Identification of Nonlinear Systems

点击翻译按钮获取中文摘要

Automatic control systems are ubiquitous in the modern world. They can range in scale from the position controller that regulates the position of the read/write head in a hard disk, to the autopilots that control the movement of aircraft, to the advanced traffic control systems that regulate the flow of traffic in major cities. Since the controller is designed based on a mathematical model of the system that will be controlled, the system’s performance will be limited by the accuracy of the model used in its design. Thus, having an accurate mathematical model of the target system is a necessary pre-requisite to the design of a high-performance control system. While models of comparatively simple systems can be derived based on first principles, this approach quickly becomes intractable as the system’s scale and/or complexity increases. System identification, on the other hand, is a data-driven process that creates mathematical models of dynamic systems from measurements of their inputs and outputs. This research programme will develop system identification techniques that can create accurate mathematical models of nonlinear systems from data that have been gathered under normal operating conditions and when some form of control system is operating. The system identification process is greatly simplified if the input to the system is independent of any disturbances that affect the system, or of any noise that may be contained in the measurements. Unfortunately, many systems can only be operated in a “closed-loop” configuration, where the system output is measured, compared to a desired trajectory, and this error is then used to adjust the input signal. This type of feedback control may be necessary for safety, performance or economic reasons. The nonlinear system identification algorithms developed in this research programme will be designed to operate on data that have been gathered under closed-loop conditions, as this will allow them to be applied without the need for a separate, often impossible to perform, open-loop experiment. Instead, we will concentrate on constructing models from normal operating data, as this will minimize the economic impact of generating the model.