"Iterative Learning Control": From Academia to Industry

1
"Iterative Learning Control" From Academia to
Industry

• YangQuan Chen
• Department of Electrical and Computer Engineering
• Utah State University
• A Seminar at The University of Windsor
• June 14, 2001

2
Outline

• What is Iterative Learning Control (ILC)
• Historical Comments
• From Analysis to Design
• Industrial Application I (ABB robots)
• Industrial Application II (Seagate HDD)
• To Probe Further and My Recent Results
• Concluding Remarks

3
Intuitions

• What can we human beings get from doing
• the same thing over and over? Yes, skill".
• When a machine is operated to perform the
• same task repeatedly, can it do the job better
• and better?
• This is "iterative learning control (ILC)".

4
Control Design Problem
5
Systems that Execute the Same Trajectory Over
and Over
6
Errors Are Repeated WhenTrajectories are Repeated

• A typical joint angle trajectory for the example
  might look like this
• Each time the system is operated it will see the
  same overshoot, settling
• time and steady-state error. They did NOT make
  use the repetitiveness!
• Iterative learning control attempts to improve
  the transient response by
• adjusting the input to the plant during future
  system operation based on
• the errors observed during past operation.

7
Memory based

• Iterative Learning Control Scheme is memory-based.

System
Memory
Memory
Memory
Learning Controller
8
ILC vs. FBC

• A typical ILC algorithm has the form

Whereas a feedback control (FBC) has the form

• The subscript k indicates the trial or the
  repetition number.
• The subscript t indicates the time.
• All signals shown are assumed to be defined on a
  finite interval t ,and t ?0,

is the input applied to the system during the k
-th trial.

is the output of the system during the k -th
trial.

is the desired output of the system.

, is the error observed between the
actual output and the desired output during the k
-th trial.
9
Trial (k-1)
Trial k
Trial (k1)
Error
Input
(a) ILC
Error
Input
(b) Conventional feedback
10
Feedback-Feedforward Configuration
11
Arimotos 6 Postulations on ILC

• P1. Every cycle (pass, trial, batch, iteration,
  repetition) ends in a fixed time of duration Tgt0.
• P2. A desired output yd(t) is given a priori over
  0,T.
• P3. Repetition of the initial setting is
  satisfied.
• P4. Invariance of the system dynamics is ensured
  throughout these repeated iterations.
• P5. Output can be measured and the tracking error
  can be utilized in the construction of the next
  input.
• P6. The system dynamics are invertible, that is,
  for a given desired output yd(t) with a piecewise
  continuous derivative, there exists a unique
  input ud(t) that drives the system to produce
  yd(t)

12
(No Transcript)
13
(No Transcript)
14
(No Transcript)
15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
Outline

• What is Iterative Learning Control (ILC)
• Historical Comments
• From Analysis to Design
• Industrial Application I (ABB robots)
• Industrial Application II (Seagate HDD)
• To Probe Further and My Recent Results
• Concluding Remarks

22
ILC historical review (1)

• Historical Roots of ILC go back about 25 years.
• Idea of a multipass system studied by Owens and
  Rogers in mid- to late-1970's, with several
  resulting monographs.
• Learning control concept introduced (in Japanese)
  by Uchiyama in 1978.
• Pioneering work of Arimoto, et al. 1984-present.
• Related research in repetitive and periodic
  control.
• 1993 Springer monograph had about 90 ILC
  references. (Kevin L. Moore)

23
ILC historical review (2)

• 1997 Asian Control Conference had 30 papers on
  ILC (out of 600 papers presented at the meeting)
  and the first panel session on this topic.
• 1998 survey paper has about 250 ILC references.
• Web-based online, searchable bibliographic
  database maintained by Yangquan Chen has about
  500 references (see http//cicserver.ee.nus.edu.sg
  /ilc).
• ILC Workshop and Roundtable and three devoted
  sessions at 1998 CDC.
• Edited book by Bien and Xu resulting from 1997
  ASCC
• Springer-Verlag monograph by Chen and Wen, 1999.

24
ILC historical review (3)

• 4 invited sessions at 2000 ASCC (Shanghai) with
  an Invited Panel Discussion on ILC.
• 3 invited sessions at ICARCV 2000 (Singapore),
• The 2nd Int. Conference on nD Systems. (Poland)
• Tutorial at ICARCV 2000 and first IEEE CDC
  Tutorial Workshop 2000, Sydney.
• Special Issues in Int. J. of Control (2000),
  Asian J. of Control (2001) and J. of Intelligent
  Automation and Soft Computing (2001).
• Industrial use, e.g., Seagate and ABB (Sweden)

25
ILC historical review (4)

• Murray Garden (1967). Learning control of
  actuators in control systems. United States
  Patent 3,555,252.
• Chen, YangQuan and Kevin L. Moore. Comments on
  US Patent 3555252 LEARNING CONTROL OF ACTUATORS
  IN CONTROL SYSTEMS. ILC Invited Sessions. In
  Proc. of the ICARCV'2000 (The Sixth
  International Conference on Control, Automation,
  Robotics and Vision). (archeological
  contribution!)

26
Past efforts

• Past work in the field demonstrated the
  usefulness and applicability of the concept of
  ILC
• Linear systems.
• Classes of nonlinear systems.
• Applications to robotic systems.

27
Current efforts

• Present status of the field reflects the
  continuing efforts of researchers to
• Develop design tools.
• Extend earlier results to broader classes of
  systems.
• Realize a wider range of applications.
• Understand and interpret ILC in terms of other
  control paradigms and in the larger context of
  learning in general.

28
A Partial Classification of ILC Research

• Systems
• Open-loop vs. closed-loop.
• Discrete-time vs. continuous-time.
• Linear vs. nonlinear.
• Time-invariant or time-varying.
• Relative degree 1 vs. higher relative degree.
• Same initial state vs. variable initial state.
• Presence of disturbances.
• Update algorithm
• Linear ILC vs. nonlinear ILC.

29
A Partial Classification of ILC Research

• First-order ILC vs. higher-order.
• Current cycle vs. past cycle.
• Fixed ILC or adaptive ILC.
• Time-domain vs. frequency analysis.
• Analysis vs. design.
• Assumptions on plant knowledge.
• Applications
• Robotics.
• Chemical processing.
• Mechatronic systems (HDD, CD/DVD).

30
Trial (k-1)
Trial k
Trial (k1)
Error
Input
(a) ILC with Current Cycle Feedback
Error
Input
(b) Higher-Order ILC
31
Outline

• What is Iterative Learning Control (ILC)
• Historical Comments
• From Analysis to Design
• Industrial Application I (ABB robots)
• Industrial Application II (Seagate HDD)
• To Probe Further and My Recent Results
• Concluding Remarks

32
ILC Panel Discussion at ASCC2000

• General Trend from Analysis to Design
• Analysis
• Attack the Arimotos classical 6 Postulates for
  ILC.
• Structurally known uncertain nonlinear systems.
  System class Combined Feedforward-Feedback
  analysis!
• Add practical constraints in analysis changing
  delay, anti-windup
• Spatial ILC (state-dependent repetitiveness),
  distributed parameter system, redundancy in
  control authorities...
• Design
• How to explicitly use the available (assumed)
  prior knowledge?
• Systematic design method - e.g. via noncausal
  filtering, Local Symmetrical Integration (LSI)
  etc.
• Supervisory Iterative Learning Control (e.g.
  planning while tracking via ILC)

33
ILC Design as easy as PID?

• Yamamoto, S. and Hashimoto, I. (1991). Recent
  status and future needs The view from Japanese
  industry. In Arkun and Ray, editors, Proceedings
  of the fourth International Conference on
  Chemical Process Control, Texas. Chemical Process
  Control -CPCIV.
• Survey by Japan Electric Measuring Instrument
  Manufacturer's Association, more than 90 of the
  control loops were of the PID type.
• Bialkowski, W. L. (1993). Dreams versus reality
  A view from both sides of the gap. Pulp and Paper
  Canada, 94(11).
• A typical paper mill in Canada has more than
  2000 control loops and that 97 use PI control.

34
Tuning knobs of ILC

• Only two tuning knobs
• learning gain
• bandwidth of the learning filter
• an example my ASCC2000 paper
• Chen, YangQuan and Kevin L. Moore, Improved
  Path Following for an Omni-Directional Vehicle
  Via Practical Iterative Learning Control Using
  Local Symmetrical Double-Integration,'' Asian
  Control Conference 2000, July 5-7, 2000,
  Shanghai, China. pp. 1878-1883.
• Note Full version of this paper will appear in
  the Special Issue of ILC in Asian Journal of
  Control, 2001

35
LSI2-ILC Scheme
LSI2 -ILC Block Diagram
Overall control signal
LSI2 -ILC Speical Case TL2 0
LSI2
LSI2 -ILC Speical Case TL 0
ILC feedforward updating law
In the sequel, TL1TL2 TL
36
LSI2-ILC Analysis Design

• Discrete-time form

• Frequency domain

a
37
LSI2-ILC Design Procedures

• Convergence Condition

• Design of TL

• Design of

For given TL , the optimal choice of
38
Performance Limit Rule Based Learning

• Performance limit and heuristics

• Best achievable convergence rate

• Heuristics for better ILC performance (Rule
  Based Learning)

1. re-evaluate TL at the end of every iteration.
2. start ILC with a smaller and increase
when the tracking error keeps decreasing and
decrease while the tracking error keeps
increasing. 3. use a cautious (larger) TL at the
beginning of ILC iteration and then decrease TL
when the ILC scheme converges to a stage with
little improvement. ...
39
USU-ODV Simulation for LSI2-ILC

• USU ODV

Three parts to the control problem Outer-Loop
Control Compute the center-of-gravity motion
required to follow the desired path. Wheel
Coordination Determine appropriate commands for
each individual wheel to produce the desired
overall vehicle motion. Smart Drive Control
Generate input signals for the actuators in each
wheel (steering motor, speed motor).
6 smart wheels
40
USU-ODV Simulation for LSI2-ILC

• PI-control and LSI2-ILC

41
Standard deviation of tracking errors
observations
It. 0 0.1852 0.1436 0.1057 1 0.1086 0.0573
0.0810 2 0.0878 0.0508 0.0629 3 0.0768 0.0378
0.0535 4 0.0679 0.0323 0.0471 5 0.0596 0.0314
0.0438
1. Simple ILC scheme 2. Simple design steps 3.
Stable monotone convergence 4. Less modeling
efforts 5. Add-on to existing controller 6.
Effective in ODV path-following 7. Rule-based ILC
possible 8. Practically applicable.
42
Outline

• What is Iterative Learning Control (ILC)
• Historical Comments
• From Analysis to Design
• Industrial Application I (ABB robots)
• Industrial Application II (Seagate HDD)
• To Probe Further and My Recent Results
• Concluding Remarks

43
ABB Robotics

• Swiss - Swedish company (part of the ABB Group)
• Production and most of the RD in Västerås,
  Sweden
• 600 employees (at ABB Robotics)
• Produces 10,000 robots/year
• Installed a total of 90,000 robots in the world
• Leading producer of industrial robots

44
Motivation for ILC in ABB robots

• Highly repetitive dynamics
• In production (laser cutting) the same procedure
  is repeated by the robot many times
• Easy to implement in an already existing control
  structure
• Can easily co-exist with other improvements of
  the control system

45
Previous solution in ABB robots

• Traditional feedback and feedforward control
• Model based feedforward control (non adaptive but
  user configurable)
• Resulting absolute accuracy (approx) 0.5 - 5 mm

46
ILC implementation for ABB Laser Cutting Robots
1. Measure the position
2. Compensate in cartesian coordinates
3. Run the program again
47
After using ILC (in laser cutting)

• After the second ILC Path Errors 0.10mm
• NOTE previous error range 0.5 to 5 mm
• Tuned in approximately one minute
• Minor improvements after 2 iterations
• Improvements in the example
• No ILC
• ILC 1st Iteration 50
• ILC 2nd Iteration 61

48
Outline

• What is Iterative Learning Control (ILC)
• Historical Comments
• From Analysis to Design
• Industrial Application I (ABB robots)
• Industrial Application II (Seagate HDD)
• To Probe Further and My Recent Results
• Concluding Remarks

49
Typical Hard Disk Drive
50
Typical Hard Disk Drive
51
TPI how high now?

• TPI track density (tracks per inch) in radial
  direction.
• High capacity high TPI.
• 80Gb HDD at 60,000 TPI
• track pitch 25.4 mm/60,000 423 nano
• tracking accuracy /- 10 425 lt 50 nano.
• Note
• no position sensor
• no velocity sensor
• no acceleration sensor

HDD servo control is a magic.
52
Embedded Servo
53
Why ZAP (Zero-Acceleration-Path)?
54
ILC Industrial Application II My patent on ZAP

• Seagates solutions to written-in repeatable
• runout due to STW (Servo track-writer)
• J. Mooris et al. Compensations of written-in
  errors in servo. US Patent 6,069,764 (2000)
• B. Qiang, K. Gomez, Y. Chen, K. Ooi, Repeatable
  runout compensation using iterative learning
  control in a disc storage system. US Patent
  Pending Serial No. 60/132,992. (1999)
• Y. Chen et al. Repeatable runout compensation
  using a learning algorithm with scheduled
  parameters. US Patent Pending Serial No.
  60/145,499. (1999)

55
My Pending Patents

• 16 patent disclosures. All evaluated as pursue.
• Under processing of patent lawyers in USA
• US patent takes 3 years, first one in 2002?
• All implemented on actual hard disk drives in
  assembly language (Siemens C166, 16
  bits/fixed-point) in Seagate Singapore Design
  Centre.
• Some used in Seagate products like U8/U10
  (15/30Gb) and U6 (40/80G).
• Some taken as trade secret or technological
  inventory.
• Received 10,000 patent awards in 2000.

56
Before and After ZAP
57
Before and After ZAPSpectrum
58
Summary of My Patent on ZAP

• Benefits
• Increase TPI and double the HDD capacity. Or, for
  the same TPI, increase the reliability
• Purely algorithm/code change
• Reduce STW cost
• Show the power of advanced control ideas

• Price to pay
• Extra time to learn the compensation table during
  factory process
• Better servo demodulator chip to embed the
  learned compensation table

Used now in U6 (40/80Gb) 60KTPI product line
59
Outline

• What is Iterative Learning Control (ILC)
• Historical Comments
• From Analysis to Design
• Industrial Application I (ABB robots)
• Industrial Application II (Seagate HDD)
• To Probe Further and My Recent Results
• Concluding Remarks

60
To probe further

• ILC web server http//cicserver.ee.nus.edu.sg/ilc
  http//www.crosswinds.net/yqchen (updates,
  reference library, links to other researchers
  etc)
• Another site http//www.ilcworld.net

61
My ILC results01 (1)

• YangQuan Chen and Kevin L. Moore. On
  -type Iterative Learning Control''. Submitted to
  IEEE CDC'2001.
• In between P-type
• and D-type.

62
My ILC results01 (2)

• YangQuan Chen and Kevin L. Moore. Frequency
  Domain Adaptive Learning Feedforward Control''.
  The 2001 IEEE International Symposium on
  Computational Intelligence in Robotics and
  Automation (IEEE CIRA 2001), July 29 - August 1,
  2001, Banff, Alberta, Canada. (contributed)

63
My ILC results01 (3)

• YangQuan Chen, Kevin L. Moore and Vikas Bahl.
  Improved Path Following of USU ODIS By
  Learning Feedforward Controller Using Dilated
  B-Spline Network". The 2001 IEEE International
  Symposium on Computational Intelligence in
  Robotics and Automation (IEEE CIRA 2001), July
  29 - August 1, 2001, Banff, Alberta, Canada
  (invited)
• Ping Jiang and YangQuan Chen. Repetitive Robot
  Visual Servoing Via Segmented Trained Neural
  Network Controller''. The 2001 IEEE International
  Symposium on Computational Intelligence in
  Robotics and Automation (IEEE CIRA 2001), July
  29 - August 1, 2001, Banff, Alberta, Canada
  (invited)

64
My ILC results01 (4)

• YangQuan Chen and Kevin L. Moore. Frequency
  Domain Analysis and Design of Learning
  Feedforward Controller Using The Second Order
  B-Spline Network". Automatica, 9 Feb 2001.
  (Brief paper under review)
• YangQuan Chen, Kevin L. Moore and Vikas Bahl,
  "Learning Feedforward Controller Using Dilated
  B-Spline Network Analysis and Design in
  Frequency Domain". IEEE Trans. on Neural
  Networks. (full paper under review May 4, 2001)

65
My ILC results01 (5)

• Y. Chen and K. L. Moore, A Practical Iterative
  Learning Path-Following Control of an
  Omni-Directional Vehicle''. Special Issue on
  Iterative Learning Control, Asian Journal of
  Control. (accepted, to appear) 2001. (full paper)
• Y. Q. Chen, H. F. Dou and K. K. Tan, Iterative
  Learning Control Via Weighted Local-Symmetrical-I
  ntegration'', Asian Journal of Control, Accepted
  and scheduled in vol. 3, no. 4, 2001. (short
  paper)
• K. K. Tan, H. F. Dou, Y. Q. Chen and T. H. Lee,
  High Precision Linear Motor Control Via
  Relay-Tuned Iterative Learning Based On
  Zero-Phase Filtering'', IEEE Transactions of
  Control Systems Technology, vol. 9, no. 2 pp.
  244-253, 2001. (full paper)

66
My ILC results02

• IFAC02 (Spain), ILC Invited Session.
• High-order in time ILC design.
• ASCC02 (Singapore), ILC Invited Session.
• Monotonic (H_2 ) ILC design via feedback.
• ACC02 (Alaska), Contributed.
• Spatial ILC for autonomous ground vehicle
• an Automatica paper under preparation ...

67
Outline

• What is Iterative Learning Control (ILC)
• Historical Comments
• From Analysis to Design
• Industrial Application I (ABB robots)
• Industrial Application II (Seagate HDD)
• To Probe Further and My Recent Results
• Concluding Remarks

68
Concluding Remarks

• Repetition improves skill, for both man and
  machine.

--- Chen,Yangquan and Wen,Changyun.
Iterative Learning Control Convergence,
Robustness and Applications', Springer-
Verlag, 1999. Lecture Notes series on Control and
Information Science. Vol. LNCIS-248. (199
pages. ISBN1-85233-190-9)

• Will be as popular and effective as PID.

69
Concluding Remarks

• Academia
• fusion with other existing feedback controls
• time-frequency domain
• 2-D systems etc.
• real time SPC?

• Industry
• more applications
• auto industry, mechatronic systems etc.
• chemical reactors, semiconductor processes etc.
• more design schemes

70
Second-order crime
http//cicserver.ee.nus.edu.sg/ilc/control/humor/
71
Arts a professor wants

• Balancing between science and engineering
• Balancing between research and development
• Balancing between academia and industry
• Balancing between teaching and research and
  service
• etc. etc.

72
Thank you! Q/A Session Please visit ILC website
http//cicserver.ee.nus.edu.sg/ilc or, http//ww
w.crosswinds.net/learningcontrol
73
Acknowledgments

• Dr Kevin L. Moore for his leadership in ILC
  research, especially, for his organization of the
  First ILC Roundtable Discussion in IEEE CDC98
  and IEEE CDC00 Tutorial Workshop. Some slides
  are from his CDC00 presentation.
• Dr Mikael Norrlöf of Linköping University for
  providing some slides on the successful story on
  ILC applications in ABB Robots.
• My ex-colleagues and co-inventors in Seagate
  Singapore Science Park, especially, Mr. K K Ooi
  and Mr. M Z Ding for unforgettable creative OTs.
• The University of Windsor for inviting me to
  deliver this talk.