Professional Documents
Culture Documents
Abstract- We propose a new Iterative Learning Controller (ILC) uncertainties that are exhibited during actual system operation
algorithm based on M. Pandit and D. Andress [3] and Niels or to the lack of suitable design techniques. Thus, it is not easy
Johannes Maria Bosch [10] considering the current cycle
to achieve perfect tracking using traditional control theories.
parameters and some part of the system dynamics. We have
selected a Single Inverted Pendulum - which is multivariable and ILC is a design tool that can be used to overcome the
highly unstable system, as a control object to demonstrate the shortcomings of traditional controller design, especially for
effectiveness of the proposed ILC algorithm. The test results of obtaining a desired transient response, for the special case when
proposed ILC algorithm are compared with the other the system of interest operates repetitively. For such systems,
implementations such as LQR (non-learning), Lifted ILC and ILC
ILC can often be used to achieve perfect tracking, even when
algorithm suggested by M. Pandit and D. Andress [3]. The results
show that the proposed algorithm gives much better control as the model is uncertain or unknown and we have no information
compared to LQR, Lifted ILC algorithm and there is a significant about the system structure in details and nonlinearity.
improvement in the settling time as compared to ILC algorithm There are many application of ILC such as, industrial robots,
given by M. Pandit and D. Andress [3]. computer numerical control (CNC) machine tools, wafer stage
motion systems, injection-molding machines, aluminum
I. INTRODUCTION extruders, cold rolling mills, induction motors, chain conveyor
systems, camless engine valves, rapid thermal processing, semi
ILC is based on the notion that the performance of a system
batch chemical reactors etc [2]. There are many variations of
that executes the same task multiple times can be improved by
the basic ILC scheme and are discussed in [2]. ILC algorithms
learning from previous executions (trials, iterations, passes) [1].
can be divided in two basic forms Continuous-time ILC and
It differs from most existing control methods in the sense that it
Discrete time ILC algorithm [9]. In our work, we focus
exploits every possibility to incorporate past control
especially on discrete-time ILC.
information: the past tracking error signal and in particular the
For discrete time system, it uses lifted system representation.
past control input signal into the construction of present control
All signals are lifted to discrete time domain and update rule is
action.
formed considering controller output and error at previous
ILC is a relatively recent but well-established area of
iteration. One of the author M. Pandit and D. Andress suggested
study in control theory, and can be categorized as branch of an
ILC update rule in [3], which is based on reformulation of ILC
intelligent control methodology. From 1984, ILC started to
problem as multivariable control problem. This has enabled the
become an active research area. In 1984 [6]-[8] are
possibility of designing an iterative learning controller using
independently published describing a method that iteratively
well known methods from multi variable control theory. Here,
could compensate for model errors and disturbances. The
the controller K is designed such that it will generate correction
development of ILC stems originally from the robotics area
to the actual control input vector so as to minimize performance
where repetitive motion shows up naturally in many
measure.
applications [5]. It is an approach for improving the transient
In case of general discrete time ILC algorithm, convergence
performance of systems that operate repetitively over a fixed
condition includes only direct transmission part of the system to
time interval.
be controlled. In the method given by M. Pandit and D. Andress
Although control theory provides numerous design tools for
[3], the controller K is designed using Riccati equation
improving the response of a dynamic system, it is not always
considering system dynamics. This results in better control with
possible to achieve desired performance requirements, due to
faster convergence (less number of learning cycles) as
the presence of un-modeled dynamics or parametric
A. Introduction
6
New Update rules for the proposed algorithm
Mx + N = F A new update rule for the proposed algorithm is give as
N = mx + mlθ cosθ - mlθ 2 sin θ
If one substitutes this equation into the first equation, we get the u k +1 = u k + ∆ u k + Ke k
first equation of motion for this system:
Proposed Iterative Learning rule: Problem statement is to add a
(M + m) x + mlθ cos θ − mlθ 2 sin θ = F correction to update rules so as to improve the convergence
To get the second equation of motion, speed of the system and to get a good tracking effect.
P sin θ + N cos θ − mg sin θ = mlθ + mx cos θ ∆ u k = L ek
− Pl sin θ − Nl cos θ = Iθ = L(Yk - Yd )
Combining these last two equations, we get the second dynamic Here Yd = 0, and Yk = Huk then
equation: ek = Yk
( I + ml 2 )θ − mgl sinθ = −mlx cosθ u k +1 = u k + LHu k + Ke k
= (I + LH)u k + Kek
Since Matlab can only work with linear functions, this set of
equations should be linearized about = 0. Considering tiny The SVD of H as discussed in [10] is given as:
vibration of the pendulum in the given point when =0, local
linearize the above equation, i.e. cos( ) 1, sin( ) 0 are used. T
After linearization the two equations of motion become (where Σ1 0 V1
u represents the input):
H = UΣV T
= [U1 U 2 ]
0 Σ 2 V2 T
as V T1 w2 = V T2 w1 = 0. Now, w1 is chosen as
The Eigen values of discrete system version of this system w1 = V1u
matrix are at 1.0000, 1.0000, 1.0568 and 0.9463. We can be see
that two poles are at 1.000 i.e. on the unit circle and one outside For the closed-loop system the expression
the unit circle at 1.0568. This confirms the intuition that the
u k +1 = u k + L * HV1u k + Ke k
system is unstable in open loop.
III. PROPOSED ILC ALGORITHM = (I + L * HV1 )u k + Kek (1)
is obtained.
In this section, a correction is proposed to ILC update rule
suggested by M. Pandit and D. Andress [3] based on the
By omitting 2, model reduction is applied (indicated by *).
approach discussed by Niels Johannes Maria Bosch in [10], for
Let
improving the response time of the system. The proposed new
algorithm is then implemented to SIP model discussed in L * HV1 = L * U 1 Σ 1V1T V1 = α I
section B and results along with observations are presented. −1
L* = α Σ 1 U 1T (2)
7
The proposed algorithm gives better transient response as
The parameter is chosen which decides the percentage of compared to other three techniques. The percentage reduction
correction applied to extension rule. in peak overshoot, undershoot and settling time as compared to
LQR is 60%, 28% and 58%, respectively.
Steps of implementation
In comparison with the lifted ILC, the reduction in peak
1. Build a system as in form of equation given in [3]. overshoot, undershoot and in settling time is 35%, 9% and 68%,
2. Design the controller K as given in [3]. respectively and as compared to the algorithm suggested by M.
3. Neglecting singular values which are nearly equal to Pandit and D. Andress [3], we obtained the percentage
zero or in other words we can say that by considering reduction in peak overshoot by -89%, undershoot by 4% and
only l columns of , build matrices Ul, Vl and l. settling time by 73%.
4. Then build learning gain matrix from equation (2). Finally, it is found that as compared to ILC algorithm by
5. Generate control input uk. Pandit and Andress [3], the proposed ILC algorithm with new
6. Apply the control input to system. update rules gives much faster response by reducing settling
7. Store control input and error of current iteration. time 73%.
8. From this current iteration data, generate control input It is seen that, there is a considerable improvement in the
from equation (1) to be applied to the system when settling time at the expense of increase in peak overshoot.
operated in the next iteration. Since, peak overshoot is within +/- 0.05 radiance, this does not
affect the overall performance of the system, at the same time
To demonstrate the effectiveness of the proposed algorithm, this is much less than the LQR and Lifted ILC.
we also design and implement the controller on SIP system
based on LQR, Lifted ILC and algorithm suggested by M. 0.1
8
Table 2 Percentage reduction in transient response
parameters obtained by proposed algorithm over the
existing methods
REFERENCES
[1] Kevin L. Moore and Jian-Xin Xu (Eds). Special Issue Iterative Learning
Control. Int. J. of Control, 73(10):819–999, 2000.
[2] Hyo Sung Ahn,YangQuan Chen and Kevin Moore,”Iterative Learning
Control:Brief Survey and Categorization”, IEEE Trans. Systems, Man and
Cybernetics vol.37 .no.6, Nov2007, pp1099-1118.
[3] D.Andress,M.Pandit,”Iterative Learning Control for Discrete Time
System as MultiLoop Control”, CCSP2000 International Conference on
Communication Controls and Signal Processing, July2000.
[4] Xianhe Zhang, Jie Wu, XishengZhan, ”Iterative Learning control on
Single Inverted Pendulum”, International symposium on Intelligent
Tech.Application Workshop, 2008.
[5] YangQuan Chen,Chagyun Wen, Iterative Learning Control Convergenc,,
Robustness and Application, Springer-Verlag, London, 1999.
[6] S. Arimoto, S. Kawamura, and F. Miyazaki, “Bettering operation of
robots by learning,” J. Robot. Syst., vol. 1, no. 2, pp. 123–140, 1984.
[7] G. Casalino and G. Bartolini, “A learning procedure for the control of
movements of robotic manipulators,” in Proc. IASTED Symp. Robotics
and Automation, San Francisco, CA, May 23–24, 1984, pp. 108–111.
[8] J. J. Craig, “Adaptive control of manipulators through repeated trials,” in
Proc. ACC, San Diego, CA, June 1984, pp. 1566–1572.
[9] C. Leondes, Discrete Time Control System analysis and design,
Academic Press,1995
[10] Niels Johannes Maria Bosch, “Iterative Learning Control on an Inkjet
Printhead”, Research Report, Eindhoven University of Technology , June
2005.
[11] John Stang, “The Inverted Pendulum” , Design Project, Cornell
University, May, 2005
[12] Jianqiang Yi, Naoyoshi yu,bazaki and Kaoru Hirota,” Systematically
Stabilization Fuzzy Controller for Single and Double Pendulum
Systems”, IEEE, 2000.
[13] Masaki Yamakita, Tasuku hoshino, Katsuhisa furuta,” Control Practice
Using Pendulum”, Proceedings of the American Control Conference San
Diego, California June 1999
[14] M. Yamakita, K.Nonaka, K.Furuta,”Swing Up Coiitrol of Double
Pendulum”, Proc. of ACC’93, pp. 2229/2234, (1993)
[15] M.Yamakita and K.Furuta,”Stabilization of a Limit Cycle in TITech
Double Pendulum ”, Proc. IFAC Congrcss (1999)
[16] Kent H. Lundberg James K. Roberge,“Classical Dual-Inverted-Pendulum
Control”, presented at IEEE conference on DECISION AND CONTROL
4399, 2003.
[17] Carina Hansen, Cecilia Svensson, “Construction and Control of an
Inverted Pendulum”, Master Thesis, Lund Institute of Technology, July
2000.
[18] Fatai Olalekan, Ayman Abdallah, SE 514 OPTIMAL CONTROL FOR
SINGLE AND DOUBLE INVERTED PENDULUM, project report.
[19] Daoying Pi, D. Seborg, Jianxia Shou, Youxian Sun, and Qing Lin.
Analysis of current cycle error assisted iterative learning control for
discrete nonlinear time-varying systems. In IEEE International
Conference on Systems, Man, and Cybernetics, pages 3508 – 3513,
Nashville, TN USA, Oct. 2000.