Professional Documents
Culture Documents
&$$
1PSUP
1PSUVHBM
4FQUFNCFS
Keywords: Variable Structure Control, Trajectory Tracking parameter needs to be adjusted [8, 10]. The control can be
in Non-linear Systems, Stabilization of Non-linear Systems, made an arbitrarily smooth function of time, totally removing
Robust Filtering, Non-linear Observers. the chattering effect and providing for ultimate accuracy in
realization.
Abstract
Hence, the real-time robust exact differentiation becomes the
An arbitrary-order finite-time-convergent exact robust main problem of output-feedback HOSM control design. Even
differentiator is constructed based on higher-order sliding the most modern differentiators [3, 2, 14] do not provide for
mode (HOSM) technique. Being used in a feedback together exact differentiation with finite-time convergence or require
with previously proposed HOSM controllers, it produces a some knowledge on the noise and would actually eliminate the
universal controller, formulated in input-output terms only, above-listed advantages of the HOSM controllers [10]. The
which causes the output of any uncertain smooth SISO derivatives may be calculated by successive implementation of
minimum-phase dynamic system with known relative degree a robust exact first-order differentiator [7] with finite-time
to vanish in finite time. That allows exact tracking of convergence. That differentiator is based on 2-sliding mode
arbitrary real-time smooth signals. The control can be made and is proved to feature the best possible asymptotics in the
arbitrarily smooth, providing for arbitrarily-high tracking- presence of infinitesimal Lebesgue-measurable measurement
accuracy order with respect to the sampling step. Only one noises if the second time derivative of the unknown base
parameter is to be tuned. signal is bounded.
*4#/
1/2
control under heavy uncertainty conditions. Nevertheless, he z 1 = v1 , v1 = -λ1 | z1 - v0| sign(z1 - v1) + w1,
believes that the proposed theoretical solution indicates a w 1 = -α1 sign(z1 - v0); z2 = w1.
way promising future practical success. At the same time the
proposed differentiator is possibly the best solution for the Similarly, a 2nd-order differentiator from each of these
single problem of real-time numerical differentiation. sequences may be used as a base for a new recursive scheme.
An infinite number of differentiator schemes may be
constructed in this way. The only requirement is that the
2 Arbitrary-order differentiator resulting systems be homogeneous in a sense described
further. While the author has checked only the above two
Let input signal f(t) be a function defined on [0, ∞) consisting
schemes (1), (2) and (3), (4), the hypothesis is that all such
of a bounded Lebesgue-measurable noise with unknown
schemes produce working differentiators, provided suitable
features and an unknown base signal f0(t) with the nth
parameter choice. Differentiator (1) takes on the form
derivative having a known Lipschitz constant C > 0. The
n/( n + 1)
problem is to find real-time robust estimations of f0 (t), z 0 = v0 ,v0 = -λ0 | z0 - f(t)| sign(z0 - f(t)) + z1,
(n-1)/ n
f (n)
(t), ... , f0 (t) being exact in the absence of measurement z 1 = v1 ,v1 = -λ1 | z1 - v0| sign(z0 - v0) + z2,
0
...
noise. (n - i)/ ( n - i + 1)
z i = vi , vi = -λi | zi - vi-1| sign(zi - vi-1) + zi+1, (5)
Two similar recursive schemes of the differentiator are ...
1/ 2
proposed here. Let an (n-1)th-order differentiator Dn-1(f(⋅), C) z n-1 = vn-1 , vn-1 = -λn-1 | zn-1 - vn-2| sign(zn-1 - vn-2) + zn,
i
produce outputs Dn-1 , i = 0, 1, ..., n - 1, being estimations of z n = -λn sign(zn - vn-1),
(n-1) (n-1)
f0, f0 , f0 , ..., f0 for any input f(t) with f0 having Theorem 1. The parameters being properly chosen, the
Lipschitz constant C > 0. Then the nth-order differentiator following equalities are true in the absence of input noises
i
has the outputs zi = Dn , i = 0, 1, ..., n, defined as follows: after a finite time of a transient process for both
differentiator schemes:
n/( n + 1)
z 0 = v , v = -λ0 | z0 - f(t)| sign(z0 - f(t)) + z1, (i)
z0 = f0(t); zi = vi-1 = f0 (t), i = 1, ..., n.
(1)
0 n-1
z1 = Dn-1 (v(⋅), C), ..., zn = Dn-1 (v(⋅), C). Moreover, the corresponding solutions of the dynamic
systems are Lyapunov stable, i.e. finite-time stable [12]. The
Here D0(f(⋅), C) is a simple nonlinear filter (i)
Theorem means that the equalities zi = f0 (t) are kept in
D0: z = -λ sign(z - f(t)), λ > C. (2) 2-sliding mode, i = 0, ..., n-1.
Thus, the first-order differentiator coincides here with the Theorem 2. Let the input noise satisfy the inequality | f(t) -
differentiator published in [7]: f0(t)| ≤ ε. Then the following inequalities are established in
z 0 = v , finite time for some positive constants μi, νi depending
1/2
v = -λ0 | z0 - f(t)| sign(z0 - f(t)) + z1, (3) exclusively on the parameters of differentiators (1) or (4):
z 1 = -λ1 sign(z1 - v) = -λ1 sign(z0 - f(t)). (i)
|zi - f0 (t)| ≤ μi ε
(n - i +1)/(n + 1)
, i = 0, ..., n;
(i+1) (n - i )/(n + 1)
|vi - f0 (t)| ≤ νi ε , i = 0, ..., n-1.
Another recursive scheme is based on the differentiator (3)
~ Consider the discrete-sampling case, when z0(tj) - f(tj) is
as the basic one. Let Dn −1 (f(⋅), C) be such a new (n-1)th-
~ substituted for z0 - f(t) with tj ≤ t < tj+1, tj+1 - tj = τ > 0.
order differentiator, n ≥ 1, D1 (f(⋅), C) coinciding with the
differentiator D1(f(⋅), C) given by (3). Then the new scheme Theorem 3. Let τ > 0 be the constant input sampling
is defined as interval in the absence of noises. Then the following
z 0 = v , inequalities are established in finite time for some positive
v = -λ0 | z0 - f(t)|
n/( n + 1)
sign(z0 - f(t)) + w0 + z1 , constants μi, νi depending exclusively on the parameters of
(4) differentiators (1) or (4):
(n - 1)/( n +1)
w 0 = -α0 | z0 - f(t)| sign(z0 - f(t)); (i) n-i+1
~ ~ |zi - f0 (t)| ≤ μi τ , i = 0, ..., n;
z1 = Dn0−1 (v1(⋅), C), ..., zn = Dnn−−11 (v(⋅), C). (i+1) n-i
|vi - f0 (t)| ≤ νi τ , i = 0, ..., n - 1.
The resulting 2nd-order differentiator was presented at
In particular, the nth derivative error is proportional to τ. The
ECC’99, CDC’00 [8, 9]:
latter Theorem means that there are a number of real sliding
2/3
z 0 = v0, v0 = -λ0 | z0 - f(t)| sign(z0 - f(t)) + w0 + z1, modes of different orders [6]. Nevertheless, nothing can be
1/3
w 0 = -α0 | z0 - f(t)| sign(z0 - f(t)); said on the derivatives of vi and of z i, because they are not
continuously differentiable functions.
(n-1)/ n
The Lemma is a simple consequence of the invariance of (6),
σ 1 = -λ1 | σ1 - σ 0 | sign(σ1 - σ 0 ) + σ2, (7) with respect to transformation (8). The convergence time
... is estimated as a sum of a geometric series.
(7)
1/ 2
σ n −1 = -λn-1 | σn-1 - σ n −2 | sign(σn-1 - σ n −2 ) + σn, Proof of Proposition 3. Consider first the case εM = 0.
σ n ∈ -λn sign(σn - σ n −1 ) + [-C, C], Choose some larger region |σi| ≤ S′i, Si < S′i , i = 0, ..., n, and
let Σi > S′i, i = 0, ..., n - 1, be some upper bounds chosen with
where ξ(t) ∈ [-ε, ε] is a Lebesgue-measurable noise function.
respect to Proposition 3 for (7) and that region. It is easy to
It is important to mark that (6), (7) do not “remember”
check that for any q > 1 with sufficiently large λ0 the
anything on the unknown input basic signal f0(t). ( n + 1)/n
trajectory enters the region |σ0| ≤ (qΣ1/ λ0) in arbitrarily
System (6) is homogeneous in sense [12] with ε = 0, its small time. During that time σ 0 does not change its sign.
trajectories are invariant with respect to the transformation Therefore, ∫| σ 0 |dt ≤ S′0 for sufficiently large λ0. Thus, the
Gη: (t, σi, ξ, ε) ( ηt, η
n-i+1
σi, η
n+1 n+1
ξ, η ε). (8) disturbance - σ 0 entering the subsystem (7) satisfies
Let now εM > 0. As follows from the continuous dependence Third order differentiator. The measurement step τ = 5⋅10
-4
of the solutions of a differential inclusion on the right-hand was taken, noises are absent. The attained accuracies are
side [5] with sufficiently small εM all trajectories concentrate -11 -7 -4
9.40⋅10 , 2.97⋅10 , 2.40⋅10 , 0.0401 for the signal
within a finite time in a small vicinity of the origin. The tracking, the first, second and third derivatives respectively.
asymptotic features of that vicinity with εM changing follow -15
now from the homogeneity of the system. The derivative tracking deviations changed to 9.44⋅10 ,
-10 -6
2.97⋅10 , 2.42⋅10 and 0.00410 respectively after τ was
-5
Theorems 1 and 2 are simple consequences of Proposition 3 reduced to 5⋅10 . That corresponds to Theorem 3.
and the homogeneity of the system. To prove Theorem 3 it is
sufficient to consider Fifth order differentiator.. The attained accuracies are
-13 -10 -7 -4
n/( n + 1) 1.46⋅10 , 7.16⋅10 , 9.86⋅10 , 3.76⋅10 , 0.0306 and 0.449
σ 0 = -λ0 |σ0 (tj)| sign(σ0(tj)) + σ1 , (9) for tracking the signal, the first, second, third, fourth and
-4
instead of (6) with tj ≤ t < tj+1 = tj + τ. The resulting hybrid fifth derivatives respectively with τ = 5⋅10 . The derivative
-16 -12
system (9), (7) is invariant with respect to the transformation tracking deviations changed to 4.44⋅10 , 6.34⋅10 ,
-8 -5
2.24⋅10 , 2.44⋅10 , 0.00599 and 0.265 respectively after τ
( ηt, ητ, η
n-i+1
(t, τ, σi) σi). (10) -5
was reduced to 5⋅10 . The latter results are demonstrated in
On the other hand, it may be considered as system (6), (7) Fig. 1. Reduction of τ no longer leads to noticeable
with arbitrarily small ε for any fixed region of the initial improvement of the fifth-derivative tracking. Following is
values when τ is sufficiently small. Applying successively the explanation.
Proposition 3 and the homogeneity reasoning with
transformation (10) achieve Theorem 3. The demonstrated asymptotics is not trivially obtained by
computer simulation. For example, the computer realization
Remark. It is easy to see that the above proof may be accuracy of the harmonic functions is not sufficient here. In
transformed in order to obtain a constructive upper other words, in computer GWG sin t ≠ cos t, and even the long
estimation of the convergence time. Also the accuracy may double precision does not help here. For that reason the
be estimated by means of the suggested inductive approach. input was modeled as a solution of the corresponding
differential equation and most calculations were carried out
4. Numeric differentiation results with the long double precision. Nevertheless, there is such a
value of the sampling interval τ = τ∗ that any further
Following are equations of the 5th-order differentiator with reduction of τ no longer leads to any considerable accuracy
-4
simulation-tested coefficients for C = 1. As mentioned, it improvement. In particular, τ∗ is about 10 with the
contains also differentiators of the lower orders: -3
differentiation order 4 and τ∗ ≈ 10 with the differentiation
5/6 order 5, when the Lipschitz constant C of the 4th and 5th
z 0 = v0 , v0 = -50| z0 - f(t)| sign(z0 - f(t)) + z1 , (11)
4/5 derivative respectively is C = 1. While preparing this paper
z 1 = v1 , v1 = -30 | z1 - v0| sign(z1 - v0) + z2 , (12)
3/4 the author wanted to demonstrate 10th order differentiation,
z 2 = v2 , v2 = -16 | z2 - v1| sign(z2 - v1) + z3 , (13) but found that differentiation of the order exceeding 5 is
2/3
z 3 = v4 , v4 = -8 | z3 - v2| sign(z4 - v3) + z4 , (14) unlikely to be performed with the standard software. Further
1/2 calculations are to be carried out with precision higher than
z 4 = v4 , v4 = -4 | z4 - v3| sign(z4 - v4) + z5 , (15)
z 5 = -2 sign(z5 - v4); (16) the standard long double precision (128 bits per number).
References
[1] G. Bartolini, A. Ferrara, and E. Usai, “Chattering
avoidance by second-order sliding mode control”, IEEE
Trans. Automat. Control, vol. 43, no. 2, pp.241-246,
(1998).
[2] G. Bartolini, A. Pisano and E. Usai, “First and second
Fig. 3: Trajectory of the car derivative estimation by sliding mode technique”,
Journal of Signal Processing, vol. 4, no. 2, pp. 167-
176, March 2000, Research Institute of Signal
Processing, Tokyo, (2000).
[3] A. Dabroom, H.K. Khalil, “Numerical Differentiation
Using High-Gain Observers”, Proc. of the 37th Conf.
on Decision and Control (CDC'97), San Diego,
California, December '97, pp. 4790-4795, (1997).
[4] S.V. Emelyanov, S.K. Korovin, and L.V Levantovsky,
“Higher order sliding regimes in the binary control
systems”, Soviet Physics, Doklady, vol. 31, no. 4, pp.
291-293, (1986).
[5] A.F. Filippov, Differential Equations with
Discontinuous Right-Hand Side, Kluwer, Dordrecht,
Fig. 4: Steering angle the Netherlands, (1988).
[6] A. Levant (L.V. Levantovsky), “Sliding order and
sliding accuracy in sliding mode control”, International
6. Conclusions Journal of Control, vol. 58 , no.6, pp.1247-1263,
(1993).
[7] A. Levant, “Robust exact differentiation via sliding
Arbitrary order real-time exact differentiation with the
mode technique”, Automatica, vol. 34, no.3, pp. 379-
optimal input-output noise asymptotics is demonstrated.
384, (1998).
Together with the arbitrary order sliding controllers it
[8] A. Levant, “Controlling output variables via higher
provides for full SISO control based on the input
order sliding modes”, Proc. of the European Control
measurements only, when the only information on the
Conference (ECC’99), September 1999, Karlsruhe,
controlled process is actually its relative degree. The
Germany, (1999).
proposed differentiator as well as the resulting closed-loop
[9] A. Levant, “Higher order sliding: differentiation and
controller are robust with respect to measurement noises. At
black-box control”, Proc. of the 39th IEEE Conf. on
the same time this robustness drastically decreases with the
Decision and Control (CDC2000), December 2000,
growing differentiation order. The reason is not an
Sydney, Australia, pp. 1703-1708, (2000).
unsuccessful differentiator structure, but the very nature of
[10] A. Levant, “Universal SISO sliding-mode controllers
higher order differentiation.
with finite-time convergence”, to appear in IEEE
Transactions on Automatic Control, (2001).
It is easy to see that most practically important problems of
[11] R. Murray and S.Sastry, “Nonholonomic motion
output control are covered by the cases when relative degree
planning: steering using sinusoids”, IEEE Trans.
r equals 2, 3 and 4. Indeed, according to the Newton law, the
Automat. Control, vol. 38, no. 5, pp.700-716, (1993).
relative degree of a spatial variable with respect to a force,
[12] L. Rosier, “Homogeneous Lyapunov function for
being understood as a control, is r = 2. Taking into account
homogeneous continuous vector field,” Syst. Control
some dynamic actuator, achieve relative degree 3. If the
Lett., vol. 19, no. 4, pp. 467–473, (1992).
actuator input is needed to be a continuous Lipschitz
[13] V.I. Utkin, Sliding Modes in Optimization and Control
function, the relative degree has to be artificially increased to
Problems, Springer Verlag, New York, (1992).
4. As was mentioned above, the rth order sliding controller
(r-1) [14] X. Yu and J.X. Xu, “Nonlinear derivative estimator”,
requires availability of σ, σ , ..., σ . Thus, in practice the Electronic Letters, vol. 32, no. 16, (1996).
need for more than 3 successive output derivatives is very
(r-1)
rare. In fact, often only the sign of σ is actually required.