You are on page 1of 33

Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.

php

Chapter 9

Digital Implementation

This chapter deals with digital implementation of high-gain observers and properties
of closed-loop systems under digital control. The nonlinear system under considera-
tion is

ẇ = f0 (w, x, u), (9.1)


ẋ = Ax + Bφ(w, x, u), (9.2)
y = C x, (9.3)

where u ∈ R m is the control input, y ∈ R is the measured output, and w ∈ R` and


x ∈ Rρ constitute the state vector. The ρ × ρ matrix A, the ρ × 1 matrix B, and the
1 × ρ matrix C are given by

0 1 ··· ··· 0 0
   
 0 0 1 ··· 0   0 
.. .. ..
   
A=  B = C=
 
, , 1 0 ··· ··· 0 .
   
 . .  . 
 0 ··· ··· 0 1   0 
0 ··· ··· ··· 0 1

The high-gain observer for the system (9.1)–(9.3) is taken as

x̂˙ = Ax̂ + Bφ0 (x̂, u) + H (y − C x̂), (9.4)

where
α1 /"
 
 α2 /"2 
..
 
H = , (9.5)
 
.
 αρ−1 /"ρ−1
 

αρ /"ρ

φ0 (x, u) is a nominal model of φ(w, x, u), " is a small positive constant, and the posi-
tive constants αi are chosen such that the polynomial

s ρ + α1 s ρ−1 + · · · + αρ−1 s + αρ

279
280 CHAPTER 9. DIGITAL IMPLEMENTATION

is Hurwitz. The function φ0 (x̂, u) is locally Lipschitz in (x, u) over the domain of
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

interest and globally bounded in x.


Section 9.1 describes the discretization of the high-gain observer for digital imple-
mentation. Both linear and nonlinear observers are discussed. The nonlinear observer
is discretized using the Forward Difference method. For the linear observer, three dif-
ferent discretization methods are discussed. Section 9.2 deals with output feedback
stabilization. Properties similar to what we have seen for continuous-time controllers
are established for discrete-time controller when the observer parameter " and the
sampling period are sufficiently small. The effect of measurement noise is studied in
Section 9.3, where, once again, results similar to the continuous-time case are shown.
Section 9.4 presents a multirate digital control scheme in which the control sampling
period is larger than the observer sampling period. This scheme is useful for computa-
tionally demanding controllers where discrete-time state feedback can be implemented
with an appropriate sampling period, but such period cannot be reduced because of
the computation time. Such situations arise, for example, when the controller com-
pensates for a nonlinearity, such as hysteresis, by calculating its inverse at each time
step, or in feedback linearization when the computed nonlinear terms are complicated.
Another example is model predictive control where the controller solves an optimal
control problem within each sampling period.75 Such a slow sampling period might
not be small enough to implement the high-gain observer, which has fast dynamics.
The multirate scheme allows the observer to run with a sampling period shorter than
the controller’s sampling period.

9.1 Observer Discretization


The first step in discretizing the observer equation (9.4) is to scale the state estimates
so as to remove the negative powers of " in the observer gain H . Let

q = D x̂, (9.6)

where
D = diag(1, ", "2 , . . . , "ρ−1 ).
The change of variables (9.6) transforms the observer equation (9.4) into

"q̇ = F q + E y + "ρ Bφ0 (D −1 q, u), (9.7)


−1
x̂ = D q, (9.8)

where
−α1 1 0 ··· 0 α1
   
 −α2 0 1 ··· 0  α2
.. .. ..  ..
   
F = . and E = .
  
. . .
αρ−1 
   
−αρ−1 0 1
−αρ 0 ··· ··· 0 αρ

While negative powers of " appear in (9.8), the terms F q and E y on the right-hand
side of (9.7) have no negative powers of ". The function φ0 depends on D −1 q, but it
is globally bounded. The appearance of " multiplying q̇ on the left-hand side of (9.7)
75 See, for example, [29] and [124].
9.1. OBSERVER DISCRETIZATION 281

indicates the time scale of the observer. The change of time variable from t to τ = t /"
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

puts (9.7) into the form

dq
= F q + E y + "ρ Bφ0 (D −1 q, u). (9.9)

If α is an appropriate sampling period for (9.9), then the sampling period for (9.7) is
taken as T = α".
Depending on whether or not the nominal function φ0 is zero, the observer (9.7)–
(9.8) could be linear or nonlinear. In the nonlinear case when φ0 6= 0, the most prac-
tical way to discretize the observer is to use the Forward Difference method, which
yields

ξ (k + 1) = Ad ξ (k) + +Bd y(k) + α"ρ Bφ0 (D −1 ξ (k), u(k)), (9.10)


−1
x̂(k) = D Cd ξ (k), (9.11)

where ξ = q and ξ (k), x̂(k), u(k), and y(k) denote the signals ξ , x̂, u, and y at the
sampling points tk = kT . The matrices Ad , Bd , and Cd are defined in Table 9.1. When
φ0 = 0, the observer (9.7)–(9.8) reduces to the linear equation

"q̇ = F q + E y, (9.12)
−1
x̂ = D q (9.13)

for which several discretization methods are available.76 We limit our discussion to
the Forward Difference, Backward Difference, and Bilinear Transformation methods,
for which the discretized observer is given by

ξ (k + 1) = Ad ξ (k) + +Bd y(k), (9.14)


−1
x̂(k) = D [Cd ξ (k) + Dd y(y)]. (9.15)

The matrices Ad , Bd , Cd , Dd and the state ξ are given in Table 9.1 for the three meth-
ods. For the Forward Difference method, α is chosen small enough to ensure that the
eigenvalues of Ad are in the interior the unit circle, while for the other two methods
the eigenvalues of Ad will be in the interior the unit circle for any α > 0 since F is
Hurwitz.
An extensive simulation study was carried out to compare the performance of
the digital implementation of the linear high-gain observer for different discretization
methods, with and without measurement noise.77 The best performance was obtained
with the Bilinear Transformation method when both the transient and steady-state re-
sponses were taken into consideration.
The same discretization methods can be used with reduced-order observers. The
previously mentioned simulation study [31] found similar results for linear reduced-
order observers. In particular, the best performance was obtained with the Bilinear
Transformation method. It was also found that in the presence of measurement noise
the full-order high-gain observer outperforms the reduced-order one, which is not sur-
prising in view of the simulation of Example 1.4. If the reduced-order high-gain ob-
server
ż = −(β/")(z + y), x̂2 = (β/")(y + z),
76 See, for example, [16], [45], or any digital control textbook.
77 See [31]. The comparison included also the zero-order hold and first-order hold discretization methods.
282 CHAPTER 9. DIGITAL IMPLEMENTATION

Table 9.1. Coefficients of the discrete-time implementation of the high-gain observer.


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Forward Difference Backward Difference Bilinear Transformation

def α α def
Ad (I + αF ) (I − αF )−1 = M1 (I + 2 F )(I − 2 F )−1 = N2 M2

Bd αE αM1 E αM2 E

Cd I M1 M2
α
Dd 0 αM1 E M E
2 2

ξ q (M1−1 q − αE y) (M2−1 q − (α/2)E y)

is used to estimate the derivative of y by x̂2 , then Forward Difference discretization


with T = "/β yields Euler’s formula
1
x̂2 (k) = [y(k) − y(k − 1)],
T
which is commonly used in engineering applications to estimate the derivative of a
signal. The fact that Euler’s formula is just one of the possible choices with high-gain
observers shows that better results can be obtained by designing a high-gain observer
rather than resorting to the simple Euler’s formula.78

Example 9.1. Reconsider the system

ẋ1 = x2 , ẋ2 = −x1 − 2x2 + 0.25x12 x2 + 0.2u, y = x1

from Example 1.1, where u(t ) = sin 2t . A continuous-time high-gain observer can be
taken as
2 1
x̂˙1 = x̂2 + (y − x̂1 ), x̂˙2 = φ0 (x̂, u) + (y − x̂1 ).
" "2
We consider two choices of φ0 :

φ0 = −x̂1 − 2x̂2 + 0.25x̂12 x̂2 + 0.2u and φ0 = 0.

With the scaling


1 0
•
˜
q = D x̂ = ,
0 "
the observer is given by

"q̇ = F q + E y + "2 Bφ0 (x̂, u),

where
−2 1 2 0
• ˜ • ˜ • ˜
F= , E= , and B= .
−1 0 1 1
78 This point is demonstrated experimentally in [32].
9.1. OBSERVER DISCRETIZATION 283

When φ0 6= 0, the observer is discretized using the Forward Difference method to


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

obtain

ξ (k + 1) = Ad ξ (k) + Bd y(k) + α"2 Bφ0 (x̂(k), u(k)), x̂(k) = D −1 ξ (k),

where Ad and Bd are defined in Table 9.1. When φ0 = 0, the observer is discretized
using the Forward Difference and the Bilinear Transformation methods. In the first
case, the observer is given by

ξ (k + 1) = Ad ξ (k) + Bd y(k), x̂(k) = D −1 ξ (k),

while in the second one it is given by

ξ (k + 1) = Ad ξ (k) + Bd y(k), x̂(k) = D −1 [Cd ξ (k) + Dd y(k)],

where the matrices are defined in Table 9.1. In particular, for the Forward Difference
method,
(1 − 2α) α 2α
• ˜ • ˜
Ad = and Bd = ,
−α 1 α

and for the Bilinear Transformation method,

1 (1 − α − α2 /4) α
• ˜
Ad = ,
1 + α + α2 /4 −α (1 + α − α2 /4)
α 2 + α/2 1 1 α/2
• ˜ • ˜
Bd = , Cd = ,
1 + α + α /4
2 1 1 + α + α /4
2 −α/2 (1 + α)

1
and Dd = 2 Bd . Simulation is carried out with initial conditions x(0) = col(1, −1) and
ξ (0) = 0. Figure 9.1 shows x2 and its estimate x̂2 as determined by the continuous-time
observer with φ0 6= 0, the Forward Difference discrete-time observer with φ0 6= 0, the
Forward Difference discrete-time observer with φ0 = 0, and the Bilinear Transfor-
mation discrete-time observer with φ0 = 0. The observers with φ0 6= 0 are nonlin-
ear, while those with φ0 = 0 are linear. In all cases " = 0.1 and T = 0.01; hence,
α = T /" = 0.1. The performance of the four observers is comparable, with the
nonlinear ones slightly better than the linear ones. Figure 9.2 shows the steady-state es-
timation errors of the three discrete-time observers. The errors are comparable for the
two linear observers, but a smaller error is achieved with the nonlinear one. Figure 9.3
shows the impact on the estimation error x̃2 when " is reduced from 0.1 to 0.01 while
keeping α = 0.1; hence, T = 0.01 when " = 0.1 and T = 0.001 when " = 0.01. Similar
to the continuous-time case, reducing " causes a larger peak and faster decay. The figure
shows results for the two discrete-time linear observers, but similar results are obtained
for the nonlinear observer. The effect on the steady-state error is shown in Figure 9.4
for the Forward Difference linear observer, where reducing " leads to smaller error. Fi-
nally, Figure 9.5 compares the Forward Difference and Bilinear Transformation linear
observers when α = 0.5 and " = 0.01. The steady-state estimation errors are compara-
ble, but there are differences in the transient response. The Bilinear Transformation
observer has less peaking at the expense of a slightly larger settling time. 4
284 CHAPTER 9. DIGITAL IMPLEMENTATION

Continuous Nonlinear Discrete FD Nonlinear


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

x2
3 x̂2 3

x2 and x̂2

x2 and x̂2
2 2

1 1

0 0

−1 −1
0 1 2 3 0 1 2 3
Time Time

Discrete FD Linear Discrete BT Linear

3 3
x2 and x̂2

x2 and x̂2
2 2

1 1

0 0

−1 −1
0 1 2 3 0 1 2 3
Time Time

Figure 9.1. Simulation of Example 9.1. The estimation of x2 using four different ob-
servers: continuous nonlinear (φ0 6= 0) observer, discrete Forward Difference nonlinear observer,
discrete Forward Difference linear (φ0 = 0) observer, and discrete Bilinear Transformation linear
observer.

Figure 9.2. Simulation of Example 9.1. The steady-state estimation error x̃2 of the discrete
Forward Difference nonlinear observer, the discrete Forward Difference linear observer, and the
discrete Bilinear Transformation linear observer.

9.2 Stabilization
Consider the single-input–single-output nonlinear system

ẇ = f0 (w, x, u), (9.16)


ẋ = Ax + Bφ(w, x, u), (9.17)
y = C x, (9.18)
z = ψ(w, x), (9.19)
9.2. STABILIZATION 285
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Figure 9.3. Simulation of Example 9.1. The impact of reducing " on the transient re-
sponse of x̃2 for the Forward Difference and Bilinear Transformation linear observers.

Figure 9.4. Simulation of Example 9.1. The impact of reducing " on the steady-state
error x̃2 for the Forward Difference linear observer.

Figure 9.5. Simulation of Example 9.1. Comparison of the transient and steady-state
estimation error x̃2 between the Forward Difference and Bilinear Transformation linear observers
when α = 0.5 and " = 0.01.
286 CHAPTER 9. DIGITAL IMPLEMENTATION

where u ∈ R is the control input, y ∈ R and z ∈ R s are measured outputs, and w ∈ R`


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

and x ∈ Rρ constitute the state vector. The matrices A, B, C are defined after equation
(9.3). The functions f0 , φ, and ψ are locally Lipschitz in their arguments for (w, x, u) ∈
Dw × D x × R, where Dw ⊂ R` and D x ⊂ Rρ are domains that contain their respective
origins. Moreover, f0 (0, 0, 0) = 0, φ(0, 0, 0) = 0, and ψ(0, 0) = 0.
The goal is to design a discrete-time output feedback controller to stabilize the
origin (w = 0, x = 0) using only the measured outputs y and z at the sampling points.
It is assumed that the digital control system uses zero-order hold (ZOH); that is, the
control signal is held constant in between the sampling points. We start by designing
a continuous-time partial state feedback controller that uses measurements of x and
z. A high-gain observer is designed to estimate x from y. Finally, the observer and
control signal are discretized to arrive at the discrete-time controller.
The continuous-time state feedback controller takes the form

u = γ (x, z), (9.20)

where γ is locally Lipschitz in (x, z) over the domain of interest and globally bounded
in x, with γ (0, 0) = 0. It is designed such that the origin χ = 0 of the closed-loop
system
χ̇ = f r (χ , γ (x, z)) (9.21)
is asymptotically stable, where

w f0 (w, x, u)
• ˜ • ˜
χ= , f r (χ , u) = .
x Ax + Bφ(w, x, u)

The continuous-time output feedback controller is given by

u = γ (x̂, z), (9.22)

where x̂ is generated by the high-gain observer

x̂˙ = Ax̂ + Bφ0 (x̂, z, u) + H (y − C x̂), (9.23)

where H is given by (9.5) and φ0 (x, z, u) is a nominal model of φ(w, x, u), which is
locally Lipschitz in its arguments over the domain of interest and globally bounded in
x, with φ0 (0, 0, 0) = 0.
Assuming that the sampling period T = α", the observer is discretized as described
in the previous section. In particular, when φ0 6= 0, the observer is discretized using
the Forward Difference method to obtain the discrete-time observer

ξ (k + 1) = Ad ξ (k) + +Bd y(k) + α"ρ Bφ0 (D −1 ξ (k), z(k), u(k)), (9.24)


−1
x̂(k) = D Cd ξ (k), (9.25)

while when φ0 = 0, a linear discrete-time observer is given by

ξ (k + 1) = Ad ξ (k) + +Bd y(k), (9.26)


−1
x̂(k) = D [Cd ξ (k) + Dd y(k)], (9.27)

where the matrices Ad , Bd , Cd , Dd and the state ξ are given in Table 9.1 for the three
discretization methods. In the case of the Forward Difference method, it is assumed
9.2. STABILIZATION 287

that α is chosen small enough such that the eigenvalues of Ad are in the interior of the
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

unit circle. The discrete-time output feedback controller is given by

u(k) = γ (x̂(k), z(k)). (9.28)

To analyze the closed-loop system, we derive a discrete time model that describes
the state variables at the sampling points. A key step in deriving the model is the
representation of the system in a two-time-scale form that reflects the fact that the
observer dynamics are faster than the plant dynamics. The state χ of the plant satisfies
the equation
χ̇ = f r (χ , u(k)) (9.29)
over the sampling period [kT , kT + T ], during which the control is kept constant at
u(k). The solution of (9.29) is given by
Z t
χ (t ) = χ (k) + (t − kT ) f r (χ (k), u(k)) + [ f r (χ (τ), u(k)) − f r (χ (k), u(k))] d τ
kT

for t ∈ [kT , kT + T ]. Using the Lipschitz property of f r and the Gronwall–Bellman


inequality, it can be shown that

1 ” (t −kT )L1 —
kχ (t ) − χ (k)k ≤ e − 1 k f r (χ (k), u(k))k ∀ t ∈ [kT , kT + T ], (9.30)
L1

where L1 is a Lipschitz constant of f r with respect to χ . Therefore,

χ (k + 1) = χ (k) + "α f r (χ (k), u(k)) + "2 Φ(χ (k), u(k), "), (9.31)

where Φ is locally Lipschitz in (χ , u), uniformly bounded in ", for sufficiently small
", and Φ(0, 0, ") = 0. This equation and inequality (9.30) are sufficient to characterize
the plant dynamics. However, to develop a model that describes the dynamics of the
estimation error, we need a more detailed model of x that makes use of the special
structure of (9.17) and the properties of the matrices A and B. The solution of (9.17)
is given by
Z t Z t
A(t −kT ) A(t −τ)
x(t ) = e x(k) + e B d τ φ(w(k), x(k), u(k)) + e A(t −τ) B∆1 (τ) d τ,
kT kT

where ∆1 (τ) = φ(w(τ), x(τ), u(k)) − φ(w(k), x(k), u(k)). Using the fact that the ma-
trices A and B satisfy the relations

Aρ = 0, "i DAi = Ai D, DB = "ρ−1 B, (9.32)

it can be shown that


ρ−1 i ρ−1
t ti i
A DB = "ρ−1 e A(t /") B.
X X
D e At B = D Ai B = (9.33)
i =0
i! i=0
"i i !

Hence,
Z kT +T Z kT +T
A(kT +T −τ) ρ−1
e B∆1 (τ) d τ = " D −1
e A(kT +T −τ)/" B∆1 (τ) d τ. (9.34)
kT kT
288 CHAPTER 9. DIGITAL IMPLEMENTATION

Using the Lipschitz property of φ, (9.30) and the fact that


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Z kT +T
1 L1 T
e (τ−kT )L1 − 1 d τ =
€ Š
− 1 − T = O(T 2 ) = O("2 ),

e
kT L1

it can be shown that the integral on the right-hand side of (9.34) is O("2 ). Therefore,
x(k) can be modeled by
Z T
x(k + 1) = e AT
x(k) + e Aτ B d τ φ(w(k), x(k), u(k)) + "ρ+1 D −1 R(χ (k), u(k), "),
0
(9.35)

where R is locally Lipschitz in (χ , u), uniformly bounded in ", for sufficiently small
", and R(0, 0, ") = 0.
As in the continuous-time case, define the scaled estimation errors
xi − x̂i
ϕi = for 1 ≤ i ≤ ρ,
"ρ−i
which can be written as
1 1
ϕ= D(x − x̂) = (D x − Cd ξ − Dd y), (9.36)
"ρ−1 "ρ−1
where ϕ = col(ϕ1 , . . . , ϕρ ). Using the relation C = C D, it can be shown that ϕ satisfies
the equation
1
ϕ(k + 1) = Ad ϕ(k) + M D x(k) + "h(χ (k), u(k), ")
"ρ−1
− "αCd Bφ0 (x̂(k), z(k), u(k)), (9.37)

where
M = (I − Dd C )e Aα − Ad (I − Dd C ) − Cd Bd C
and Z α 
h = (I − Dd C ) e Aσ B d σφ + "R .
0

For the linear observer φ0 = 0. In arriving at (9.37) we used the property Cd Ad Cd−1 =
Ad , which can be verified for the three discretization methods using Table 9.1. The
term (1/"ρ−1 )M D x(k) in (9.37) is eliminated by the change of variables
1
η=ϕ− LD x, (9.38)
"ρ−1
where L is the solution of the Sylvester equation

Ad L + M − Le Aα = 0. (9.39)

Because the eigenvalues of Ad are in the interior of the unit circle while those of e Aα
are at one, (9.39) has a unique solution [156, Problem 9.6]. The new variable η satisfies
the equation

η(k + 1) = Ad η(k) + "(h − αCd Bφ0 ) − "L e Aσ d σ Bφ − "2 LR,
0
9.2. STABILIZATION 289

Table 9.2. D −1 J D and D −1 J M D in terms of X = −(1/α)D −1 F −1 D.


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

D −1 J D D −1 J M D

Forward Difference X X (e AT − I − T A)

Backward Difference I +X X [(I − T A)e AT − I ]

1 T T
”€ Š —
Bilinear Transformation 2
I +X X I − 2 A e AT − I − 2 A

and the state estimate x̂ is given by

x̂(k) = (I − D −1 LD)x(k) + Q1 η(k),

where Q1 (") = −"ρ−1 D −1 is an analytic function of ".

Lemma 9.1. D −1 LD = "Q2 ("), where Q2 (") is an analytic function of ".

Proof: Using the property Aρ = 0, it can be verified that


ρ
X
L= J i M (I − e Aα )i −1 , (9.40)
i =1

where J = (I − Ad )−1 . In particular,


ρ
X ρ
X
(I − Ad )L − L(I − e Aα ) = J i −1 M (I − e Aα )i −1 − J i M (I − e Aα )i
i =1 i=1
ρ−1
X ρ
X
= J k M (I − e Aα )k − J i M (I − e Aα )i
k=0 i =1
ρ Aα ρ
= M − J M (I − e ) = M

since (I − e Aα )ρ = 0. Using (9.40), D −1 LD can be written as


ρ
X
D −1 LD = (D −1 J D)i−1 (D −1 J M D)[D −1 (I − e Aα )D]i −1 .
i =1

Table 9.2 gives expressions of D −1 J D and D −1 J M D for the three discretization meth-
ods in terms of X = −(1/α)D −1 F −1 D. It can be shown that

D −1 (I − e Aα )D = I − e AT = O(T ) = O("),
e AT − I − T A = O(T 2 ) = O("2 ),
(I − T A)e AT − I = O(T 2 ) = O("2 ),
T T
 ‹
I − A e AT − I − A = O(T 2 ) = O("2 ),
2 2
290 CHAPTER 9. DIGITAL IMPLEMENTATION

and
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

−"ρ−1 /αρ
 
0 0 ··· 0
1/"
 0 ··· 0 −"ρ−2 α1 /αρ 

D −1 F −1 D = 
 1/" −"ρ−3 α2 /αρ 
 = O(1/").
 .. ..
 

 . . 
0 0 ··· 1/" −αρ−1 /αρ

Using Table 9.2 and the foregoing expressions we conclude that D −1 J D = O(1/"),
D −1 J M D = O("), and D −1 (I − e Aα )D = O("). Hence,
ρ 
K1 i −1
X ‹
kD −1 LDk ≤ K2 "(K3 ")i −1 = O("). 2
i=1
"

In summary, the closed-loop digital control system can be represented at the sam-
pling points by the discrete-time model

χ (k + 1) = χ (k) + "α f r (χ (k), u(k)) + "2 Φ(χ (k), u(k), "), (9.41)
η(k + 1) = Ad η(k) + "Γ (χ (k), u(k), x̂(k), "), (9.42)
u(k) = γ (x̂(k), z(k)), (9.43)
x̂(k) = [I − "Q2 (")]x(k) + Q1 (")η(k), (9.44)

where the functions f r , Φ, γ , and Γ are locally Lipschitz, Φ and Γ are globally bounded
in x̂, and the matrices Q1 and Q2 are analytic functions of ".

Theorem 9.1. Consider the closed-loop system of the plant (9.16)–(9.19) and the output
feedback digital controller (9.28) with the observer (9.24)–(9.25) or the observer (9.26)–
(9.27). Let R be the region of attraction of (9.21), let S be any compact set in the interior
of R, and let Q be any compact subset of Rρ . Suppose (χ (0), x̂(0)) ∈ S × Q. Then

• there exists "∗1 > 0 such that for every " ∈ (0, "∗1 ], χ (t ) is bounded for all t ≥ 0 and
η(k) is bounded for all k ≥ 0;

• given any µ > 0, there exist "∗2 > 0, T1 > 0, and k ∗ > 0, all dependent on µ, such
that for every " ∈ (0, "∗2 ],

kχ (t )k ≤ µ ∀ t ≥ T1 and kη(k)k ≤ µ ∀ k ≥ k ∗; (9.45)

• given any µ > 0, there exists "∗3 > 0, dependent on µ, such that for every " ∈ (0, "∗3 ],

kχ (t ) − χ r (t )k ≤ µ ∀ t ≥ 0, (9.46)

where χ r is the solution of (9.21) with χ r (0) = χ (0);

• if the origin of (9.21) is exponentially stable and f r (χ , γ (x, z)) is twice continuously
differentiable in the neighborhood of the origin, then there exists "∗4 > 0 such that
for every " ∈ (0, "∗4 ], the origin of (9.41)–(9.44) is exponentially stable and S × Q
is a subset of its region of attraction. Moreover, the continuous-time trajectory χ (t )
decays to zero exponentially fast. 3
9.2. STABILIZATION 291

Proof: Rewrite (9.41)–(9.42) as


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

χ (k + 1) = χ (k) + "F (χ (k), η(k), "), (9.47)


η(k + 1) = Ad η(k) + "G (χ (k), η(k), "), (9.48)

where

F (χ , η, ") = α f r (χ , u) + "Φ(χ , u, "), G (χ (k), η(k), ") = Γ (χ , u, x̂, "),

and u and x̂ are substituted using (9.43) and (9.44), respectively. The functions F and
G are locally Lipschitz, globally bounded in η, and vanish at the origin (χ = 0, η = 0).
Since the origin of (9.21) is asymptotically stable and R is its region of attraction, by
the converse Lyapunov theorem of asymptotic stability [78, Theorem 4.17], there is a
smooth, positive definite function V (χ ) and a continuous, positive definite function
U (χ ), both defined for all χ ∈ R, such that

V (χ ) → ∞ as χ → ∂ R,
∂V
f (χ , γ (x, z)) ≤ −U (χ ) ∀ χ ∈ R
∂χ r

and for any c > 0, {V (χ ) ≤ c} is a compact subset of R. Choose positive constants b


and c such that c > b > maxχ ∈S V (χ ). Then

S ⊂ Ω b = {V (χ ) ≤ b } ⊂ Ωc = {V (χ ) ≤ c} ⊂ R.

We use V (χ ) as a Lyapunov function for the slow subsystem (9.47) and W (η) = ηT P η
as a Lyapunov function for the fast subsystem (9.48), where P is the positive definite
solution of the discrete-time Lyapunov equation ATd PAd − P = −I . It can be seen that
P= ∞ (Akd )T (Akd ) > I . Therefore, kP k = λmax (P ) > λmin (P ) > 1 and
P
k=0

λmin (P )kηk2 ≤ W (η) ≤ λmax (P )kηk2 .

Let Λ = Ωc × {W (η) ≤ c1 "2 }, where c1 > 0 will be chosen. For every " ∈ (0, "˜], where
"˜ < 1, there are positive constants K1 , K2 , K3 , and L1 , independent of ", such that

kF (χ , η, ")k ≤ K1 , kΦ(χ , γ (x, z), ")k ≤ K2 , kG (χ , η, ")k ≤ K3

for all (χ , η) ∈ Ω × Rρ and

kF (χ , η, ") − F (χ , 0, ")k ≤ L1 kηk

for all (χ , η) ∈ Λ.
We start by showing that there exist positive constants c1 > 0 and "1 > 0 (depen-
dent on c1 ) such that the compact set Λ is positively invariant for " ∈ (0, "1 ]. This
is done by showing that for η ∈ {W (η) ≤ c1 "2 }, Ωc is positively invariant, and for
χ ∈ Ωc , {W (η) ≤ c1 "2 } is positively invariant:
292 CHAPTER 9. DIGITAL IMPLEMENTATION

∂V
∆V (χ (k)) = V (χ (k + 1)) − V (χ (k)) = (χ̄ (k))[χ (k + 1) − χ (k)]
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

∂χ
∂V
= (χ (k))[χ (k + 1) − χ (k)]
∂χ
∂V ∂V
 
+ (χ̄ (k)) − (χ (k)) [χ (k + 1) − χ (k)]
∂χ ∂χ
∂V
≤ (χ (k))"F (χ (k), η(k), ") + L2 kχ (k + 1) − χ (k)k2
∂χ

∂ V
≤ −"αU (χ (k)) + " (χ (k)) kF (χ (k), η(k), ") − F (χ (k), 0, ")k

∂ χ

∂ V
+ "2 (χ (k)) kΦ(χ (k), γ (x(k), z(k)), ")k + L2 kχ (k + 1) − χ (k)k2

∂ χ
≤ −"αU (χ (k)) + "L1 L3 kη(k)k + "2 L3 K2 + "2 L2 K12 ,

where χ̄ (k) is a point on the line segment joining χ (k) to χ (k + 1), and L2 and L3 are
Lipschitz constant and upper bound for [∂ V /∂ χ ] over Ωc . In Λ,
v
u c
2 2 1
λmin (P )kηk ≤ W (η) ≤ c1 " ⇒ kηk ≤ "t .
λmin (P )

Thus,
∆V (χ (k)) ≤ "α[−U (χ (k)) + "K4 ], (9.49)
where K4 = [L1 L3 c1 /λmin (P ) + L2 K12 + L3 K2 ]/α. For any positive constant c̃ < c,
p

trajectories starting inside {V (χ ) ≤ c̃} cannot leave {V (χ ) ≤ c}, provided " is suffi-
ciently small. This is seen from

V (χ (k + 1)) ≤ V (χ (k)) − "αU (χ (k)) + "2 αK4 ≤ c̃ + "2 αK4 < c


def
whenever "2 < "22 = (c − c̃)/(αK4 ). Trajectories starting inside {V (χ ) ≤ c} but outside
{V (χ ) ≤ c̃} remain inside {V (χ ) ≤ c}, provided " is sufficiently small. This is seen
from

V (χ (k + 1)) ≤ V (χ (k)) − "αU (χ (k)) + "2 αK4 ≤ c − "α%1 + "2 αK4 < c
def
whenever " < "3 = %1 /K4 , where %1 = minc̃≤V (χ )≤c U (χ ). Thus, for "1 = min{"2 , "3 },
the set {V (χ ) ≤ c} is positively invariant for " < "1 . On the other hand,

∆W (η(k)) = W (η(k + 1)) − W (η(k)) = ηT (k + 1)P η(k + 1) − ηT (k)P η(k)


≤ ηT (k)[ATd PAd − P ]η(k) + 2"kη(k)k kAd k kP k kG (χ (k), η(k), ")k
+ "2 kG (χ (k), η(k), ")k2 kP k
≤ −kη(k)k2 + "(2K3 kη(k)k kAd k kP k + "K32 kP k).

Hence,
 
1
W (η(k + 1)) ≤ 1 − W (η(k)) + "K3 kP k(2kη(k)k kAd k + "K3 ).
kP k
9.2. STABILIZATION 293

For η(k) ∈ {W (η) ≤ c1 "2 },


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

  v
1 u c
2 2 1
W (η(k + 1)) ≤ 1 − c1 " + 2" K3 kP k kAd kt + "2 K32 kP k.
kP k λmin (P )

Choosing c1 large enough ensures that W (η(k +1)) ≤ "2 c1 , which shows that {W (η) ≤
c1 "2 } is positively invariant. This completes the proof that Λ is positively invariant.
Now consider the initial state (χ (0), x̂(0)) ∈ S × Q. It can be verified that the
corresponding initial state η(0) satisfies kη(0)k ≤ l /"ρ−1 , where l depends on S and
Q. Using the fact that χ (0) is in the interior of Ωc , it can be shown from (9.47) that

Xk−1
kχ (k) − χ (0)k = " F (χ (i), η(i), ") ≤ "kK1 (9.50)

i =0

as long as χ (k) ∈ Ωc . Therefore, there exists a positive constant K5 , independent of


", such that χ (k) ∈ Ωc for all k ≤ K5 /". Let k̃0 be the first time η(k) enters the set
{W (η) ≤ c1 "2 } and take k̄0 ≤ min{k̃0 − 1, K5 /"}. During the time interval [0, k̄0 ],
W (η(k)) > c1 "2 and χ (k) ∈ Ωc . Therefore,
 v !
u W (η(k))
1
W (η(k + 1)) ≤ − W (η(k)) + "K3 kP k 2kAd k + "K3 
t
2kP k λmin (P )
 
1
+ 1− W (η(k)).
2kP k

Choosing c1 large enough ensures that the bracketed term in the foregoing inequality
is nonpositive. Hence,

l2
W (η(k + 1)) ≤ λW (η(k)) ⇒ W (η(k)) ≤ λk W (η(0)) ⇒ W (η(k)) ≤ λk kP k ,
"2ρ−2
where λ = 1 − 1/(2kP k) ∈ (0, 1). Therefore, W (η(k)) is strictly decreasing toward
zero, which shows that η(k) must enter the set {W (η) ≤ c1 "2 } within the finite time
[0, ln(K6 /"2ρ )/ ln(1/λ)], where K6 = l 2 kP k/c1 . To ensure that η(k) enters {W (η) ≤
c1 "2 } before χ (k) leaves Ωc , "4 is chosen small enough so that for " ∈ (0, "4 ],

ln(K6 /"2ρ ) K5
< −1
ln(1/λ) "

or, equivalently,
1 K 1
 ‹  ‹  ‹
" ln + " ln 2ρ6 < K5 ln .
λ " λ
Such choice of "4 is possible because the left-hand side of the foregoing inequality
tends to zero as " tends to zero. Taking "∗1 = min{˜ ", "1 , "4 } guarantees that for ev-
ery " ∈ (0, "∗1 ], (χ (k), η(k)) enters Λ during the interval [0, ∆2 (")] and remains inside
thereafter, where
ln(K6 /"2ρ ) K5
∆2 (") ≤ 1 + < .
ln(1/λ) "
294 CHAPTER 9. DIGITAL IMPLEMENTATION

Thus, the trajectory is bounded for all k ≥ ∆2 ("). It is also bounded for k ∈ [0, ∆2 (")]
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

since during this interval

kχ (k) − χ (0)k ≤ "kK1 and W (η(k)) ≤ λk kP k(l 2 /"2ρ−2 ).

Next we show (9.45). We already know that (χ (k), η(k)) ∈ Λ for all k ≥ k̃ for some
k̃. We can find "5 ∈ (0, "∗1 ], dependent on µ, such that for every " ∈ (0, "5 ], kη(k)k ≤ µ
for all k ≥ k̃. The ultimate bound on χ is proved in two steps. First, it is shown that
trajectories inside Λ enter the set

Λ̃ = {V (χ ) ≤ c2 (")} × {W (η) ≤ c1 "2 }

in finite time k ∗ , where lim"→0 c2 (") = 0. Second, it is shown that there is a positive
constant c3 (") > c2 ("), with lim"→0 c3 (") = 0, such that trajectories in Λ̃ at time k ∗
belong to the set
Λ̄ = {V (χ ) ≤ c3 (")} × {W (η) ≤ c1 "2 }
for all k ≥ k ∗ + 1. From (9.49), it is seen that for all (χ , η) ∈ Λ,
"α ”1 —
V (χ (k + 1)) ≤ V (χ (k)) − U (χ (k)) − "α 2 U (χ (k)) − "K4 . (9.51)
2
Since U (χ ) is positive definite and continuous, the set {U (χ ) ≤ 2"K4 } is compact for
sufficiently small ". Let c2 (") = maxU (χ )≤2"K4 V (χ ). The inequality V (χ (k)) > c2 (")
implies that U (χ (k)) > 2"K4 and

V (χ (k + 1)) ≤ V (χ (k)) − U (χ (k)).
2
Starting with V (χ (0)) > c2 ("), it can be shown by mathematical induction that, as
long as V (χ (k)) > c2 ("),
k
"α X
V (χ (k + 1)) ≤ V (χ (0)) − U (χ (i))
2 i=0
k
"α X
< V (χ (0)) − 2"K4 = V (χ (0)) − "2 αK4 (1 + k),
2 i=0

which shows that there is finite time k ∗ such that V (χ (k ∗ )) ≤ c2 ("). When V (χ (k)) ≤
c2 (") one of two cases could happen: U (χ (k)) > 2"K4 or U (χ (k)) ≤ 2"K4 . It is seen
from (9.51) that if U (χ (k)) > 2"K4 ,

V (χ (k + 1)) ≤ V (χ (k)) − U (χ (k)) ≤ c2 ("),
2
and if U (χ (k)) ≤ 2"K4
def
V (χ (k + 1)) ≤ V (χ (k)) + "2 αK4 ≤ c2 (") + "2 αK4 = c3 (").

In the latter case, one of two things could happen: either U (χ (k + 1)) > 2"K4 or
U (χ (k + 1)) ≤ 2"K4 . If U (χ (k + 1)) ≤ 2"K4 , then V (χ (k + 1)) ≤ c2 ("). If U (χ (k +
1)) > 2"K4 , then V (χ (k + 2)) < V (χ (k + 1)), and if V (χ (k + 1)) > c2 ("), V (χ ) will
decrease until eventually V (χ (k)) ≤ c2 (") in a finite number of steps. Therefore,

V (χ (k + 1)) ≤ c3 (") ∀ k ≥ k ∗.
9.2. STABILIZATION 295

From (9.30) and the fact that lim"→0 c3 (") = 0, we conclude that there is a finite time
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

T1 such that, for sufficiently small ", kχ (t )k ≤ µ for all t ≥ T1 .


To show (9.46), note that from (9.50) and asymptotic stability of the origin of
(9.21), there exists T2 = T2 (µ) > 0 such that

kχ (t ) − χ r (t )k ≤ µ ∀ t ≥ T2 . (9.52)

From (9.30) and (9.52) it is sufficient to prove that kχ (k)−χ r (k)k ≤ µ for k ∈ [0, K7 /"]
for some K7 (independent of "). Let e(k) = χ (k) − χ r (k) and divide the interval
[0, K7 /"] into [0, ∆2 (")] and [∆2 ("), K7 /"]. Over [0, ∆2 (")] we use (9.50), a similar
expression for χ r , and lim"→0 ∆2 (") = 0 to show that e(k) = O("∆2 (")) for k ∈
[0, ∆2 (")]. Over the interval [∆2 ("), K7 /"],

e(k + 1) = e(k) + "α[ f r (χ (k), γ (x(k), z(k))) − f r (χ r (k), γ (x r (k), z r (k)))] + O("2 ).

Hence,
ke(k + 1)k ≤ (1 + "αL4 )ke(k)k + O("2 ),
where L4 is a Lipschitz constant for f r over Ωc . Solving the inequality results in

k−1
ke(k)k ≤ (1 + "αL4 )k−∆2 (") O("∆2 (")) +
X
(1 + "αL4 )k−`−1 O("2 )
`=∆2 (")

– ™
k−∆2 (")
X 1 2
≤ (1 + "αL4 ) O("∆2 (")) + O(" )
r =0 (1 + "αL4 ) r +1
K7 /"
≤ (1 + "αL4 ) [O("∆2 (")) + O(")].

Using
lim "∆2 (") = 0 and lim(1 + "αL4 )K7 /" = e αL4 K7
"→0 "→0

completes the proof of (9.46).


Finally, if the origin of (9.21) is exponentially stable, by the converse Lyapunov the-
orem of exponential stability,79 there is a twice continuously differentiable Lyapunov
function V1 (χ ), defined over a ball {kχ k ≤ r } ⊂ R, and four positive constants δ1 to
δ4 such that

2 2 ∂ V1 2
∂ V
1
δ1 kχ k ≤ V1 (χ ) ≤ δ2 kχ k , f (χ , γ (x, z)) ≤ −δ3 kχ k , ≤ δ4 kχ k

∂χ r ∂χ

for all χ ∈ {kχ k ≤ r }. Using the local Lipschitz properties of F and G and the fact
that F (0, 0, ") = 0 and G (0, 0, ") = 0, it can be shown that the composite Lyapunov
function Vc = V1 + W satisfies

∆Vc ≤ −"Y T QY ,

where
(αδ3 − "β1 ) −(β2 + "β3 ) kχ (k)k
• ˜ • ˜
Q= , Y =
−(β2 + "β3 ) ((1/") − β4 − "β5 kη(k)k
79 See [78, Theorem 4.14]. The fact that V is twice continuously differentiable when f is twice contin-
1 r
uously differentiable can be seen from the proof of the theorem.
296 CHAPTER 9. DIGITAL IMPLEMENTATION

for some nonnegative constants βi . The matrix Q is positive definite for sufficiently
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

small ". Hence, the origin of (9.47)–(9.48) is exponentially stable, and there is a neigh-
borhood N of the origin (independent of ") such that all trajectories in N converge to
the origin. By choosing " small enough, Λ̄ = {V (χ ) ≤ c3 (")} × {W (η) ≤ c1 "2 } ⊂ N .
Thus, all trajectories starting in S × Q enter N in finite time and converge to zero as
k tends to infinity. The fact that χ (t ) decays to zero exponentially fast follows from
(9.30). 2

Example 9.2. Reconsider the pendulum equation


ẋ1 = x2 , ẋ2 = sin x1 − c1 x2 + c2 u, y = x1
from Example 3.4, where 0 ≤ c1 ≤ 0.2 and 0.5 ≤ c2 ≤ 2. It is shown there that the
output feedback controller
x̂˙1 = x̂2 + (2/")(y − x̂1 ),
x̂˙2 = φ0 (x̂, u) + (1/"2 )(y − x̂1 ),
x̂ + x̂2
       
|x̂1 | |x̂2 |
u = −2 1.25π sat + 2.25π sat + 1 sat 1
1.25π 2.25π 0.1
stabilizes the origin (x = 0, x̂ = 0), for sufficiently small ", and the set Ω = {|x1 | ≤
1.25π} × {|x1 + x2 | ≤ π} is included in the region of attraction. Two choices of φ0
are considered in Example 3.4. The choice φ0 = 0 yields a linear observer, while
φ0 = sin x̂1 −0.1x̂2 +1.25u yields a nonlinear observer. The continuous-time observer
is discretized using the Forward Difference method to obtain
ξ (k + 1) = Ad ξ (k) + Bd y(k) + α"2 Bφ0 (x̂(k), u(k)), x̂(k) = D −1 ξ (k),
where the matrices Ad , Bd , and B are as defined in Example 9.1. Simulation results
are shown in Figure 9.6 and 9.7 for the initial conditions x1 (0) = −π, x2 (0) = ξ1 (0) =
ξ2 (0) = 0. Figure 9.6 shows the state trajectories under continuous-time state feed-
back and discrete-time output feedback for the linear and nonlinear observers. The
parameter " is taken as 0.05 and 0.01, with α = 0.1 in both cases; hence, T = 0.005 and
0.001, respectively. The figure shows a lot of similarity with Figure 3.8 for continuous-
time output feedback. In particular, as " decreases, the trajectories under discrete-time
output feedback approach the trajectories under continuous-time state feedback, as
stated in Theorem 9.1. The comparison of the linear and nonlinear observers shows
an advantage for the nonlinear one when " = 0.05, which is more prominent in the
trajectory of x2 . The choice of α plays an important role because it determines the
relation between the sampling period T and the observer time constant ". It is rea-
sonable to have T < ", and that is why α = 0.1 is chosen. However, sometimes we
might not be able to push T very small due to hardware constraints. In such cases we
may have to work with α close to one. Figure 9.7 shows the trajectories of x1 and x2
and the estimation errors x̃1 and x̃2 for α = 1, at which T = " = 0.01, and compares
them with the trajectories and estimation errors for α = 0.1, at which T = 0.001 and
" = 0.01. The larger α causes a larger peak in the estimation errors, but they quickly
recover the errors for the smaller α. The state trajectories, on the other hand, are al-
most unaffected by the change in the sampling period. This is likely because the larger
sampling period is still adequate for controlling the system. The fact that the control
and estimation problems may require different sampling periods is explored further
in the Section 9.4, where a multirate digital control scheme is developed. 4
9.2. STABILIZATION 297

(a) (b)
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

3.5
0
3
−0.5
2.5
−1
2

x1

x2
−1.5
1.5
−2
SF 1
−2.5 OF ε = 0.05
OF ε = 0.01 0.5
−3
0
0 1 2 3 0 1 2 3
Time Time

(c) (d)
3
0
2.5
−0.5

−1 2

x2
x1

−1.5 1.5

−2 1
−2.5
0.5
−3
0
0 1 2 3 0 1 2 3
Time Time

Figure 9.6. Simulation of Example 9.2. Comparison of continuous-time state feedback


(SF) with discrete-time output feedback (OF). (a) and (b) are for the linear observer while (c) and (d)
are for the nonlinear one.

(a) (b)
4 350
α = 0.1
α=1 300
2 250
200
x̃1

x̃2

0 150
100
−2 50
0
−4 −50
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Time Time

(c) (d)
3
0
2.5
−0.5

−1 2
x2
x1

−1.5 1.5

−2 1
−2.5
0.5
−3
0
0 1 2 3 0 1 2 3
Time Time

Figure 9.7. Simulation of Example 9.2. The effect of changing α = T /" on the estimation
errors and trajectories under discrete-time output feedback with nonlinear observer.
298 CHAPTER 9. DIGITAL IMPLEMENTATION

9.3 Measurement Noise


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Reconsider the closed-loop system of the plant (9.16)–(9.19) and the output feedback
digital controller (9.28) with the observer (9.24)–(9.25) or the observer (9.26)–(9.27)
from the previous section. Suppose the measured output y is corrupted by noise v,
that is, y = C x + v, where |v(k)| ≤ N . To arrive at the representation (9.41)–(9.44) of
the closed-loop system, the transformations (9.36) and (9.38) were used. These trans-
formations can be combined into the change of variables

1
η= [(I − L − Dd C )D x − Cd ξ ].
"ρ−1
To include the effect of measurement noise, consider the system

θ(k + 1) = Ad θ(k) + Bd v(k), θ(0) = 0, (9.53)


w(k) = Cd θ(k) + Dd v(k) (9.54)

and apply the change of variables

1
η= [(I − L − Dd C )D x − Cd ξ + Cd θ].
"ρ−1
By repeating the derivations of Section 9.2, it can be shown that the closed-loop digital
control system can be represented at the sampling points by the discrete-time model

χ (k + 1) = χ (k) + "α f r (χ (k), u(k)) + "2 Φ(χ (k), u(k), "), (9.55)
η(k + 1) = Ad η(k) + "Γ (χ (k), u(k), x̂(k), "), (9.56)
u(k) = γ (x̂(k), z(k)), (9.57)
−1
x̂(k) = [I − "Q2 (")]x(k) + Q1 (")η(k) + D w(k). (9.58)

From (9.58) it is seen that

x(k) − x̂(k) = "Q2 (")x(k) − Q1 (")ξ (k) − D −1 w(k).

Over compact sets of χ , Γ is be bounded; hence, η(k) is ultimately bounded by an O(")


bound. The same is true for the term "Q2 (")x(k). On the other hand, if |v(k)| ≤ N for
all k ≥ 0, then using the fact that the eigenvalues of Ad are in the interior of the unit
circle, we see from (9.53)–(9.54) that w(k) is O(N ); hence, D −1 w(k) is O(N /"ρ−1 ).
Thus, we can establish an ultimate bound on kx(k)− x̂(k)k of the form c1 "+c2 N /"ρ−1
for some positive constants c1 and c2 , independent of " and N . Properties of this
bounded are established in Lemma 8.1.

Theorem 9.2. Consider the closed-loop system of the plant (9.16)–(9.19) and the output
feedback digital controller (9.28) with the observer (9.24)–(9.25) or the observer (9.26)–
(9.27). Suppose y(k) = C x(k) + v(k), where |v(k)| ≤ N for all k ≥ 0. Let R be the
region of attraction of (9.21), let S be any compact set in the interior of R, and let Q be
any compact subset of Rρ . Suppose (χ (0), x̂(0)) ∈ S × Q. Then

• there exist positive constants ca and N ∗ such that for each N < N ∗ there is a con-
stant "a = "a (N ) > ca N 1/ρ , with limN →0 "a (N ) = "∗a > 0, such that for each
" ∈ ca N 1/ρ , "a the solutions (χ (k), η(k)) of the closed-loop system are bounded
for all k ≥ 0, and χ (t ) is bounded for all t ≥ 0;
9.3. MEASUREMENT NOISE 299

• there exist N1∗ > 0 and a class K function %1 such that for every N < N1∗ and every
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

µ1 > %1 (N ), there are constants T1 = T1 (µ1 ) ≥ 0 and " b = " b (N , µ1 ) > ca N 1/ρ ,
with limN →0 " b (N , µ1 ) = "∗b (µ1 ) > 0, and an integer k ∗ = k ∗ (µ1 ) such that for
each " ∈ ca N 1/ρ , " b we have


kx(k) − x̂(k)k ≤ µ1 ∀ k ≥ k ∗, (9.59)


kχ (t )k ≤ µ1 ∀ t ≥ T1 ; (9.60)

• there exist N2∗ > 0 and a class K function %2 such that for every N < N2∗ and every
µ2 > %2 (N ), there is a constant "c = "c (N , µ2) > ca N 1/ρ , with limN →0 "c (N , µ2 ) =
"∗c (µ2 ) > 0, such that for each " ∈ ca N 1/ρ , "c we have

kχ (t ) − χ r (t )k ≤ µ2 ∀ t ≥ 0, (9.61)

where χ r (t ) is the solution of (9.21) with χ s (0) = χ (0).


3

The proof of Theorem 9.2 is a merging of the proofs of Theorems 8.4 and 9.1.80
Remark 8.2 on Theorem 8.4 applies almost verbatim to Theorem 9.2.

Example 9.3. Consider the pendulum equation of Examples 3.4 and 9.2. A digital con-
troller is derived in Example 9.2 with Forward Difference discrete-time implementa-
tion of two high-gain observers: a linear observer and a nonlinear one. In this example
we consider the linear observer, and the results are representative of the nonlinear one.
Since all the equations are given in Example 9.2, only simulation results are presented
here. Simulation is carried out with α = 1 and T = " = 0.01. The initial conditions
are x1 (0) = −π, x2 (0) = ξ1 (0) = ξ2 (0) = 0. The measurement noise is generated by
the Simulink block “Uniform Random Number” with amplitude between ±0.01 and
sample time 0.001. The noise waveform is shown in Figure 9.8. Figure 9.9 compares
the state trajectories of the system under digital control with and without measure-
ment noise. The trajectories are fairly close, but the effect of measurement noise is
seen in some jittering in x2 . Figure 9.10 focuses on that jittering part. It is typical in
digital control systems to use an antialiasing filter to reduce the effect of aliasing due
to signal sampling.81 It is an analog low-pass filter with cutoff frequency equal to half
the sampling frequency. An analog signal is passed through the filter before sampling.
In view of our earlier discussion on low-pass filters in Chapter 8, it is reasonable to
expect the antialiasing filter to reduce the effect of measurement noise. This is shown
in Figures 9.11 and 9.12, which compare the state trajectories with and without filter.
The antialiasing filter is a Butterworth one of order 8 and cutoff frequency 50 Hz (half
the sampling frequency of 100 Hz). The state trajectories are fairly close to each other,
although the filter increases the settling time a little bit, which is not surprising in view
of its phase lag. Focusing on the steady state of x2 in Figure 9.12 shows reduction in
the error due to noise. 4

80 See [79] for the details of the proof.


81 See [16], [45], or any digital control textbook.
300 CHAPTER 9. DIGITAL IMPLEMENTATION
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

0.01

0.005

Noise
0

−0.005

−0.01
0 0.1 0.2 0.3 0.4 0.5
Time

Figure 9.8. Measurement noise of Example 9.3.

No Noise Noise

0 0

−0.5 −0.5

−1 −1
x1

x1
−1.5 −1.5

−2 −2

−2.5 −2.5

−3 −3
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

3 3

2.5 2.5

2 2

1.5 1.5
x2

x2

1 1

0.5 0.5

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

Figure 9.9. Comparison of the trajectories of Example 9.3 under digital control with and
without measurement noise.

No Noise Noise
0.1 0.1

0.05 0.05
x2

x2

0 0

−0.05 −0.05

−0.1 −0.1
9 9.2 9.4 9.6 9.8 10 9 9.2 9.4 9.6 9.8 10
Time Time

Figure 9.10. Comparison of the x2 trajectories of Example 9.3 under digital control with
and without measurement noise.
9.4. MULTIRATE DIGITAL CONTROL 301

Without Antialiasing Filter With Antialiasing Filter


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

0 0

−0.5 −0.5

−1 −1

x1

x1
−1.5 −1.5

−2 −2

−2.5 −2.5

−3 −3
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

3 3

2.5 2.5

2 2

1.5 1.5
x2

x2
1 1

0.5 0.5

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

Figure 9.11. Comparison of the trajectories of Example 9.3 under digital control with
and without antialiasing filter.

Without Antialiasing Filter With Antialiasing Filter


0.08 0.08

0.06 0.06

0.04 0.04

0.02 0.02
x2

x2

0 0

−0.02 −0.02

−0.04 −0.04
9 9.2 9.4 9.6 9.8 10 9 9.2 9.4 9.6 9.8 10
Time Time

Figure 9.12. Comparison of the x2 trajectories of Example 9.3 under digital control with
and without antialiasing filter.

9.4 Multirate Digital Control


Reconsider the stabilization problem of Section 9.2 for the system
ẇ = f0 (w, x, u), (9.62)
ẋ = Ax + Bφ(w, x, u), (9.63)
y = C x, (9.64)
z = ψ(w, x), (9.65)
where u ∈ R is the control input, y ∈ R and z ∈ R s are measured outputs, and w ∈
R` and x ∈ Rρ constitute the state vector. The matrices A, B, C are defined after
equation (9.3). The functions f0 , φ, and ψ are locally Lipschitz in their arguments for
(w, x, u) ∈ Dw × D x × R, where Dw ⊂ R` and D x ⊂ Rρ are domains that contain their
302 CHAPTER 9. DIGITAL IMPLEMENTATION

respective origins. Moreover, f0 (0, 0, 0) = 0, φ(0, 0, 0) = 0, and ψ(0, 0) = 0. Suppose


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

the continuous-time state feedback controller

u = γ (x, z) (9.66)

stabilizes the origin χ = 0 of the closed-loop system

χ̇ = f r (χ , γ (x, z)), (9.67)

where
w f0 (w, x, u)
• ˜ • ˜
χ= , f r (χ , u) = ,
x Ax + Bφ(w, x, u)
and γ is locally Lipschitz in (x, z) and globally bounded in x, with γ (0, 0) = 0.
Assuming that with the sampling period T s , measurements of x and z are available
at the sampling points, the discrete-time state feedback controller is taken as

u(k) = γ (x(k), z(k)), (9.68)

where u(k) is constant over the period [kT s , (k + 1)T s ), x(k) = x(kT s ), and z(k) =
z(kT s ). As in Section 9.2, it can be shown that in between the sampling points,
1 ” (t −kTs )L1 —
kχ (t ) − χ (k)k ≤ e − 1 k f r (χ (k), u(k))k, (9.69)
L1

where L1 is a Lipschitz constant of f r with respect to χ . The state of the closed-loop


system at the sampling points satisfies the difference equation

χ (k + 1) = χ (k) + T s f r (χ (k), u(k)) + T s2 Φ s (χ (k), u(k), T s ), (9.70)

where
Z (k+1)T s
1
Φ s (χ (k), u(k), T s ) = 2 [ f r (χ (τ), u(k)) − f r (χ (k), u(k))] d τ
Ts kT s

is locally Lipschitz in (χ , u) and, using (9.69),


1  Ts L1
kΦ s (χ (k), u(k), T s )k ≤ k f r (χ (k), u(k))k.
 
e − 1 − T s
T s2
Since  
1 1 T s L1 1
s = L1

lim 2
e − 1 − T
T s →0 T s L1 2
there are positive constants K and T s∗ such that for all T s ≤ T s∗ , kΦ s (χ (k), u(k))k ≤
Kk f r (χ (k), u(k))k.
The discretized high-gain observer is implemented with the fast sampling period
T f = α", where h = T s /T f is a positive integer greater than one. The observer is
discretized as in Section 9.1 using the Forward Difference method to obtain

ξ (n + 1) = Ad ξ (n) + +Bd y(n) + α"ρ Bφ0 (D −1 ξ (n), z(k), u(k)), (9.71)


−1
x̂(n) = D ξ (n), (9.72)

where Ad = A+ αF , Bd = αE, and the matrices D, E, and F are defined after equation
(9.6). It is assumed that α is chosen such that the eigenvalues of Ad satisfy |λ| < 1. The
9.4. MULTIRATE DIGITAL CONTROL 303
1
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

y(t)
u(kTs )
- ZOH - Plant z(t)

Ts
z(kTs )

Ts
Controller x̂(kTs ) x̂(nTf ) y(nTf )
 Observer 
h Tf
Down Sampling

Figure 9.13. Block diagram of the multirate digital control scheme.

index n is used to indicate the sampling points that are equally spaced with period T f .
The index k indicates the sampling points that are equally spaced with period T s . For
each value of k, n takes h values from n = k h to n = k h + (h − 1). The observer
estimates are down-sampled for use in the output feedback controller. If x̂(n) is the
estimate calculated by the observer for n ∈ [k h, k h + h − 1], then the down-sampled
estimate x̂ s is given by x̂ s (k) = x̂(h k) and is kept constant over the slow sampling
period T s . For clarity, x f (n) and χ f (n) are used to denote the samples of x(t ) and
χ (t ) in the fast sampling period T f . They are related to x(k) and χ (k) by the relations
x(k) = x f (h k) and χ (k) = χ f (h k). The output feedback controller is given by

u(k) = γ (x̂ s (k), z(k)), (9.73)


where z(k) are the samples of z(t ) in the slow sampling period T s . A block diagram
of the multirate scheme is shown in Figure 9.13.
By repeating the derivation of Section 9.2, it can be shown that
x̂(n) = [I − "Q2 (")]x f (n) + Q1 (")η(n), (9.74)
where Q1 and Q2 are analytic functions of " while η, which is defined by the changes
of variables (9.36) and (9.38), satisfies the equation
η(n + 1) = Ad η(n) + "Γ (χ f (n), u(k), x̂(n), "), (9.75)

where Γ is locally Lipschitz in (χ , u, x̂), globally bounded in x̂, and uniformly bounded
in " for sufficiently small ". The down-sampled estimate x̂ s (k) is given by
x̂ s (k) = x̂(h k) = [I − "Q2 (")]x f (h k) + Q1 (")η(h k) = [I − "Q2 (")]x(k) + Q1 (")η s (k),
where η s (k) = η(h k) is down-sampling of η(n). Thus, the closed-loop multirate digital
control system can be represented at the slow sampling points k by the discrete-time
model
χ (k + 1) = χ (k) + T s f r (χ (k), u(k)) + T s2 Φ s (χ (k), u(k), T s ), (9.76)
u(k) = γ (x̂ s (k), z(k)), (9.77)
x̂ s (k) = [I − "Q2 (")]x(k) + Q1 (")η s (k). (9.78)

Theorem 9.3. Consider the closed-loop system of the plant (9.62)–(9.65) and the output
feedback multirate digital controller (9.73) with the observer (9.71)–(9.72). Let R be the
304 CHAPTER 9. DIGITAL IMPLEMENTATION

region of attraction of (9.67), let S be any compact set in the interior of R, and let Q be
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

any compact subset of Rρ . Suppose (χ (0), x̂(0)) ∈ S × Q. Then

• there exists T s∗1 > 0 and for each T s ∈ (0, T s∗1 ], there is "∗1 > 0 such that for every
" ∈ (0, "∗1 ], χ (t ) is bounded for all t ≥ 0 and η(n) is bounded for all n ≥ 0;

• given any µ > 0 there exists T s∗2 > 0, and for each T s ∈ (0, T s∗2 ], there is "∗2 > 0,
T1 > 0, and n ∗ > 0, all dependent on µ, such that for every " ∈ (0, "∗2 ],

kχ (t )k ≤ µ ∀ t ≥ T1 and kη(n)k ≤ µ ∀ n ≥ n∗; (9.79)

• given any µ > 0 there exists T s∗3 > 0, and for each T s ∈ (0, T s∗3 ], there is "∗3 > 0,
dependent on µ, such that for every " ∈ (0, "∗3 ],

kχ (t ) − χ r (t )k ≤ µ ∀ t ≥ 0, (9.80)

where χ r is the solution of (9.67) with χ r (0) = χ (0);

• if the origin of (9.67) is exponentially stable and f r (χ , γ (x, z)) is twice continuously
differentiable in the neighborhood of the origin, then there exists T s∗4 > 0, and for
each T s ∈ (0, T s4∗ ] there is "∗4 > 0 such that for every " ∈ (0, "∗4 ] the discrete-time
trajectories χ (k) and η s (k) converge to zero exponentially fast as k → ∞ and the
continuous-time trajectory χ (t ) decays to zero exponentially fast as t → ∞. 3

Proof: Rewrite (9.76) and (9.75) as

χ (k + 1) = χ (k) + T s F (χ (k), η s (k), T s , "), (9.81)


η(n + 1) = Ad η(n) + "G (χ f (n), η(n), x(k), η s (k), z(k), "), (9.82)

where

F (χ , η s , T s , ") = f r (χ , γ (x̂ s , z)) + T s Φ s (χ , γ (x̂ s , z), T s )


= f r (χ , γ ((I − "Q2 )x + Q1 η s , z))
+ T s Φ s (χ , γ ((I − "Q2 )x + Q1 η s , z), T s )

and

G (χ f , η, x, η s , z, ") = Γ (χ f , γ (x̂ s , z), x̂, ")


= Γ (χ f , γ ((I − "Q2 )x + Q1 η s , z), (I − "Q2 )x f + Q1 η, ").

The functions F and G are locally Lipschitz, globally bounded in η s and η, and vanish
at the origin (χ = 0, η = 0). Since the origin of (9.67) is asymptotically stable and R is
its region of attraction, by the converse Lyapunov theorem of asymptotic stability [78,
Theorem 4.17], there is a smooth, positive definite function V (χ ) and a continuous,
positive definite function U (χ ), both defined for all χ ∈ R, such that

V (χ ) → ∞ as χ → ∂ R,

∂V
f (χ , γ (x, z)) ≤ −U (χ ) ∀ χ ∈ R
∂χ r
9.4. MULTIRATE DIGITAL CONTROL 305

and for any c > 0, {V (χ ) ≤ c} is a compact subset of R. Choose positive constants b


Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

and c such that c > b > maxχ ∈S V (χ ). Then

S ⊂ Ω b = {V (χ ) ≤ b } ⊂ Ωc = {V (χ ) ≤ c} ⊂ R.

As in the proof of Theorem 9.1, we use V (χ ) as a Lyapunov function for the slow
subsystem (9.81) and W (η) = ηT P η as a Lyapunov function for the fast subsystem
(9.82), where P is the positive definite solution of the discrete-time Lyapunov equation
ATd PAd − P = −I . Let Λ = Ωc × {W (η) ≤ c1 "2 }, where c1 > 0 will be chosen. For
every T s ∈ (0, T̃ s ] and " ∈ (0, "˜], where T̃ s and "˜ < 1 are positive constants, there are
positive constants K1 to K6 , independent of T s and ", such that

kF (χ , η s , T s , ")k ≤ K1 , kΦ s (χ , γ (x, z), T s )k ≤ K2 , kG (χ f , η, x, η s , z, ")k ≤ K3

for all (χ , η) ∈ Ω × Rρ and

kF (χ , η s , T s , ") − F (χ , 0, T s , 0)k ≤ K4 kη s k + K5 "kxk ≤ K6 "

for all (χ , η) ∈ Λ. It will now be shown that Λ is positively invariant. Since

F (χ , 0, T s , 0) = f r (χ , γ (x, z)) + T s Φ s (χ , γ (x, z), T s ),

equation (9.81) can be written as a perturbation of (9.70), that is,

χ (k + 1) = χ (k) + T s f r (χ , γ (x, z)) + T s2 Φ s (χ , γ (x, z), T s ) + T s ∆1 (·),

where k∆1 (·)k ≤ K6 " for all (χ , η) ∈ Λ. Using this equation in calculating ∆V (χ (k)) =
V (χ (k + 1)) − V (χ (k)) yields

∂V
∆V (χ (k)) = (χ̄ (k)) [χ (k + 1) − χ (k)]
∂χ
∂V
= (χ (k))[χ (k + 1) − χ (k)] (9.83)
∂χ
∂V ∂V
 
+ (χ̄ (k)) − (χ (k)) [χ (k + 1) − χ (k)]
∂χ ∂χ
∂V
(χ (k)) T s f r (χ , γ (x, z)) + T s2 Φ s (χ , γ (x, z), T s ) + T s ∆1 (·) (9.84)
 

∂χ
+ L2 kχ (k + 1) − χ (k)k2

∂ V
≤ −T s U (χ (k)) + (χ (k)) T s2 kΦ s (χ , γ (x, z), T s )k + T s k∆1 (·)k

∂ χ
+ L2 T s2 K12 (9.85)
≤ −T s U (χ (k)) + T s2 K7 + T s "K8 , (9.86)

where L2 , K7 , and K8 are positive constants independent of T s and ". Repeating argu-
ments from the proof of Theorem 9.1, the foregoing inequality can be used to show
that for all η ∈ {W (η) ≤ c1 "2 }, Ωc is positively invariant for sufficiently small T s and
". It can be also shown that by choosing c1 large enough, the set {W (η) ≤ c1 "2 } is
positively invariant for all χ ∈ Ωc . Hence, Λ = Ωc × {W (η) ≤ c1 "2 } is positively
invariant.
306 CHAPTER 9. DIGITAL IMPLEMENTATION

With the initial state (χ (0), x̂(0)) ∈ S × Q, we have kη(0)k ≤ l /"ρ−1 , where l
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

depends on S and Q. Since S is in the interior of Ωc , it can be shown from (9.81)


that
Xk−1
kχ (k) − χ (0)k = T s F (χ (k), η s (k), T s , ") ≤ T s kK1


i=0

as long as χ (k) ∈ Ωc . Therefore, there exists a positive constant K9 , independent of T s


and ", such that χ (k) ∈ Ωc for all k ≤ K9 /T s . With χ (k) ∈ Ωc and η(n) ∈ / {W (η) ≤
c1 "2 }, it can be shown that by choosing c1 large enough

l2
W (η(n)) ≤ λn W (η(0)) ⇒ W (η(n)) ≤ λn kP k ,
"2ρ−2

where λ ∈ (0, 1). Hence, η enters {W (η) ≤ c1 "2 } and stays therein for all
€ k pkl 2 Š
def
ln c "2ρ
n ≥ n̄(") = €1 1 Š .
ln λ

To ensure that η(n) enters {W (η) ≤ c1 "2 } before χ (k) leaves Ωc , " should be chosen
small enough that n̄(") < hK9 /T s . Since T f = T s /h = α", " should be chosen small
enough that n̄(") < K9 /(α"), which can be done since lim"→0 "n̄(") = 0.
The proof of the ultimate boundedness property (9.79) follows an argument similar
to the proof of Theorem 9.1. In particular, using (9.86) it can be shown that after finite
time the trajectories inside Λ will be in the set

Λ̄ = {V (χ ) ≤ c3 (T s + ")} × {W (η) ≤ c1 "2 },

where lim r →0 c3 (r ) = 0. Choosing T s and " small enough ensures that (9.79) is satis-
fied. The proof of (9.80) is the same as the corresponding argument in the proof of
Theorem 9.1 with the error e(k) = χ (k) − χ r (k) satisfying an equation of the form

e(k + 1) = e(k) + T s [ f r (χ (k), γ (x(k), z(k))) − f r (χ r (k), γ (x r (k), z r (k)))]


+ O(T s2 ) + O(T s ").

Finally, we prove the last bullet of the theorem. Since the origin of χ̇ =
f r (χ , γ (x, z)) is exponentially stable and f r (χ , γ (x, z)) is twice continuously differ-
entiable, by the converse Lyapunov theorem for exponential stability [78, Theorem
4.14], there is a twice continuously differentiable Lyapunov function V1 (χ ) that sat-
isfies the following inequality in some neighborhood of the origin χ = 0 for some
positive constants β1 to β4 :

2 2 ∂V 2
∂ V
β1 kχ k ≤ V1 (χ ) ≤ β2 kχ k , f (χ , γ (x, z)) ≤ −β3 kχ k , ≤ β4 kχ k.

∂χ r ∂ χ

Let B be ball centered at the origin (χ , η) = (0, 0) such that the foregoing inequalities
hold in B. The second bullet of the theorem guarantees that, for sufficiently small T s
and ", the trajectories will enter B in finite time. The forthcoming analysis is limited
to the ball B. We prove exponential stability of the origin of the discrete-time system
in the slow sample time k. This requires a description of η s (k) = η(h k), which is
9.4. MULTIRATE DIGITAL CONTROL 307

obtained by studying (χ f (n), η(n)) over one slow time period [h k, h k + 1]. In the fast
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

sample time n, the discrete-time system is represented by


χ f (n + 1) = χ f (n) + α" f r (χ f (n), u(k)) + α2 "2 Φ f (χ f (n), u(k), "), (9.87)
η(n + 1) = Ad η(n) + "G (χ f (n), η(n), x(k), η s (k), z(k), "). (9.88)

Even though |λ(Ad )| < 1, it is not clear that performing accumulation of the solu-
tion of (9.88) over the interval [h k, h k + 1] will add up to O("). This is due to the
appearance of η in the function G . We perform a change of variables to weaken the
dependence of G on η. The system (9.87)–(9.88) has two-time-scale behavior with χ f
as the slow variable and η as the fast one. Treating χ f as a constant input to the fast
equation (9.88), its quasi-steady-state solution η̄(n) satisfies the equation
η̄(n) = Ad η̄(n) + "G (χ f (n), η̄(n), x(k), η s (k), z(k), "),

where x(k), η s (k), and z(k) are constant during the interval [h k, h k + 1]. We seek a
solution to the foregoing equation in the form η̄(n) = "θ(χ f (n), x(k), η s (k), z(k), "),
where θ is a continuously differentiable function of its arguments and θ(0, 0, 0, 0, 0) =
0. Evaluating η̄(n) = "θ(χ f (n), x(k), η s (k), z(k), ") at n+1 and substituting for χ f (n+
1) and η̄(n + 1) using (9.87) and (9.88) result in
Ad θ(χ f (n), . . . , ") + G (χ f (n), "θ(·), . . . , ") = θ(χ f (n) + α" f r (·) + α2 "2 Φ f (·), . . . , ").
(9.89)
Setting " = 0 yields
(I − Ad )θ(χ f (n), x(k), η s (k), z(k), 0) = G (χ f (n), 0, x(k), η s (k), z(k), 0).

Since |λ(Ad )| < 1, (I − Ad ) is nonsingular. Hence, the foregoing equation has a unique
solution. By the implicit function theorem, there is a continuously differentiable
function θ(χ f (n), x(k), η s (k), z(k), ") that satisfies (9.89) for sufficiently small ". The
change of variable
ζ (n) = η(n) − "θ(χ f (n), x(k), η s (k), z(k), ") (9.90)

transforms (9.88) into


ζ (n + 1) = Ad ζ (n) + "G˜(χ f (n), ζ (n), x(k), η s (k), z(k), "), (9.91)

where G˜ is continuously differentiable and G˜(χ f (n), 0, x(k), η s (k), z(k), ") = 0. There-
fore, ζ = 0 is an equilibrium point of (9.91). In the ball B, G˜ can be represented as
[80, Equation (B.4)].
Z1 ˜
˜ ∂G
G (χ f , ζ , x, η s , z, ") = (χ f , σζ , x, η s , z, ") d σ ζ
0 ∂ζ
def
= B̃(χ f , ζ , x, η s , z, ") ζ ,

and (9.91) can be rewritten as the linear time-varying system


def
” —
ζ (n + 1) = Ad + "B̃(χ f (n), ζ (n), x(k), η s (k), z(k), ") ζ (n) = Ãd (n) ζ (n), (9.92)
308 CHAPTER 9. DIGITAL IMPLEMENTATION

which is valid because χ f (n) and η(n) are bounded and belong to B. Since |λ(Ad )| ≤
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

λ1 < 1 for some λ1 > 0, it follows that for sufficiently small " the transition matrix Φ̃
of Ãd satisfies kΦ̃(n, m)k ≤κ1 λ2n−m , where 0 < λ2 < 1 and κ1 > 0 [126, Theorem 24.7].
Because h = T s /T f = T s /(α"),

kΦ̃(h k + h, h k)k ≤κ1 λ2h = O(").


def
Therefore, ζ s (k) = ζ (h k) satisfies an equation of the form
ζ s (k + 1) = "G(χ (k), ζ s (k), ")ζ s (k),
where G is a continuous function of its arguments. Writing (9.90) at n = h k, we obtain
ζ s (k) = η s (k) − "θ(χ (k), x(k), η s (k), z(k), ").
At " = 0, η s (k) = ζ s (k). By the implicit function theorem, it can be concluded that,
for sufficiently small ", there is a continuously differentiable function ϑ such that
η s (k) = ζ s (k) + "ϑ(χ (k), ζ s (k), ").
Thus, the discrete-time system can be represented at the slow sample times k by the
equations
χ (k + 1) = χ (k) + T s f r (χ (k), γ (x(k), z(k)))
+ T s2 Φ s (χ (k), γ (x(k), z(k)), T s ) + T s ∆2 (·), (9.93)
ζ s (k + 1) = "G(χ (k), ζ s (k), ")ζ s (k), (9.94)
where k∆2 (·)k ≤ K10 (kζ s k + "kχ k) and kG(·)k ≤ K11 for some positive constants K10
and K11 independent of T s and ". Equation (9.93) is a perturbation of the system
χ (k + 1) = χ (k) + T s f r (χ (k), γ (x(k), z(k))) + T s2 Φ s (χ (k), γ (x(k), z(k)), T s ), (9.95)
which is the closed-loop system under state feedback. Using V1 (χ ) as a Lyapunov
function for this system, we obtain
∂ V1
V1 (χ (k + 1)) − V1 (χ (k)) ≤ (χ (k))[χ (k + 1) − χ (x)] + L2 kχ (k + 1) − χ (k)k2
∂χ
∂ V1
≤ Ts (χ (k)) f r (χ (k), γ (x(k), z(k)))
∂χ
∂ V1
+ T s2 (χ (k))Φ s (χ (k), γ (x(k), z(k)), T s )
∂χ
+ L2 T s2 k f r (·) + T s Φ s (·)k2
≤ −T s β3 kχ (k)k2 + T s2 Ka kχ (k)k2 ,
where Ka > 0 is independent of T s . Thus,

1 β3
V1 (χ (k + 1)) − V1 (χ (k)) ≤ − 2 T s β3 kχ (k)k2 for T s ≤ .
2Ka

Taking V2 (χ , ζ s ) = V1 (χ )+γ ζ sT ζ s , with γ > 0, as a Lyapunov function for the system


(9.93)–(9.94), it can be shown that for T s ≤ β3 /(2Ka ),

V2 (χ (k + 1), ζ s (k + 1)) − V2 (χ (k), ζ s (k)) ≤ −Y T QY ,


9.5. NOTES AND REFERENCES 309

where
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

1
χ (k)

T s β3 − "T s δ1 −T s δ2
• ˜
Y = , Q= 2 ,
ζ s (k) −T s δ2 γ (1 − "2 δ3 ) − δ4 T s2

and δ1 to δ4 are nonnegative constants independent of T s and ". Take " small enough
1
so that 2 T s β3 − "T s δ1 > 0 and 1 − "2 δ3 > 0; then choose γ large enough to make
Q positive definite. Hence, the origin (χ , ζ s ) = (0, 0) of (9.93)–(9.94) is exponen-
tially stable. Consequently, there are positive constants C1 and λ3 < 1 such that
k−k
kχ (k)k ≤ C1 λ3 0 kχ (k0 )k for k ≥ k0 . Set a = (1/T s ) ln(1/λ3 ) so that e −aTs = λ3 .
From inequality (9.69), there is a positive constant C2 such that kχ (t )k ≤ C2 kχ (kT s )k
for t ∈ [kT s , (k + 1)T s ). Hence,

kχ (t )k ≤ C2 C1 e −a(kTs −k0 Ts ) kχ (k0 T s )k ≤ C3 e −a(t −k0 Ts ) kχ (k0 T s )k

for some C3 > 0, which shows that χ (t ) decays to zero exponentially fast. 2

Example 9.4. Reconsider the pendulum equation

ẋ1 = x2 , ẋ2 = sin x1 − c1 x2 + c2 u, y = x1

from Example 3.4, where 0 ≤ c1 ≤ 0.2 and 0.5 ≤ c2 ≤ 2. A robust stabilizing state
feedback controller is given by
 
|x1 |
 
|x2 |
  x +x 
u = −2 1.25π sat + 2.25π sat + 1 sat 1 2
.
1.25π 2.25π 0.1

Implementing this controller using zero-order hold and sampling period T s = 0.05 re-
sults in adequate response, as shown in Figure 9.14. The simulation is carried out with
c1 = 0.01, c2 = 0.5, x1 (0) = −π, and x2 (0) = 0. Assuming that the output y = x1 is
the only measured signal, a high-gain observer is used to estimate the states. Suppose
we want, for some practical reason, to implement the output feedback controller such
that the control is updated with the same sampling period T s = 0.05. A single-rate out-
put feedback controller will have to use the same sampling period for the control and
state estimation. This corresponds to choosing " = 0.05 and α = 1 so that α" = 0.05.
Although we see in Example 3.4 that a continuous-time output feedback controller
with " = 0.05 gives adequate response, in discrete-time implementation the single-rate
output feedback controller fails to stabilize the system, as shown in Figure 9.15. The
observer used here is the same linear observer with Forward Difference discretization
that is used in Example 9.2, with x̂1 (0) = x̂2 (0) = 0. A multirate discrete-time out-
put feedback controller is implemented with T s = 0.05 and T f = 0.005. The choice
of T f corresponds to " = 0.05 and α = 0.1. Simulation is carried out using the same
type of observer and the same parameters and initial conditions. Figure 9.16 shows
the down-sampling of the estimates x̂1 and x̂2 , while Figure 9.17 shows the response of
the system, which is fairly close to the discrete-time state feedback response. 4

9.5 Notes and References


The observer discretization of Section 9.1 is based on [31], which considers also the
zero-order hold and first-order hold discretization methods in the addition to the three
310 CHAPTER 9. DIGITAL IMPLEMENTATION
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

Continuous SF Discrete SF

0 0

−0.5 −0.5

−1 −1

x1

x1
−1.5 −1.5

−2 −2

−2.5 −2.5

−3 −3
0 2 4 6 0 2 4 6
Time Time

3 3

2.5 2.5

2 2
x2

x2
1.5 1.5

1 1

0.5 0.5

0 0
0 2 4 6 0 2 4 6
Time Time

Figure 9.14. Simulation of Example 9.4. Comparison of the closed-loop trajectories under
continuous-time state feedback (SF) and discrete-time state feedback with sampling period 0.05.

6 8
6
4
4
2 2
x1

x2

0
0
−2
−2 −4

−4 −6
−8
0 2 4 6 8 10 0 2 4 6 8 10
Time Time

Figure 9.15. Simulation of Example 9.4. The closed-loop trajectories under single-rate
discrete-time output feedback controller with sampling period 0.05.

−2.8 2
x1 x2
T = 0.005 1.8 T = 0.005
−2.85 T = 0.05 T = 0.05
x1 and x̂1

x2 and x̂2

1.6
−2.9
1.4

−2.95
1.2

−3 1
0.3 0.32 0.34 0.36 0.38 0.4 0.3 0.32 0.34 0.36 0.38 0.4
Time Time

Figure 9.16. Simulation of Example 9.4. Down sampling of the state estimates from the
observer sampling period 0.005 to the control sampling period 0.05.
9.5. NOTES AND REFERENCES 311

Discrete SF Multi−rate OF
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php

0 0

−0.5 −0.5

−1 −1

x1

x1
−1.5 −1.5

−2 −2

−2.5 −2.5

−3 −3
0 2 4 6 0 2 4 6
Time Time

3 3

2.5 2.5

2 2
x2

x2
1.5 1.5

1 1

0.5 0.5

0 0
0 2 4 6 0 2 4 6
Time Time

Figure 9.17. Simulation of Example 9.4. Comparison of the closed-loop trajectories un-
der discrete-time state feedback (SF) with sampling period 0.05 and multirate discrete-time output
feedback (OF) with control sampling period 0.05 and observer sampling period 0.005.

methods discussed here. Section 9.2 is based on [32]. This paper includes experimen-
tal testing of various discretization methods of the high-gain observer. It demonstrates
the advantage of using a high-gain observer to estimate derivatives compared with the
simple, yet routinely used, Euler’s method. Section 9.3 is based on [79]. Section 9.4
is based on [3]. This paper deals with a more general control problem that allows
for time-varying external signals and stabilization of a closed set, as opposed to stabi-
lization of an equilibrium point. It includes also experimental application to a smart-
material-actuated system in which the control computes the inverse of a hysteresis
nonlinearity.

You might also like