Abstract: In this paper a new interaction measure for stable multivariable systems is
introduced. The interaction measure, the Hankel Interaction Index, is a dynamic extension
of the underlying idea in the classic Relative Gain Array (RGA) and its various extensions.
The new index is based upon the Hankel norm of the SISO elementary subsystems built
from the original MIMO system. The main advantage of the Hankel Interaction Index
is its ability to quantify frequency dependent interactions and that it can be used for
input-output pairing. Several examples are included to illustrate the new ideas presented.
Copyright ©2002 IFAC
The Hankel norm of a system with transfer function where the operator × denotes the element-by-element
G(z) is defined as multiplication of the two matrices. This normalization
p results in a matrix where the sum of all elements in
kG(z)kH = ΓcΓ o ) = σ1H
λmax (Γ a row or column is equal to one, although (7) does
not have the same nice physical interpretation as for
The Hankel norm is thus the maximum HSV. The the RGA. Another way of making the normalization is
Hankel norm can also be interpreted in the following to express every input and every output in percentage
way. Consider a system input u[k] = 0 ∀k > 0 and a of full scale. This normalization supports the idea of
system output y[k], then scaling all elements in Σ¯H so that their sum is equal
to one. The latter normalization is preferred and we
kyk` choose the following normalized Hankel Interaction
2 (0,∞)
kG(z)kH = sup (6) Index Array
u∈`2 (−∞,0) kuk`
2 (−∞,0)
kGi j (z)kH
where `2 (0, ∞) is the space of square summable vector Σ H ]i j =
[Σ (8)
sequences in the interval [0, ∞). An analogous defini- ∑i, j kGi j (z)kH
tion applies to `2 (−∞, 0).
Using the same methodology as for the RGA we
The Hankel norm gives the `2 gain from past inputs determine the input-output pairing by finding in each
to future outputs. Equation (6) can be interpreted as row i the largest element (i, j). Input j is then selected
a measure of how significant is the effect of an input to control output i.
on the state and how much that effect is reflected in
To compute the Hankel Interaction Array we have
the output. The Hankel norm can thus be interpreted
to solve the 2p Lyapunov equations (4) and (5) and
as a controllability and observability measure of the
compute the eigenvalues of the p2 products Γ (c j)Γ (i)
o to
system. Summarizing, the Hankel norm has two fun-
obtain the Hankel norm.
damental properties: is a controllability-observability
measure and, secondly is independent of the state If Gi j (z) = 0 for a given pair (i, j), then Γ c( j)Γ o(i) =
space representation of the system. 0, leading to [Σ ΣH ]i j = 0. This implies that a block
diagonal G(z) gives a block diagonal Σ H matrix, with The Hankel Interaction Index Array (8) is
the same structure. This is consistent with intuition,
since, in those cases, the Σ H matrix will suggest the
0.3755 0.1872
ΣH =
right controller structure. It is important to observe that 0.1300 0.3073
the Σ H takes the full dynamic effects of the system into
account and not only the steady-state performance as This result suggests that a diagonal controller should
the RGA or the behavior at a single frequency. be chosen with the pairs (u1 , y1 ) and (u2 , y2 ), since the
largest elements are the elements (1, 1) and (2, 2).
For the MIMO process (9), the RGA is
5. CLOSED LOOP BANDWIDTH EFFECT
When a measure of dynamic interaction is built, atten-
0.8242 0.1758
RGA(G) =
tion should be paid to the relevant frequency range. 0.1758 0.8242
This has, for instance, been proposed in Witcher and
McAvoy (1977) and Gagnepain and Seborg (1982). In this example the RGA leads to the same conclusion
Specifically, interactions are meaningful in control de- regarding the pairing of inputs and outputs.
sign only in a frequency band where the plant in-
put has significant energy. One way to introduce this E XAMPLE 2
element into the building of the interaction index is The system in this case has a transfer function G(z)
to observe that the (vector) plant input, u[k] is con- given by
nected to the reference signals and to the output distur-
0.1021 0.3707z − 0.3535
bances through the control sensitivity Su (z), see Good-
win et al. (2000), through the expression
z − 0.9048 z2 − 1.724z + 0.7408
−0.192z + 0.1826 0.09516
Su (z) = (G(z))−1 T(z) z2 − 1.869z + 0.8781 z − 0.9048
(10)
The frequency response of Su (z) depends on the re- If the RGA is computed we obtain
lationship between the plant bandwidth and the closed
loop bandwidth. To gain insight, we consider the SISO
0.5033 0.4967
RGA(G) =
case with a biproper controller. In that simpler situa- 0.4967 0.5033
tion, we observe that, when the closed loop bandwidth
is larger than the plant bandwidth, the control sensi- We observe that the RGA suggests, very weakly, that
tivity has larger magnitude at high frequencies than the pairing should be (u1 , y1 ) and (u2 , y2 ). On the other
at low frequencies. In the reverse situation, the con- hand, the Hankel Interaction Index Array is
trol sensitivity has a low-pass characteristic. This also
applies to MIMO systems. However, it has to be ap- 0.1936 0.3281
ΣH =
(11)
plied with caution since there may be a non-unique 0.2978 0.1805
closed loop bandwidth. Even with this caution, there
are many cases when this approach will provide a use- The Hankel Interaction Index provides a clear sugges-
ful information if we apply the Hankel Interaction In- tion. It directs the designer to pair (u1 , y2 ) and (u2 , y1 ).
dex to a modified system, G̃(z), obtained by filtering The reason for this difference is that the Hankel In-
the plant through a (scalar) case-dependent filter, F(z), teraction Index Array takes into account the dynamic
i.e. features of the interaction.
G̃(z) = G(z)F(z) A deeper insight can be gained with a simple control
The effect of this filtering will be illustrated in the design. Assume that we want that y1 [k] tracks a step
following section. reference signal, r1 [k] = µ [k], and that y2 [k] tracks
another step reference signal, r2 [k] = −µ [k − 5]. We
want to synthesize a dead-beat control which drives
6. EXAMPLES the error to zero after two time units. The synthesis
method is the Youla parameterization of all stabilizing
Some examples are used to illustrate the usefulness of controllers (Goodwin et al. (2000)), i.e. the controller
the Hankel Interaction Index. has a transfer function, C(z), given by
E XAMPLE 1 C(z) = (I − Go (z)Q(z))−1 Q(z)
Consider the system with the pulse transfer function
0.1 0.08 where Q(z) is any stable transfer function matrix, and
where Go (z) is the nominal model. The dead-beat
(z − 0.8)(z − 0.5) z − 0.8
G(z) =
performance is nominally achieved if
−0.24 0.1(z − 0.1)
z − 0.5 (z − 0.8)(z − 0.5) 1
(9) Q(z) = (Go (z))−1
z2
1.5 This is very close to the case when no filter was used
y [k] (as seen in equation (11)). If intermediate filters were
1
1 used one could observe a transition from the RGA to
the Hankel Interaction Index Array given in (11), as
Step response
0.5
we increase the frequency response of the filter at high
0 frequencies.
y [k]
2
−0.5
E XAMPLE 3
−1 In this case we deal with a 3 × 3 MIMO system with
the transfer function
−1.5
0 5 10 15 20 G(z) = G1 (z) G2 (z) G3 (z)
Time [s]