Some results on Iterative Learning Control with disturbances
M. Norrlof and S. Gunnarsson Department of Electrical Engineering Linkoping University, S-581 83 Linkoping, Sweden
WWW: http://www.control.isy.liu.se
Email: mino@isy.liu.se, svante@isy.liu.se
March 17, 1998
REGLERTEKNIK
AUTOMATIC CONTROL LINKÖPING
Report no.: LiTH-ISY-R-2020
Submitted to CCSSE'98 (First Conference on Computer Science and Systems Engineering in Linkoping)
Technical reports from the Automatic Control group in Linkoping are available
by anonymous ftp at the address
ftp.control.isy.liu.se. This report is
contained in the compressed postscript le
2020.ps.Z.
Some results on Iterative Learning Control with disturbances
M. Norrlof and S. Gunnarsson
Department of Electrical Engineering, Linkoping University, Linkoping, Sweden
Email: mino@isy.liu.se, svante@isy.liu.se
Abstract
Iterative Learning Control is presented briey together with a review of some results on convergence and some new results on Iterative Learning Control with load disturbances and measurement disturbances. In the example presented, Iterative Learning Control is used in combination with conventional feed-back and feed-forward control, and it is shown that learning control is highly a ected by the di erent disturbances. Some ideas on how the lters in the ILC algorithm should be chosen are also discussed.
Keywords
: Robot control, iterative learning control, disturbance a ects
1 Introduction
In many applications manipulators repeat their motions over and over in cycles, this is the case in for example laser cutting. Often the robot manufacturer gives a measure of the accuracy of the robots repeatability, this measure tells how well the robot repeats the same motion. If there is an error in the absolute accuracy, i.e. a dierence between the programmed path and the actual path, this error is repeated with the same accuracy as the repeatability. The basic idea behind iterative learning control (ILC) is that the controller should learn from previous cycles and perform better every cycle
1. ILC was rst introduced by Arimoto et al 2] in 1984 and since than many papers have been addressing robot control in combination with iterative learning control, e.g. 4], 11], 13], 7], and 9]. The convergence properties of the algorithm were addressed already in 2], and are further covered in many articles e.g. 10], 1], and 8].
In this paper the fundamental properties of the ILC algorithm applied as a complement to conventional feed-forward and feed-back control will be illus- trated. Focus will be on disturbance aects but also convergence and robustness will be discussed.
1
Better means here with respect to the absolute accuracy.
2 Problem Statement
Consider the representation of a control system shown in Figure 1. After we have introduced a load disturbance and a measurement disturbance we get
Y
=
G(
U+
D) +
N(1)
where
2 U,
Y,
D, and
Nrepresents input, output, load disturbance, and mea- surement disturbance respectively. G is the transfer function of the system, in the example in this paper a linear model of a one degree of freedom robot manipulator. The system is controlled using a combination of feed-forward and feed-back control. The control law is given by
U
=
FfYD+
F(
YD;Y) (2) where
YDdenote the reference signal.
Ffand
Fdenote the transfer function of the feed-forward and feed-back regulators.
+ + + +
+ -
F
f
F G
y
d
(
t)
d
(
t)
y
(
t)
n
(
t)
u
(
t)
z
(
t)
Figure 1: A system with load disturbance and measurement disturbance, con- trolled with feed-forward and feed-back regulators
To evaluate the properties of the applied algorithm we use the control sig- nal generated by the feed-back regulator when no measurement disturbance is applied,
E
=
F(
YD;Z) (3)
This error signal will of course not be used in the feedback for the algorithm because it is not available in the real system. If the feed-back controller is of PD type the error will be a linear combination of the position error and the velocity error, which is reasonable, the control objective is to minimize the position error and the velocity error. Using equations (1), (2), and (3), the error can be described as
E
=
GC((
G;1;Ff)
YD;D+
FN) (4) where
GCis the transfer function of the closed loop system given by
G
C
= 1 +
FGFG(5)
Note that the error the controller will use in the feedback is dened by,
E
con
=
GC((
G;1;Ff)
YD;D) (6) how this will aect the properties will be discussed in section 5.
2
When capital letters are used in the following it indicates that the signals are transformed,
the discussion covers both continuous and discrete time signals unless otherwise stated.
3 Outline of the ILC Method
As stated in the introduction the input signal,
yD, in many applications is repet- itive. This means that if the control system is time invariant
3and there is an error in the trajectory following in the rst iteration, this error will be repeated cycle after cycle. If the dynamics of the system is largely repeatable a control algorithm that improves performance from trial to trial can be constructed. A new control signal
Ukis added to the control signal
Uin Figure 1 and the input signal to the system will thus be given by
U
k
=
FfYD+
F(
YD;Yk) +
Uk(7) The index k indicates iteration or cycle number. Considering only linear oper- ations the updating of the correction signal can, in the frequency domain, be expressed as
Uk +1=
Xkj=0
H
j
(
Ej;FNj) (8) where
Hj j= 0
:::kare linear lters. Note that the denition of
Ein equa- tion (3) implies that we have to compensate for the measurement disturbance by subtracting
FNj. The updating formula used in this paper, though, is
Uk +1=
H1Uk+
H2(
Ek;FNk) (9) where
H1and
H2are linear lters. The choice of the lters
H1and
H2is the main task when designing an iterative learning control algorithm, since the lters determine the convergence and robustness properties. One method for choosing appropriate lters in the update equation is presented in 5] where methods from design of robust controllers are applied and the lters are designed to give a convergent ILC algorithm despite uncertainties in the process model. In 6] the problem is considered from a dierent viewpoint and the choice of the ILC input signal is formulated as an optimization problem, resulting in a time domain updating equation for the input signal. A variant of this approach is also discussed in 8].
4 Convergence Properties
We shall now investigate how the error signal behaves when the update equation (9) is used. If we dene ~
Eas the disturbance free error signal obtained in the
rst iteration when
U00 we get
~
E
=
GC(
G;1;Ff)
YD(10) From equations (4), (7), and (9) the following can be derived
E
k +1
= ~
E;GCH1Uk;GCH2(
Ek;FNk)
;GCDk +1+
FGCNk +1(11) By adding and subtracting relevant terms on the right hand side we arrive at
E
k +1
= (1
;H1) ~
E+
H1( ~
E;GCUk;GCDk+
FGCNk) (12)
; G
C H
2
(
Ek;FNk) +
H1GCDk;GCDk +1;H1FGCNk+
FGCNk +13
no adaption
3
by expanding
Ukusing equation (9) and then use equation (11) we get the following error update equation
E
k +1
= (1
;H1) ~
E+ (
H1;GCH2)
Ek(13) +
GC(
H1Dk;Dk +1) +
FGC(
Nk +1+ (
H2;H1)
Nk)
A corresponding equation is presented in 13] for the open loop case and for load disturbances only.
The convergence properties are determined by the homogeneous part of the dierence equation (13) and referring to 4] the convergence condition, in the continuous-time case, is that
jH
1
(
i!)
;GC(
i!)
H2(
i!)
j<1
8 !(14) This criterion can be viewed in a Nyquist diagram, and the condition is that the Nyquist curve of
GCH2should be held within the circle with center in +1 and radius 1. To the right in Figure 2 the amplitude curve of a
H2lter is shown above and the resulting convergence property for the ILC algorithm, 1
;GcH2, is shown below. It is obvious that the convergence is best for low frequencies and that it is very slow for frequencies close to the Nyquist frequency.
−0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
−1
−0.5 0 0.5 1
Nyquist curve
Real
Imag
GcH
Gc learning circle
stab. circle
100 101 102 103 104
0 2 4 6 8 10 12 14
Amplitude curve for the filter H
100 101 102 103 104
0 0.2 0.4 0.6 0.8 1
Amplitude curve for the error updating filter
Figure 2: Learning circle and stabilizing circle (left). Example of lters,
H11,
H
2
(upper right), the resulting updating lter 1
;GCH2(lower right) Provided that the learning procedure converges the error signal becomes
E
= 1
;H11
;H1+
GCH2E~ (15)
We see that by using
H16= 1 we are not able to eliminate the error completely, but as will be seen later other advantages are obtained by this choice. An alternative parameterization of the lters in the learning law was presented in 5], where
H
1
=
Q H2=
QL(16)
Q
and
Lare linear lters. In 12] this formulation is used with
Qdened as
Q
= 1 1 +
V(17)
and
La scalar. The condition for convergence, based on 5], becomes
k
1
;LGCk1<kQ;1k1=
k1 +
Vk1(18)
and it is obvious that the stability region can be extended by a suitable choice of the lter
Q, resulting in a so called stabilizing circle (see Figure 2).
By letting
Qbe frequency dependant the stability region can be extended in a frequency dependant way and equation (18) shows that if we choose
Qas a low-pass lter,
Q;1will extend the stability region for high frequencies. In 5] it is also shown that the lter
Lcan be found through a 'model matching problem', i.e. solving the
H1problem
L
= arg min
L2H1
kQ
(1
;LGC)
k1(19) This minimization will result in
kQ
(1
;LGC)
k1=
<1 (20) It should be noted that the smaller
is the faster the convergence of
uand
e
.
5 Disturbance Aects
We will now study input and output disturbances separately and we will start by exploring the load disturbance. First we recall the equation describing how the error is updated
E
k +1
= (1
;Q) ~
E+
Q(1
;GCL)
Ek(21) +
GC(
QDk;Dk +1) +
FGC(
Nk +1+
Q(
L;1)
Nk)
here with the denition of the lters
H1and
H2as is found in equation (16).
5.1 Load disturbance
A number of observations can be made using equation (21). Let us rst con- sider the case
Q1 and
N0, which means that we have no measurement disturbance. This implies the update equation
E
k +1
= (1
;GCL)
Ek+
GC(
Dk;Dk +1) (22) The disturbances contribute to the error equation by their dierences between the iterations. If a disturbance is of repetitive nature in the sense that the disturbance signals
dk(
t) =
dk +1(
t) for all
k, the contribution to the error dier- ence equation is zero. This assumption is likely for the load disturbance where for example load disturbances due to gravitational forces can be expected to be rather similar during dierent iterations.
Let us also consider the situation with
Q 6= 1, neglect measurement dis- turbances and assume that
dk(
t) =
d(
t)
8 k. This corresponds to the error dierence equation
E
k +1
= (1
;Q) ~
E+
Q(1
;GCL)
Ek(23)
; G
C
D
(1
;Q)
The load disturbance will act as a driving term similar to the initial error ~
E, and we can draw some conclusions about the frequency content of this driving term. We know that the closed loop system
GC, and the lter
Qare both of low pass type and this means that 1
;Qis of high pass type, multiplying these two will give a bandpass lter with a passband around the bandwidth of the closed loop system. However, if the load disturbance is of low frequency type or the
lter
Qcan be chosen with a higher crossover frequency than the bandwidth of the system, this will not be a problem.
5
Frequency (rad/sec)
Singular Values (dB)
Singular Values
10−4 10−2 100 102 104
−300
−250
−200
−150
−100
−50 0 50
1 − G c L
Frequency (rad/sec)
Singular Values (dB)
Singular Values
10−4 10−2 100 102 104
−300
−250
−200
−150
−100
−50 0 50
Q(1 − G c L)
Figure 3: The amplitude plot of
jQ(1
;GCL)
jwith
Q1 (left), and the same amplitude plot but with
Qof low pass type (right)
5.2 Measurement disturbance
Measurement disturbances, on the other hand, are more likely to be of random character which means that
nk +1(
t)
6=
nk(
t) in general, and there will hence always be a driving term on the right hand side of equation (21) that prevents
E
k
from tending to zero. Consider the case when we only have a measurement disturbance, this means that
YD0 and
D0. The fact that
YD0 implies
~
E
0, and we get
E
k +1
=
Q(1
;GCL)
Ek+
FGC(
Nk +1+
Q(
L;1)
Nk) (24) this equation can also be written as a nite sum
E
k +1
=
FGC(
Nk +1+
L(1
;GC)
Xkj=0 Q
k ;j;1
(1
;GCL)
k ;j;1Nj) (25) A sucient condition for stability is
lim
k !1 kE
k k
1
<M
(26)
which can be proved to hold when
kQ
(1
;GCL)
k1=
<1
and kNjk1K<1(27) because then the sum has an upper bound
k
X
j=0
kQ
(1
;GCL)
kk ;j;11 kNjk1(28)
k
X
j=0
j
KK
1
;1
(29)
Let us choose
Q1 and
L= ^
G;1C(1
;HB) where
HBis a high-pass lter
which decides the desired convergence properties of the ILC algorithm. ^
GCis a
model of the closed loop system including
Fand a model of
Greferred to as ^
G.
In the left plot in Figure 3 the principal behavior of
j1
;GCLjis shown and we
see that in this case we will have problems at high frequencies. If the
Qlter is
introduced the high frequency disturbance impact can be reduced. In the right
plot of Figure 3 the aect when Q is a low pass lter is shown.
We see that by introducing
Qwe can reduce the measurement disturbance impact on the ILC algorithm, but with the price that the error in the trajectory following will not tend to zero. The demand of trajectory following and distur- bance reduction will in this sense be contradictory and we have to have this in mind when designing the lters
Qand
L.
−0.5 0 0.5 1 1.5 2 2.5 3
−30
−20
−10 0 10 20 30
sec
rad
yD − reference signal
−0.5 0 0.5 1 1.5 2 2.5 3
−60
−40
−20 0 20 40 60 80 100
sec
rad/s
dyD/dt − velocity reference signal
Figure 4: The reference signal,
yD(left) and the derivative, _
yD(right).
6 Simulation Example
We shall consider a simplied description of a single robot joint modeled as a double integrator, i.e.
G
(
s) = 1
Js
2
(30)
Since the system is computer controlled we shall use the discrete time represen- tation given by the transfer function
G
(
z) =
T2(
z+ 1)
2
J(
z;1)
2(31)
where
J= 0
:0094 is the moment of inertia. The system is controlled by a discrete time PD-regulator given by
F
(
z) =
KP+
KDT
(
z;1)
z
(32) with
KP= 12
:7 and
KD= 0
:4. The feed-forward lter is chosen as a double dierentiation represented by
F
f
=
J(
z;1)
2T 2
z
2
(33)
where
Jis the estimated moment of inertia. The correction signal will be updated according to equation (9) where
H1(
z) =
Q(
z) and
H2(
z) =
Q(
z)
L(
z) are lters that both may be non-causal. The model is simulated using 1 kHz sampling frequency. For evaluation of the algorithm we shall apply the reference trajectory shown in Figure 4.
6.1 Design of the
Llter
For
Q(
z) = 1 the ideal choice of
Lwould be to choose it as the inverse of
G
C
(
z), which, theoretically, would result in convergence to zero in one step.
7
This is however an unrealistic choice since it requires exact knowledge of the system and results in a lter with very high gain for high frequencies. Instead we consider
L
(
z) = ^
G;1C(
z)(1
;HB(
z)) (34) where ^
GC(
z) denotes the closed loop transfer function we obtain by using the model of the open loop system. We will consider the case when we have a 30%
error in
J, the estimated moment of inertia in the model of the system. The
lter
HB(
z) is a Butterworth high pass lter (here of second order) for which the gain tends to one for high frequencies. Choosing
L(
z) according to this design rule, with cut-o frequency of the high pass lter equal to 0
:4 times the Nyquist frequency, gives the Nyquist curve depicted in the left plot of Figure 2. Figure 2 also shows
GC(
z) for comparison. The whole Nyquist curve is now inside the learning circle while it for large frequencies tends to the origin.
−0.5 0 0.5 1 1.5 2 2.5 3
−20
−15
−10
−5 0 5 10 15 20
sec
Nm
d − load disturbance
0 1 2 3 4 5 6 7 8 9 10
10−12 10−10 10−8 10−6 10−4 10−2 100
Iteration
Energy (arb. unit), log−scale
Energy in the error signal
with Q filter without Q filter
0 200
400
600 0 2 4 6 8 10
−20
−18
−16
−14
−12
−10
−8
−6
−4
−2
Iteration number Error signal spectrum (Ek) without Q
Hz
log10 of power (arb. unit)
0 200
400
600 0 2 4 6 8 10
−20
−18
−16
−14
−12
−10
−8
−6
−4
−2
Iteration number Error signal spectrum (Ek) with Q filter
Hz
log10 of power (arb. unit)
Figure 5: Load disturbance (upper left), error signal energy (upper right), error signal spectrum without Q lter (lower left), and error signal spectrum with Q
lter (lower right)
6.2 Load disturbance
The rst goal is to investigate how the load disturbance in!uences the ILC algorithm. In Figure 5, the upper left plot shows the applied load disturbance as a function of time. This disturbance signal is applied at the same time every cycle as a repetitive disturbance. Recall equation (23)
E
k +1
= (1
;Q) ~
E+
Q(1
;GCL)
Ek; G
C
D
(1
;Q)
If
Q1 the load disturbance will not have any aect on the result, except in the
rst iteration. By introducing the lter
Q6= 1 we can see that the disturbance
will have an aect on the resulting error. Normally, as discussed in section 5, the
lter
Qis chosen as a low-pass lter which means that 1
;Qwill have high-pass characteristics. In Figure 5 the result when we have introduced a lter
Qcan be seen. In the upper right plot the energy in
Ekas function of iteration is shown and in the lower plots the spectrum of
Ekis shown as a function of iteration.
Clearly the introduction of a lter
Qreduces the convergence speed of the ILC algorithm and there will also not be convergence to zero of the error. Already after one iteration when the lter
Qis used the energy in the error has almost reached its nal value and also in the spectrum it is possible to see that there is not much change after the rst iteration.
0 100
200 300 400
500 0
2 4
6 8
10
−14
−13
−12
−11
−10
−9
−8
−7
−6
−5
Iteration number Position error signal spectrum (Epk) without Q filter
Hz
log10 of power (arb. unit)
0 100
200 300 400
500 0
2 4
6 8
10
−14
−13
−12
−11
−10
−9
−8
−7
−6
−5
Iteration number Position error signal spectrum (Epk) with Q filter
Hz
log10 of power (arb. unit)
0 1 2 3 4 5 6 7 8 9 10
100 101
Iteration
Energy (arb. unit), log−scale
with Q filter without Q filter
0 1 2 3 4 5 6 7 8 9 10
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
Iteration
max(abs(ek)) − [rad]
Speed of convergence
with Q filter without Q filter
Figure 6: Position error signal spectrum, without Q lter (upper left), position error signal spectrum with Q lter (upper right), position error signal energy (lower left), and maximum absolute position error (lower right)
6.3 Measurement disturbance
The measurement disturbance,
Nk, introduced is a time discrete white noise process with standard deviation 3
:2
10
;4. The aect of introducing a measure- ment disturbance is very important because, when the algorithm is tested on a real robot, measurement disturbances will always be present. Therefore it is of great importance to really investigate how the lters should be chosen to reduce the in!uence of the measurement disturbance. First recall equation (25)
E
k
=
FGC(
Nk+
L(1
;GC)
k ;1Xj=0 Q
k ;j;1
(1
;GCL)
k ;j;1Nj)
If the lters are chosen as in equation (16) it is possible to shape the transfer function
Q(1
;GCL) by introducing
Q6= 1, as shown in Figure 3. By looking
9
at equation (25) it is possible to see that the in!uence of the measurement disturbance at high frequencies can be reduced by using
Qas a low-pass lter.
In Figure 6 the spectrum of the position error signal
YD;Zis shown without Q lter (upper left) and with Q lter (upper right). In the lower left plot the energy of the position error signal is shown as a function of iteration, and in the lower right plot the maximum absolute error in the time domain is shown as a function of iteration. It is obvious in Figure 3 that introducing the
Qlter will make the aect of measurement disturbances decrease at certain frequencies. In the convergence plot of Figure 6, it is possible to see again that nearly all the aect of the size of the error is in the rst iteration. This is clearly an aect of introducing the measurement disturbance. In the previous section it was clear that if
Q1 the error would decrease to zero, but here we can see that this will not be the case when there is a measurement disturbance.
7 Conclusions
The Iterative Learning Control algorithm has been studied when there are dis- turbances in the system and it is shown that the choice of lters in the algorithm is very much dependent on what kind of disturbance we have. In the example it is shown that nearly static values in the error are already reached in the rst iteration of the algorithm. This result has also been found in experiments, e.g.
in 3] where an ILC algorithm is applied on a DELTA robot. Some new results on ILC with measurement disturbances are presented and some ideas of how the lters in the ILC algorithm should be chosen are also discussed.
8 Acknowledgments
This work was supported by CENIIT at Linkoping University and ABB Robotics within ISIS at Linkoping University.
References
1] N. Amann, D. H. Owens, and E. Rogers. Iterative learning control for dis- crete time systems with exponential rate of convergence. Technical Report 95/14, Centre for Systems and Control Engineering, University of Exeter, 1995.
2] S. Arimoto, S. Kawamura, and F. Miyazaki. Bettering operation of robots by learning. Journal of Robotic Systems, 1(2):123{140, 1984.
3] E. Burdet, L. Rey, and A. Codourey. A trivial method of learning control.
In Preprints of the 5th IFAC symposium on robot control, volume 2, Nantes, France, September 1997.
4] J. J. Craig. Adaptive Control of Mechanical Manipulators. Addison-Wesley Publishing Company, 1988.
5] D. de Roover. Synthesis of a robust iterative learning controller using an
H
1