NEURAL NETWORK INVERSE CONTROL

Abstract

This thesis investigates the application of artifical neural networks, in short, neural networks, for model-based control, in particular in a direct Neural Network Inverse Control (NNIC) scheme.

Neural networks have been known to be able to be trained to approximate any complex function to any desired degree of accuracy. This characteristic is made use of in model-based control in which a neural network trained to map the dynamic characteristics of the plant-under-control is used to achieve good control. One of the pre-conditions of using direct NNIC is the existence of a desired actuating (control) signal which can be determined from measured input­ output data of the plant. In this thesis, such a signal is shown to exist for the class of non-linear systems which can be written in the NARMA (Nonlinear Auto­ Regressive Moving Average) form, or simpler models of this. The control of such systems using the direct NNIC scheme is thus achievable.

This thesis then shows that neural networks can very quickly, with a few training steps, achieve good approximating characteristics around the local region in which it was last trained. This ability allows it to be able to achieve good trajectory-following control, in the direct NNIC scheme, with on-line re-training. This ability will also allow it to adapt to changes in the plant and to maintain good control with on-line re-training.

When a neural network is used in control with on-line re-training, there is the possibility of system instability if the learning parameters chosen is not suitable or under certain circumstances. This thesis investigates further the pre-training of the neural network controller for direct NNIC without-on-line re-training. Under this control scheme, if the neural network controller can be accurately and adequately trained to map the inverse of the plant's dynamic characteristics within the desired operating region, good control could be achieved. Of course, without on-line re-training, the control system will not be able to adapt to changes in the plant's dynamic characteristics during operation.

An approach for experimentally obtaining suitable training data sets from measured input-output plant data using a swept sine with noise forcing function was developed and demonstrated in this work. Very good trajectory-following control was achieved in the simulation example on a second-order linear plant without any prior knowledge of plant characteristics. The importance of adding noise to the swept sine forcing function to better cover the required training input space for the neural network controller was also demonstrated by comparing the performance with that when using a swept sine without noise forcing function.