International Journal of Control, Automation, and Systems (2010) 8(5):994-1002 http://www.springer.

com/12555
DOI 10.1007/s12555-010-0508-x

Disturbance Observer for Non-minimum Phase Linear Systems

Nam Hoon Jo, Hyungbo Shim, and Young Ik Son*

Abstract: Disturbance observer (DOB) approach has been widely employed in the industry thanks to
its powerful ability to attenuate disturbances and compensate plant uncertainties. Motivated by the fact
that the application of the DOB approach has been limited to minimum phase systems, we propose a
new DOB configuration for non-minimum phase systems. Rigorous analysis for robust stability and
performance recovery is presented in terms of a new filter Θ (s). By restricting the plant uncertainty to
a multiplicative perturbation, we also present a systematic design methodology for the filter Θ (s) by
virtue of the H∞ synthesis technique. An illustrative example of autopilot design is presented to show
the effectiveness of the proposed method for which it is not easy to design H∞ controller to achieve
some control goals. The simulation results show that the response of the perturbed system in the pres-
ence of disturbance can be recovered to the nominal system response in the absence of disturbance.

Keywords: Disturbance observer, H∞ synthesis, linear system, non-minimum phase systems.

1. INTRODUCTION minimum phase systems are often met.

The goal of the paper is to propose a new DOB
Disturbance observer (DOB) approach [2-9] as a tool method that can be applied to non-minimum phase linear
for disturbance attenuation has been widely employed in systems. Although there exist some research works on
the industry. Versatility of DOB comes from, without DOB for non-minimum phase systems [11-14], they
doubt, its simplicity and powerful ability to attenuate have difficulties in designing controllers for general
disturbances and compensate plant uncertainties. An cases. In fact, the application of [11] is limited to those
interesting feature of DOB is that it can be used as an disturbances that come from a known external system
inner-loop controller so that the real uncertain plant in (i.e., exosystem). The authors of [13] presented no
the presence of disturbances is forced to behave like the systematic method to determine design parameters for
nominal plant in disturbance-free environment. In other robust stability and disturbance attenuation. Research
words, DOB can be combined with a pre-existing output works by [12] and [14] were focused on the estimation of
feedback controller in order to compensate for model disturbances only and did not deal with the controller
uncertainties and attenuate disturbances. However, synthesis using their DOBs. In contrast to those previous
applications of the DOB approach have been limited to works, we present a new DOB method that can be
the minimum phase systems or systems having no zero applied to general non-minimum phase linear systems.
dynamics. (It was proved in [10] that the minimum Rigorous analysis for robust stability and performance
phaseness of the plant is one of the necessary conditions recovery is presented in terms of a new filter Θ (s). One
for internal stability in case of the classical DOB might think that it is not necessary to develop a new
approach.) This prohibits applying the powerful DOB DOB approach that can be applied to non-minimum
technique into, e.g., the aerospace area where the non- phase systems, because there already exist several well-
__________
established robust control techniques. To illustrate the
benefit of the DOB method, suppose that we need to
Manuscript received August 31, 2009; revised February 17,
achieve two control goals simultaneously: 1) a certain
2010; accepted March 15, 2010. Recommended by Editorial
Board member Duk-Sun Shim under the direction of Editor Young control goal, and 2) plant uncertainty compensation and
Il lee. The preliminary version [1] of this paper was presented at disturbance attenuation. For the purpose of illustration,
American Control Conference, 2008. This work was financially we assume the former is given by time domain
supported by the advanced human resource development program specifications in terms of overshoot, settling time, and so
of MKE (Ministry of Knowledge and Economy) through the Re- on. Since H∞ method may be a good candidate for robust
search Center for Intelligent Microgrid in Myongji University.
Nam Hoon Jo is with the Department of Electrical Engineering,
control problem, we may try H∞ design method.
Soongsil University, 1-1 Sangdo-dong, Dongjak-ju, Seoul 156- However, the conventional H∞ controller designed to
743, Korea (e-mail: [email protected]). achieve the latter goal often exhibits a large overshoot.
Hyungbo Shim is with ASRI, School of Electrical Engineering, (Using an example of autopilot design, this will be
Seoul National University, Seoul 151-742, Korea (e-mail: hshim demonstrated in Section 4.) In such a case, by using the
@snu.ac.kr). H∞ method, it is not easy to achieve both goals
Young Ik Son is with the Department of Electrical Engineering,
Myongji University, Gyeonggi 449-728, Korea (e-mail: sonyi@ simultaneously. On the contrary, since the proposed
mju.ac.kr). DOB method has two degree of freedoms (one in an
* Corresponding author. outer-loop controller and another in the filter Θ( s )), it

© ICROS, KIEE and Springer 2010

Disturbance Observer for Non-minimum Phase Linear Systems 995

is not hard to achieve both at the same time. Moreover,

although it has two degree of freedoms, the design
procedure is simple and easy in the sense that it can be
separated.
The paper is organized as follows. In Section 2, a new
DOB configuration is presented and robust stability and
performance recovery criterion are given in terms of the
new filter Θ( s ). Section 3 provides a design procedure
for Θ( s ) based on the H∞ synthesis technique. Section
Fig. 1. The classical DOB configuration. The signals r,
4 illustrates an application of the proposed DOB to an d, and n represent the reference command, the
autopilot design and demonstrates the novelty of the unknown external disturbance, and the measure-
proposed DOB compared to other robust control ment noise, respectively. y is the output of the
techniques. Finally, some conclusions are given in closed-loop system while u is the feedback
Section 5. control to the plant.
2. A NEW DOB CONFIGURATION FOR NON-
MINIMUM PHASE SYSTEMS Nn (s)/Dn (s), and C(s)= Nc (s)/Dc (s). Then, regarding
robust internal stability of Fig. 1, it has been pointed out
In this section, we first give some intuition for a new in [16,10] that one of the necessary conditions for
DOB configuration and then present rigorous analysis for internal stability is that the polynomial
robust stability and performance recovery. ps ( s ) := N ( s ) ( Dc ( s ) Dn ( s ) + N c ( s ) N n ( s ) ) (2)
2.1. Limitation of classical DOB and new DOB config- is Hurwitz. This in turn implies that the (uncertain) plant
uration must be of minimum phase because N(s) is the numerator
The classical DOB configuration is illustrated in Fig. 1, of P(s).
which has been actively studied in, e.g., [2-6,10,15,16]. In order to overcome the restriction of the minimum
In the figure, P(s) is the uncertain plant, Pn(s) is its phaseness, a new configuration is devised as in Fig. 2(a).
nominal model, and C(s) is a controller designed a priori
for the nominal model Pn(s). The transfer function Q(s)
is called a ‘Q-filter’, which is a stable low-pass filter
with the unity DC gain. The role of the Q(s) is to make
the transfer function Q( s ) Pn−1 ( s ) proper so that it is
implementable and to avoid the algebraic loop in Fig.
1. From Fig. 1, the plant output y is obtained as
y ( s ) = Tyr ( s )r ( s ) + Tyd ( s )d ( s ) − Tyn ( s )n( s ), (1)
where (a)
Pn PC
Tyr = ,
Pn (1 + PC ) + Q( P − Pn )
Pn P(1 − Q )
Tyd = ,
Pn (1 + PC ) + Q ( P − Pn )
P(Q + Pn C )
Tyd = .
Pn (1 + PC ) + Q( P − Pn )

Since the low-pass filter Q( jω ) ≈ 1 and the measure- (b)

ment noise n( jω ) ≈ 0 in the low frequency range
(0, ω L ) with a certain constant ω L > 0, (1) becomes
Pn C
y ( jω ) ≈ ( jω )r ( jω ), ∀ω ∈ (0, ω L ).
1 + Pn C
This implies that, assuming that all the transfer functions
are stable, the closed-loop system with the DOB behaves
as if it were the disturbance-free nominal one in the
frequency range (0, ω L ). (c)
Let P, Pn, and C be represented by some ratios of Fig. 2. Intermediate diagrams towards the proposed
coprime polynomials, that is, P(s) = N(s)/D(s), Pn(s)= DOB configuration of Fig. 3.
996 Nam Hoon Jo, Hyungbo Shim, and Young Ik Son

In the figure, we imagine that the plant is P(s) +V(s) that u = ur, i.e., the inner-loop controller does nothing as
rather than P(s), with a certain filter V(s), which is if it were not there.
motivated by the fact that the zero dynamics of P(s) can From the proposed DOB configuration (Fig. 3), we
be affected by the parallel connection of V(s). Therefore, have
under the assumption that P(s) +V(s) is of minimum
phase, we may construct the classical DOB like in Fig. y ( s ) = Tyr ( s )r ( s ) + Tyd ( s )d ( s ) − Tyn ( s )n( s ), (3)
2(a) as if the plant were P(s) +V(s). In fact, this approach where
makes sense only when the signal d enters V(s) as well,
which is not the case. Although d cannot be injected into ( Pn Θ + 1) PC
Tyr = ,
V(s) in practice, we can approximately assume that d 1 + PC + (1 + Pn C ) PΘ
enters V(s) (as represented by the dotted line in Fig. 2(a))
Tyd = P ,
if | V ( jω ) | is small in the low frequency range where d 1 + PC + (1 + Pn C ) PΘ
is significant. Assuming this, we consider the
PC + (1 + Pn C ) PΘ
configuration in Fig. 2(b) instead of that in Fig. 2(a), Tyn = .
both of which are approximately equivalent. Then, Fig. 1 + PC + (1 + Pn C ) PΘ
2(c) is obtained from Fig. 2(b). Finally, by letting Note that Tyd must be kept small to minimize the effects
V ( s ) = 1/Θ( s) and by eliminating the Q-filter Q(s) in of disturbances. For the time being, let us assume that all
Fig. 2(c), we obtain the proposed configuration of Fig. 3. the transfer functions above are stable. In addition, it is
(Here, we could remove Q(s) because Θ( s ) in Fig. 3 natural to assume that there exists a frequency ω L such
plays the same role as Q(s), that is, making the transfer that the measurement noise n( jω ) is negligible on
function 1/ (1 + Pn ( s )Θ( s )) proper and avoiding the (0, ω L ) while the disturbance d ( jω ) and the reference
algebraic loop.) Therefore, Θ( s ) should be designed to r ( jω ) are significant on (0, ω L ).
be a proper transfer function such that | Θ( jω ) | is large Now, suppose that the filter Θ( s ) is designed so that,
at low frequencies.
for all ω ∈ (0, ω L ),
2.2. Performance recovery and robust stability | Θ( jω ) |≈ ∞, (4)
We assume that P(s) is unknown but is an element of
the set P that is a collection of strictly proper rational or
transfer functions. The nominal model Pn(s) is also
assumed to belong to P. An outer-loop controller C(s) is 1 C ( jω ) 1
≈ 0, ≈ 0, ≈ 0. (5)
assumed to be designed a priori for the nominal model Θ( jω ) Θ( jω ) P ( jω )Θ( jω )
Pn(s) so that the nominal closed-loop system (i.e.,
Pn ( s )C ( s ) Pn C + C Pn C
) is stable and has satisfactory perform- Then, we get Tyr ( jω ) = Θ ≈ ,
1 + Pn ( s)C ( s )
(1 + Pn C ) + 1 + C (1 + Pn C )
ances for a certain control goal. However, due to the PΘ Θ
uncertainty of the real plant and the external disturbance Tyd ( jω ) ≈ 0, and Tyr ( jω ) ≈ 1, which implies that
d, the actual closed-loop performance could be degraded
or, even worse, the stability could be lost. The role of the Pn ( jω )C ( jω )
inner-loop controller (i.e., systems represented by y ( jω ) ≈ r ( jω ), ∀ω ∈ (0, ω L ). (6)
1 + Pn ( jω )C ( jω )
Pn Θ Θ , with a certain filter Θ( s ), in
and Remark 1 (Performance recovery): Since Tyd ( jω )
1 + Pn Θ 1 + Pn Θ
Fig. 3) is to maintain the stability of the actual closed- ≈ 0, the effects of disturbances d on the plant output y
loop system and preserve the nominal performance even can be negligible. Moreover, (6) indicates that, even
in the presence of the plant uncertainty and disturbance d. when there exist plant uncertainty and input disturbance,
Note that, if Pn= P and d = n = 0, then it is easily seen the steady-state performance of the actual closed-loop
system can be recovered to the nominal one in the
absence of disturbance. (More rigorous justification of
performance recovery for the conventional DOB can be
found in [10,16].)
In order to enjoy this performance recovery, the filter
Θ( s ) should be designed such that robust internal
stability of the closed-loop system (Fig. 3) is ensured
while | Θ( jω ) | is sufficiently large in the low
frequency range.
To inspect internal stability, nine transfer functions
Fig. 3. The proposed DOB configuration (the boxed
block) with an outer-loop controller C(s). from [r, d , n]T to [e , u , y ]T in Fig. 3 are computed by
Disturbance Observer for Non-minimum Phase Linear Systems 997

 1 + PΘ, − P, −1  thus, δ ( s ) is Hurwitz for all P( s) ∈ P.

1  ,
( P Θ + 1)C , 1, − (1 + P C ) Θ − C Assumption (H1) is nothing but the stability
∆ cl ( s )  
n n
( Pn Θ + 1) PC , P,  requirement of the nominal closed-loop system with an
1
outer-loop controller C(s). Therefore, with such a C(s)
where ∆ cl ( s ) = (1 + PC ) + (1 + Pn C ) PΘ. If the above designed a priori, the remaining task is to design Θ( s )
nine transfer functions are stable for all P( s) ∈ P, then such that the assumptions (H2) and (H3), which play the
role of robust stability criterion, are satisfied together
the closed-loop system is said to be robustly internally
stable. with the following performance recovery criterion:
Now let us write Θ as ratios of coprime polynomials (H4) | Θ( jω ) | 1, ∀ω ∈ (0, ω L ).
Θ( s ) = Nθ ( s ) /Dθ ( s ), and recall that P ( s ) = N ( s ) /D( s ),
However, the conditions (H2) and (H3) should be
Pn ( s ) = N n ( s ) /Dn ( s ), and C ( s ) = N c ( s ) /Dc ( s ). Then,
satisfied for all P( s) ∈ P, which may make it still
it can be shown that the closed-loop system is internally
difficult to obtain Θ( s ). In the next section, by
stable if and only if the characteristic polynomial
restricting our interest to a multiplicative perturbation
δ ( s ) := Dθ ( Dn Dc D + Dn N c N ) model of the plant, we present a systematic way to
(7)
+ Nθ ( Dn Dc N + N n N c N ) design Θ( s ) by virtue of the H∞ synthesis tool.
is Hurwitz. We summarize the discussions so far in the
following. 3. CONSTRUCTION OF Θ( s ) FOR THE PLANT
Theorem 1: Let P be a set of strictly proper rational HAVING MULTIPLICATIVE UNCERTAINTY
transfer functions. The closed-loop system in Fig. 3 is
robustly internally stable if and only if δ ( s ) is Hurwitz Throughout this section, we assume that the uncertain
plant transfer function belongs to a set P := {P = (1 +
for all P( s) ∈ P.
Unfortunately, Theorem 1 is not convenient to design ∆W2 ) Pn : ∆( s ) ∞ ≤ 1, ∆( s) is stable and rational},
Θ( s ) for robust internal stability. To overcome this where Pn(s) is a strictly proper rational transfer function,
difficulty, a more viable but sufficient condition is given and W2(s) is a fixed stable (possibly improper) transfer
as follows. function such that every P(s) in P is strictly proper.
Theorem 2: The closed-loop system in Fig. 3 is Pn ( s )
Theorem 3: Let W1 ( s ) = 2W1∞ , and
robustly internally stable if the following conditions hold. ( s /ω L + 1)2
(H1) Pn C / (1 + Pn C ) is stable, Pn C
W 2 = 1 + P C W2 , where W1∞ is a sufficiently large
(H2) PΘ/ (1 + PΘ) is stable for all P( s) ∈ P, n

C ( P − Pn ) positive constant such that min W1∞ | Pn ( jω ) | 1.

(H3) < 1 for all P ( s) ∈ P. ω ∈[0,ω L )
(1 + Pn C )(1 + PΘ) ∞ Then, the conditions (H2), (H3), and (H4) hold if Θ( s )
Proof: By manipulating (7), we obtain stabilizes1 Pn and satisfies

δ ( s ) = ( Dn Dc + N n N c )( Dθ D + Nθ N ) 1
W1 < 1, (9)
+ Dθ N c ( NDn − N n D) (8) 1 + Pn Θ ∞
= ( Dn Dc + N n N c )( Dθ D + Nθ N ) (1 + Φ( s) ) , 1 PΘ
W 2 + W2 n < 1. (10)
Dθ N c ( NDn − N n D) 1 + Pn Θ 1 + Pn Θ
∞
where Φ( s ) := . Note that,
( Dn Dc + N n N c )( Dθ D + Nθ N ) Although the conditions (9) and (10) may look more
from assumptions (H1) and (H2), ( Dn Dc + N n N c ) and complicated than (H2), (H3), and (H4), these conditions
( Dθ D + Nθ N ) are Hurwitz, which implies that Φ( s ) do not include the statement ‘for all P( s) ∈ P ’.
is stable. Moreover, assumption (H3) implies that Moreover, they are in the standard form frequently used
in the H∞ synthesis solver (e.g., MATLAB Robust
Nc  N N n  Control Toolbox). For convenience, define
−
Dc  D Dn 
= Φ ∞ < 1. 1 PΘ
 N n N c  N Nθ  L = Pn Θ, SΘ = , TΘ = n .
1 + D D  1 + D D  1 + Pn Θ 1 + Pn Θ
 n c  θ 
∞
1
Thus, it follows from the Nyquist criterion or the small- Θ( s ) is said to stabilize Pn(s) if the unity feedback system
composed of Pn(s) and Θ( s ) is stable, i.e., the transfer function
gain theorem that (1 + Φ( s)) −1 is stable. This implies
Pn Θ
that the denominator of (1 + Φ( s )) −1 is Hurwitz, and 1 + Pn Θ
is stable.
998 Nam Hoon Jo, Hyungbo Shim, and Young Ik Son

Then, (9) and (10) are rewritten as low frequencies because TΘ ( jω ) ≈ 1 there and (12)
W1SΘ ∞ < 1, (11) should be satisfied. Secondly, W2(s) should be chosen
such that Pn ( s )C ( s )W2 ( s ) is either strictly proper, or
| W2 SΘ | + | W2TΘ |∞ < 1, (12)
biproper and less than unity at high frequencies. To see
respectively. It will be shown in the proof that (11) this, since Pn ( s ) is strictly proper, observe that
implies (H4) while (12) implies (H2) and (H3). Thus, SΘ ( jω ) ≈ 1 and W 2( jω ) ≈ Pn ( jω )C ( jω )W2 ( jω ) at
(11) and (12) are called ‘performance recovery
high frequencies. This, together with (14), implies that
condition’ and ‘robust stability condition’, respectively.
| Pn ( jω )C ( jω )W2 ( jω ) |< 1 at high frequencies.
Proof of Theorem 3: (H4) follows from (11), that is,
Choosing W2 as in Remark 2, the standard H∞ solver
W1 can be used for finding a suitable Θ( s ) that stabilizes
W1SΘ ∞ < 1 ⇔ < 1, ∀ω
1+ L the nominal plant Pn and meets (11) and (12). (This will
⇔ 1+ | L |≥| 1 + L |>| W1 |, ∀ω be illustrated further in Section 4.) Once such a Θ( s ) is
⇒| Pn || Θ |≥| L |> W1 − 1 ≥ W1∞ Pn − 1, found, then, under the condition (H1), the closed-loop
system in Fig. 3 is robustly internally stable and
≈ W1∞ Pn , on (0, ω L ) performance recovery criterion (H4) is met.
⇒| Θ |≥ W1∞ 1, on (0, ω L ). Remark 3: Although the set P has been limited to a
collection of strictly proper P(s), all the developments so
On the other hand, (10) implies that W2TΘ ∞ < 1 and far are also applicable to the set of proper transfer
functions P(s) if both C(s) and Θ( s ) are strictly proper.
W2 SΘ
< 1. (13) (The reason for the restriction of strictly proper P(s) is to
1− | W2TΘ | ∞ guarantee the well-posedness of the control system in Fig.
3. However, since the strictly proper C(s) and Θ( s )
Since 1 + PΘ = 1 + (1 + ∆W2 ) L = (1 + L)(1 + ∆W2TΘ ), it
also guarantee the well-posedness, the set P could be just
follows from W2TΘ ∞ < 1 that the net change in the proper although it is hard to check the strict properness
angle of 1 + PΘ( s ) equals the net change in the angle of of Θ( s ) a priori.) This fact allows a larger class of
1 + Pn Θ( s ), as s goes around the Nyquist contour. W2(s) to enter in the expression of the set P in the sense
Hence, by Nyquist argument, Θ stabilizes every plant that P = (1 + ∆W2 ) Pn could be biproper as well as
in P, which implies the assumption (H2). Now, from (13), strictly proper.
it follows that
4. CASE STUDY: AUTOPILOT DESIGN FOR
W2 SΘ UNMANNED AERIAL VEHICLE
< 1.
1 + ∆W2TΘ ∞
In order to illustrate the proposed DOB technique, we
Thus, we have consider an altitude-hold autopilot design problem for
the Tower Trainer 60 unmanned aerial vehicle studied in
W2 1 [17] and [18]. The nominal transfer function is given by
1 + Pn Θ
<1
PΘ h( s )
1 + ∆W2 n Pn ( s ) :=
1 + Pn Θ ∞
δ e (s)
1 −34.16s 3 − 144.4 s 2 + 7047 s + 557.2
⇔ W2 <1 = ,
1 + Pn Θ + ∆W2 Pn Θ ∞
s + 13.18s 4 + 95.93s 3 + 14.61s 2 + 31.94s
5

Pn C W2 where h( s) and δ e ( s ) represent an altitude and an

⇔ <1 elevator angle, respectively. We assume that all possible
1 + Pn C 1 + Pn (1 + ∆W2 )Θ ∞
disturbance signals | d ( jω ) | are relatively small over
C PnW2 the frequency range (0.01,∞), from which ω L can be
⇔ < 1.
1 + Pn C 1 + PΘ ∞ chosen as ω L = 0.01. Note that the plant is of non-
Since ∆ ∞ ≤ 1, we obtain minimum phase, and thus, the classical DOB approach
cannot be applied.
C Pn ∆W2
< 1. 4.1. A comparison with H∞ design method
1 + Pn C 1 + PΘ ∞ Now, suppose that our control goals are as follows: 1)
Thus, since ∆ is arbitrary, this implies assumption (H3). overshoot is less than 1% and | δ e (t ) |< π/ 2 for all
Remark 2: A remark on the feasibility of (12) is in t ≥ 0, and 2) the plant output should not be affected too
order. Firstly, | W2 ( jω ) | should be less than unity at much by plant uncertainty and input disturbance. We
Disturbance Observer for Non-minimum Phase Linear Systems 999

first try H∞ design method by focusing on the second

control goal. The weights for H∞ mixed sensitivity
synthesis are chosen as W 1( s ) = 1 ,
( s + 10−6 )( s + 0.01) 2
s
W 2( s ) = 0.65, W 3( s ) = 5 and a controller C1(s) is sought
that maximizes γ such that [γ W1S ,W2 R , W3T ] < 1,
∞
where S ( s ) = 1/ (1 + Pn ( s)C1 ( s )), R ( s) = C1 ( s ) S ( s), and
T ( s ) = Pn ( s )C1 ( s) / (1 + Pn ( s )C1 ( s)). The resulting H ∞
controller is given by C1 ( s ) = N1 ( s ) /D1 ( s ), where
N1 ( s ) = 1773s 7 + 2.60 ⋅104 s 6 + 2.06 ⋅105 s 5 + 3.02 ⋅105 s 4 +
2.76 ⋅105 s 3 + 1.11 ⋅105 s 2 + 6.03 ⋅104 s + 1.89 ⋅10−3 and
D1(s)=s8+1187s7+ 4.05 ⋅104 s 6 + 7.05 ⋅105 s 5 + 4.92 ⋅106 s 4 Fig. 4. Step response in the face of plant uncertainty.
Although not shown here, elevator angle for
+ 4.81⋅105 s 3 + 8145s 2 + 38.28s + 3.83 ⋅10−5. The per-
C1(s) results in a large surge, i.e., max | δ e (t ) |
formance of C1(s) is tested for the perturbed plant 0≤ t ≤ 30
=1.43.
h( s )
P ( s ) :=
δ e (s)
(14)
−34.16s 3 − 62.64 s 2 + 8252 s + 715.9
= ,
s5 + 10.79 s 4 + 48.61s3 + 7.852 s 2 + 15.96s
which was used for the simulation study in [18]. Fig. 4
shows the step response and Fig. 5 shows the system
response in the presence of input disturbance d1 (t ) =
0.005[us (t − 30) − us (t − 50)], where us (t ) represents
the unit step function. Although the disturbance is
effectively attenuated on t ∈ [30,100] (Fig. 5, blue
dash-dot line), C1(s) results in a large overshoot in h(t)
and some initial surge in δ e (t ) (Fig. 4, blue dash-dot
line). Of course, we may reduce the overshoot in h(t) by
choosing a smaller W2 ( s ), e.g., W2 ( s) = 0.1. But, this Fig. 5. System response against input disturbance d1(t)
in turn leads to a larger initial surge in δ e (t ), which in the presence of plant uncertainty.
violates | δ e (t ) |< π/ 2. On the other hand, to achieve the
first goal, we design a new H∞ controller C2(s) by controller C2(s). Then, C2(s) can accomplish the first
choosing another weights. By selecting W 1( s ) = ( s 2 + goal and the added DOB structure provide an additional
ability of achieving the second goal. (See Remark 1.) In
1.84s + 0.846) / (0.001s 3 + 1.0018s 2 + 1.84s + 1.84 ⋅10−6 ),
order to apply the proposed DOB, we need to compute
W 2( s ) = 1, W 3( s ) = s / 5, we obtain C2(s)=N2(s)/D2(s), Θ( s ) that stabilizes Pn ( s ) and maximizes γ such
where N2(s)=10.9s7+2265s6+ 4.35 ⋅104 s 5 + 3.95 ⋅105 s 4 + that [γ W S , W S , W T ] < 1. By letting W ( s ) =
1 Θ 2 Θ 2 Θ ∞ 1
1.43 ⋅106 s 3 + 2.80 ⋅105 s 2 + 4.64 ⋅105 s +0.01453 and D2(s)
1 and W2 ( s) = s , we obtain Θ( s ) = Nθ ( s )
=s8+1053s7+ 5.38 ⋅104 s 6 + 1.32 ⋅106 s 5 + 2.04 ⋅107 s 4 + s ( s + 0.01) 2 5
1.58 ⋅108 s 3 + 2.38 ⋅108 s 2 + 1.78 ⋅107 s + 17.85. As seen in
/Dθ ( s ), where Nθ ( s ) = 6.22 ⋅106 s 7 + 9.22 ⋅107 s 6 +
Fig. 4, the step response of the nominal closed-loop
system in the absence of disturbance (green dotted line) 7.39 ⋅108 s 5 + 1.18 ⋅109 s 4 + 1.13 ⋅109 s 3 + 4.46 ⋅108 s 2 +
is quite excellent. However, it is seen from Fig. 5 that the 2.61⋅108 s +14.53, and Dθ ( s ) =s8+1.31·104s7+4.29·107s6
plant output under C2(s) is severely deteriorated by the
disturbance d1(t). Thus, neither C1 ( s ) nor C2(s) can + 1.62 ⋅109 s 5 + 1.51 ⋅1010 s 4 + 1.49 ⋅109 s 3 + 2.52 ⋅107 s 2 +
achieve both goals simultaneously, and it seems that it is 1.18 ⋅105 s − 3.12 ⋅10−7.
difficult to design H ∞ controller that achieves both. Remark 4: From a practical point of view, the choice
of W1 is not important as long as (5) or (H4) is met. So,
Now, we turn to the proposed DOB controller. In fact,
the proposed DOB method provides a very simple it is usual to try several W1(s), simpler than one in
solution: Just add the proposed DOB structure to Theorem 3, so that the resulting Θ( jω ) makes
1000 Nam Hoon Jo, Hyungbo Shim, and Young Ik Son

Fig. 7. Simulation results in the face of plant uncer-

Fig. 6. The magnitude Bode plots of 1 , C ( s) , tainty.
Θ( s) Θ( s )
1 and | W2 SΘ | + | W2TΘ | .
Pn ( s)Θ( s)

min | Θ( jω ) | as large as possible while the robust

ω ∈(0,ω L )
stability condition (12) is preserved.
1 , C ( s)
The magnitude Bode plots of ,
Θ( s ) Θ( s )
1 and | W2 SΘ | + | W2TΘ | are depicted in Fig.
Pn ( s )Θ( s )
6, which confirms (5) and indicates that | Θ( jω ) |>
60.65dB =1077 for all ω ∈ (0, 0.01) Moreover, it is also
seen from Fig. 6 that | W ( jω ) S ( jω ) | + W ( jω )
2 Θ 2

TΘ ( jω ) < 1 for all ω ∈ (0,∞). Therefore, because of Fig. 8. Simulation results in the presence of plant
uncertainty and input disturbance d1(t).
Theorem 2 and Theorem 3, both robust stability criterion
and performance recovery criterion are satisfied.
order to apply the proposed DOB, we need to compute
The responses of the controller C2 ( s ) combined with
Θ( s ). By letting W1 ( s ) = 1 and W2 ( s )
the proposed DOB are again shown in Figs. 4 and 5. As ( s + 1)( s + 0.01)
suggested by Remark 1, the plant output under the
proposed controller (Fig. 4, red solid line) is almost = s , we obtain Θ( s ) = Nθ ( s ) /Dθ ( s ), where Nθ ( s ) =
5
indistinguishable from that of the nominal closed-loop
system (Fig. 4, green dotted line). Moreover, compared 2.56 ⋅106 s 6 + 3.92 ⋅107 s 5 + 3.17 ⋅108 s 4 + 5.59 ⋅108 s 3 +
to the plant output under C2 ( s ) only, the disturbance 1.61 ⋅108 s 2 + 1.74 ⋅108 s + 4.04, and Dθ ( s ) =s7+9425s6
attenuation performance is greatly improved by using the + 1.78 ⋅107 s 5 + 6.88 ⋅108 s 4 + 6.97 ⋅109 s 3 + 6.86 ⋅109 s 2 +
proposed DOB (Fig. 5). In fact, even though the
5.61⋅108 s + 4.93 ⋅106.
disturbance and plant uncertainty are present, the
For the purpose of performance comparison with [18],
proposed DOB controller achieves both control goal
the plant is assumed to asymptotically track the output of
simultaneously.
a simple reference model with the transfer function
4.2. A comparison with adaptive controller M ( s ) = 1/ (1 + s / 2). The simulations are carried out using
In this subsection, the performance of the proposed the perturbed plant (14). Fig. 7 shows the simulation
DOB controller is compared with simple-adaptive- results when unit step input is applied to the reference
control (SAC) scheme presented in [18]. For a fair model M ( s ). Because of plant uncertainty, C3 ( s )
comparison, we choose the outer-loop controller as same controller only (blue dash-dot line) and SAC scheme
as in [17] and [18], i.e., C3(s)=N3(s)/D3(s) with N3(s)= (blue dashed line) yield some overshoots over the time
0.00842(s +7.895)(s2 +0.108s+0.3393) and D3(s) = (s + interval [0,8]. On the contrary, the plant output under the
0.07895)(s2 + 4s +8). We will show that the proposed proposed DOB controller (red solid line) is almost
DOB can provide C3 ( s ) with an additional ability of indistinguishable from that of the nominal closed-loop
uncertainty compensation and disturbance attenuation. In system (green dotted line).
Disturbance Observer for Non-minimum Phase Linear Systems 1001

Fig. 8 shows the simulation results when input disturb- [9] X. Wei, H. Zhang, and L. Guo, ‘‘Saturating compo-
ance d1(t) is present. Although the disturbance and plant site disturbance-observer-based control and H ∞
uncertainty are present, the response of the proposed control for discrete time-delay systems with nonli-
DOB controller is almost similar to the nominal response nearity,” Int. J. of Control, Automation, and Sys-
in the absence of disturbance. Moreover, it is seen that tems, vol. 7, pp. 691-701, 2009.
the proposed DOB controller suppresses the disturbance [10] H. Shim and N. H. Jo, ‘‘An almost necessary and
much more effectively than the SAC scheme. sufficient condition for robust stability of closed-
loop systems with disturbance observer,” Automati-
5. CONCLUSIONS ca, vol. 45, pp. 296-299, 2009.
[11] W. C. Yang and M. Tomizuka, ‘‘Disturbance rejec-
We have proposed a new DOB method that can deal tion through an external model for nonminimum
with the non-minimum phase linear systems. Robust phase systems,” ASME J. of Dynamic Systems,
stability and performance recovery are analyzed in terms Measurement and Control, vol. 116, no. 1, pp. 39-
of Θ( s ). We also presented a design methodology for 44, 1994.
Θ( s ) using the H∞ synthesis technique. For any [12] X. Chen, G. Zhai, and T. Fukuda, ‘‘An approximate
controller C(s) designed a priori, the proposed DOB inverse system for nonminimum-phase systems and
controller provides C(s) with an additional ability of its application to disturbance observer,” Systems &
disturbance attenuation and uncertainty compensation Control Letters, vol. 52, pp. 193-207, 2004.
without hurting the role of C ( s ). This enables the [13] N. Bajcinca and T. Bunte, ‘‘A novel control struc-
ture for dynamic inversion and tracking tasks,”
following simple design steps: First, just design a Proc. of 16th IFAC World Congress, 2005.
controller that achieves a certain control goal in the [14] J. Chang, ‘‘Applying discrete-time proportional
absence of uncertainties and disturbances. Next, simply integral observers for state and disturbance estima-
add the proposed DOB structure to the controller. tions,” IEEE Trans. Automat. Contr., vol. 51, no. 5,
pp. 814-818, 2006.
REFERENCES [15] T. Umeno, T. Kaneko, and Y. Hori, ‘‘Robust servo-
[1] H. Shim, N. H. Jo, and Y. I. Son, ‘‘A new distur- system design with two degrees of freedom and its
bance observer for non-minimum phase linear sys- application to novel motion control of robot mani-
tems,” Proc. of American Control Conference, pp. pulators,” IEEE Trans. Ind. Electron., vol. 40, pp.
3385-3389, 2008. 473-485, 1993.
[2] K. Ohishi, K. Ohnishi, and K. Miyachi, ‘‘Torque- [16] H. Shim and Y. Joo, ‘‘State space analysis of dis-
speed regulation of dc motor based on load estima- turbance observer and a robust stability condition,”
tion method,” Proc. of the International Power Proc. of Conf. on Decision and Control, pp. 2193-
Electronics Conference, 1983. 2198, 2007.
[3] H. S. Lee and M. Tomizuka, ‘‘Robust motion con- [17] K. Cohen and D. E. Bossert, ‘‘Fuzzy logic non-
troller design for high-accuracy positioning sys- minimum phase autopilot design,” Proc. of AIAA
tems,” IEEE Trans. Ind. Electron., vol. 43, pp. 48- Guidance, Navigation, and Control Conference ,
55, 1996. AIAA Paper 2003-5550.
[4] C. J. Kempf and S. Kobayashi, ‘‘Disturbance ob- [18] I. Barkana, ‘‘Classical and simple adaptive control
server and feedforward design for a high-speed di- for nonminimum phase autopilot design,” J. Guid-
rect-drive positioning table,” IEEE Trans. on Contr. ance, Control, and Dynamics, vol. 28, no. 4, pp.
Sys. Tech., vol. 7, no. 5, pp. 513-526, 1999. 631-638, 2005.
[5] E. Schrijver and J. van Dijk, ‘‘Disturbance observ-
ers for rigid mechanical systems: equivalence, sta-
bility, and design,” J. of Dynamic Systems, Mea-
Nam Hoon Jo received his B.S., M.S.,
surement, and Control, vol. 124, pp. 539-548, 2002.
and Ph.D. degrees from School of Elec-
[6] Y. Choi, K. Yang, W. K. Chung, H. R. Kim, and I. trical Engineering, Seoul National Uni-
H. Suh, ‘‘On the robustness and performance of versity, Seoul, Korea, in 1992, 1994 and
disturbance observers for second-order systems,” 2000, respectively. From 2000 to 2001,
IEEE Trans. Automat. Contr., vol. 48, pp. 315-320, he was a Post-doctoral Research Asso-
2003. ciate at Automation and Systems Re-
[7] J. Back and H. Shim, ‘‘Adding robustness to no- search Institute (ASRI) at Seoul National
minal output feedback controllers for uncertain University, Seoul, Korea. From 2001 to
nonlinear systems: a nonlinear version of distur- 2002, he worked as a Senior Research Engineer at Samsung
Electronics, Suwon, Korea. Since March 2002, he has been
bance observer,” Automatica, vol. 44, pp. 2528-
with the Department of Electrical Engineering at Soongsil Uni-
2537, 2008. versity, Seoul, Korea, where he is currently an associate profes-
[8] K.-W. Song and G.-S. Choi, and G.-H. Choi, sor. His research interests include nonlinear systems control
‘‘Tracking position control of DC servo motor in theory and its applications to systems biology, and artificial
LonWorks/IP network,” Int. J. of Control, Automa- intelligence with its application to electric load forecasting. For
tion, and Systems, vol. 6, pp. 186-193, 2008. more details please visit http://ee.ssu.ac.kr/~nhjo
1002 Nam Hoon Jo, Hyungbo Shim, and Young Ik Son

Hyungbo Shim received his B.S., M.S., Young Ik Son received his B.S., M.S.,
and Ph.D. degrees from Seoul National and Ph.D. degrees from Seoul National
University, Korea, in 1993, 1995 and University, Korea, in 1995, 1997 and
2000, respectively. From 2000 to 2001 he 2002, respectively. In 2002, he joined the
was a post-doctoral researcher at the faculty of the Department of Electronic
Center for Control Engineering and Engineering, Dong-A University, Korea.
Computation, University of California, From 2007 to 2008 he was a visiting
Santa Barbara. In 2002, he joined the scholar at Cornell University, New York.
faculty of the Division of Electrical and From 2003, he has been with the De-
Computer Engineering, Hanyang University, Seoul, Korea. partment of Electrical Engineering at Myongji University, Ko-
From 2003, he has been with the School of Electrical Engineer- rea, where he is currently an associate professor. His research
ing at Seoul National University, Korea, where he is currently interests include robust controller design and its application to
an associate professor. His research interests include analysis industrial electronics.
and control of nonlinear systems. He is serving as an Associate
Editor on European Journal of Control and International Jour-
nal of Robust and Nonlinear Control from 2008 and 2009, re-
spectively.