Plenary Lecture 2	From Data to Model - A Learning Approach for Motion Control Systems
Date/Time	Tuesday, May 31, 2016 10:00-11:00
Venue	International Leture Hall of the 2nd floor
Presenter	Prof. Jian-Xin Xu , National University of Singapore

Abstract

Motion systems range from precision servo at nanometer scale to space exploration at mega-kilometer scale. Traditionally motion control systems benefit significantly from model-based controller design methodology in a systematic manner. When dealing with extreme precision or complex systems, however, control-oriented models no longer tally with physical models that are derived from the first principles.In such circumstances, a synergy of data-driven modeling and model-based design offers new solutions for controller design. We first learn from industries how they make use of data to facilitate controller designs. From data analysis,a data-driven model can be established, in which the main objective is not on approximating the physical process accurately, but on capturing time/frequency domain features that are useful in motion control design. Different from industrialists who prefer data-driven models that are application-specific and established off-line, control theorists also undergo exploration from principle-driven to semi-data-driven models that could be more generically used for a variety of applications and updated on-line. Thus, we next review several adaptive and learning algorithms that do not require the complete model knowledge, discuss their advantages and limitations in data usage, model dependency, as well as the suitability in motion control. In order to reduce the gap between industrial practice and academic research, we focus on two major modeling issues - system uncertainty and non-affine mapping. Aiming at approximating inverse mapping for a motion system with uncertainty and non-affine structure, basis functions in an appropriate space will be constructed first. Accordingly, learning or adaptation will be carried out in another appropriate parametric space with comprehensive data. Learning or adaptation can further be made optimal by minimizing an objective function defined in a feature space that may constitute temporal, spatial, frequency, ensemble or time-frequency (e.g. wavelet) features obtained through data analysis. Data-driven modeling with ergodicity and automatic construction will also be briefly explored. A number of practical examples will be presented to illustrate the concepts and methods introduced in this talk.

This talk presents syntheses and implementations of a smart optimal control system for the energy intensive processing equipment. The talk will focus on three main functions of the proposed smart optimal control system: (i) process control; (ii) operational optimization control; and (ii) operational conditions diagnostics and self-healing control. The design of a novel data-driven dual closed-loop intelligent optimal operational control will be described for realizing these primary functions.

The data-driven dual closed-loop control employs a two-layered structure: (i) an intelligent optimal control layer for identification of optimal set points of control loops which takes functions of target indices associated with energy saving, product yield, product quality and efficiency as optimization index, and the set points as the decision variables; and (ii) a set points tracking intelligent control layer focusing on virtual unmodeled dynamics compensation based controller.

This talk introduces a hybrid simulation system for operational optimization and control of complex industrial processes developed by our team. Simulations to electric magnesium melting furnace for magnesia production industry are used to demonstrate the effectiveness of the proposed method.

This talk also introduces the smart embedding control system of electric magnesium melting furnace developed by our team adopting the novel data-driven dual closed-loop intelligent optimal operational control algorithm proposed. It has been successfully applied to the largest magnesia production enterprise in China, resulting in great returns. `Issues for future research on the smart optimization control system are outlined in the final section.

Biography

Jian-Xin Xu Dr Xu Jian-Xin received PhD degree in 1989 from University of Tokyo, currently is professor at the Department of Electrical Engineering, National University of Singapore. His research interests lies in the fields of learning theory, intelligent control, nonlinear and robust control, robotics, precision motion control.Up to now he has published more that 200 journal papers and 7 books in system control theory and applications, supervised 30 PhD students, completed more than 20 research projects with grants and currently works on? biomimetic underwater robots. He is a fellow of IEEE.