Academic Commons

Theses Doctoral

Eliminating the Internal Instability in Iterative Learning Control for Non-minimum Phase Systems

Li, Te

Iterative Learning Control (ILC) iterates with a real world control system repeatedly performing the same task. It adjusts the control action based on error history from the previous iteration, aiming to converge to zero tracking error. ILC has been widely used in various applications due to its high precision in trajectory tracking, e.g. semiconductor manufacturing sensors that repeatedly perform scanning maneuvers.
Designing effective feedback controllers for non-minimum phase (NMP) systems can be challenging. Applying Iterative Learning Control (ILC) to NMP systems is particularly problematic. Asking for zero error at sample times usually involves inverting the control system. However, the inverse process is unstable when the system has NMP zeros. The control action will grow exponentially every time step, and the error between time steps also grows exponentially. If there are NMP zeros on the negative real axis, the control action will alternate its sign every time step.
ILC must be digital to use previous run data to improve the tracking error in the current run. There are two kinds of NMP digital systems, ones having intrinsic NMP zeros as images of continuous time NMP zeros, and NMP sampling zeros introduced by discretization. Two ILC design methods have been investigated in this thesis to handle NMP sampling zeros, producing zero tracking error at addressed sample times: (1) One can simply start asking for zero error after a few initial time steps, like using multiple zero order holds for the first addressed time step only (2) Or increase the sample rate, ask for zero error at the original rate, making two or more zero order holds per addressed time step.
The internal instability can be manifested by the singular value decomposition of the input-output matrix. Non-minimum phase systems have particularly small singular values which are related to the NMP zeros. The aim is to eliminate these anomalous singular values. However, when applying the second approach, there are cases that the original anomalous singular values are gone, but some new anomalous singular values appear in the system matrix that cause difficulties to the inverse problem. Not asking for zero error for a small number of initial addressed time steps is shown to eliminate all anomalous singular values. This suggests that a more accurate statement of the second approach is: using multiple zero order holds per addressed time step, and eliminating a few initial addressed time steps if there are new anomalous singular values.
We also extend the use of these methods to systems having intrinsic NMP zeros. By modifying ILC laws to perform pole-zero cancellation inside the unit circle, we observe that all of the rules for sampling zeros are effective for intrinsic zeros. Hence, one can now achieve convergence to zero tracking error at addressed time steps in ILC of NMP systems with a well behaved control action.
In addition, this thesis studies the robustness of the two approaches along with several other candidate approaches with respect to model parameter uncertainty. Three classes of ILC laws are used. Both approaches show great robustness. Quadratic cost ILC is seen to have substantially better robustness to parameter uncertainty than the other laws.


  • thumnail for Li_columbia_0054D_13763.pdf Li_columbia_0054D_13763.pdf binary/octet-stream 2.38 MB Download File

More About This Work

Academic Units
Mechanical Engineering
Thesis Advisors
Longman, Richard W.
Ph.D., Columbia University
Published Here
February 1, 2017