2013 Theses Doctoral

Robustification in Repetitive and Iterative Learning Control

Shi, Yunde

Repetitive Control (RC) and Iterative Learning Control (ILC) are control methods that specifically deal with periodic signals or systems with repetitive operations. They have wide applications in diverse areas from high-precision manufacturing to high-speed assembly, and nowadays these algorithms have even been applied to biomimetic walking robots, where tracking a periodic reference signal or rejecting periodic disturbances is desired. Compared to conventional feedback control designs (including the inverse dynamics method), RC and ILC improve the control performance over repetitions -- by learning from the previous input-output data, RC and ILC adaptively update the control input for the next run, aiming for zero tracking error in the hardware instead of in a model, as time goes to infinity. The stability robustness to model uncertainty however remains a fundamental topic as it determines the successful implementation of RC and ILC on any real-world system whose model dynamics cannot normally be determined precisely over all frequencies up to Nyquist. In the control field, there are various existing methods of robustification, such as Linear Matrix Inequality (LMI), mu-synthesis and H-infinity, but few of these methods offer intuitive information about how the stability robustness is achieved. In addition, many of these existing algorithms produce conservative stability boundaries, leaving room for further optimization and enhancement. In this study, several robustification approaches are developed, where better insight into the robustification design process and a tighter stability boundary are established. The first method presents an algorithm for RC compensator design that not only uses phase adjustments, but also adjusts the learning rate as a function of frequency to obtain improved robustification to model parameter uncertainty. The basic objective of this algorithm is to make the system learn at each frequency at the maximum rate consistent with the need for robustness at that frequency. The second method, on the other hand, explores the benefits of compromising on the zero tracking error requirement for frequencies that require extra robustness, making RC tolerate larger model errors. The third topic focuses on the development of robustification algorithms for Iterative Learning Control that is analogous to the above two RC robustification designs, extending frequency response concepts to finite time problems. The final approach to robustification treated in this dissertation is based on Matched Basic Function Repetitive Control (MBFRC), which individually addresses each frequency, eliminating the need for a robustifying zero phase low pass filter and the need for interpolation in data as required in conventional RC design. Furthermore, this algorithm only uses the frequency response knowledge at the frequencies addressed, and as long as the phase uncertainties at those frequencies are within +/- 90 deg the system is guaranteed stable for all sufficiently small projection gains.