AbstractThe research area of this Ph.D. project is the control design for active magnetic bearings (AMBs). The aim of this project is to develop optimal strategies for the control of several aspects of AMBs and to gain a deeper insight into their related algorithms. The AMB system has been applied in many areas. It is a mechatronic system related to various research areas. In this thesis, two system models will be built based on the electromagnetic analysis of the AMB system. Basic LQR control will then be introduced and applied to the proposed AMB model to guarantee closed-loop stability. To estimate and reject exogenous disturbances, disturbance observers are used in many real-life systems. It is known that there are two types of AMBs which can be defined according to their controlling mode: voltage type and current type, while voltage controlled AMBs are subject to inevitable mismatched disturbances. In this thesis, several kinds of disturbance observer, when applied under different circumstances, are discussed. These include constant disturbance observers, general exogenous disturbance observers, and their enhanced formulations. A universal extended state observer (ESO) has also been investigated. For systems which work in a repetitive mode, iterative learning control (ILC) is a method of tracking control which collects error information in every cycle. While standard ILC approaches can achieve perfect tracking for active magnetic bearing systems under external disturbances, the disturbances are required to be iteration-invariant. A disturbance observer-based ILC scheme is proposed which consists of a universal ESO and a classic ILC law. By using output feedback only, the proposed control estimates and attenuates the disturbances in every iteration. The convergence of the closed-loop system is guaranteed by analysing the contraction behaviour of the tracking error. Simulation and comparison studies demonstrate the superior tracking performances of the proposed controller based on the AMB model. The neural networks are complex computing algorithms inspired by the human brains. The connecting weights can be adjusted by a learning rule. The weighting coefficients can be updated by specific methods, such as backpropagation (BP) or the radial basis function method (RBF). Such characteristics mean that the neural networks have the ability to approximate arbitrary nonlinear functions. Thus, in this thesis, we combine a traditional ILC with two different neural networks in order to obtain a better control performance. The parameters of a composite ILC are tuned and updated during the processes, and optimised results can be obtained.
|Date of Award||1 Aug 2019|
|Supervisor||Haiyu Li (Supervisor) & Zhengtao Ding (Supervisor)|
- Neural network
- Active magnetic bearings
- Iterative learning control
- Disturbance observer based control