IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO.

3, MARCH 2007 493

task-space setpoints (in [m]) and corresponding desired joint-space set- [13] R. Kelly and V. Santibáñez, “A class of global regulators with bounded
points (in [deg]) were selected as control actions for robot manipulators,” in Proc. IEEE Conf. Decision
Control, 1996, pp. 3382–3387.
[14] R. Kelly, J. Moreno, and R. Campa, “Visual servoing of planar robots
via velocity fields,” in Proc. IEEE Conf. Decision Control, 2004, pp.
xd = [2:6 0 0:08] qd = [0 0 5:7] 4028–4033.
[15] F. Lewis, C. Abdallah, and D. Dawson, Control of Robot Manipula-
where tors. New York: MacMillan, 1993.
[16] A. Loria, R. Kelly, and R. Ortega, “On global output feedback regu-

x(0) = [2:38 0:84] q (0) = [28:65 0 28:65]: lation of Euler-lagrange systems with bounded inputs,” in Proc. IFAC
World Congr., 1996, pp. 161–166.
[17] A. Loria and H. Nijmeijer, “Bounded output feedback tracking control
By adjusting the control gains as of fully-actuated Euler-lagrange systems,” Syst. Control Lett., vol. 33,
no. 3, pp. 151–161, 1998.
= diagf120; 80g kv = diagf1; 1g " = 0:005
[18] R. Lozano and B. Brogliato, “Adaptive control of robot manipulators
kp with flexible joints,” IEEE Trans. Autom. Conrol, vol. 37, no. 2, pp.
00 = diagf0:35; 0:025g 01J = diagf0:015; 0:0015g: 174–181, Feb. 1992.
[19] R. Ortega, A. Loria, and R. Kelly, “A semiglobally stable output feed-
the results given in Figs. 1 and 2 can be obtained. Specifically, the back PI D regulator for robot manipulators,” IEEE Trans. Autom. Con-
task-space error is depicted in Fig. 1. Fig. 2 shows the control input trol, vol. 40, no. 8, pp. 1432–1436, Aug. 1995.
[20] S. Perlis, Theory of Matrices. New York: Dover, 1991.
torques. Fig. 1 illustrates that the developed adaptive controller yields
[21] J. A. Piepmeier, G. V. McMurray, and H. Lipkin, “A dynamic quasi-
asymptotic regulation of the robot to a desired constant task-space set- Newton method for uncalibrated visual servoing,” in Proc. IEEE Int.
point despite uncertainties throughout the manipulator kinematics and Conf. Robot. Autom., 1999, pp. 1595–1600.
dynamics. [22] ——, “A dynamic Jacobian estimation method for uncalibrated visual
servoing,” in Proc. IEEE Int. Conf. Adv. Int. Mech., 1999, pp. 944–949.
[23] J. A. Piepmeier, B. A. Gumpert, and H. Lipkin, “Uncalibrated eye-in-
VI. CONCLUSION hand visual servoing,” in Proc. IEEE Int. Conf. Robot. Automat., 2002,
pp. 568–573.
An amplitude limited controller was developed for robot manipu- [24] J. A. Piepmeier, “Experimental results for uncalibrated eye-in-hand vi-
lators despite uncertainty in the dynamic and kinematic models. The sual servoing,” in South. Symp. Syst. Theory, 2003, pp. 335–339.
[25] J. A. Piepmeier, G. V. McMurray, and H. Lipkin, “Uncalibrated dy-
adaptive controller yields semi-global asymptotic regulation of the
namic visual servoing,” IEEE Trans. Robot. Automat., vol. 20, no. 1,
task-space error. The advantages of the proposed controller include pp. 143–147, Jan. 2004.
the ability to actively compensate for unknown parametric effects in [26] V. Santibáñez and R. Kelly, “Global regulation for robot manipula-
the dynamic and kinematic model and the ability to a priori calculate tors under SP-SD feedback,” in Proc. IEEE Int. Conf. Robot. Automat.,
the maximum required torque. 1996, pp. 927–932.
[27] S. Sastry and M. Bodson, Adaptive Control: Stability, Convergence,
and Robustness. Englewood Cliffs, NJ: Prentice-Hall, 1989.
[28] M. Shahamiri and M. Jagersand, “Uncalibrated visual servoing
REFERENCES using a biased Newton method for on-line singularity detection and
[1] S. Arimoto, “Robotics research toward explication of everyday avoidance,” in Proc. IEEE/RSJ Int. Conf. Int. Rob. Syst., 2005, pp.
physics,” Int. J. Rob. Res., vol. 18, no. 11, pp. 1056–1063, 1999. 3953–3958.
[2] M. Bridges, D. M. Dawson, and C. Abdallah, “Control of rigid-link [29] J. J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood
flexible-joint robots: A survey of backstepping approaches,” J. Rob. Cliffs, NJ: Prentice-Hall, 1991.
Syst., vol. 12, pp. 199–216, 1995. [30] E. Zergeroglu, W. E. Dixon, A. Behal, and D. M. Dawson, “Adaptive
[3] C. Cheah, M. Hirano, S. Kawamura, and S. Arimoto, “Approximate set-point control of robotic manipulators with amplitude-limited con-
Jacobian control for robots with uncertain kinematics and dynamics,” trol inputs,” Robotica, vol. 18, no. 2, pp. 171–181, 2000.
IEEE Trans. Robot. Automat., vol. 19, no. 4, pp. 692–702, Jul. 2003.
[4] C. Cheah, S. Kawamura, and S. Arimoto, “Feedback control for robotic
manipulators with uncertain kinematics and dynamics,” in Proc. IEEE
Int. Conf. Robot. Automat., 1998, pp. 3607–3612. Estimation of Sinusoidal Frequencies and Amplitudes
[5] ——, “Feedback control for robotic manipulators with an uncertain Using Adaptive Identifier and Observer
Jacobian matrix,” Int. J. Rob. Syst., vol. 12, no. 2, pp. 119–134, 1999.
[6] C. Cheah, S. Kawamura, S. Arimoto, and K. Lee, “PID control for
robotic manipulator with uncertain Jacobian matrix,” in Proc. IEEE Int.
M. Hou
Conf. Robot. Automat., 1999, pp. 494–499.
[7] J. Chen, D. M. Dawson, W. E. Dixon, and V. Chitrakaran, “An opti-
mization-based approach for fusing image-based trajectory generation
with position-based visual servo control,” in Proc. IEEE Conf. Deci- Abstract—A simple method is proposed for estimation of amplitudes of
sion Conttrol, 2004, pp. 4034–4039. multiple sinusoids. The estimation is based on the existing adaptive identi-
[8] R. Colbaugh, E. Barany, and K. Glass, “Global regulation of uncertain fier which offers the globally convergent estimate of sinusoidal frequencies.
manipulators using bounded controls,” in Proc. IEEE Int. Conf. Robot. To deal with possible singularities of the amplitude estimation, adaptive
Automat., 1997, pp. 1148–1155. observers are also proposed for estimation of sinusoidal amplitudes. Simu-
[9] ——, “Global stabilization of uncertain manipulators using bounded lations illustrate the results.
controls,” in Proc. IEEE Amer. Control Conf., 1997, pp. 86–91.
[10] W. E. Dixon, M. S. de Queiroz, D. M. Dawson, and F. Zhang, Index Terms—Adaptive identifier, adaptive observer, frequency and am-
“Tracking control of robot manipulators with bounded torque inputs,” plitude estimation, signal processing.
Robotica, vol. 17, pp. 121–129, 1999.
[11] W. E. Dixon, A. Behal, D. M. Dawson, and S. Nagarkatti, Non-
linear Control of Engineering Systems: A Lyapunov-Based Approach. Manuscript received October 18, 2005; revised May 18, 2005 and September
Boston, MA: Birkhäuser, 2003. 4, 2006. Recommended by Associate Editor W. X. Zheng.
[12] W. E. Dixon, “Adaptive regulation of amplitude limited robot manipu- The author is with the Department of Engineering, the University of Hull,
lators with uncertain kinematics and dynamics,” in Proc. IEEE Amer. Hull HU6 7RX, U.K. (e-mail: [email protected]).
Control Conf., 2004, pp. 3839–3844. Digital Object Identifier 10.1109/TAC.2006.890389

0018-9286/$25.00 © 2007 IEEE

494 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 3, MARCH 2007

I. INTRODUCTION
Consider a sinusoidal signal described by

y = A1 sin(!1 t + 1 ) + A2 sin(!2 t + 2 ) + 111 Fig. 1. Second-order system.

+ An sin(!n t + n ) (1)
where the constant values of amplitudes Ai , frequency !i and phase i
are all unknown. Assume !i > 0 and !i 6= !j for i 6= j . This note is
concerned with the problem of reconstructing the unknown values !i
and Ai for arbitrary constant i from the measurable signal y .
The reconstruction of the frequencies and amplitudes of sinusoids is
a basic and important problem finding various applications in diverse
scientific and engineering disciplines [1]. Traditionally, frequencies
and amplitudes can be estimated using the classical Fourier analysis
requiring batch processing of data, which is not suitable for real
time applications. In digital signal processing [2], the least-square
method and statistical analysis have been used for frequency and
amplitude estimation. In contrast to these traditional estimations with
Fig. 2. Composite system.
local convergence in nature, several globally convergent estimators of
multiple frequencies have been proposed in [3]–[5] stimulated by the
notch filter for single frequency estimation in [6].
A quadratic transform of a sinusoidal signal has led to a globally con- where the polynomial
vergent estimator simultaneously reconstructing frequency and ampli- n

tude of a sinusoid in [7]. However, a direct extension of the treatment q (s ) = s2 + !i2

to the case of multiple sinusoids seems to be difficult if not impossible. i=1

This note shows that the existing state–space based methods for =s
2n
+ n01 s
2(n 01) + 1 1 1 + 1 s2 + 0 (4)
multifrequency estimation of sinusoids contain information about the has the coefficients i s corresponding to !i s, and the 2(n 0 1)th-order
component yi = Ai sin(!i t + i ) and its first-order time derivative polynomial p(s) is of no interest.
y_ i = Ai !i cos(!i t + i ). A method for calculating yi and y_ i is Choose a Hurwitz polynomial
proposed and further sinusoidal amplitudes can be easily obtained
as Ai = yi2 + y_ i2 =!i2 , given convergent estimates of multiple
frequencies. This method referred to as the amplitude estimators is (s) = s2n + 2n01 s2n01 + 111 +  s+  :
1 0 (5)
developed on the basis of the adaptive identifier proposed in [4]. The
method can also be adopted to the adaptive observers in [3] and [5] to From
have both frequency and amplitude estimates of sinusoids.
As an alternative, a method of adaptive observers will also be pre-
p(s )
y (s ) =
0 (s)) u(s)
sented for estimation of sinusoidal amplitudes. The global asymptotic (6)
convergence of amplitude observers being connected to the frequency  (s ) + ( q (s )
identifier in cascade will be proved. Compared with the amplitude esti- it follows:
mators where calculations of the inverse of a Vandermonde may incur
numerical problems, the amplitude observers can overcome the diffi-
culties.
y (s ) =
0
 (s ) q (s )
y (s ) +
p(s )
u (s ): (7)
 (s )  (s )
II. STATE–SPACE MODEL AND ADAPTIVE IDENTIFIER With u = 0 a state–space realization of the s-domain expression (7) is
Rewrite the sinusoidal signal in (1) as given by

x_ = Ax + by (8)
y ( t) = y 1 ( t) + y 2 ( t) + 111 + y n( t) (2) y = cx (9)
where
with yi (t) = Ai sin(!i t+i ). Each component yi satisfies the second- 0 1 0 111 0 0
order differential equation yi = 0!i2 yi . In terms of the transfer func- .. .. .. .. .. ..
tion, the second-order system is depicted in Fig. 1. The second-order . . . . . .
A= .. .. .. b= ..
system has the input u = 0 and is with some initial conditions yi (0) . . . 0 . (10)
and y_ i (0) determining the amplitude Ai and phase i of sinusoid yi . 0 111 111 0 1 0
In the s-domain, signal y in (2) is the summation of the outputs of the
n second-order systems connected in parallel as illustrated in Fig. 2.
0 0 1 1 1 1 1 1 0 0
0 1 1 2n 1
c = [ 0    0 
0  111  0 0  0
0 1 2 1 3 2n 2 n 1 2n01 ]:
In terms of the transfer function, the output of the composite system
is expressed as (11)
Equations (8) and (9) reproduce y (t) only if x(t) is properly initial-
n n
ized and i s are known. This requirement is equivalent to knowing all
1 p(s ) the components of frequencies, amplitudes and phases of the sinusoidal
y (s ) = y i (s ) = u (s ) = u (s ) (3)
i=1 i=1
s2 + !i2 q (s ) wave (2). Nevertheless, as will be seen, without any prior knowledge of
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 3, MARCH 2007 495

initial conditions, (8) and (9) can be used to estimate frequencies and On the other hand, y = n i=1 yi and yi = (0!i2)j yi for j
(2j )
=
amplitudes of sinusoids. 0; 1; . . ., generate the matrix equation yd = V yc with
Let ^i s be estimates of i s and denote 1 1 111 111 1
c^ = [0 0 ^0 1 2 0 ^1 2n02 0 ^n01 2n01 ]:
3 111 0! 1
2
0! 2111
2
1 1 1 0!n 2

(12) V = ! 4
1 ! 4
2 111 111 !n 4

.. .. ..
Based on a well-established adaptive system theory in [8], it has been . . 111 111 .
shown in [4] that an adaptive identifier of i s is formed by 0! n0 2
1
1
0! n0 1 1 1
2
2
1
1 1 1 0!n n0 2 1

y1
x^_ = Ax ^ + by (13)
y2
y^ = c^x^ (14) .
yc = .. (19)
with the parameter updating ..
.
yn
_ = x^2i+1 (^y 0 y)
^i (15) where Vandermonde matrix V is invertible because of the assumption
!i 6= !j for i 6= j .
for i = 0; 1; . . . ; n 0 1, where > 0 is an arbitrary real number and Equations yd = Qx and yd = V yc give rise to
x^2i+1 is the (2i + 1)th element of x^. The initial conditions x^(0) and
^i (0) can be arbitrarily chosen. The parameter estimation has the prop-
erty of global exponential convergence [8]. In addition, it may be worth yc = V 01 (!)Q()x (20)
pointing out that, although x(0) is unknown, since A is Hurwitz, x ^ in
(13) exponentially converges to x, and y^ to y = m with ! indicating !1 ; !2 ; . . . ; !n and  representing 0 ; 1 ; . . . ; n01 .
i=1 Ai sin(!i t +
i ), as proved in [8]. Similarly, the time derivative of yc is given by
The estimates ! ^i s of frequencies !i s can be determined by
^0 ; ^1 ; . . . ; ^n01 , through the relationship implied in the polynomial
(4). For a small integer n, explicit formulas can be derived, while
y_ c = V 01 (!)Q()3()x: (21)
in general a numerical procedure for finding roots of a polynomial Note that if define V (! ) = 1 for n = 1, the formulas (20) and (21)
may be needed. It should be pointed out that in the latter case, the remain valid when a single sinusoid is considered. In such a case, yc =
advantage of real-time operations of the proposed method would be y and y_ c = y_ .
compromised. If denote the ith row of V 01 by vi , (20) and (21) can be rewritten as

III. AMPLITUDE ESTIMATOR yi = vi Qx and y_ i = vi Q3x (22)

The adaptive identifier discussed in Section II not only provides for i = 1; 2; . . . n. Now, the amplitude of the ith sinusoidal component
globally convergent estimates of unknown frequencies !i ’s through is given by
^i s, but also contains sufficient information about yi s and y_ i s. Recall
that given yi and y_ i along with !i , the amplitude Ai of the ith sinu-
soidal component can be determined. A simple method for obtaining Ai = yi2 + y_ i2 / !i2 ; i = 1; 2; . . . ; n: (23)
yi s and y_ i s is described in the following.
To establish a relationship between the state x in (8) and yi s and y_ i s, For online estimation of amplitudes, formula (23) needs to be re-
note the derivatives placed by

y (j ) = c3j x; j = 0 ; 1; . . . (16) A^i = y^i2 + y^_ 2i !^i2 ; i = 1 ; 2; . . . ; n (24)

where y^i and y^_ i are, respectively, the ith component of y^c and y^_ c at-
as simple outcomes of (8), where
tainable from
0 1 0 111 0 y^c = V 01 (^!)Q(^)^x (25)
.. .. .. .. ..
. . . . . and
3 = A + bc = ..
.
..
.
..
. 0 : (17)
= V 01 (^!)Q(^)3(^)^x:
y^_ c (26)
111 111 0
0 1 ^ i 6= 0 has to be assumed. In real calculations, !^ i
Obviously, in (24) !
0 0 1 1 1 0n0 0 1 0 can be replaced by the lowest frequency bound whenever ! ^ i is below
the bound.
Equations in (16) for j = 0; 2; . . . ; 2(n 0 1) can be denoted as yd = Clearly, since the asymptotic convergence of ! ^ i to !i is guaranteed
Qx with by the adaptive identifier, to ensure that A^i converges to Ai for i =
1; 2; . . . n, it needs to show convergence of y^c to yc and y^_ c to y_c .
y Use the notation V 01 = V 01 (! ) 0 V 01 (^ ! ); Q = Q() 0 Q(^)
c
and x = x 0 x ^. It is easy to have
y c3 2
yd = .. Q= .. : (18)
yc 0 y^c = V 01 Q()x + V 01 (^!)Qx + V 01 (^!)Q(^)x
. .
y (2n02) c3 01)
2(n
(27)
496 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 3, MARCH 2007

Fig. 3. Scheme of frequency and amplitude estimation.

which implies that y^c converges to yc when all V 01 ; Q, and x where diag(d!^ ; . . . ; d!^ ) denotes the diagonal matrix with d!^ being on
converge to zero, provided kQk; kxk and kV 01 k are bounded. Under the diagonal line. In a similar way as for (27), convergence of d!^ y^i to
the assumption on boundedness, the convergence of V 01 ! 0 and d! yi and of d!^ y^_ i to d! y_ i can be verified.
Q ! 0 is guaranteed by the parameter convergence of the adaptive Opposite to being very close to each other, two frequencies could be
identifier due to continuity of all the components in V 01 and Q with significantly apart from each other, say ! ^ i = 1 but !^ j = 105 . In such a
respect to their arguments, while x ! 0 is ensured by the stable error case or similarly in the case of !i  1, although calculation of d!^ does
dynamics  x_ = Ax because A is a Hurwitz matrix. This verifies, not cause any problems, singularity of V (^ ! ) lies in possible numerical
under the required boundedness, the convergence of y^c to yc , and in a rank deficiency of adjV (^ ! ). This implies that the information about
similar way, y^_ c to y_ c . Clearly, kQ()k; kQ(^)k; kxk and kV 01 (! )k amplitudes cannot easily pass through from x ^ to d!^ y^i s and d!^ y^_ i s. In
are all bounded, however it cannot be said so about kV 01 (^ ! )k . A general, if two frequencies to be estimated are very close or signifi-
remedy for this could be to accept the current V 01 (^ ! ) if its norm is cantly apart from each other, the amplitude estimation is a numerically
less than a prespecified number, otherwise retain the previous V 01 (^ !) ill-posed problem. Suppose the underlying estimation problem is nu-
in the online calculations of (25) and (26). This keeps running of the merically well posed. Any kind of singularities of V (^ ! ) will not affect
amplitude estimation and meanwhile preserves global asymptotic con- the final convergence of amplitude estimation due to eventual conver-
vergence of y^c to yc and y^_ c to y_ c . A more sophisticated method will be gence of V (^! ) to V (! ) and V 01 (^ ! ) to V 01 (! ) followed by ! ^ ! !.
given in Section IV to overcome the problem of singularities of V (^ ! ). Nevertheless, during the period of singularity of V (^ ! ), updating of am-
The discussion so far suggests a two-step procedure for estimating plitude estimation completely or partially stops, which shows freezing
frequencies and amplitudes. In the first step, frequencies are estimated phenomena that can occur in the estimators of Section III and in ob-
by the adaptive identifier (13)–(15), [4]. This is possible because the servers to be developed in the following.
frequency estimation does not depend on sinusoidal amplitudes. In the Fig. 3 shows the scheme of estimation of frequencies and amplitudes
second step, amplitudes are calculated from (24) on the basis of the by using the adaptive identifier, estimators of d!^ y^i and d!^ y^_ i , and am-
state estimate and frequency estimate offered by the adaptive identifier. plitude observers to be developed in the following.
As a remark, the frequency estimators in [3], [5] have also a state- Suppose temporarily that !i s, yi , and y_ i are available. Recall that
space representation similar to (8)–(9). Moreover, the estimators pro- yi = Ai sin(!it + i ). Define a state variable as zi = d!2 yi y_ i . It is
vide globally asymptotically convergent state estimate x ^ directly in [3] immediate to have

= d!2 !i2 ai 0 2!i2d!2 yi2

and indirectly in [5]. It is, therefore, straightforward to combine the
method developed here with the frequency estimators proposed in [3] z_i (31)
and [5] to obtain amplitude estimates. with ai = A2i . On the strength of the adaptive observer theory in [9],
IV. AMPLITUDE OBSERVERS it has been shown in [7] that the adaptive observer

The amplitude estimators introduced in Section III may incur diffi- = d!2 !i2 ai 0 2!i2d!2 yi2 + li d!2 yi y_i 0 zi
z_ i (32)
culties in calculating the inverse of Vandermonde matrix V involved in _ i = i d2! !i2 d!2 yi y_i 0 zi
a (33)
(20) and (21). This is particularly true in the approximations (25) and
(26) when two or more frequencies are very close to each other. The with constant gains li > 0 and i > 0 and arbitrary initial conditions
problem is caused by possible small values of the determinant zi (0) and a
i (0), ensures the global asymptotic convergence of zi to zi
and a i to ai . With slight alterations, the amplitude observer (32) and
(33) is a special case of a more general result in [7].
d!^ = jV (^!)j = (01) 01 ^i2 0 !^j2
! (28) By intuition, a practical observer could be obtained from (32) and
1i<j n (33) by replacing d! ; !i ; d! yi and d! y_ i by their estimates d!^ ; !
^i;
d!^ y^i and d!^ y^_ i . As a matter of fact, the following equations:
when ! ^ i is close to !^ j . To allow a single sinusoid to be equally con-
sidered, note that d!^ = 1 for n = 1 because of V (^ !) = 1 as defined z^_ i= d!2^ !^i2 a^i 0 2^!i2 d!2^ y^i2 + li d!2^ y^i y^_ i 0 z^i (34)
before.
Recall the well-known formula for the inverse of a matrix V : V 01 = ^_ i = i d2!^ !^i2 d!2^ y^i y^_ i 0 z^i
a (35)
adjV =jV j, where matrix adjV is the adjugate of V . To avoid divisions
by d!^ , instead of y^i and y^_ i ; d!^ y^i and d!^ y^_ i should be used to estimate
are proposed to be the amplitude observers with the global asymptotic
convergence of z^i to zi and a^i to ai , for i = 1; 2; . . . ; n. A verification
Ai . In comparison with (25) and (26), d!^ y^i and d!^ y^_ i can be obtained of the convergence is given as follows.
from Obviously, it suffices to prove the observer convergence for any spe-
diag(d!^ ; . . . ; d!^ )^yc = adjV (^!)Q(^)^x (29) cific i. Denote ez;i = zi 0 z^i and ea;i = ai 0 a ^i . Tedious yet straight-
forward manipulations yield
and
diag(d!^ ; . . . ; d!^ )y^_ c = adjV (^!)Q(^)3(^)^x (30) e_ z;i = 0li ez;i + d!2^ !^i2 ea;i + i;1 (36)
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 3, MARCH 2007 497

Fig. 4. Frequency and amplitude estimation of sinusoids y = 2 sin + 4sin 3 .

t t

Fig. 5. Frequency and amplitude estimation of sinusoids y = 2sin + 4sin 1 2 .

t : t

e_ a;i = 0 i d2!^ !^ i2 ez;i + i;2 (37) where

with
ei =
ez;i
A!^ ;i =
0li d! !^i 2
^
2

ea;i 0 i d! !^i 02 2

i;1 = ai d!2 !i2 + !^ i2 d2! 0 2!i2 (d! yi + d!^ y^i )dyi

^

0 2d!2^ y^i2 !i2 + li (d!^ y^i dy_i + d! y_i dyi ) i = i;1 : (43)
(38) i;2
i;2 = i d!^ !^ i (d!^ y^i dy_ i + d! y_i dyi )
2 2
(39) Note that A!^ ;i is a time-varying matrix because ! ^ i s are produced by
the frequency identifier. Let 8i (t;  ) be the transition matrix corre-
where !i = !i 0 !
2 2
^i ; d! = d! 0 d!^ ; dyi = d! yi 0 d!^ y^i and
2 2 2 2
sponding to A!^ ;i . It is well known that the solution of (42) can be ex-
dy_ i = d! y_ i 0 d!^ y^_ i . The proof proceeds in two steps. First, it is to
prove ez;i ! 0 and ea;i ! 0 as t ! 1 if i;1 = 0 and i;2 = 0.
pressed as
Then, the same will be proved if i;1 ! 0 and i;2 ! 0 as t ! 1.
Now assume i;1 = i;2 = 0 in (36) and (37). Take t
ei (t) = 8i (t; 0)ei (0) + 8i (t;  )i ( ) d: (44)
0

v = 1 ez;i
2
+ 1 ea;i
2 The global asymptotic convergence of (36) and (37) under the assump-
2 2 i
(40)
tion i;1 = i;2 = 0 implies that
as a Lyapunov function candidate. Its time derivative is
k8i (t;  )k  i e0 t0 ( )
8t  0 (45)
v_ = 0li ez;i  0
2
(41) for some positive constants i and i . In fact, (45) is a necessary and
sufficient condition for global asymptotic stability of unforced linear
where the equality holds for ez;i = 0 and arbitrary ea;i . However, any time-varying systems [10, Th. 4.11]. This inequality is used to estimate
such a solution cannot identically satisfy the error dynamic equations the solution of (44) by
(36) and (37) for i;1 = i;2 = 0 unless ez;i = ex;i = 0. By virtue of
LaSalle’s invariance principle [10], this implies the global asymptotic k e i ( t) k
convergence of the error dynamic equations (36) and (37) under the t
assumption i;1 = i;2 = 0.  k8i (t; 0)kkei (0)k + k8i (t;  )kki ( )k d
Return to (36) and (37) without the assumption i;1 = i;2 = 0 and
0
t
by rewriting  i e0 t kei (0)k + i sup ki ( )k e0 (t0 ) d
 t
0 0

e_ i = A!^ ;i ei + i  i e0 t kei (0)k + i sup ki ( )k: (46)

(42) i 0  t
498 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 3, MARCH 2007

Fig. 6. Plots of y = 2 sin t + 4 sin 3t and y + v with v  N (0; 0:2 ).

Fig. 7. Frequency and amplitude estimation of sinusoids y = 2 sin t + 4 sin 3t + v with v  N (0; 0:2 ).

If i is considered as an input to system (42), inequality (46) implies 2 sin t + 4 sin 1:2t. For the frequency identifier, the coefficients i in
that (42) possesses the input-to-state stability with respect to i , a well- (5) are chosen from (s + 2)4 = s4 + 8s3 + 24s2 + 32s + 16, and the
known stability notion [11]. Furthermore, the error dynamic system gain in (15) is taken as = 500. For the amplitude observers, the gains
(42) of the adaptive observer (34) and (35) can be considered as being in (34) and (35) are chosen as 1 = 2 = 0:01 for the first example
connected in cascade to the dynamic system of i . Note that i does and 1 = 2 = 200 for the second example, and l1 = l2 = 50 for
not depend on ei , and i is locally Lipschitz in all  ( 1 ) defined there. both cases.
Denote the right hand-side equation of (42) as f (t; ei ; "i ). It is ready to Figs. 4 and 5 show the responses of the frequency identifier, ampli-
see that f is piecewise continuous in t, and locally Lipschitz in (ei ; "i ). tude estimators and amplitude observers, respectively, to the two inputs
Because i ! 0 is guaranteed by the global asymptotic convergence of the sinusoids. For the first example in Fig. 4, the amplitude estima-
of !
^ ; d!^y
^i and d! _ respectively to !; d! yi and d! y_ i , the entire cas-
^y
^ i tors and amplitude observers provide good estimates of the amplitudes,
cade dynamic system of ei and i is globally asymptotically stable [10, and the former have better transit performance. For the second example
Lemma 4.7]; [12, Lemma C.4]. This verifies ei ! 0 as t ! 1 in (42). in Fig. 5, since the frequencies are close to each other, the frequency
Similar as in the case of amplitude estimators discussed in identifier has the degraded performance. In such a case, the amplitude
Section III, globally asymptotic convergent estimates of d! ; !i ; d! yi observers outperform the amplitude estimators considerably.
and d! y_ i can also be derived from the adaptive frequency observers Finally, it may be interesting to see how the amplitude estimators
in [3] and [5]. The amplitude observers (34) and (35) can therefore be and observers may behave in the presence of noise. The example y =
applied to the methods in [3] and [5] as well. 2 sin t + 4 sin 3t is taken again, but with an additive Gaussian noise v
of zero mean and variance 0:22 . Fig. 6 shows the plots of the sinusoidal
V. ILLUSTRATIVE SIMULATION signal with and without noise, and Fig. 7 exhibits the responses of
MatLab/SIMULINK simulations have been carried out for several the frequency identifier, amplitude estimators and amplitude observers
cases of frequency and amplitude estimations by using the methods with respect to the noisy sinusoidal input. The parameters used in the
developed in this note. In the simulations, all the integral initial condi- simulation are = 50; 1 = 2 = 0:01, and l1 = l2 = 500. Fur-
tions were set to be zero. The solver used for integration of differential ther simulations on this example have indicated that in the case of large
equations is ode23tb with relative tolerance of 1003 . In order to avoid noise the amplitude estimators oscillate significantly, while the ampli-
singularities in (25) and (26) caused by the zero initial conditions on tude observers can still work but with a notable bias in the A1 estima-
!i ’s, a computational zero " = 10018 is added to jV (^ !)j in the calcu- tion.
lations, and the first outputs of the amplitude estimators for t = 0 will
not be shown in the corresponding plots. VI. CONCLUDING REMARKS
The simulation results of both amplitude estimators and observers Amplitude estimators and amplitude observers are proposed in this
are obtained for the two examples y = 2 sin t + 4 sin 3t and y = note as two solutions to the problem of estimating both frequencies
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 52, NO. 3, MARCH 2007 499

and amplitudes of sinusoids. As proved, both types of the estimators Multirate Output Feedback Based Robust Quasi-Sliding
and observers possess the property of global convergence. Although Mode Control of Discrete-Time Systems
the amplitude estimators are conceptually and computationally simpler
than the amplitude observers, the latter avoids a critical step in calcu- S. Janardhanan and B. Bandyopadhyay
lations of the inverse of the Vandermonde matrix and hence can better
deal with the cases when two or more estimated sinusoidal frequencies
are close to each other. Nevertheless, the amplitude observers have the
drawback of slowing down the estimation process, and still freezing Abstract—Over the last few years, the research on discrete-time sliding
when the determinant of the Vandermonde matrix approaches zero. mode control has received a considerable attention. Unlike its continuous-
Essentially, the proposed method is a two-step approach to the un- time counterpart, discrete-time sliding mode control is not invariant in gen-
eral. In this note, an algorithm is presented for robust discrete-time sliding
derlying problem. In the first step, frequency estimation is carried out, mode control using the concept of multirate output feedback.
while in the second step, amplitudes are estimated on the basis of the
outcomes of the frequency estimation. Because of possible numerical Index Terms—Discrete-time systems, multirate output feedback, sliding-
problems in estimating individual sinusoidal components, the two-step mode control (SMC), uncertain systems.
method is perhaps less efficient than the frequency and amplitude esti-
mator in [7] which applies to a single sinusoid only. An analogy of the
treatment of a single sinusoid is still missing for the case of multiple I. INTRODUCTION
sinusoids.
The proposed amplitude estimators and amplitude observers appear
to be the first solutions of the underlying estimation problem with the The use of digital computers and samplers in the control circuitry
property of global asymptotic convergence, whilst applications of the in the recent years, has made the use of discrete-time system repre-
methods in a wide range of scientific and engineering disciplines de- sentation more justifiable for controller design than continuous-time
mand exploitations. representation. In the recent years, considerable amount of research
has been done on discrete-time sliding mode controller design [1]–[4].
In discrete-time sliding mode control, the control input is calculated
once in every sampling interval and is held constant during this period.
REFERENCES Due to the finite sampling frequency, it may happen in discrete-time
sliding mode, that the system state trajectory is unable to move along
[1] D. W. Clarke, “On the design of adaptive notch filters,” Int. J. Adapt. the sliding surface. It may move about the surface, thus giving a sliding-
Control, vol. 15, pp. 715–744, 2001. like mode or quasi-sliding mode motion (QSM).
[2] P. Stoica, H. B. Li, and J. Li, “Amplitude estimation of sinusoidal
signals: Survey, new results, and an application,” IEEE Trans. Signal The aforementioned sliding mode control strategies are based on
Process., vol. 48, no. 2, pp. 338–352, Feb. 2000. state feedback. Since all the systems states may not be available for
[3] G. Obregon-Pulido, B. Castillo-Toledo, and A. Loukianov, “A globally measurement in most systems, such control strategies have a problem
convergent estimator for n-frequencies,” IEEE Trans. Autom. Control,
from the implementation point of view. This prompted the develop-
vol. 47, no. 5, pp. 857–863, May 2002.
[4] X. Xia, “Global frequency estimation using adaptive identifiers,” IEEE ment of output feedback sliding mode control strategies [5]–[7]. How-
Trans. Autom. Control, vol. 47, no. 7, pp. 1188–1193, Jul. 2002. ever, these control strategies also have certain shortcomings. Sliding
[5] R. Marino and P. Tomei, “Global estimation of n unknown frequen- mode control strategies based on static output feedback may not exist
cies,” IEEE Trans. Autom. Control, vol. 47, no. 8, pp. 1324–1328, Aug.
2002. for all controllable and observable linear systems, whereas dynamic
[6] L. Hsu, R. Ortega, and G. Damm, “A globally convergent frequency controllers would increase the complexity of the system.
estimator,” IEEE Trans. Autom. Control, vol. 44, no. 4, pp. 698–713, In this note, a robust sliding mode control strategy based on multirate
Apr. 1999.
[7] M. Hou, “Amplitude and frequency estimator of a sinusoid,” IEEE
output feedback [8] for uncertain discrete-time systems is proposed. In
Trans. Autom. Control, vol. 50, no. 6, pp. 855–858, Jun. 2005. this technique, the system output is sampled at a rate faster than the
[8] S. Sastry and M. Bodson, Adaptive Control—Stability, Convergence, control input. Consequently, the control algorithm is based on output
and Robustness. Upper Saddle River, NJ: Prentice-Hall, 1989. feedback and at the same-time, is applicable to all controllable and
[9] R. Marino and P. Tomei, Nonlinear Control Design—Geometric, Adap-
tive, and Robust. Upper Saddle River, NJ: Prentice-Hall, 1995. observable systems. Thus, it has the advantages of both state feedback
[10] H. K. Khalil, Nonlinear Systems, 3rd ed. Upper Saddle River, NJ: and output feedback control philosophies.
Prentice-Hall, 2002. The organization of the note is thus. First, the concept of multi-
[11] E. D. Sontag, “Smooth stabilization implies coprime factorization,”
IEEE Trans. Autom. Control, vol. 34, no. 4, pp. 435–443, Apr. 1989.
rate output feedback for uncertain systems is explained. A brief review
[12] M. Krstic, I. Kanellakopoulos, and P. Kokotović, Nonlinear and Adap- of state feedback based quasi-sliding mode control for uncertain dis-
tive Control Design. New York: Wiley-Interscience, 1995. crete-time system is discussed [2]. The robust multirate output feed-
back based sliding mode control of uncertain LTI systems is proposed
next. The proposed control strategy is then illustrated via a numerical
example, followed by the conclusions.

Manuscript received March 22, 2006; revised August 3, 2006. Recommended

by Associate Editor L. Magni.
The authors are with the Interdisciplinary Programme in Systems and Control
Engineering, Indian Institute of Technology Bombay, Mumbai-400 076, India
(e-mail: [email protected]; [email protected]).
Digital Object Identifier 10.1109/TAC.2006.890391

0018-9286/$25.00 © 2007 IEEE