A NON-LINEAR SYSTEM DISCRETE ROBUST TRACKING ALGORITHM

Luiz Carlos Campos Pedroza

Federal Center of Technology, Department of Electrical Engineer , Av. Maracanã, 229, Rio de Janeiro, Brazil
[email protected]

Bruno Campos Pedroza

Universidade Federal Fluminense, Rua Passos da Pátria, 156, Bloco D, Sala 505, Niterói, RJ, Brazil, CEP 24210-240
[email protected]

Abstract - In this paper, a new formulation is proposed to the tracking problems in the Discrete-Time Nonlinear SISO systems in
presence of large uncertainty. Based on this formulation a Robust Adaptive Controller is proposed. This controller is based in the
generation of a viable trajectory which take in account the desired trajectory and the parameters uncertainties due to the estimation
problems.

Keywords - Robust Control, Adaptive Control, Variable Structure Control, Tracking Problems.

1. Introduction

One of the requirements of any good control system is accuracy in tracking trajectories. Achieving this in complex
systems, in presence of large uncertainty concerning the process, is a challenge to a good control.
Adaptive control is an appealing approach for controlling uncertain dynamics. In principle the systems can be
uncertain in terms of its dynamics structures or in its parameters. However, adaptive control can deal better with
parameters uncertainty (Narenda and Annaswamy, 1989). In the early years, a considerable amount of research has been
developed to the so-called Robust Adaptive Control, where close-loop stability properties are retained in the presence
not only of parametric uncertainty, but also modeling errors such as additive disturbances and unmodeled dynamics
(Ioannou and Sun, 1995), (Ioannoou and Datha, 1991).
The sliding mode control was originated in Soviet Union (Utkin, 1997), (Young and Drakunov, 1992), (Fillipov,
1960), (Aizermann and Gantmakher, 1957). Classical sliding mode control has drawbacks limiting its applicability such
as chattering and large control authority (Slotine and Li, 1991). The use of state observers is one form to minimize the
first problem (Bondarev et al., 1985) together with saturation type relay that originate a boundary layer of accuracy
(Slotine and Coetsee, 1986).
The main concern of this paper is to introduce a new approach in discrete-time robust control (Pedroza and
Pedroza, 2001), which stay between the adaptive and sliding control. As we will see, the main difference of this
approach is that the wished trajectory to be tracked gives only a shape. Over this shape a new trajectory is computed
concerning the maximum error obtained in parameters estimation.

2. Theoretical Foundation

Let the nonlinear single-input single-output system described by

, (1)

where: are vectors dependents of time and system’s state; ; and

As discretezed, this system may be expressed as

, (2)

where:
Our tracking problem can be posed as: possessing estimated values ’(k) and ’(k) of (k) and (k) find the value
of u(k) that make the system track a wished trajectory yD(k). Because the parameters estimated are seldom accurate there
will exist an error E(k) in the tracking of the desired trajectory defined as:

.
(3)

To guarantee that the system converges to the wished trajectory it is sufficient that

. (4)
 Handling the Equation (4) together with Equation (2) we can arrive at

(5)

We can see from Equation (5) that the value , when chosen as u(k), cancels the tracking

error E(k+1). However, as the true values of the (k) and (k) are not known, theirs estimations are used to compute
u(k), conducting to a value uc(k) different from uR(k):

(6)

We can note from Equation (5) that Equation (4) would be satisfied by this uC(k) only if

(7)

or, in others worlds, if

(8)

3. Trajectory Tracking

From the results already shown, we can see that save in situation where the estimation errors are nulls there will
ever be a nonzero tracking error when the use of Equation (6). Using probabilistic tools is possible to define maximum
estimation errors such that

(9)

and

(10)

where: ,   +.
To guarantee that under these uncertainties, we propose evaluating, in every interaction k, a set of
possible points of being reached, yP(k+1), and from these points define a target point yT(k+1) which is in according with
yD(k+1). All points yP(k+1) must satisfy the condition given by the Equation (8) with (k) and (k) changed for theirs
worst estimations. So that, these points need to satisfy the set of four equations

(11)

resulting of signal combinations of  and .

After the choice of yT(k+1), the input uC(k) to be applied in the system is calculated as

, (12)

from the change of yD(k+1) by yT (k+1) in Equation (6).

When the calculus of yP(k+1) two things may happen. At first, yD(k+1) may not belongs to the set of yP(k+1). In
this case, the point which belongs to yP(k+1) and is closer to yD(k+1) is chosen as the target yT(k+1). At second, in
accordance with the errors ,  and E(k), the set of Equations (11) may not have real solution. In this case the imaginary
part of the complex solution must be disdained and the real part of these solutions used as yP(k+1). In both cases, the
uC(k) would probably make . However, this increase in the tracking error would give more freedom to
the proper solution of Equations (11) in the next step.
We can note that the solution of Equations (11) needs a heavy computational effort. A simplification in that
calculus is hereafter proposed based in the following theorem.

3.1. Theorem

If
(a) , so that ;
(13)

(b) , with .

(14)

then

(15)


The necessary conditions (a) and (b) are not difficult to be satisfied in practice because the signal of (k) is
generally known a priori and  and  are generally small.
From Theorem and Equation (11), we can see that to guarantee we only have to solve the
quadratic inequality

, (16)

what needs a low computational effort problem. After that, we can choose yT(k+1) in accordance with have already
been commented.
To simplify our algorithm, in the cases where conditions (b) are not satisfied we will go on computing yP(k+1) by
Equation (16). In this case the tracking error might be increased leading to more freedom to satisfy the theorem
conditions in the next step of the algorithm.

3.2. Algorithm

1. Let k=1;
2. Choose values to , ;
3. Calculate ;
4. Estimate and ;
5. Find, using (16) the set of possible outputs ;
6. Choose as the element computed in step 5 which is closer to ;
7. Calculate, using (12) the uc(k) to be introduced in the system;
8. Let k=k+1 and go to step 3;

4. Trajectory Tracking

In this section we will show an example of our tracking method applied to a nonlinear SISO system. To take
advantages of the algorithm simplicity, at first, we introduce a naive estimation method that will be used in this
example.

4.1. A Naive Estimation

 Beginning with a good hint of min e MAX we estimate (k) as
.
(17)

Note that this is a backward estimation based in past value of u and actual value of y.
To be able to guarantee the existence of the maximum estimations error in the spirit of Equations (9) and (10) some
condition have to be satisfied.

i) First of all, there must be a sufficiently small  such that

,
(18)

a reasonable situation for systems with high sampling rate.

ii) At second, there must be a sufficiently small  such that
, (19)

a reasonable situation to a nonlinear system with slow dynamics.

ii) At third, the system must have limited input,

, (20)

this is a true condition for real system.

Handling Equations (1), (17), (18), (19), (20) we conclude that

(21)

where: (l+)= and = are the maximum estimation errors of (k) and (k).

4.2. Numerical Example

In this example we study an application of the robust tracking algorithm to the pendulum inverted problem. This
system has the following structural model

(22)

where m(pendulum mass)=1kg, M(car mass)=2kg, g(gravitational acceleration)=9.81m/s 2, l(pendulum length)=3m. In

every state of the system is introduced a white noise perturbation with normal distribution and variance 0.001. Notice
that the output of the system is a weighted sum by 1 and 2 of position and velocity of the pendulum. This was done to
improve the performance of the tracking system by the use of more information of the states.
This continuos system was simulated in computer and sampled at every h=0.001s. At the sampling time the system
may be written in a difference equation form as:
 (23)

The function comes from the discretization of system (22) with sampling time h=0.001s and it is dependent of
the state of the system at time k and perturbations.
Due to our system’s output choice the correspondent wished trajectory is chosen as

(24)

where initial=0.49rd and final=0rd and =2,5s. Notice that the pendulum initial position is almost horizontal, the wished
final position is vertical and the system is initially at rest. Note also that 1 and 2 are necessarily the same constants
used to weigh the system output.
At this point, the next step consists in obtaining a good estimation of and using
the naive estimator described previously. Doing some approximations on Equation (22) we can conclude that

. (25)

taking to conclusion that

(26)

and

(27)

are good estimations of the bound of .

 The maximum input allowed to the system is 300N; it is , guarantying the Equations (20) and (21).
In a practical case, this maximum input might be decreased with the inverse of cos(), but this was not done in this
example.
Beginning with a choice of 1=0.7 and 2=0.3 and running the robust track algorithm the system tracks the
and parts of the wished trajectory as shown in the Figures 1 and 2. Notice that there is an average error of 0.5
rd in the final position trajectory and an average error of –0.9 rd/s in the final velocity. However the sum of these two
average errors weighted by 1, 2 is near zero showing that the algorithm works well.

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

Figure 1: Wished position trajectory and position tracked.

-1

-2

-3
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

Figure 2: Wished velocity trajectory and velocity tracked.

For a good tracking of position and velocity individually, we can change in every algorithm interaction the values
of 1, 2, choosing at random any value to 1 between [0 1] a computing 2=1-1. As we can see the continuos change
in theses values nullify neither the convergence of the algorithm nor the use of the naive estimator. On other side, this
proceeding improves the performance of the position and velocity tracks as shows figures 3 a 4.

1.5

0.5

-0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

Figure 3: Position wished and position performed.

-2

-4

-6
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Figure 4: Velocity wished and velocity performed.

5. Final Remarks

In this paper a new method to tacking trajectories in discrete-time systems was introduced. As the main advantages
of this method we can mention: a low computational effort when compared with the traditional adaptive control; a
possibility of generating smoother signal inputs when compared with sliding modes; the possibility of using simplified
method of identification. In section 4, it was shown that based in a naive estimation method the robust tracking
algorithm was able to tracking with a good performance a wished trajectory in a complex system.
When using a very refined identification method the performance of the robust tracking will probable be improved,
however the performance of the naive estimator pointed out, as shown in the example, quite good. On other side the
naive estimator allow that, together with the robust tracking algorithm, to change the sampling frequency during the
process.
The performance of the algorithm, nevertheless, has yet to be better investigated in real complex non-linear SISO
systems and extended to MIMO systems.

6. Appendix

Proof of Theorem

At first, let , then from condition (b)

(a1)

From Equation (9), we have

.
(a2)

Supposing , using equations (a1) and (a2) we have,

and

.
(a3)

Now, supposing , we have,

and

.
(a4)

Then, from Equations (a3) and (a4), we conclude that

(a5)

By the same way, is easy to check that

(a6)

Hence, from Equations (a5) and (a6),

(a7)

Now, suppose and . From result (a7) and Equation (9), we have

(a8)

that bring us to

(a9)

From condition (b), we have

(a10)

Using Equations (a9) and (a10), we can write

(a11)

Hence, considering conditions (a) and (b), we conclude that

 (a12)

Similarly, we can prove that

(a13)

(a14)

(a15)

Hence, from results (a12) to (a15), and conditions (a) and (b), we have

(a16)

Now, for conclude the proof of theorem, from condition (a) and the theorem hypothesis, we have

Using result (a16), we have

(a17)

proving the theorem.

7. References

Aizerman, M.A., and Gantmakher, F.R., 1957, “On some features of Switching in Nonlinear Control with Piecewise-
Smooth Response”, Avtomatika i Telemekhanika, Vol.18, No.11.
Bondarev, A.G., Bondarev, S.A., Kostyleva, N.Y., and Utkin, V.I., 1985, “Sliding Modes in Systems with Asymptotic
State Observers”, Automation and Remote Control, Vol.6, pp. 679-684.
Fillipov, A.F., 1964, “Differential Equations with Discontinuous Right-hand Sides”, Mathematichiskii Sbornik, Vol.51,
No.1 (1960) in Russian. Translated in English, Am. Math. Soc. Trans. , Vol.62, No.199.
Ioannou, P.A., and Sun, J., 1995, ”Robust Adaptive Control”, Prentice-Hall, Englewood Cliffs ,NJ.
Ioannou, P.A., and Datha, A., 1991, “Robust Adaptive Control; A unified approach”, Proc. IEEE, Vol.79, No.12, pp.
1736-1768.
Narendra, K.S., and Annaswamy, A.M., 1989, ”Stable Adaptive Systems”, Prentice-Hall,Englewood Cliffs, NJ.
Pedroza, L.C.C., and Pedroza, B.C., 2001, “A Study in Discrete-Time Robust Control”, Proc. of the IASTED Int. Conf.
Modeling, Identification and Control, Vol. 1, Innsbruck, Austria, pp. 12-16
Slotine, J-J., and Coetsee, J.A., 1986, “Adaptive Sliding Controller Synthesis for Nonlinear Systems”, Int. J. Control,
Vol.43, No.6.
Slotine, J-J., and Li, W., 1991, ”Applied Nonlinear Control”, Prentice-Hall, Englewood Cliffs, NJ.
Utkin, V.I., 1997, “Variable Structure Systems With Sliding Modes”, IEEE Trans. Automat. Contr., Vol.AC-22, pp.
212-222.
Young, K.D., and Drakunov, S.V., 1992, “Sliding Mode Control with Chattering Redution”, Proc. 1992 Amer. Contr.
Conf., Chigago, IL, pp. 1291-1292.