ABSTRACT
ith block diagonal entry, where H i ∈ IR(s)ni × ni. For are left coprime and D –pl1 Npl = P. We need the following
f, g ∈ IR[s], deg(f) := degree of f, the relative degree of assumptions:
f/g ∈ IR(s) is defined as deg(f) – deg(g). For H = [hij] ∈
IR(s)n × m, the relative degree of H is defined as the largest A1) The plant P has full normal row rank, and
relative degree of hij, 1 ≤ i ≤ n and 1 ≤ j ≤ m. Two matrices A2) P+[P] ∩ Z+[P] = Ø.
A ∈ S n × n and B ∈ S n × n are equivalent if there exist
S-unimodular matrices U and V such that A = UBV. Assumption A1) implies that m ≥ n, and it is neces-
sary for block decoupling [7]; assumption A2) means
II. PRELIMINARY that the CI+-poles and CI+-zeros of P do not coincide.
These assumptions together are sufficient for the exist-
Consider the unity-feedback system S(P, C) shown in ence of block decoupling controllers for P. Although A2)
Fig. 1, where P ∈ IRpo(s)n × m is the plant, C ∈ IRp(s)m × n is is not necessary [6], it is generically satisfied, and it well
the controller, (u1, u2) is the input and (y1, y2) is the output. simplifies the derivations to be described below.
Let u := [u 1T u 2T ] T and y := [y1T y2T ] T . The closed-loop
transfer matrix Hyu ∈ IRp(s)(n + m) × (n + m) is given by III. PARAMETERIZATIONS OF BLOCK
DECOUPLING CONTROLLERS
H y1u1 H y1u2 C(I + PC) – 1 – CP(I + CP) – 1 Consider the plant P ∈ IRpo(s)n × m, which is a strictly
H yu = = .
H y2u1 H y2u2 PC(I + PC) – 1 P(I + CP) – 1 proper rectangular rational matrix. Since P ∈ IRpo(s)n × m
has full normal row rank, by assumption A1), there exists
(2.1)
a unimodular matrix U ∈ Sm × m such that
Definition 2.1. The system S(P, C) is stable if and only if
PU = [P 0], (3.1)
Hyu is stable. We say that C stabilizes P or C is a stabilizing
controller for P if S(P, C) is stable.
where P ∈ IRpo(s)n × m is nonsingular. Clearly P satisfies
assumptions A1) and A2). Let (Nr, Dr) be a right coprime
Definition 2.2. The system S(P, C) is block decoupled
factorization of P, and let (Dl, Nl) be a left coprime factor-
with respect to the partition (n1, …, nk)1 if C stabilizes P and
ization of P, where Nr, Dr, Nl, Dl ∈ Sn × n. We note that
the I/O map Hy2u1 is nonsingular and block diagonal with
these coprime factorizations can be computed through
respect to the partition (n1, …, nk). We say that C is a
standard state space procedure [8]. Let
block decoupling controller for P if S(P, C) is block
decoupled with respect to the partition (n1, …, nk).
Dr 0
N pr = N r 0 , D pr = U , (3.2)
Remark. We note that decoupling is a special case of 0 I
block decoupling with each block size ni = 1.
N pl = N l 0 U – 1, D pl = D l , (3.3)
n×n
Definition 2.3. A stable rational matrix M ∈ S is an
achievable I/O map of the system S(P, C) if there exists a where Npr, Npl ∈ Sn × m, Dpr ∈ Sm × m and Dpl ∈ Sn × n. Then,
stabilizing controller C such that Hy2u1 = M. (Npr, Dpr) and (Dpl, Npl) are, respectively, a right coprime
factorization of P and a left coprime factorization of P. In
We say that (Npr, Dpr) is a right coprime factorization the following, we construct a special generalized Bezout
of P if Npr ∈ Sn × m and Dpr ∈ Sm × m are right coprime and identity which is suitable for the description of block
NprD –pr1 = P. Similarly, we say that (D pl, Npl) is a left decoupling controllers. We can write
coprime factorization of P if Dpl ∈ Sn × n and Npl ∈ Sm × m
Z ij
N r– 1 = , (3.4)
Pij–Pij+
N ij
1
We consider block decoupling with respect to a fixed but D l– 1 = , (3.5)
arbitrary partition (n1, …, nk) throughout.
D ij–D ij+
148 Asian Journal of Control, Vol. 1, No. 3, September 1999
αλN l– 1
U cr = , V cr = βψD l– 1, (3.7)
0
αλN l– 1 + D rQ 1 –1
C Pd = U βψD l– 1 – N rQ 1
Q2
αλN r– 1 βψD r– 1 0
U cl = , V cl = U – 1 , (3.8)
0 0 I
V cl U cl D pr – U cr
= I0 . (3.9)
– N pl D pl N pr V cr 0I
In the descriptions of HPd and CPd, (3.13) and (3.14),
respectively, the parameter Q1 is not free but must satisfy
The set of all stabilizing controllers for P is, thus, that NrQ1D l is block diagonal. In order to obtain an
given by, unconstrained parameterization, we partition N r– 1 as
N r– 1 = N 1 Nk , (3.15)
–1
Q1 Q1
CP = U cr + D pr V cr – N pr
Q2 Q2 where Ni ∈ IR(s)n × ni, 1 ≤ i ≤ k. Let (Γi, Θi) be a right
coprime factorization of Ni, 1 ≤ i ≤ k, where Γi ∈ Sn × ni
and Θi ∈ Sni × ni. Thus, N r– 1 has the following property:
N r– 1 = Γ 1Θ –1 1 Γ kΘ –k 1 = ΓΘ – 1 , (3.16)
1
Q 1 ∈ S n × n and Q 2 ∈ S (m – n) × n . (3.10)
where Γ = [Γ1 … Γk] and Θ = diag{Θ i} ki = 1. Since, for each
i, P+[Ni] ⊂ P+[N r– 1 ] = Z+[P] and P+[Ni] = Z+[Qi], it follows
that
Since NprUcl = αλI, the set of all achievable I/O maps
Hy2u1 is Z+[Θ] ⊂ Z+[P]. (3.17)
C.-A. Lin and C.-M. Wu: Block Decoupling Control of Linear Multivariable Systems 149
For 1 ≤ i ≤ k, P+[Di] ⊂ P+[D l– 1 ] = P+[P] and P+[Di] 1 Q d is block diagonal and stable
= Z+[∆i]. It then follows that . (3.23)
Q 2 ∈ S (m – n) × n
Z+[∆] ⊂ P+[P]. (3.20)
HPd = {αλI + ΘQd∆|Qd is block diagonal and stable}. Another way to design block decoupling controllers
(3.21) is to first find a block decoupling precompensator which
makes the transfer matrix of the cascade connection (of the
(b) Since NrQ1Dl = ΘQd∆ if and only if precompensator and the plant) block diagonal and then
150 Asian Journal of Control, Vol. 1, No. 3, September 1999
A PF = (αΘ + ∆Q d)(βψ∆ – 1 – ΘQ d) – 1
Let ∆ i = diag{∆ i} i = 1 , and let
k
d
F = P Θ∆
–1 –1
. (4.2)
1
Since P–1Θ = DrΓ and ∆–1 are proper, F in (4.2) is a proper Q d is block diagonal and stable . (4.5)
rational matrix, and the cascade connection of F and P has
a block diagonal transfer matrix PF =: PF. Clearly, PF =
Θ∆ i –1. To show that the cascade connection of F and P is Hence,
stabilizable, we only have to show that there is a controller
A(s) such that the system S(P, F , A) shown in Fig. 2 is A = (αΘ + ∆Q d)(βψ∆ – 1 – ΘQ d) – 1 (4.6)
stable. We will do this by showing that if A(s) is a block
diagonal controller which stabilizes the plant PF , that is, for some Qd which is block diagonal and stable. Since
the system S(PF , A) shown in Fig. 3 is stable, then the S(PF, A) is stable, Hyiuj, i = 2, 3 and j = 1, 3, are stable. In
system S(P, F , A) is also stable. Let u := [u 1T u 2T u 3T ]T and the Appendix, we will show that the transfer matrices
y := [y1T y2T y3T ]T. The closed-loop transfer matrix H yu of Hy1u1, Hy1u2, Hy1u3, Hy2u2 and Hy3u2 are also stable. Hence,
the system S(P, F , A) is the system S(P, F , A) is stable and F defined in (4.2) is
C.-A. Lin and C.-M. Wu: Block Decoupling Control of Linear Multivariable Systems 151
F(I + PFA) – 1 FAP(I + FAP) – 1 F(I + APF) – 1 By Proposition 4.1, ψ∆ i ∈ Sni × ni; thus, ψ∆ i ∈ Sni × ni and
0 –1 –1
–
F(I + PFA) – 1 FAP(I + FAP) – 1 0 F(I + APF) – 1
ψ∆ is stable.
–1
–1
= I + Θ∆ (αΘ + ∆Q d)(βψ∆ – ΘQ d)
–1 –1 –1 –1
A(I + PFA) –1
– AP(I + FAP) –1
0 – APF(I + APF) –1
(I + PFA)
(A.1)
152 Asian Journal of Control, Vol. 1, No. 3, September 1999