Professional Documents
Culture Documents
php
Chapter 8
Measurement Noise
8.1 Observer
Reconsider the system
from Section 2.2, where f0 , ψ1 ,. . . , ψρ−1 , and φ are locally Lipschitz in their argu-
ments; u(t ) and v(t ) are piecewise continuous functions of t ; and w(t ), x(t ), u(t ),
237
238 CHAPTER 8. MEASUREMENT NOISE
and v(t ) are bounded for all t ≥ 0. In particular, let |v(t )| ≤ N , w(t ) ∈ 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
x(t ) ∈ X ⊂ Rρ , and u(t ) ∈ U ⊂ R m for all t ≥ 0, for some compact sets W , X , and
U . The new feature here is the presence of bounded measurement noise v(t ). It is as-
sumed that ψ1 to ψρ−1 are known and, for any compact set S ⊂ Rρ , there are positive
constants L1 to Lρ−1 such that
i
X
|ψi (x1 , . . . , xi , u) − ψi (z1 , . . . , zi , u)| ≤ Li |xk − zk | (8.5)
k=1
and
i
X
|ψis (x1 , . . . , xi , u) − ψis (z1 , . . . , zi , u)| ≤ Li |xk − zk | (8.10)
k=1
Theorem 8.1. Consider the system (8.1)–(8.4) and the observer (8.6)–(8.7). Suppose (8.9)
to (8.11) are satisfied. Then there is "∗ ∈ (0, 1] such that for 0 < " ≤ "∗ , the estimation
error x̃i = xi − x̂i , for 1 ≤ i ≤ ρ, satisfies the bound
b Nd
§ ª
|x̃i | ≤ max e −a t /" kx̃(0)k, "ρ+1−i c M + (8.12)
"i −1 "i −1
for some positive constants a, b , c, d . 3
xi − x̂i
ηi = for 1 ≤ i ≤ ρ.
"ρ−i
It can be shown that η satisfies the equation
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
−α1 1 0 ··· 0 α1
−α2 0 1 ··· 0 α2
.. .. .. ..
F = . , E = ,
. . .
αρ−1
−αρ−1 0 1
−αρ 0 ··· ··· 0 αρ
δ = col(δ1 , δ2 , . . . , δρ ),
1
δi = [ψi (x1 , . . . , xi , u) − ψis (x̂1 , . . . , x̂i , u)] for 1 ≤ i ≤ ρ − 1,
"ρ−i
δρ = φ(w, x, u) − φ0 (x̂, u).
kδk ≤ Lδ kηk + M ,
1 N
"V̇ ≤ − kηk2 + 2 "M kP k + ρ−1 kP Ek kηk,
2 "
1 N
"V̇ ≤ − kηk2 ∀ kηk ≥ 8 "M kP k + ρ−1 kP Ek .
4 "
Nd
§ ª
kη(t )k ≤ max b e −a t /" kη(0)k, "c M + ρ−1
"
yields (8.12). 2
Similar to the bound (2.24) of Theorem 2.1, the bound (8.12) has a transient term
due to the initial estimation error x̃(0) and an ultimate bound term. The transient term
decays to O(") values within a time period [0, T (")], where lim"→0 T (") = 0. The ul-
timate bound is the sum of two terms: "ρ+1−i c M , due to uncertainty in modeling φ,
and N d /"i −1 , due to measurement noise. Except for x̃1 , the term N d /"i−1 grows as
" decreases. A sketch of the ultimate bound as function of " is shown in Figure 1.11
for ρ = 2. Decreasing " reduces the bound until a minimum point where " is pro-
portional to (N /M )1/ρ . Reducing " beyond this point increases the ultimate bound.
Thus, the presence of measurement noise puts a lower bound on ". Another trade-off
exists between the speed of convergence of the observer and the ultimate bound on
the estimation error.
240 CHAPTER 8. MEASUREMENT NOISE
The ultimate bound given by (8.12) is conservative because it does not take into
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
x̂˙1 = x̂2 + (α1 /")(y − x̂1 ), x̂˙2 = (α2 /"2 )(y − x̂1 ),
For the sinusoidal noise v(t ) = N sin(ωt + θ), the amplitude of the steady-state esti-
mation error due to noise is given by
α2 ωN
N |G( j ω)| = p .
(α2 − "2 ω 2 )2 + (α1 "ω)2
p
At ω = α2 /", N |G( j ω)| = N α1 /(α2 "), which agrees with the ultimate bound of
p
(8.12). However, for ω α2 /" the low-pass feature of |G( j ω)| will diminish the
steady-state error since limω→∞ |G( j ω)| = 0. Therefore, the high-frequency compo-
nents of the noise will be attenuated. The low-pass filtering characteristics of the high-
gain observer are captured in the following theorem when v(t ) is the sum of sinusoidal
signals generated by the model
τ ż = S z, v = H z, (8.14)
Theorem 8.2. Consider the system (8.1)–(8.4) and the observer (8.6)–(8.7). Suppose (8.9)
to (8.11) are satisfied and the measurement noise v is generated by (8.14). Then there are
"∗ ∈ (0, 1], λ > 0, and T > 0 such that for 0 < " ≤ "∗ and 0 < τ/" ≤ λ the estimation
error x̃i = xi − x̂i , for 1 ≤ i ≤ ρ, satisfies the bound
τN d1
|x̃i (t )| ≤ "ρ+1−i c1 M + ∀ t ≥T (8.15)
"i
for some positive constants c1 and d1 . 3
"q̇ = F q + "δ,
8.1. OBSERVER 241
where kδk ≤ Lδ (kqk + k pk) + M . Using (8.14), it can be seen that p is the solution 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
" ṗ = F p − (1/"ρ−1 )E H z, τ ż = S z.
Seeking the solution of (8.16) as a power series in τ/", it can be seen that
∞ k
τ τX τ −(k+1) τ
Π = − ρ EH + F k+1
EHS S −1 = − ρ [E H + O(τ/")] S −1 ,
" " k=0 " "
where the infinite series converges for sufficiently small τ/". Let W = q T P q, where
p
we arrive at
1 kP k"
"Ẇ ≤ − W+p (Lδ kΠzk + Lδ k p − Πzk + M ) .
4λmax (P ) λmin (P )
k1 τN
Lδ kΠz(t )k ≤ ∀ t ≥ 0.
"ρ
Hence,
k1 τN
"Ẇ ≤ −`1W + "`2 + M + γ (t ) ,
"ρ
where `1 and `2 are positive constants and
γ (t ) = Lδ k p(t ) − Πz(t )k → 0 as t → ∞.
242 CHAPTER 8. MEASUREMENT NOISE
k τN
"U̇ = −`1 U + "`2 1 ρ + M + γ (t ) , U (0) = W (0).
"
Therefore,
t
k1 τN
Z
−`1 t /" −`1 (t −σ)/"
U (t ) = e W (0) + e `2 + M + γ (σ) d σ
0 "ρ
"` k1 τN
= e −`1 t /"W (0) + 2 −`1 t /"
+ M 1 − e
`1 "ρ
Zt
+ e −`1 (t −σ)/" `2 γ (σ) d σ.
0
Since lim t →∞ γ (t ) = 0, the integral term in the preceding expression tends to zero as
t → ∞. Hence, there is T1 > 0 such that
k1 τN
kq(t )k ≤ k2 " +M ∀ t ≥ T1
"ρ
for some positive constant k2 . On the other hand, since lim t →∞ k p(t ) − Πz(t )k = 0
and kΠz(t )k ≤ k1 τN /"ρ , there are positive constants k3 and T2 such
k3 τN
k p(t )k ≤ ∀ t ≥ T2 .
"ρ
Thus, for all t ≥ max{T1 , T2 },
kη(t )k = kq(t ) + p(t )k
k1 τN k τN
≤ k2 " ρ
+M + 3 ρ
" "
τN
= ("k1 k2 + k3 ) ρ + "k2 M
"
τN
≤ 2k3 ρ + "k2 M
"
for "k1 k2 ≤ k3 . Using x̃i = "ρ−i ηi yields (8.15). 2
Remark 8.1. The ultimate bound on the estimation error due to measurement noise in
Theorem 8.1 is N d /"ρ−1 , while in Theorem 8.2 it is τN d1 /"ρ . When τ/" 1, the latter
bound will be much smaller, which demonstrates the low-pass filtering characteristics of the
high-gain observer. 3
The emphasis in this chapter is on measurement noise that causes the estimation
error to be of the order of a negative power of ". It is useful, however, to note that
such a situation does not arise if the measurement noise is low frequency. Consider
again the second-order high-gain observer with the transfer function G(s) of (8.13)
and suppose v(t ) = N sin(ωt + θ), where ω 1/". The amplitude of the steady-state
estimation error due to noise is given by
α2 ωN
N |G( j ω)| = p ≈ ωN .
(α2 − "2 ω 2 )2 + (α1 "ω)2
8.1. OBSERVER 243
Theorem 8.3. Consider the system (8.17)–(8.20) and the observer (8.21)–(8.22). Suppose
(8.11) and (8.23) are satisfied. Then there is "∗ ∈ (0, 1] such that for 0 < " ≤ "∗ the
estimation error x̃i = xi − x̂i , for 1 ≤ i ≤ ρ, satisfies the bound
b −a t /"
§ ª
ρ+1−i
|x̃i (t )| ≤ Ni + max e (kx̃(0)k + N ), " c(LN + M + Nρ+1 ) (8.24)
"i−1
for some positive constants a, b , c. 3
where
−α1 1 0 ··· 0 0
−α2 0 1 ··· 0 0
.. .. ..
B = ... ,
F = . ,
. .
−αρ−1 0 1 0
−αρ 0 ··· ··· 0 1
δ = φ(w, x, u) − φ0 (x̂, u).
The matrix F is Hurwitz by design and
for " ≤ 1. Let V = ηT P η, where P = P T > 0 is the solution of the Lyapunov equation
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
P F + F T P = −I . Then
1
For "LkP Bk ≤ 4 ,
1
"V̇ ≤ − kηk2 + 2"kP Bk(LN + M + Nρ+1 ) kηk,
2
1
"V̇ ≤ − kηk2 ∀ kηk ≥ 8"kP Bk(LN + M + Nρ+1 ).
4
Using
1 1
kη(0)k ≤ (kx̃(0)k + kν(0)k) ≤ (kx̃(0)k + N )
"ρ−1 "ρ−1
and
x̃i = "ρ−i ηi − v (i −1) ,
we arrive at (8.24). 2
Example 8.1. Consider the van der Pol oscillator with corrupted measurement
We use simulation to study the effect of reducing " on the steady-state estimation error
for two cases of the measurement noise v. In the first case, v is wide-band random
signal generated by the Simulink block “Uniform Random Number” with amplitude
between ±0.1 and sample time 0.0001. In the second case, v is a low-frequency signal,
which is obtained by passing
p the noise of the first case through a filter with the transfer
function 8000/(s 2 + 10 2s + 100), which filters out the high-frequency noise. The
numerator of the transfer function is chosen such v has almost the same amplitude in
both cases. The noise and filtered noise are shown in Figure 8.1. Simulation is carried
out with the initial conditions x1 (0) = 1, x2 (0) = x̂1 (0) = x̂2 (0) = 0 and two values of
": 0.01 and 0.001. The estimation error x̃2 at steady state is shown in Figure 8.2. The
upper graphs are for the wide-band noise and the lower ones for the low-frequency
noise. For the wide-band noise, reducing " from 0.01 to 0.001 increases the steady-
state error by an order of magnitude. This is not the case for the low-frequency noise,
where the estimation error is of the same order of magnitude for the two values of ".
4
8.2. CLOSED-LOOP CONTROL 245
0.15 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
0.1
0.1
Filtered Noise
0.05
Noise
0 0
−0.05
−0.1
−0.1
−0.2
0 0.02 0.04 0.06 0.08 0.1 0 1 2 3 4 5
Time Time
Figure 8.1. The measurement noise of Example 8.1. The noise is generated by a random
number generator, and the filtered noise is obtained by passing the random signal through a low-pass
filter to remove its high-frequency content.
ε = 0.01 ε = 0.001
1 30
20
0.5
10
x̃2
x̃2
0 0
−10
−0.5
−20
−1 −30
19 19.2 19.4 19.6 19.8 20 19.8 19.85 19.9 19.95 20
Time Time
ε = 0.01 ε = 0.001
2 2
1 1
x̃2
x̃2
0 0
−1 −1
−2 −2
19 19.2 19.4 19.6 19.8 20 19 19.2 19.4 19.6 19.8 20
Time Time
Figure 8.2. The estimation error x̃2 = x2 − x̂2 of Example 8.1. The upper graphs are for
the wide-band noise and the lower ones for the low-frequency noise.
0 1 ··· ··· 0 0
0 0 1 ··· 0 0
.. .. ..
A= . B = C=
, , 1 0 ··· ··· 0 ,
. .
0 ··· ··· 0 1 0
0 ··· ··· ··· 0 1
where γ and Γ are locally Lipschitz functions in their arguments over the domain of in-
terest and globally bounded functions of x. Moreover, γ (0, 0, 0) = 0 and Γ (0, 0, 0) = 0.
A static state feedback controller u = γ (x, z) is a special case of the foregoing equation
where the ϑ̇-equation is dropped. For convenience, the closed-loop system under state
feedback is written as
χ̇ = f r (χ ), (8.30)
where χ = col(w, x, ϑ). The output feedback controller is given by
α1 /"
α2 /"2
..
H = , (8.33)
.
αρ−1 /"ρ−1
αρ /"ρ
where " is a small positive constant and the positive constants αi are chosen such that
the polynomial
s ρ + α1 s ρ−1 + · · · + αρ−1 s + αρ
is Hurwitz. The function φ0 (z, x, 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.
Moreover, φ0 (0, 0, 0) = 0.
Theorem 8.4. Consider the closed-loop system of the plant (8.25)–(8.28) and the output
feedback controller (8.31)–(8.32). Suppose the origin of (8.30) is asymptotically stable with
8.2. CLOSED-LOOP CONTROL 247
R as its region of attraction and |v(t )| ≤ N for all t ≥ 0. Let S be any compact set in
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 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 constant
"a = "a (N ) > ca N 1/ρ , with l i mN →0 "a (N ) = "∗a > 0, such that for each " ∈
(ca N 1/ρ , "a ] the solutions (χ (t ), x̂(t )) of the closed-loop system are bounded for all
t ≥ 0;
• there exist N1∗ > 0 and class K function %1 such that for every N < N1∗ and every
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, such that for " ∈ (ca N 1/ρ , " b ], the solutions
of the closed-loop system satisfy
• there exist N2∗ > 0 and 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 ], the solutions of the closed-loop system
satisfy
kχ (t ) − χ r (t )k ≤ µ2 ∀ t ≥ 0, (8.35)
where χ r is the solution of (8.30) with χ r (0) = χ (0).
Remark 8.2. The three bullets of the theorem show, respectively, boundedness of all tra-
jectories, ultimate boundedness where the trajectories come close to the equilibrium point
{x = 0} × {x − x̂ = 0} as time progresses, and closeness of the trajectories under out-
put feedback to the ones under state feedback. In the absence of measurement noise, it is
shown in Theorem 3.1 that the output feedback controller recovers asymptotic stability of
the origin if either φ0 = φ or the origin of (8.30) is exponentially stable. Comparison of
Theorem 8.4 with Theorem 3.1 shows that the presence of measurement noise is manifested
in three points, which are intuitively expected:
• The constants µ1 and µ2 , which measure ultimate boundedness and closeness of tra-
jectories, respectively, cannot be made arbitrarily small. Instead, they are bounded
from below by class K functions of N .
χ̇ = f (χ , D −1 (")η), (8.36)
ρ −1
"η̇ = F η + " B g (χ , D (")η) − E v, (8.37)
248 CHAPTER 8. MEASUREMENT NOISE
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
−α1 1 0 ··· 0 α1
−α2 0 1 ··· 0 α2
.. .. .. ..
F = . , E = ,
. . .
αρ−1
−αρ−1 0 1
−αρ 0 ··· ··· 0 αρ
where σ3 = 1/(2kP k). From the scaling ηi = "i −1 (xi − x̂i ), we see that whenever
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
x(0), x̂(0), and " are bounded, there is a constant kw > 0, independent of ", such that
W (η(0)) ≤ kw2 . It follows from (8.41) and (8.42) that η reaches the set Σ within the
time interval [0, T (")], where
2" kw
T (") = ln → 0 as " → 0. (8.43)
σ3 σ1 " ρ
Inside Σ, η is bounded by
kηk ≤ c1 "ρ + c2 N , (8.44)
where c1 = σ1 / λmin (P ) and c2 = σ2 / λmin (P ). Therefore,
p p
N def
kx(t ) − x̂(t )k = kD −1 (")η(t )k ≤ "c1 + c = h(", N ). (8.45)
"ρ−1 2
Lemma 8.1. h(", N ) has the following properties for " > 0 and N ≥ 0:
def
• h(", N ) has a global minimum at " = [(ρ − 1)c2 N /c1 ](1/ρ) = ca N 1/ρ and
ρ−1 def
min">0 h(", N ) = (c1 ca + c2 /ca )N 1/ρ = ka N 1/ρ .
• For " > ca N 1/ρ , h(", N ) is a strictly increasing function of " and h(", N ) ≤ k b ",
ρ
where k b = c1 + c2 /ca .
• Given k > 0, for every N ∈ [0, (k/ka )ρ ) there exist " m = " m (N , k) ≥ 0 and "M =
"M (N , k) > 0, with " m ≤ ca N 1/ρ < "M , limN →0 " m = 0, and limN →0 "M = k/c1 ,
such that h(", N ) ≤ k for " ∈ (" m , "M ]. 3
for all χ ∈ Ωc and η ∈ Rρ . Let L2 be an upper bound for k∂ V /∂ χ k over Ωc and set
L = L1 L2 . Then
kχ (t ) − χ (0)k ≤ k1 t (8.47)
Since U (χ ) is positive definite and continuous, there is a positive constant τ ∗ < L3 such
that the set {U (χ ) ≤ 2Lτ} is compact for τ ≤ τ ∗ . Let c0 (τ) = maxU (χ )≤2Lτ {V (χ )};
c0 (τ) is nondecreasing, and limτ→0 c0 (τ) = 0. Let ϕ(τ) be a class K function such that
ϕ(τ) ≥ c0 (τ). Then
1
V̇ ≤ − 2 U (χ ) for ϕ(τ) ≤ V (χ ) ≤ c.
V (χ (t )) ≤ ϕ(τ) ∀ t ≥ T ∗ + Ta (τ).
It can be shown that there is a class K∞ function ϕ1 such that ϕ1 (kχ k) ≤ V (χ ) for all
χ ∈ R.73 Therefore,
def
kχ (t )k ≤ ϕ1−1 (ϕ(τ)) = %a (τ) ∀ t ≥ T ∗ + Ta (τ).
Thus,
def
max{kχ (t )k, kx(t ) − x̂(t )k} ≤ max{τ, %a (τ)} = % b (τ).
The function % b is a class K∞ function. Define the class K∞ function %1 by %1 (N ) =
% b (ka N 1/ρ ) and suppose N < N1∗ = min{N ∗ , (τ ∗ /ka )ρ }. Then, given µ1 > %1 (N ), take
τ = min{τ ∗ , %−1 b
(µ1 )} and set T1 = T ∗ + Ta (τ), to obtain (8.34).
Finally, we show (8.35). From (8.34) and asymptotic stability of the origin of (8.38),
given µ2 > 2%1 (N ), there is T2 = T2 (µ2 ) > 0, independent of ", such that
kχ (t )k ≤ µ2 /2, kχ r (t )k ≤ µ2 /2 ∀ t ≥ T2 (8.48)
kχ (t ) − χ r (t )k ≤ kχ (t )k + kχ r (t )k ≤ µ2 ∀ t ≥ T2 . (8.49)
73 See the proof of Theorem 4.17 of [78].
8.2. CLOSED-LOOP CONTROL 251
Hence,
kχ (t ) − χ r (t )k ≤ 2k1 T (") ∀ t ∈ [0, T (")]. (8.50)
View (8.36) as a perturbation of (8.38). From continuous dependence of the solutions
on initial conditions and parameters [78, Theorem 3.4], it follows that
for some constants c3 ≥ 1 and c4 > 0, independent of ". From (8.50) and (8.51) we
conclude that
def
min {2k1 c3 T (") + c4 h(", N )} = 2k1 c3 T (ca N 1/ρ ) + c4 ka N 1/ρ = %3 (N ).
"∈(ca N 1/ρ , "g ]
It can be verified that %3 (N ) is a class K function. For each µ2 > %3 (N ), there exists
"¯c = "¯c (N , µ2 ) > ca N 1/ρ , with limN →0 "¯c (N , µ2 ) = "¯∗c (µ2 ) > 0, such that
for all " ∈ ca N 1/ρ , "¯c . Taking %2 (N ) = max{2%1 (N ), %3 (N )}, N2∗ = min{N1∗ , N5∗ },
"c = min{" b , "¯c , " g }, and using (8.49), (8.52), and (8.53), we arrive at (8.35). 2
from Example 3.4, where 0 ≤ c1 ≤ 0.2 and 0.5 ≤ c2 ≤ 2. It is shown there that, with
y = x1 as the measured output, the output feedback controller
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. Suppose y = x1 + v,
where |v(t )| ≤ 0.001, which is consistent with the roundoff error in typical optical
encoders. Simulation is carried out with the initial conditions x1 (0) = −π, x2 (0) =
252 CHAPTER 8. MEASUREMENT NOISE
−4
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
x 10
−1
0
−2
−1
x1
x1
−3
−2
−4
−3
−5
0 1 2 3 4 5 9 9.2 9.4 9.6 9.8 10
Time Time
−3
x 10
3 1.5
1
2
0.5
x2
x2
1 0
0 −0.5
−1
0 1 2 3 4 5 9 9.2 9.4 9.6 9.8 10
Time Time
0.2
20
0.1
u
u
10 0
−0.1
0
−0.2
0 1 2 3 4 5 9 9.2 9.4 9.6 9.8 10
Time Time
Table 8.1. The maximum over t ∈ [9, 10] of |x1 |, |x2 |, |x̃1 |, |x̃2 |, and |u|.
x̂1 (0) = x̂2 (0) = 0, the parameters c1 = 0.01, c2 = 0.5, and with three values of ":
0.05, 0.01, and 0.002. The noise v is random signal generated by the Simulink block
“Uniform Random Number” with amplitude between ±0.001 and sample time 0.0001.
Figure 8.3 shows the trajectories of x1 , x2 , and u for " = 0.01. It can be seen that the
measurement noise has little effect on the transient behavior of the system. Zooming
on the interval t ∈ [9, 10] shows the effect on the steady-state behavior. To examine
how the effect of measurement noise changes with ", Table 8.1 shows the maximum
over t ∈ [9, 10] of the absolute values of x1 , x2 , x̃1 = x1 − x̂1 , x̃2 = x2 − x̂2 , and u.
The table demonstrates that, in general, the steady state of these signals increases as "
decreases. There is, however, an important difference between the various variables.
To illustrate this difference, Figure 8.4 displays the numbers for x1 , x̃1 , x̃2 , and u, which
are shown as “dots.” The same figure displays, in “+”, points proportional to 1/". It
can be seen that the increase in x̃2 and u is of the order of 1/", while the increase in
x1 and x̃1 is not of that order. The difference between x̃1 and x̃2 is consistent with
Theorem 8.1, which shows that the ultimate bound on x̃2 due to measurement noise
is O(1/") while the ultimate bound on x̃1 is independent of ". Since the control u
depends on x̂2 , it is not surprising that its change with " will be similar to the change
of x̃2 . The behavior of x1 is not unique to this example. It is a general trend that is
explored in the next section. 4
8.3. TRACKING 253
−1 −2
10 10
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
−2
10
max|x1 |
max|x̃1 |
−3
10
−3
10
−4 −4
10 10
−3 −2 −1 −3 −2 −1
10 10 10 10 10 10
ε ε
0 1
10 10
−1
10
0
10
max|x̃2 |
max|u|
−2
10
−1
10
−3
10
−4 −2
10 10
−3 −2 −1 −3 −2 −1
10 10 10 10 10 10
ε ε
8.3 Tracking
It is observed in Example 8.2 that the steady state of x1 in the presence of measurement
noise is not of the order of 1/". This is particularly significant in tracking problems
where x1 is required to track a given reference signal. In this section, we elaborate on
this observation for a special case of the tracking problem of Section 3.3 where the
plant is linear and has no zero dynamics.
Consider the single-input–single-output linear system
ei = xi − r (i −1) for 1 ≤ i ≤ n,
1
u = − (k1 e1 + · · · + kn−1 en−1 + en ),
µ
where K = k1
... kn−1 1 .
Lemma 8.2. There exist µ∗ > 0 and T > 0 such that the matrix A + Ba − (b /µ)BK is
Hurwitz for all 0 < µ ≤ µ∗ and the solution of (8.58) satisfies e = O(µ) for all t ≥ T . 3
Proof: Let
ζ˙ = A1 ζ + B1 s , (8.59)
µṡ = µC1 ζ − (b − µan − µkn−1 )s + µ(θ + δ), (8.60)
where
0 1 ··· ··· 0 0
0 0 1 ··· 0 0
.. .. B1 = ... ,
A1 = ,
. .
0 ···
··· 0 1 0
−k1 ···
· · · · · · −kn−1 1
and C1 = c1 c2 . . . cn−2 cn−1 , in which c1 = a1 − (an + kn−1 )k1 and ci = ai +
ki−1 − (an + kn−1 )ki for 2 ≤ i ≤ n − 1. The matrix A1 is Hurwitz by design. Let L be
the solution of
Equation (8.61) has a unique solution for sufficiently small µ and L = −(1/b )C1 +
O(µ) [87, Section 2.2]. The change of variables ψ = s + µLζ transforms the system
(8.59)–(8.60) into the block triangular form
which shows that the closed-loop eigenvalues are −(b − µan − µkn−1 − µ2 LB1 )/µ and
the eigenvalues of (A1 − µB1 L). Hence, the closed-loop matrix is Hurwitz for suffi-
ciently small µ. Since θ + δ is bounded, it is seen from (8.63) that ψ is ultimately
8.3. TRACKING 255
bounded by O(µ). Then (8.62) shows that ζ is ultimately bounded by O(µ). There-
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
In preparation for output feedback, the control is saturated outside a compact set
of interest. Since A + Ba − (b /µ)BK is Hurwitz, let P0 = P0T > 0 be the solution of
the Lyapunov equation
V = e T P0 e, and Ω = {e T P0 e ≤ c}, where c > 0 is chosen large enough that the initial
error e(0) is in the interior of Ω and V̇ < 0 on the boundary {V = c}, which is possible
because θ + δ is bounded. Take
1
U > max (k1 e1 + · · · + kn−1 en−1 + en ) .
e∈Ω µ
where the estimates ê1 to ên are provided by the high-gain observer
˙ê = ê + αi (y − r − ê ), 1 ≤ i ≤ n − 1, (8.66)
i i +1 1
"i
α
˙ê = n (y − r − ê ),
n 1 (8.67)
"n
where " is a sufficiently small positive constant, and α1 to αn are chosen such that the
polynomial
λn + α1 λn−1 + · · · + αn−1 λ + αn
is Hurwitz.
Theorem 8.5. Consider the closed-loop system of the plant (8.54)–(8.56) and the output
feedback controller (8.65)–(8.67). Let S be any compact set in the interior of Ω, and let
Q be any compact subset of Rn . Suppose (e(0), ê(0)) ∈ S × Q. Let ē(t ) be the tracking
error under output feedback without noise. Then there is µ∗ > 0, and for each µ ∈ (0, µ∗ ],
there are N ∗ > 0 and T ∗ > 0, dependent on µ, such that for each N < N ∗ there is "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 (e(t ), ê(t )) of the closed-loop system are bounded for all t ≥ 0 and
for all t ≥ T ∗ . 3
Proof: With
ϕi = "i−1 (ei − êi ) for 1 ≤ i ≤ n,
256 CHAPTER 8. MEASUREMENT NOISE
K(e − D −1 (")ϕ)
ė = Ae + B ae + θ + δ − b U sat ,
µU
K(e − D −1 (")ϕ)
n
"ϕ̇ = F ϕ + " B ae + θ + δ − b U sat − E v,
µU
where
−α1 1 0 ··· 0 α1
−α2 0 1 ··· 0 α2
.. .. , ..
F =
.. E =
,
. . . .
−α 0 1 α
n−1 n−1
−αn 0 ··· ··· 0 αn
and
D(") = diag(1, ", . . . , "n−1 ).
The matrix F is Hurwitz. In analyzing this system, as " tends to zero, µ is maintained
at a value small enough for the conclusions of Lemma 8.2 to hold. Similar to the proof
of Theorem 8.4, it can be shown that there is T (") > 0, with lim"→0 T (") = 0, such that
for all t ≥ T (") the trajectories of the closed-loop system are confined to the positively
invariant set Ω × Σ, where Σ is defined by (8.41). Inside Ω × Σ the control saturation
is not active, and the closed-loop system is linear. The closed-loop system without
noise is represented by the same linear equations with v = 0. Therefore ẽ = e − ē is
determined by the linear singularly perturbed system
where η is the difference between ϕ and the corresponding variable in the absence of
noise. Since F + ("n b /µ)BK D −1 (") = F + O("), it is Hurwitz for sufficiently small
". Next, a change of variables is used to transform the system (8.70)–(8.71) into a
block triangular form where the fast variable η is removed from the slow equation
[87, Section 2.2]. Let M be the solution of the equation
Noting that "n D −1 (") = O("), it can be seen that (8.72) has a unique solution. The
change of variables
1
ξ = ẽ − Mη (8.73)
" n−2
from (8.74), then using (8.73). In both cases it is important to reveal how M depends on
". This can be done by seeking the solution of (8.72) using successive approximation.
Rewrite (8.72) as
where
S(") = "n−1 D −1 (") = diag("n−1 , "n−2 , . . . , ", 1).
For sufficiently small ", the right-hand side of (8.76) is a contraction mapping. There-
fore, we can solve for M using the successive approximation [101]
M0 = (· · · )S F −1 ,
M1 = (· · · )S F −1 + "(· · · )S F −2 ,
M2 = (· · · )S F −1 + "(· · · )S F −2 + "2 (· · · )S F −3 ,
M3 = (· · · )S F −1 + "(· · · )S F −2 + "2 (· · · )S F −3 + "3 (· · · )S F −4 ,
.. ..
. .
M n−2 = (· · · )S F −1 + "(· · · )S F −2 + · · · + "n−2 (· · · )S F −(n−1) + "n−1 (· · · ).
−"n−1
0 0
0 n−2
−" ..
..
.
S F −1 E = . , SF E = 0 ,
−2
..., SF −(n−1)
E = 0 .
.
.
..
.. −"
0 0 0
Therefore, M E = O("n−1 ), and the term (1/"n−1 )M E v on the right-hand side of (8.74)
is O(N ), uniformly in " for sufficiently small ". Similar to the proof of Lemma 8.2, it
can be shown that the ultimate bound of ξ is O(N ). Alternatively, it can be seen that
M0 = B(· · · ),
M1 = B(· · · ) + "AB(· · · ),
M2 = B(· · · ) + "AB(· · · ) + "2 A2 B(· · · ),
.. ..
. .
M n−2 = B(· · · ) + "AB(· · · ) + · · · + "n−2 An−2 B(· · · ) + "n−1 (· · · ),
258 CHAPTER 8. MEASUREMENT NOISE
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
0 0 0
0
.. .. .. 1
. . . 0
.. ..
B = AB = A B = 0 ,
2 n−2
A B = .
., ., ..., .. .
0 0 1
.
0 ..
0 1
1 0 0 0
Therefore,
O("n−1 )
O("n−2 )
..
M = M n−2 + O("n−1 ) = .
.
O(")
O(1)
Since the matrix F + ("n b /µ)BK D −1 (") + "2 B(a − (b /µ)K)M is Hurwitz for suffi-
ciently small ", it follows from (8.75) that the ultimate bound of η is O(N ). Using
the ultimate bounds of ξ and η, (8.73), and the foregoing orders of magnitude of the
components of M , we arrive at (8.68) and (8.69). 2
Remark 8.3. The significant finding of Theorem 8.5 is that, unlike the estimation error,
the component of the tracking error due to measurement noise does not increase as a nega-
tive power of ". This is important because it is the tracking error that usually has to meet
stringent accuracy requirements. This does not mean, however, that we can arbitrarily
decrease " without consequences. For high-dimensional systems, the effect of noise on e3
to en will increase as a negative power of ". More importantly, the effect of noise on the
control signal, which is a function of all the state estimates, will be O(N /"n−1 ). Another
important observation about the control signal is seen from the proof of the theorem, which
requires the noise level to be small enough such that the control signal saturates only dur-
ing the peaking period of the observer. The ultimate bound estimates (8.68) and (8.69) are
not guaranteed to hold if the noise causes the control signal to saturate beyond the peaking
period. 3
1
u = − (e1 + 2e2 + e3 )
µ
with µ = 0.05. The maximum of |u| over Ω = {e T P0 e ≤ 2}, where P0 is the solution
of the Lyapunov equation (8.64), is 177.1. The control is saturated at ±200. The third-
order observer is designed with α1 = α2 = 3 and α3 = 1. The simulation is carried out
with x(0) = 0, ê(0) = 0, and the measurement noise is generated by the uniform ran-
dom number block of Simulink with amplitude ±0.0001 and sampling time 0.0001.
Four values of " are tried: 0.01, 0.005, 0.002, and 0.001. Figure 8.5 shows that track-
ing error e1 = x1 − r under state feedback (noise free) and output feedback with two
8.3. TRACKING 259
0.5 0.04
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 0.02
e1
e1
−0.5 0
State
−1 Output ε=0.01 −0.02
Output ε=0.005
−1.5 −0.04
0 5 10 15 20 10 12 14 16 18 20
Time
Figure 8.5. The tracking error of Example 8.3 under noise-free state feedback (solid) and
noisy output feedback for two values of ".
−5 −5 −5
x 10 ε = 0.01 x 10 ε = 0.005 x 10 ε = 0.002
1 1 1
e1
e1
e1
0 0 0
−1 −1 −1
10 15 20 10 15 20 10 15 20
−4 −4 −4
x 10 x 10 x 10
2
2
2
e2
e2
e2
0 0 0
−2
−2
−2 −4
10 15 20 10 15 20 10 15 20
0.02 0.2
0.02
e3
e3
e3
0 0 0
−0.02
−0.02 −0.2
10 15 20 10 15 20 10 15 20
2 10 100
u
0 0 0
−2 −10 −100
10 15 20 10 15 20 10 15 20
Time Time Time
Figure 8.6. The difference between the trajectories under noisy output feedback and noise
free output feedback of Example 8.3. The left column is for " = 0.01, the middle column for " =
0.005, and the right column for " = 0.002.
different values of ". The error is shown over the time interval [0, 20] and also over
the interval [10, 20] to examine the transient and steady-state behavior. The simula-
tion shows that the response under output feedback approaches the noise-free response
under state feedback as " decreases. To focus on the effect of measurement noise,
Figure 8.6 shows the difference between output feedback trajectories with and with-
out noise for " = 0.01, 0.005, and 0.002. The response is shown over the time interval
[10, 20] to focus on the steady-state response. The figure demonstrates the ultimate
bounds of Theorem 8.5. As " decreases, e1 and e2 remain almost the same, e3 grows
proportional to 1/", while u grows proportional to 1/"2 . Figure 8.7 shows the control
u and the tracking error e1 when " = 0.001. In this case the control saturates beyond
the peaking period of the observer. As noted in Remark 8.3, the ultimate bounds of
Theorem 8.5 no longer hold, and that is clear by examining e1 . Let us now use transfer
functions to interpret the results we have seen so far. Let H1 , H2 , H3 , and H4 be the
260 CHAPTER 8. 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
Figure 8.7. The control signal u and the difference between tracking error e1 under noisy
output feedback and noise-free output feedback of Example 8.3 when " = 0.001.
transfer functions from the noise v to e1 , e2 , e3 , and u, respectively. They are given by
where
λ(s) = s 3 + s 2 + 1,
q(s, ") = (1 + 6" + 3"2 )s 2 + (2 + 3")s + 1,
χ (s, ") = µ p("s )λ(s ) + q(s, "),
p("s) = ("s + 1)3 .
It can be shown that maxω∈R |Hi ( j ω)| is O(1) for H1 and H2 , O(1/") for H2 , and
O(1/"2 ) for H4 . The magnitude Bode plots of these four transfer functions are shown
in Figure 8.8 for " = 0.01 and 0.002. It is clear that decreasing " does not increase the
peak of the magnitude plot of H1 and H2 . However, for the smaller " the magnitude
plot does not roll off as fast as for the larger ". Thus, while the effect of noise will not
increase as a negative power of ", more higher-frequency components of the noise will
pass to e1 and e2 . A careful examination of Figure 8.6 shows that there is an increase
in the high-frequency content of e1 and e2 as " decreases. For H3 and H4 , the figure
shows that the maximum of the magnitude plots of H3 and H4 grows by O(1/") and
O(1/"2 ), respectively. Noting that the ratio of 0.01 to 0.002 is 5, 20 log(5) = 14 dB, and
20 log(25) = 28 dB, it can been checked from the figure that this is roughly the increase
in the peaks of H3 and H4 . 4
H1 H2
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
100 50
50
Magnitude (dB)
Magnitude (dB)
0 0
−50
−100 −50
−150
−200 −100
0 2 4 0 2 4
10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)
H3 H4
100 150
100
Magnitude (dB)
Magnitude (dB)
50
50
0
0
−50 −50
0 2 4 0 2 4
10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)
Figure 8.8. Magnitude Bode plots of the transfer functions of Example 8.3 for " = 0.01
(solid) and " = 0.002 (dashed).
so that a higher gain is used during the transient period to achieve fast convergence;
then the gain is lowered as the estimation error approaches steady-state since the effect
of measurement noise is prominent when the estimation error is small. Three tech-
niques have been proposed in the literature to achieve this gain adjustment. The first
technique switches the gain between two values [2]. The second uses an adaptive law
to adjust the gain [130]. The third technique replaces the linear gain of the observer
with a nonlinear one that reduces the gain when the estimation error is small [120].
The latter technique is described in Section 8.4.2.
-
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
-
- Controller - - +h
?
Observer +h
Plant
Filter
6
Observer Filter
Filter
6
xi − x̂i
ηi = for 1 ≤ i ≤ ρ (8.86)
"ρ−i
satisfy the equation
−α1 1 0 ··· 0 α1
−α2 0 1 ··· 0 α2
.. .. .. ..
F = . , E = ,
. .
.
αρ−1
−αρ−1 0 1
−αρ 0 ··· ··· 0 αρ
8.4. REDUCING THE EFFECT OF MEASUREMENT NOISE 263
and δ = φ(w, x, u) − φ0 (x̂, u). The matrix F is Hurwitz by design. In view of (8.85)
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
"θ̇ = F θ − E v, (8.89)
and let η̃ = η − (1/"ρ−1 )θ. It can been seen that η̃ satisfies the equation
τ σ̇i = A f σi + τA−1
f
B f xi +1 for 1 ≤ i ≤ ρ − 1, (8.97)
τ σ̇ρ = A f σρ + τA−1
f
B f φ(w, x, u), (8.98)
By the boundedness of x and φ(w, x, u), it can be shown that ξ is ultimately bounded,
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
C f σi + xi − x̂i
ϕi = for 1 ≤ i ≤ ρ, (8.100)
"ρ−i
and using −C f A−1
f
B f = 1, y f = C f σ1 + x1 + v f , and σ̇ρ = ξ + A−1
f
B f ẋρ , it can be
verified that ϕ = col(ϕ1 , . . . , ϕρ ) satisfies the equation
and let ϕ̃ = ϕ − (1/"ρ−1 )ψ. It can been seen that ϕ̃ satisfies the equation
1
x̃i = −C f σi + "ρ−i ϕi = −C f σi + "ρ−i ϕ̃i + ψ. (8.104)
"i −1 i
We have already seen that the ultimate bound on σi is O(τ). Thus, the effect of mea-
surement noise on the steady state of x̃i is captured by the term (1/"i −1 )ψi .
To compare the effect of measurement noise on the steady-state estimation error
with and without filter, we need to compare θ, which satisfies (8.89), with ψ, which
satisfies (8.102). Equation (8.89) is driven by the measurement noise v, while (8.102) is
driven by the filtered noise v f . This characterization of the noise effect is the nonlinear
counterpart of the block diagram of Figure 8.10.
Consider now closed-loop control of the system (8.78)–(8.81). We design a noise-
free state feedback controller that stabilizes the origin, then consider the closed-loop
system that arises when the controller is implemented using a high-gain observer to-
gether with a low-pass filter. The state feedback controller takes the form u = γ (x).
The closed-loop system under state feedback is represented by
χ̇ = f r (χ , γ (x)), (8.105)
f0 (w, x, u)
x2
..
f r (χ , u) = .
.
xρ
φ(w, x, u)
The output feedback controller is taken as u = γ (x̂), where x̂ is provided by the high-
gain observer (8.93)–(8.94), and y f is the output of the low-pass filter (8.92). The closed-
8.4. REDUCING THE EFFECT OF MEASUREMENT NOISE 265
χ̇ = f r (χ , γ (x̂)), (8.106)
α
x̂˙i = x̂i +1 + i (C f z − x̂1 ) for 1 ≤ i ≤ ρ − 1, (8.107)
"i
αρ
x̂˙ρ = φ0 (x̂, u) + ρ (C f z − x̂1 ), (8.108)
"
τ ż = A f z + B f (x1 + v). (8.109)
Assumption 8.1.
• f0 is continuously differentiable, φ is twice continuously differentiable, f0 (0, 0, 0) =
0, and φ(0, 0, 0) = 0;
• γ is twice continuously differentiable, globally bounded, and γ (0) = 0. The origin
of (8.105) is asymptotically stable with a region of attraction R and locally expo-
nentially stable;
• φ0 is twice continuously differentiable, globally bounded, and φ0 (0, 0) = 0.
Theorem 8.6. Consider the closed-loop system (8.106)–(8.109) under Assumption 8.1. Let
L be any compact set in the interior of R, let M be any compact subset of Rρ × R r , and
suppose that χ (0) ∈ L and (x̂(0), z(0)) ∈ M . The closed-loop system has the following
properties:
• When v = 0, there exists λ∗ > 0 such that if max{", τ} < λ∗ , the origin (χ = 0,
x̂ = 0, z = 0) is an exponentially stable equilibrium point and L × M is a subset
of its region of attraction.
• When v 6= 0 with |v| ≤ N , there exist positive constants N ∗ and k ∗ and a class
K function % such that for every N < N ∗ and every µ > %(N ), there are positive
constants τ ∗ and "∗ > k ∗ N 1/ρ such that for τ ≤ τ ∗ and " ∈ [k ∗ N 1/ρ , "∗ ], the
trajectories of the closed-loop system are bounded and
kχ (t ) − χ r (t )k ≤ µ (8.110)
for all t ≥ 0, where χ r (t ) is the solution of (8.105) with χ r (0) = χ (0). 3
Remark 8.4. The theorem is stated for initial states in any compact subset of the region
of attraction under state feedback. When the state feedback achieves global or semiglobal
stabilization, the output feedback achieves semiglobal stabilization. 3
Remark 8.5. The theorem shows that exponential stability under state feedback is recov-
ered under output feedback when " and τ are sufficiently small. It is significant that there is
no restriction on the ratio τ/"; in other wards, it is not required that one of the parameters
be much smaller than the other. This point can be easily seen in linear systems, as shown
in Example 8.4, but it takes more work to show it in nonlinear systems. In our analysis
we use a nonminimal realization of the closed-loop system by introducing the variables σi
and ξ , whose equations are copies of the filter equation. 3
Remark 8.6. The second bullet of the theorem shows that the trajectories under output
feedback approach the ones under state feedback for sufficiently small ", τ, and N . As in
266 CHAPTER 8. MEASUREMENT NOISE
Theorem 8.4, in the presence of measurement noise, the constant µ, which measures 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
closeness of trajectories, cannot be made arbitrarily small because it is bounded from below
by a class K function of N . 3
Proof of Theorem 8.6: With the definitions of σ and ξ and the change of variables
(8.100), the closed-loop system is represented by
χ̇ = f r (χ , γ (x̂)), (8.111)
τ σ̇ = A f f σ + τA−1 B
f f f f
[Ax + Bφ(χ , γ (x̂))], (8.112)
where
σ = col(σ1 , . . . , σρ ),
x̂ = x + C f f σ − "ρ−1 D −1 (")ϕ,
A f f = block diag[A f , . . . , A f ],
B f f = block diag[B f , . . . , B f ],
C f f = block diag[C f , . . . , C f ],
D(") = diag[1, ", . . . , "ρ−2 , "ρ−1 ],
V̇c ≤ −Y T QY ,
8.4. REDUCING THE EFFECT OF MEASUREMENT NOISE 267
where Y = col(kχ k, kξ˜k, kσk, kϕk) and the symmetric matrix Q is given by
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 − 2"k2 )
for some positive constants k1 and k2 , independent of τ and ". Setting τ = " = 0 in Q
results in
b k1 0 −b k2 −b k2
0 b −b k2 −b k2 ,
? ?
1 0
? ? 0 1
which is positive definite for sufficiently small b . Thus, by choosing b small enough, Q
will be positive definite for sufficiently small τ and ". Consider next the case v 6= 0. By
repeating the foregoing argument it can be shown that χ (t ) ∈ Ω for all t ≥ 0, provided
N /"ρ−1 is small enough, which is achieved by requiring " to satisfy " ≥ k ∗ N 1/ρ ; such
a choice is feasible if N is sufficiently small. Under this restriction,
for all t ≥ T1 (τ)+T2 ("). Continuous dependence of the solutions on initial conditions
and parameters yields the inequality (8.110). 2
ẋ1 = x2 , ẋ2 = u, y = x1
can be stabilized by the state feedback control u = −x1 −2x2 , which assigns the closed-
loop eigenvalues at −1, −1. The output feedback controller that uses a high-gain ob-
server is given by
2 1
u = −x̂1 − 2x̂2 , x̂˙1 = x̂2 + (y f − x̂1 ), x̂˙2 = (y f − x̂1 ),
" "2
which assigns the observer eigenvalues at −1/", −1/". The observer input y f is the
output of the low-pass filter
1
,
(τs + 1)2
which is driven by y. The filter poles are at −1/τ, −1/τ. In this example it is shown
that, for linear systems, the exponential stability result of Theorem 8.6 can be shown
by examining the closed-loop poles. The closed-loop characteristic equation is
where Λ(β) = (β+1)2 . This equation has six roots. To study the asymptotic behavior
of the roots for small " and τ, let λ = max{", τ}. As λ → 0, the characteristic equation
tends to
s 2 + 2s + 1 = 0,
268 CHAPTER 8. MEASUREMENT NOISE
−5 −5
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
x 10 x 10
5 5
0.03
0.02
∆θ
∆θ
∆θ
0.01 0 0
0
−0.01 −5 −5
0 2 4 9 9.5 10 9 9.5 10
Time Time Time
−4 −4
x 10 x 10
0.1
5 5
0.05
∆ω
∆ω
∆ω
0 0 0
−0.05
−5 −5
−0.1
0 2 4 9 9.5 10 9 9.5 10
Time Time Time
20 0.2 0.2
10 0.1 0.1
∆u
∆u
∆u
0 0 0
Figure 8.11. Comparison of the trajectories of Example 8.5 with low-pass filter (LPF)
(dashed) and without LPF (solid) when " = τ = 0.01. The plotted curves are the differences from the
trajectories under output feedback with no measurement noise.
which shows that two roots of (8.115) approach the eigenvalues of the closed-loop sys-
tem under state feedback. Multiplying the characteristic equation by λ2 and changing
the frequency variable to p = λs, the equation is rewritten as
" τ
p 2Λ p Λ p + 2λ(1 + ") p + λ2 = 0.
λ λ
As λ → 0, 2λ(1 + ") p + λ2 → 0, and the equation reduces to
" τ
p 2Λ p Λ p = 0.
λ λ
Dividing by p 2 yields " τ
Λ
p Λ p = 0,
λ λ
whose roots are p = −λ/", −λ/", −λ/τ, −λ/τ. When transformed back into the
frequency variable s , it is seen that four roots of (8.115) approach the observer eigen-
values −1/", −1/", and the filter poles −1/τ, −1/τ. The limits of the six roots of
(8.115) are independent of the ratio τ/". 4
Example 8.5. Reconsider the pendulum equation of Examples 3.4 and 8.2. The effect
of adding the low-pass filter 1/(τs +1)2 is examined for the same controller, initial con-
ditions, and measurement noise parameters as in Example 8.2. Figure 8.11 compares
the trajectories of θ, ω, and u when " = τ = 0.01. Because the trajectories are fairly
close to each other, the figure plots the deviation from the trajectories under no noise.
8.4. REDUCING THE EFFECT OF MEASUREMENT NOISE 269
τ = 0.1 τ = 0.1
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
0.01
10
0.005
∆u
∆u
0
0
−0.005
−10 −0.01
0 1 2 3 4 5 9 9.2 9.4 9.6 9.8 10
Time Time
τ = 0.01 τ = 0.01
10 0.1
0.05
∆u
∆u
0
0
−0.05
−0.1
−10
0 1 2 3 4 5 9 9.2 9.4 9.6 9.8 10
Time Time
τ = 0.001 τ = 0.001
10 0.1
0.05
∆u
∆u
0
0
−0.05
−0.1
−10
0 1 2 3 4 5 9 9.2 9.4 9.6 9.8 10
Time Time
Figure 8.12. The effect of changing τ on the control trajectory of ω of Example 8.5.
It is seen that using the filter causes a minor degradation of the transient response while
significantly reducing the steady-state error due to measurement noise, which is most
obvious in the trajectories of u. To examine the effect of the filter’s time constant τ,
Figure 8.12 shows the ∆u trajectory of Figure 8.11 for three different values of τ : 0.1,
0.01, and 0.001, while " = 0.01. The large τ = 0.1 smooths the trajectories at steady
state, which is expected because more frequency content of the noise is filtered out,
but this happens at the expense of a significant degradation of the transient response.
The small τ = 0.001 has almost no effect during the transient period, but there is sig-
nificant increase in the effect of noise at steady state. Of the three values of τ, the best
results are obtained with τ = 0.01. 4
Example 8.6. The effect of adding the low-pass filter 1/(τs + 1)2 is examined for the
system of Example 8.3 with the same controller, initial conditions, and measurement
noise parameters. Figures 8.13 and 8.14 show the trajectories of e1 , e2 , e3 , and u during
the transient and steady-state periods when τ = " = 0.005. For the steady-state period,
Figure 8.14 shows the deviation from the trajectories under output feedback control
with no measurement noise and no filter. Adding the filter has a small effect during the
transient period. At steady state, it smooths out the trajectories and makes a significant
reduction in the effect of noise on the trajectories of e3 and u. As in Example 8.3,
transfer functions can be used to interpret the results. Let H1 , H2 , H3 , and H4 be the
transfer functions from the noise v to e1 , e2 , e3 , and u, respectively. They are given by
0 0.5
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
Without LPF
−0.2
With LPF
−0.4
e1
e2
−0.6 0
−0.8
−1
−0.5
0 1 2 3 0 1 2 3
Time Time
5
200
100
0
e3
u
0
−5
−100
−200
−10
0 1 2 3 0 0.2 0.4 0.6 0.8 1
Time Time
Figure 8.13. Comparison of the trajectories of Example 8.6 with and without low-pass
filter (LPF) when the " = τ = 0.005.
∆e1
0 0
−1 −5
10 12 14 16 18 20 10 12 14 16 18 20
Time Time
−4 −4
x 10 x 10
5
2
∆e2
∆e2
0 0
−2
−5
10 12 14 16 18 20 10 12 14 16 18 20
Time Time
0.02
0.02
∆e3
∆e3
0 0
−0.02
−0.02
10 12 14 16 18 20 10 12 14 16 18 20
Time Time
10 2
∆u
∆u
0 0
−10 −2
10 12 14 16 18 20 10 12 14 16 18 20
Time Time
Figure 8.14. Comparison of the trajectories of Example 8.6 with and without low-pass
filter (LPF) when the " = τ = 0.005. The plotted curves are the differences from the trajectories under
output feedback with no measurement noise.
where
λ(s) = s 3 + s 2 + 1,
q(s, ") = (1 + 6" + 3"2 )s 2 + (2 + 3")s + 1,
χ (s , ", τ) = µ p("s )λ(s)ψ(τs) + q(s, "),
p("s) = ("s + 1)3 ,
ψ(τs) = (τs + 1)2 .
8.4. REDUCING THE EFFECT OF MEASUREMENT NOISE 271
H1 H2
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
50 50
0
0
Magnitude (dB)
Magnitude (dB)
−50
−100 −50
−150
−100
−200
−250 −150
0 2 4 0 2 4
10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)
H3 H4
100 150
50 100
Magnitude (dB)
Magnitude (dB)
0 50
−50 0
−100 −50
0 2 4 0 2 4
10 10 10 10 10 10
Frequency (rad/s) Frequency (rad/s)
Figure 8.15. Magnitude Bode plots of the transfer functions of Example 8.6 without low-
pass filter (solid) and with filter (dashed) when τ = " = 0.005.
0.5 0.04
0
0.02
−0.5
e1
e1
0
−1
τ = 0.05
−0.02
−1.5 τ = 0.005
τ = 0.0005
−2 −0.04
0 2 4 6 8 10 15 16 17 18 19 20
Time Time
The magnitude Bode plots of the four transfer functions are shown in Figure 8.15. In-
serting the low-pass filter causes the frequency response to roll off at a higher rate,
compared with the case without filter. This explains the attenuation of the high-
frequency content of the noise. Finally, Figures 8.16 and 8.17 compare the trajectories
of e1 and u when τ is increased from 0.005 to 0.05 and decreased to 0.0005. Increasing
τ deteriorates the transient response, while decreasing it deteriorates the steady-state
response. 4
In the previous two examples, the observer time constant " is chosen first to achieve
the desired recovery of the performance under state feedback. Then the filter time
constant τ is chosen of the order of ". A much larger τ could deteriorate the tran-
sient response of the closed-loop system, while a much smaller τ would deteriorate
the steady-state response by allowing higher-frequency content of the noise to influ-
ence the response.
272 CHAPTER 8. MEASUREMENT NOISE
τ = 0.05 τ = 0.05
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
200 2
100 1
u
0 0
−100 −1
−200 −2
0 1 2 3 15 16 17 18 19 20
Time Time
τ = 0.005 τ = 0.005
200
2
100
u
u
0 0
−100
−2
−200
0 1 2 3 15 16 17 18 19 20
Time Time
τ = 0.0005 τ = 0.0005
200 10
100 5
u
u
0 0
−100 −5
−200 −10
0 1 2 3 15 16 17 18 19 20
Time Time
hi (·)
d y − x̂1
The elements Ji of J satisfy 0 < Ji < 1. Similar to proof of Theorem 8.1 and the
derivations of Section 8.4.1, it can be shown that the inequality
b N +c d
|x̃i (t )| ≤ "ρ+1−i aM + ∀ t ≥T (8.120)
"i −1
is satisfied for some positive constants a, b , c, and T . Since d = kN , (8.120) shows
that, for sufficiently small "1 , x̃1 is ultimately bounded by O(N ). Knowing that the
ultimate bound on x̃1 is O(N ) does not guarantee that y − x̂1 = x̃1 + v will ultimately
be confined to the zone [−d , d ], where the higher parameter "2 is effective. However,
as we shall in the next two examples, that is usually the case. It is possible to choose
the parameters α1 to αρ to ensure that y − x̂1 will be confined to the zone [−d , d ] in
finite time, but that will complicate the observer design.74
Example 8.7. Reconsider the pendulum equation of Examples 3.4 and 8.2. The per-
formance of a nonlinear-gain observer with "1 = 0.01, "2 = 0.1, and d = 0.002 is com-
pared with two linear-gain high-gain observers with " = 0.01 and 0.1, respectively.
The controller, initial conditions, and measurement noise parameters are the same as
in Example 8.2. Figures 8.19 and 8.20 compare the trajectories of θ, ω, and u during
the transient period and at steady state. It is seen that the nonlinear-gain observer cap-
tures the performance of the linear-gain observer with " = 0.01 during the transient
period and the performance of the linear-gain observer with " = 0.1 at steady state.
74 See [120] for the choice of the parameters α to α to ensure that y − x̂ will be confined to the zone
1 ρ 1
[−d , d ]. They are chosen such that the polynomial s ρ + α1 s ρ−1 + · · · + αρ can be factored as
where λ is much larger than the magnitudes of the roots of s ρ−1 + β1 s ρ−2 + · · · + βρ−1 . Moreover, β1 to
βρ−1 are chosen such that the transfer function
has nonnegative impulse response. It is shown in [120] that such choices are always possible.
274 CHAPTER 8. MEASUREMENT NOISE
3 30
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
2.5
2.5 20
2
2
1.5
u
ω
10
θ
1.5
1
1
0
0.5
0.5
0 0 −10
0 2 4 0 2 4 0 2 4
Time Time Time
Figure 8.19. Comparison of the trajectories of Example 8.7 for a nonlinear-gain observer
with "1 = 0.01 and "2 = 0.1 (solid), a linear-gain observer with " = 0.01 (dashed), and a linear-gain
observer with " = 0.1 (dash-dotted).
−5 −5 −5
x 10 LG ε = 0.01 x 10 LG ε = 0.1 x 10 NLG
2 2 2
θ−π
0 0 0
−2 −2 −2
0 0 0
−1 −1 −1
19 19.5 20 19 19.5 20 19 19.5 20
Time Time Time
0 0 0
Figure 8.20. Comparison of the steady-state trajectories of Example 8.7 for a linear-gain
observer with " = 0.01 (left column), a linear-gain observer with " = 0.1 (middle column), and a
nonlinear-gain observer with "1 = 0.01 and "2 = 0.1 (right column).
The same is true for the estimation errors x̃1 = θ − π − x̂1 and x̃2 = ω − x̂2 , as shown
in Figures 8.21 and 8.22. Figure 8.23 shows that y − x̂1 = x̃1 +v enters the zone [−d , d ]
in finite time. 4
Example 8.8. Reconsider the tracking problem of Example 8.3. The performance
of a nonlinear-gain observer with "1 = 0.002, "2 = 0.01, and d = 0.0002 is compared
with two linear-gain high-gain observers with " = 0.002 and 0.01, respectively. The
controller, initial conditions, and measurement noise parameters are the same as in
Example 8.3. Figures 8.24 and 8.25 compare the trajectories of e1 , e2 , and e3 during the
transient period and at steady state. It is seen that the nonlinear-gain observer captures
8.4. REDUCING THE EFFECT OF MEASUREMENT NOISE 275
1 150
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
100
−1
x̃1
x̃2
50
−2
0
−3
−4 −50
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
Time Time
Figure 8.21. Comparison of the estimation errors of Example 8.7 for a nonlinear-gain
observer with "1 = 0.01 and "2 = 0.1 (solid), a linear-gain observer with " = 0.01 (dashed), and a
linear-gain observer with " = 0.1 (dash-dotted).
−4 −4 −4
x 10 LG ε = 0.01 x 10 LG ε = 0.1 x 10 NLG
2 2 2
1 1 1
x̃1
0 0 0
−1 −1 −1
−2 −2 −2
19 19.5 20 19 19.5 20 19 19.5 20
Time Time Time
0.01 0.01 0.01
0 0 0
Figure 8.22. Comparison of the steady-state estimation errors of Example 8.7 for a linear-
gain observer with " = 0.01 (left column), a linear-gain observer with " = 0.1 (middle column), and
a nonlinear-gain observer with "1 = 0.01 and "2 = 0.1 (right column).
−3
x 10
3
1
y − x̂1
−1
−2
−3
2 4 6 8 10
Time
Figure 8.23. The error term y − x̂1 = x̃1 +v of Example 8.7 for a nonlinear-gain observer
with "1 = 0.01, "2 = 0.1, and d = 0.002.
the performance of the linear-gain observer with " = 0.002 during the transient period
and the performance of the linear-gain observer with " = 0.01 at steady state. The same
is true for the estimation errors, as shown in Figures 8.26 and 8.27. Figure 8.28 shows
that y − r − ê1 = ẽ1 + v enters the zone [−d , d ] in finite time. 4
276 CHAPTER 8. MEASUREMENT NOISE
0.5 5
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
−0.2
0
−0.4 0
e1
e2
e3
−0.6 −5
−0.8 −0.5
−10
−1
−1.2
−1 −15
0 2 4 0 2 4 0 1 2
Time Time Time
Figure 8.24. Comparison of the trajectories of Example 8.8 for a nonlinear-gain observer
with "1 = 0.002 and "2 = 0.01 (solid), a linear-gain observer with " = 0.002 (dashed), and a linear-
gain observer with " = 0.01 (dash-dotted).
−4 −4 −4
x 10 LG ε = 0.002 x 10 LG ε = 0.01 x 10 NLG
0
−1 −1
−2 −2
∆e1
−0.5
−3 −3
−4 −4
−1 −5 −5
18 19 20 18 19 20 18 19 20
Time Time Time
−4 −4 −4
x 10 x 10 x 10
5 5 5
∆e2
0 0 0
−5 −5 −5
18 19 20 18 19 20 18 19 20
Time Time Time
0 0 0
Figure 8.25. Comparison of the steady-state trajectories of Example 8.8 for a linear-gain
observer with " = 0.002 (left column), a linear-gain observer with " = 0.01 (middle column), and a
nonlinear-gain observer with "1 = 0.002 and "2 = 0.01 (right column). The plots are the differences
between the trajectories under noisy output feedback and the trajectories under state feedback.
4
x 10
Downloaded 07/09/17 to 132.236.27.111. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
400 6
0.2
300
0 4
200
−0.2
2
ẽ1
ẽ3
ẽ2
100
−0.4
−0.6 0 0
−0.8 −100
−2
−1 −200
0 0.02 0.04 0.06 0 0.02 0.04 0.06 0 0.01 0.02 0.03
Time Time Time
Figure 8.26. Comparison of the estimation errors of Example 8.8 for a nonlinear-gain
observer with "1 = 0.002 and "2 = 0.01 (solid), a linear-gain observer with " = 0.002 (dashed), and
a linear-gain observer with " = 0.01 (dash-dotted).
−4 −4 −4
x 10 LG ε = 0.002 x 10 LG ε = 0.01 x 10 NLG
1 1 1
ẽ1
0 0 0
−1 −1 −1
18 19 20 18 19 20 18 19 20
Time Time Time
0.04 0.04 0.04
0 0 0
0 0 0
−5 −5 −5
18 19 20 18 19 20 18 19 20
Time Time Time
Figure 8.27. Comparison of the steady-state estimation errors of Example 8.8 for a linear-
gain observer with " = 0.002 (left column), a linear-gain observer with " = 0.01 (middle column),
and a nonlinear-gain observer with "1 = 0.002 and "2 = 0.01 (right column).
−4
x 10
2
y − r − ê1
−1
−2
Figure 8.28. The error term y − r − ê1 = e1 + v − ê1 of Example 8.8 for a nonlinear-gain
observer with "1 = 0.002, "2 = 0.01, and d = 0.0002.
278 CHAPTER 8. MEASUREMENT NOISE
[87], but it is more complicated for nonlinear systems. The nonlinear result of [121]
uses a result due to [143] to achieve the desired transformation. The low-pass filter
analysis of Section 8.4 is taken from [85]. The nonlinear-gain approach of the same
section is based on [120].
This chapter dealt only with the standard high-gain observer. Similar results can be
obtained for the reduced-order, extended, and cascade high-gain observers. It is impor-
tant to note that the cascade observer of Section 7.1 has no advantage over the standard
observer when it comes to measurement noise because all the saturation functions are
not active beyond the peaking period. Therefore, the observer reduces to the linear
cascade connection of Figure 7.1, whose transfer function is similar to the transfer
function of a linear standard observer. However, the cascade observer with feedback
injection of Section 7.3 has additional dynamics that act as a low-pass filter that reduces
the effect of high-frequency noise [13].