Automatic Design of a Fuzzy Controller from a Neural Process Model

A new model-based neuro-fuzzy controller for non-linear systems is proposed. The neural network provides a non-linear model of the process being controlled. The nonlinear space of the network is divided into small linear regions. For each of these regions linear controllers can be designed automatically. To ensure a smooth transition between different regions fuzzy membership functions are used. A comparison of a neuro-fuzzy controller like this and a classical PID-controller shows the capability of the proposed approach. Applications for this kind of controllers are systems in changing environments requiring adaptive control without human help. An example is an experiment carried out at the Nuclear Research Center Karlsruhe, where a balloon transports measuring instruments to examine trace gases in the atmosphere. The orientation of the instruments (mainly a spectrometer) has to be stabilized down to 0.03° when the balloon is pitching. Other parameters vary with time and temperature (like the friction of the motor) and cannot be set properly on the ground. Results with the controller described in this paper are shown.