )) depends only on the parameter k and the fee output evolution y l (τ). Property. ThesetE (k,u( ),y( ))isa linearvariety whichcanbeobtainedaddingaparticularsolutionx() E (k,,y l ( ))totheunobservable subspace E (k) = E (k,,) in k steps. Graphical representation: E (k,u( ),y( )) = x()+e (k) E (k,u( ),y( )) E (k) x() Setting k = n in the previous relation, one obtains the set E (u( ),y( )) of all the initial states x() which are compatible with the functions u( ) and y( ) in any number of steps: E (u( ),y( )) = x()+e This set contains only one element x() for each couple of functions u( ) and y( ) is and only if E =, that is if the system is completely observable. So, the following property holds. Property. Ifalinearsystemisobservable,thenitexistsonly one initial state x() compatible with any couple of input and output functions u(k) and y(k) in the time inteval k [, n].

5 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.5 Constructability: Istheproblemoffindingthefinalstatex(k)attimek compatiblewiththe input and output functions u(),...,u(k ) and y(),...,y(k ). If the system is observable in k steps, then the set E (k,u( ),y( )) contains only one element x() and therefore the final state x(k) is unique: x(k) = A k x()+ k i= A k i Bu(i) On the contrary, if the system is unobservable in k steps, then it is E (k,u( ),y( )) = x()+e (k), and therefore the set E + (k, u( ), y( )), of all the final states x(k) compatible with the input and output functions u( ) and y( ) has the following structure: E + (k, u( ), y( )) = x(k)+a k E (k) }{{} E + (k) The set E + (k, u( ), y( )) is obtained adding a particular solution x(k) to the subspace E + (k) of all the final states non constructable in k steps: E + (k) = A k E (k) A system is constructable in k steps if E + (k) = A k E (k) = {}, that is E (k) kera k A system is constructable if it exists a k such that E (k) kera k. It takes n steps, at most, for a discrete system to be constructable. So, the constructability condition can be expressed in the following way: E = kero kera n Note: from this last relation it follows that the observabilty implies the constructability, but not vice versa: Observability Constructability

6 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.6 Example. Let us consider the following discrete linear system: x(k +) = x(k)+ u(k) y(k) = [ ] x(k) Compute the set of all the initial states x = x() compatible with the free output evolution: y() =, y() =, y() =. The observability matrix of the system is: O =, deto = 4 Since O is not singular, the system is completely observable, and therefore if the system has a solution, then it has only one solution. The solution x can be determined as follows: that is: y() y() y() = x = C CA CA x = O x x = [O ] = = y() y() y() So, the only initial state compatible with the given free output evolution is x = [,, ]. Note: the free output evolution y(τ) is also determined for τ 3: y(τ) = CA τ x

7 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.7 Example. Let us consider the following discrete linear system: x(k +) = x(k)+ u(k) y(k) = [ ] x(k) Compute the set of all the initial states x = x() compatible with the free output evolution: y() =, y() =, y() = 4. The system is not completely observable: O = E = span The set of all the initial conditions compatible with the free output evolution y() =, y() =, y() = 4 can be determined computing a particular solution x p and adding the unobservable space E. The particular solution x p can be determined solving the following equation 4 = x p y = O x p A solution exists because the vector y belongs to the image of matrix O. y Im(O ) x p = So, the set of all the initial states x compatible with the free output evolution is the following: x = x p +E = +span

8 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.8 Example. Let us consider the following discrete linear system: x(k +) = x(k)+ u(k) y(k) = [ ] x(k) Compute the set of all the initial states x = x() compatible with the free output evolution:: y() =, y() =, y() =. The observability matrix of the system is: O = 4 Since the rank of matrix O is, the system is not completely observable. The set of all the initial states x = x() compatible with the free output evolution: y() =, y() =, y() = satisfies the following relation: y = y() y() y() O x 4 The vector y cannot be obtained as linear combination of the columns of matrix O, and therefore the given problem does not have solutions. There are no initial states x which are compatible with the given free output evolution. Property. The initial state x of minimum norm which minimizes the Euclidean norm y O x is obtained as follows: x = (O ) T N(N T O (O ) T N) N T y where matrix N is a base matrix of Im(O ): Im(N) = Im(O ) x The columns of matrix N are a set of linearly independent columns of matrix O.

9 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.9 Invariant continuous-time linear systems Let S = (A, B, C, D) be an invariant continuous-time linear system: { ẋ(t) = Ax(t)+Bu(t) y(t) = Cx(t)+Du(t) The observability and constructability properties of the continuous-time systems are very similar to those described for of the discrete-time system. Observability: Definition. An initial state x() is unobservable in the time interval [, t] if it is compatible with the zero input and output functions, u(τ) = and y(τ) =, for τ [, t]: = Ce Aτ x(), for τ [, t] The set E (t) of the initial states x() unobservable in [, t] is a vectorial subspace: E (t) = kero t where O t denotes the linear operator O t : X Y(t) which provides the free output evolution y(τ) Y(t) corresponding to the initial state x() X: O t : x() Ce Aτ x(), τ t Property. The unobservable subspace E = E (t) does not depend on the length of the time interval t > and it is equal to the kernel of the same observability matrix O defied for the discrete-time systems. E = ker[o ] So, for continuous-time systems, the computation of the initial state x() compatible with the input and output functions u( ) and y( ) does not depend on the length of he time interval [, t], provided it is not zero, but it depends only on the rank of the observability matrix O.

10 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5. Constructability: For continuous-time system the observability is equivalent to the constructability. In fact, in this case, the unobservable and the unconstructable subspaces E and E + are linked by the following relation: E + = e At E Since matrix e At is always invertible, for any matrix A, the two subspaces E + and E always have the same dimension. Moreover, it can be proved that they are equal: E + = E From this last relation it follows that for continuous-time systems the osservability implies and is implied by the constructability: Observability Constructability

11 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5. Duality Given a continuousordiscrete-time systems = (A,B,C,D), the system S D = (A T,C T,B T,D T ), is called the dual system of system S. No. of inputs of system S = No. of outputs of system S D No. of outputs of system S = No. of inputs of system S D Reachability and observability properties. The reachability and observability matrices R + D and O D of the dual system S D are linked to the reachability and observability matrices R + and O of system S as follows: R + D = [CT A T C T...(A T ) n C T ] = O D = B T B T A T. B T (A T ) n C CẠ. CA n T = (O ) T = [B AB...A n B] T = (R + ) T Property. For the systems S and S D the following properties hold: S reachable S D observable S observable S D reachable S controllable S D constructable S constructable S D controllable If two equivalent systems S and S are linked by the state space transformation x = Tx, then the corresponding dual system S D and S D are linked by the state space transformation x D = Px D where P = T T : S = (A, B, C) Duality T S = (T AT, T B, CT) Duality S D = (A T, C T, B T ) P S D = (TT A T T T, T T C T, B T T T )

12 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5. Osservability standard form Property. A linear system S = {A, B, C, D} (continuous or discrete) not completely observable can always be brought into observability standard form, that is, system S is algebraically equivalent the a transformed system S = {A, B, C, D} where the matrices A = P AP, B = P B, C = CP and D have the following structure: [ ] [ ] A, B A = B = A, A, B C = [ C ] D = D Let be ρ = n dime < n. The transformation matrix P which brings the system into the observability standard form has the following structure: ] P = [ R R, x = Px where R R ρ n is composed by ρ linearly independent rows of the observability matrix O, and R R (n ρ) n is a free matrix chosen such that the transformation matrix P is non singular. Proprerty. The subsystem of dimension ρ characterized by matrices A, and C is completely observable. The subsystem (A,, B, C ), is called observable subsystem and it represents the part of the original system which is observable. The subsystem (A,, B, ), characterized by matrices A,, B and C =, is called unobservable subsystem and it represents the part of the system which does not influence in any way the system output y. Also in this case the eigenvalues of matrix A are split in two sets: the eigenvalues of the observable part of the system (the eigenvalues of matrix A, ) and the eigenvalues of the unobservable part (the eigenvalues of of matrix A, ).

13 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.3 For the discrete case, let us divide the state vector x(k) in two parts: the observable part x and the unobservable part x : x = [ ] T, x x where dimx = ρ. The equations of the system are: x (k +) = A, x (k)+ B u(k) x (k +) = A, x (k)+ A, x (k)+ B u(k) y(k) = C x (k)+ Du(k) The corresponding block scheme is: D B x (k +) z I ρ x (k) C y(k) u(k) A, A, B x (k +) z I n ρ x (k) A, A similar decomposition holds also for continuous-time systems. Property. The transfer matrix H(z) [o H(s) ] of a linear system is equal to the transfer matrix of the observable part: H(z) [o H(s) ] is influenced only by the matrices (A,, B, C ) of the observable subsystem. Proof. The transfer matrix H(z) = C(zI A) B = C(zI A) B is: H(z) = [ C ][ ] [ ] zi A, B = A, zi A, B = [ C ][ ][ ] (zi A, ) B (zi A, ) = = C (zi A, ) B B

14 Capitolo 5. OBSERVABILITY AND CONSTRUCTABILITY 5.4 Example. Let us consider the following time-continuous linear system ẋ(t) = x(t)+ u(t) y(t) = [ ] x(t) Bring the system in the observability standard form. Sol. The observability matrix of the system is: O = deto = 3 3 Matrix O is singular. The system is not completely observable and therefore it is possible to compute the transformation matrix P which brings the system into the observability standard form: P = P = The transformed system has the following form: x(t)= x(t)+ y(t)= [ ] x(t) u(t) The unobservable part of the system is simply stable because it has an eigenvalues located in the origin: s =. The transfer matrix: G(s) = C o (si A o ) B o = [ ][ s s+ if a function only of the observable part of the system. ] [ ] = (s+) s +3