Artificial Intelligences & Metamodeling

1. Optimal Tuning of Controller Parameters for a Magnetic Levitation System Using Metamodeling Approach

Dynamic systems are quite often non-linear and require a complex mathematical model. For their optimal control, it has been always a requirement to tune the controller parameters to achieve the best performance. Parameter tuning in complex systems is predominantly a time-consuming task, even with high performance computers. This project will demonstrate how metamodeling technique can be applied to efficiently tune the control parameters of a nonlinear and unstable system with fast dynamics, namely the magnetic levitation (Maglev) system. A neural network-based metamodel is applied to optimise the parameters of PID controllers. Maglev technology has a wide range of applications such as high-speed trains, seismic attenuators for gravitational wave antennas, self-bearing blood pumps for use in artificial hearts interface, and photolithography devices for semiconductor manufacturing and microrobots. In this project, work will begin with a derivation of the mathematical model of a lab scale magnetic levitation system (e.g. CE152 kit), in the transfer function and/or state space form. It follows with the construction of metamodel using experiment design to capture unknown system behavior in the large design space. The final component is to optimize the controller parameter for the PID using the constructed metamodel. A performance comparison with other conventional tuning techniques will then be conducted. Implementation will be carried out using MATLAB SIMULINK, MATLAB M-FILE or using other compatible software (such as Scilab).

2. Data Sampling Technique of Neural-network Based Metamodeling with Application to Air Pollutant Estimation

Nowadays, simulation modelling becomes a popular tool for the analysis of complex systems’ behaviour. Metamodels (or surrogate models) can be an approximate model that able to adequately represent the intrinsically non-linear and complex relationship between the system’s input and output. Before executing the function approximation in metamodelling, it is important to select the design points in the domain which is generally termed as sampling, experimental design, or design of experiment (DOE). The aim of any sampling method is to effectively cover the design space and to gather the essential information of the design space characteristics. This project will develop a new strategy for a metamodel DOE, based on the distance measure and clustering process, referred to as the weighted clustering design (WCD) throughout this paper. Here the proposed sampling method is then used to develop a radial basis function neural network (RBFNN) metamodel as a function approximator. We first test the scheme validity with a known nonlinear function. The air pollutant estimation problem is then tackled by using the improved metamodel. Implementation will be carried out using MATLAB or using other compatible software (such as Scilab).