Nonlinear control based on structurally optimized neural fuzzy networks

Nonlinear control has been crucial in many fields, such as robotics and vehicle. Most of these systems involve nonlinear modelling. This PhD research aims to obtain compact yet efficient nonlinear modelling with optimized structure and apply it in control systems.

This thesis focuses on the local linear model tree (LOLIMOT), a typical neural fuzzy model. In the conventional LOLIMOT, structural parameters such as the centres and variances of the Gaussian kernels are usually set in a fixed manner. While this provides a simple way to set the structure, the resulted network is often not accurate enough to fit in the underlying data, leading to inefficient control. In this thesis, the structure of the LOLIMOT network is carefully optimized, leading to more efficient nonlinear modelling. The optimized LOLIMOT is then applied in several control systems, including the inverse model control and model predictive control (MPC).

The proposed new control systems have been extensively verified by numerical data, Simulink platform data and engine data. Validation results show that the proposed control systems have significantly better-controlling performance than those based on the conventional LOLIMOT, making it an attractive solution in practice.