Professional Documents
Culture Documents
VOL. , NO. ,
http://dx.doi.org/./..
system. Moreover, this approach needs the knowledge of the the solutions of the signal model
actual disturbance input locations for the design. The structure
of the new compensator has some advantages when compared to
Davisons approach. By driving the observer with the saturated v = Sv, v (0) = v 0 Rnv (3a)
input, the unstable compensator can easily be stabilised in the d = Pd v (3b)
presence of input saturation. Thereby, the unstable signal model r = Pr v (3c)
is contained in the disturbance observer and therefore is also
stabilised. This completely removes the controller windup by a
very simple structural measure. Moreover, even global asymp- (see, e.g. Knobloch, Isidori, & Flockerzi, 1993, Chapter 1).
totic stability of the closed-loop system with input saturation can In order to take bounded and persistently acting signals into
be systematically ensured by a suitable design of the state feed- account, the eigenvalues of the matrix S Rnv nv are located
back gain. Another advantage of the proposed controller is that on the imaginary axis. This includes the modelling of, e.g. sinu-
a possibly worse reference tracking behaviour resulting from the soidal or step-like disturbances and reference signals.
application of the internal model principle can be improved by The considered problem is to design a compensator that sta-
a very simple separate reference channel design. bilises the closed-loop system such that
It is demonstrated that the existence conditions for the pro-
posed robust output feedback regulator coincide with Davisons
classical approach and that the corresponding design effort of lim (y(t ) r(t )) = 0, x(0) Rn , v (0) Rnv (4)
t
both methods is the same. This implies that there exist two dif-
ferent solutions of the robust output regulation problem. In par-
ticular, the resulting compensators from the usual and from the holds for all initial values of the compensator. This property has
new approach are dual to each other. However, since the new to be robust with respect to plant parameter variations which
compensator is superior with respect to windup prevention and do not lead to an unstable closed-loop system. Furthermore, the
the reference signal tracking, it should be preferred in applica- compensator design has to be independent from the disturbance
tions. input locations described by the matrix G.
The next section gives a formulation of the considered prob- In the following, the posed robust output regulation problem is
lem. Afterwards, the classical observer-based feedforward con- solved by generalising the classical observer-based feedforward
troller for SISO systems is generalised to ensure robust out- controller. This has the advantage that the undesirable effects
put regulation. This result is extended to MIMO systems in caused by the input saturation (2) can be prevented by a sys-
the following sections. To this end, a short review of Davisons tematic and very simple measure.
approach is given in Section 4. The new robust output regulator As the SISO case allows a very straightforward generalisation
design is presented in Section 5, where the duality to Davisons of the observer-based feedforward controller to ensure robust
approach is used to verify the internal model principle. Subse- output regulation, it will be considered first.
quently, it is demonstrated that a straightforward separate ref-
erence channel design is possible for the new compensator. In
3. Generalisation of observer-based feedforward
Section 6, controller windup is prevented by a straightforward
control for SISO systems
structural measure. A simple example illustrates the advantages
Consider the SISO system
of the new controller design procedure.
x = Ax + Bu + Gd, x(0) = x0 Rn (1a) with the input u(t ) R and the measurable output y(t ) R.
The disturbance d(t ) R in (5a) is modelled by (3a) with d =
y = Cx, (1b) pT v. In order to asymptotically reject this disturbance, a feedfor-
ward of the signal model states v of (3a) along with a stabilising
with the saturated input u = sat(u) Rm , m 1, in which the state feedback
input saturation is defined by
u = kTx (x v ) + T v (6)
max
ui : ui > umax
i
sat(u) = ui : umin ui umax , umax > umin (2) is designed. It is well known that (, T ) in (6) have to be the
min i
min
i i i
ui : ui < ui , i = 1, 2, . . . , m. solution of the regulator equations
(see Knobloch et al., 1993, Theorem 1.3.1) that uniquely exists yielding the modified compensator
if
v = Sv + lv (y cT x) (13a)
A sI
x = Ax + bu + g pT v + lx (y cT x)
b
det 0,
= s (S) (8) (13b)
cT 0
u= kTx (x v ) + v.
T
(13c)
(see Knobloch et al., 1993, Corollary A.1.2). Thereby, (S)
Since the input saturation of the plant and of the disturbance
denotes the spectrum of S. The feedback gain kTx in (6) is deter-
observer cancel in the corresponding observer error dynamics,
mined such that A bkTx is a Hurwitz matrix, which is possible
it reads
for (A, b) being controllable. In order to implement (6), a distur-
bance observer is designed, where the measurement coincides
ev S lv cT ev
with the output to be controlled in (5b). This leads to the (clas- = , (14a)
ex A bkTx lx cT ex
sical) observer-based feedforward controller
with ev = v v and ex = x x so that the compensator (13)
v = Sv + lv (y cT x) (9a) is asymptotically stable. Hence, controller windup is systemati-
x = Ax + bu + g pT v + lx (y cT x)
cally prevented for this compensator with a very simple measure.
(9b)
In contrast, the controller windup prevention for the classical
u = kTx (x v ) + T v. (9c) approach of Davison requires a much more involved method
(see, e.g. Hippe, 2006). Therefore, the design of an observer-
For the corresponding observer design, the conditions based feedforward controller should be preferred for the solu-
(pT , S), (cT , A) observable and tion of the robust output regulation problem. However, different
from Davisons design method, the presented approach needs
A sI g the knowledge of the disturbance input locations, i.e. of g in
det 0,
= s (S) (10) (5) (cf. (7a) and (13b)). This is no problem since (13) satisfies
cT 0
the internal model principle and therefore achieves output reg-
ulation independent from the disturbance input locations. As a
are sufficient and necessary for an arbitrary observer eigen-
consequence, one can consider the plant
value assignment (see OReilly, 1980). In order to investigate
the robustness of the output regulation, introduce the change of
coordinates x = Ax + gu + g f d + (g g f )d, x(0) = x0 Rn
(15a)
x = x v, (11) y = cT x, (15b)
has to be designed for (18). Therein, the observer gain L lead- u = Qv Kx x (25)
ing to asymptotically stable observer error dynamics exists if (C,
A) is observable. Combining (19) and (20) yields the observer- asymptotically rejects the known disturbance if A BKx is a
based compensator in the output feedback form Hurwitz matrix (cf. (17c)). This requirement can be ensured by
a suitable feedback gain Kx if (A, B) is controllable. In order to
v S 0 v B By y implement the state feedback (25), the disturbance observer
= + y
x BRv A BRx LC x L 0 r
v = S v + Lv (y Cx), v (0) = v 0 Rmnv (26a)
(21a)
x = Ax + Bu + BQv + Lx (y Cx), x(0) = x0 Rn
v
u = Rv Rx (21b) (26b)
x
for robust output regulation when introducing is used (cf. (17a) and (17b)). The next lemma states the existence
conditions for this observer.
Lemma 5.1 (Observability of the augmented plant): The
S = bdiag(S, . . . , S), By = bdiag(by1 , by2 , . . . , bym ). (22)
matrix pair
5. Output regulation using a generalised
S 0
0 C , (27)
observer-based feedforward controller BQ A
Similar to SISO systems, the solution of the disturbance rejection
problem is the starting point for the definition of the generalised is observable if and only if (C, A) is observable, (Q, S* ) is observable
observer-based feedforward controller in the MIMO case. and no invariant zero of (C, A, B) is an element of (S).
INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 799
Proof: The proof can be found in OReilly (1980, Proof: Insert (28c) in (1a) to obtain x = (A BKx )x +
Theorem 1). BKx (x x) + (GPd BQ)v + BQ(v v ) after a simple
rearrangement. With (23a)(23b) and (28a)(28b), the dynam-
Due to the special form of S* in (22) and Q in (24), this lemma ics of ev = v v and ex = x x read
implies that qTi , i = 1, 2, , m, in (24) have to be chosen such
that the pairs (qTi , S) are observable.
ev S Lv C ev 0
When implementing (25) with the state estimates of the = + v . (31)
ex BQ A LxC ex GPd BQ
observer (26), the generalised observer-based feedforward con-
troller This yields the closed-loop system with the state xcl =
col(v , x, ev , ex ) and xcl = Acl xcl , where
v = S v + Lv (y Cx), v (0) = R v 0 mnv
(28a)
x = Ax + Bu + BQv + Lx (y Cx), x(0) = x0 Rn GPd BQ
S 0
(28b) Acl = , A21 = 0 (32)
A21 A22
GPd BQ
u = Qv Kx x (28c)
and
results (cf. (17)). In order to investigate the relationship of this
compensator with the compensator (21) resulting from Davi- A BKx BQ BKx
sons approach, insert (28c) in (28b) yielding the dynamic output A22 = 0 S Lv C . (33)
feedback 0 BQ A LxC
v S Lv C v L
= + v y (29a) Thus, the separation principle follows from the lower block tri-
x 0 A BKx LxC x Lx angular structure of the matrix in (32) and the upper block tri-
angular matrix in (33).
v
u = Q Kx . (29b)
x Another important property of the compensator (28) is the
robust asymptotic rejection of the disturbances modelled by (3a)
By determining the corresponding dual system and (3b). For this, the internal model principle has to be satis-
fied (see Francis & Wonham, 1976). This is a direct consequence
v (S )T 0 v QT of the duality to the classical approach of Davison, which is the
= + y result of the next theorem.
x CT LTv AT CT LTx KxT BT x KxT
Theorem 5.3 (Internal model principle): Assume that (qTi , S),
(30a)
i = 1, 2, , m, are observable and the closed-loop system consist-
T T
v ing of the plant (1) and the compensator (28) yields an asymp-
u = Lv Lx (30b)
x totically stable closed-loop system in the nominal case. Then, the
compensator (28) ensures the internal model principle, i.e. robust
up to the minus sign (see, e.g. OReilly, 1983, Chapter 1.3.3) and asymptotic disturbance rejection is achieved.
comparing the result with (21) for r = 0, the duality to the clas- Proof: The dynamic output feedback (30) results from apply-
sical approach of Davison becomes apparent, which is the result ing Davisons approach to the dual of (18), when using the
of the next theorem. duality relations of Theorem 5.1. Furthermore, these relations
Theorem 5.1 (Duality): The design of the compensators (21) and also imply that the existence of this compensator is ensured
(29) are dual, i.e. the duality relations ST S, CT B, LTv Rv , if the existence conditions of the new approach are satisfied.
AT A, LTx Rx , KxT L and QT By hold. Thus, the compensator (30) ensures the internal model prin-
ciple in the dual case leading to the transfer behaviour u(s) =
It is easy to verify that by applying this theorem, the existence 1
N (s)D1
det(sIST ) c c (s)y(s), in which the polynomial matrices
conditions, i.e. Lemma 4.1 and 5.1, of the compensators (21) and det(sI ST )Dc (s) and Nc (s) are right coprime (see Desoer &
(29) can be converted into each other. Furthermore, the condi- Wang, 1978). Therefore, no zero of det(sI ST ) is cancelled
tions that the plant has to satisfy for the design, i.e. (A, B) con- by the transmission zeros of the compensator, i.e. the zeros
trollable, (C, A) observable and no invariant zero of (C, A, B) in of det Nc (s). A comparison of (29) and (30) shows that the
(S), are in both cases the same. transfer matrix of (29) results by transposition of (30). Con-
The duality shown in Theorem 5.1 also suggests that proper- sequently, the transfer behaviour of (29) has the general form
ties of the compensators (21) and (29) coincide. The next theo- u(s) = det(sIS)
1
(DTc (s))1 NcT (s)y(s), in which det(sI S)DTc (s)
rem demonstrates that the design of the latter can also be based
and NcT (s) are left coprime so that no zero of det(sI S) is com-
on the separation principle.
pensated by transmission zeros of the compensator. Since the
Theorem 5.2 (Separation principle): The eigenvalues of the closed-loop system with this controller is asymptotically stable,
closed-loop system comprising the plant (1) and the compensator no invariant zero of (C, A, B) is an element of (S) so that no can-
(28) are given by the eigenvalues of A BKx and of the observer cellation between zeros of det(sI S) and the invariant zeros of
(28a)(28b). the plant (1) occurs. Then, Chen (1984, Theorem 922, p. 497)
800 J. DEUTSCHER
implies that (29) ensures the internal model principle for the cor- of (Sr ) (see, e.g. Huang, 2004, Theorem 1.10) and the corre-
responding closed-loop system. sponding Kx can be found if (A, B) is controllable. The next the-
orem shows that this modification achieves an asymptotic track-
5.2. Robust reference signal tracking ing of the reference signal modelled by (34) without exciting
A direct consequence of Theorem 5.3 is that replacing y in (28a) the disturbance models (35) of the compensator in the reference
and (28b) by the tracking error y r also ensures robust asymp- behaviour.
totic tracking of the reference signals r modelled by (3a) and (3c) Theorem 5.4 (Separate reference channel for reference signal
(see Chen, 1984, Theorem 922, p. 497). Thus, robust output tracking): Assume that the homogeneous error system (31) is
regulation (4) is achieved by the resulting compensator. asymptotically stable and that A BKx is a Hurwitz matrix.
Then, the compensator
5.3. Nonrobust separate reference channel design
In some applications, it may be desirable to assign an indepen- v d = Sd v d + Lv (y Cx), v d (0) = v d,0
Rmnd (38a)
dent dynamics for the tracking of constant reference inputs with
respect to each output, i.e. to inputoutput decouple the refer- x = Ax + Bu + BQv d + Lx (y Cx), x(0) = x0 Rn
(see (34b)) so that the reference observer (40) is not needed and d
the compensator order is thus reduced by m.
If the reference transfer behaviour has to be specified, then u u x = Ax + B u + Gd y
the state feedback (25) is replaced by - y = Cx
u = Qv d Kx x + Mr. (41)
Kx
x
S = 4 0 0 , Pd = 1 0 0 , Pr = 0 0 1 .
x = Ax + B u + BQv + Lx (ey C x)
0 0 0
v v = S v + Lv (ey C x) ey - r
Q (51)
compensator
The system has an input saturation characterised by the lim-
its umax
1 = umin 1 = 87 and umax
2 = umin
2 = 80 (see (2)). For
Figure . Robust output regulation using generalised observer-based feedforward
control. the solution of this output regulation problem, a generalised
observer-based feedforward controller (28) is designed with
ey = y r (see Figure 2). By choosing qTi = [0 1 1], i = 1, 2,
input saturation, i.e. by considering a Lure type problem (see, e.g.
the pairs (qTi , S) are observable (see Theorem 5.3) so that Q can
Khalil, 2002, Section 7.1). In order to determine the correspond-
be obtained from (24). In order to complete the compensator
ing transfer function of the linear part, consider (1a), (28c) and
design, the feedback gain Kx is chosen to place all closed-loop
(31) for v* (t) 0 to obtain
eigenvalues at s = 14 and the observer gains
x A 0 0 x B
ev = 0 Lv C ev + 0 u (47a)
S 0.4464 0.0695
0.9185 0.1221
ex 0 A LxCBQ ex 0
1.2157 0.1736
Lv = 106
0.0434 0.2163
and
x
u = Kx Q Kx ev . (47b) 0.0567 0.3845
ex 0.0938 0.5434
0.0029 0.0007
Using uL = u, the transfer behaviour of the linear part is 0.1112 0.0232
2.1977 0.4119
uL (s) = Kx (sI A)1 Bu(s). (48) Lx = 104
0.0006 0.0020
(52)
0.0173 0.0712
With this, the results in Hippe and Deutscher (2009, Theorem
0.2190 1.2917
4.3) can be applied to investigate the global asymptotic stabil-
ity of the nonlinear closed-loop system. Moreover, a system-
assign the eigenvalue set = { 10, 11, 12, 13, 14, 15,
atic design of a globally stabilising Kx is possible for asymptoti-
16, 17, 18, 19, 20, 21} to the disturbance observer
cally stable plants by making use of the positive real lemma (see
(28a) and (28b). The resulting compensator achieves robust
Hippe & Deutscher, 2009, Theorem 4.3). It should be noted that
tracking and asymptotic rejection of both constant and sinu-
this approach, in general, is not possible for Davisons classical
soidal signals.
design method. This is due to the fact that the internal model
For v(0) = [0 45 0]T , the simulation of the corresponding
always leads to a linear part which is not asymptotically stable
closed-loop system leads to the disturbance behaviour shown
(see Figure 1).
in the left-hand side of Figure 3. Without windup preven-
tion, the unstable controller leads to oscillations of the outputs.
7. Example In contrast, if the controller windup is prevented by driving
Consider system (1) with the disturbance observer with the saturated inputs, then well-
damped transients result. For larger disturbance amplitudes,
0 1 0 0 0 0 0 0 input saturation causes instability of the closed-loop system
0 0 1 0 0 0 0 0
without windup prevention. In contrast, the designed controller
1 3 3 1 1 0 1 0
A=
0
, B= with windup prevention ensures a stable closed-loop dynam-
0 0 0 1 0
0
0 ics even for these disturbances. Furthermore, these results con-
0 0 0 0 0 1 0 0 firm that the resulting compensator achieves robust asymptotic
1 2 1 1 3 3 0 1 disturbance rejection, since the actual disturbance input loca-
(49) tion is not used in the design. The corresponding reference
and behaviour with this windup prevention measure leads to the
closed-loop responses at the right-hand side of Figure 3. Due
3 0 0 1 0 0
C= , to the tracking of sinusoidal reference signals, this compensator
1 0 0 4 0 0
T leads to unsatisfactory reference transients, since the reference
G = 0.1 0 0.1 0 0.1 0.2 , (50) frequency response has gain one and zero phase additionally for
INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 803
0.1 1
0.05 0.8
0 0.6
y1
y1
-0.05 0.4
-0.1 0.2
-0.15 0
-0.2 -0.2
0 1 2 3 0 1 2 3
0.06 0.5
0.04
0
0.02
0
y2
y2
-0.5
-0.02
-0.04
-1
-0.06
-0.08 -1.5
0 1 2 3 0 1 2 3
t t
Figure . Left: disturbance behaviour without windup prevention (dotted) and with windup prevention (solid). Right: reference step response to r (t) = (t) and r = (t)
with windup prevention, robust design (dotted) and nonrobust separate reference channel design (solid).
Sd = , Pd = 1 0 (54) saturation has to be taken into account. For this, the new con-
4 0
troller design method provides a systematic and simple solution,
which also contributes to its practical applicability.
so that the compensator (43) also robustly rejects the corre-
sponding sinusoidal disturbances. In order to ensure compara-
bility of the results obtained from the compensators, the state Disclosure statement
feedback gain Kx and the observer gains Lv , Lx are the same in No potential conflict of interest was reported by the author.
both designs. The second compensator achieves the same dis-
turbance behaviour as the first compensator because both com-
pensators coincide for vr (t) 0 and r(t) 0. In contrast to the References
first compensator, the separate reference channel design has the Byrnes, C.I., Priscoli, F.D., & Isidori, A. (1997). Output regulation for uncer-
advantage that the disturbance observer and thus the sinusoidal tain nonlinear systems. Boston, MA: Birkhuser.
disturbance models do not affect the tracking behaviour. Fur- Chen, C. (1984). Linear system theory and design. New York, NY: Holt,
Rinehart and Winston.
thermore, the dynamics of each reference channel can be speci- Davison, E.J., & Goldenberg, A. (1975). Robust control of a general
fied independently by (53). This yields the fast and well-damped servomechanism problem: The servo compensator. Automatica, 11,
reference response in Figure 3 on the right-hand side. 461471.
804 J. DEUTSCHER
Desoer, C.A., & Wang, Y.T. (1978). On the minimum order of a robust ser- Johnson, C.D. (1971). Accomodation of external disturbances in linear reg-
vocompensator. IEEE Transactions on Automatic Control, 23, 7073. ulator and servomechanism problems. IEEE Transactions on Automatic
Falb, P., & Wolovich, W. (1967). Decoupling in the design and synthesis of Control, 16, 635644.
multivariable control systems. IEEE Transactions on Automatic Control, Khalil, H. (2002). Nonlinear systems. Upper Saddle River, NJ: Prentice Hall.
12, 651659. Knobloch, H., Isidori, A., & Flockerzi, D. (1993). Topics in control theory.
Francis, B.A., & Wonham, W.M. (1976). The internal model principle of Basel: Birkhuser.
control theory. Automatica, 12, 457465. Olsson, H., Astrm, K., de Wit, C.C., Gfvert, M., & Lischinsky, P. (1998).
Hippe, P. (1992). Design of MIMO compensators for systems with unmea- Friction models and friction compensation. European Journal of Con-
surable disturbances: The polynomial case. Automatica, 28, 10031009. trol, 4, 179195.
Hippe, P. (2006). Windup in control its effects and their prevention. OReilly, J. (1980). Further comments on Minimal-order observers for lin-
London: Springer-Verlag. ear multivariable systems with unmeasurable disturbances. Interna-
Hippe, P., & Deutscher, J. (2009). Design of observer-based compensators. tional Journal of Control, 31, 605608.
London: Springer-Verlag. OReilly, J. (1983). Observers for linear systems. London: Academic Press.
Hippe, P., & Wurmthaler, C. (1999). Systematic closed-loop design in the Saberi, A., Stoorvogel, A., & Sannuti, O. (2000). Control of linear systems
presence of input saturations. Automatica, 35, 689695. with regulation and input constraints. London: Springer-Verlag.
Hu, R., & Mller, P.C. (1996). Independent joint control: estimation and Yao, J., Jiao, Z., & Ma, D. (2014a). Extended-state-observer-based output
compensation of coupling and friction effects in robot position control. feedback nonlinear robust control of hydraulic systems with backstep-
Journal of Intelligent & Robotic Systems, 15, 4151. ping. IEEE Transactions on Industrial Electronics, 61, 62856293.
Huang, J. (2004). Nonlinear output regulation. Philadelphia, PA: SIAM. Yao, J., Jiao, Z., & Ma, D. (2014b). Adaptive robust control of DC motors
Isidori, A., Marconi, L., & Serrani, A. (2003). Robust autonomous guidance. with extended state observer. IEEE Transactions on Industrial Electron-
London: Springer-Verlag. ics, 61, 36303637.