Multivariable System Controller Tuning Techniques Based on Sensitivity Measures

A controller tuning technique using sensitivity functions of the controller parameters is proposed which does not require a detailed model of the plant either based on physical principles or through system identification. Only simple signal processing is required in the tuning process. The sensitivity information is used by an adjustment algorithm involving a least squares type of criterion function. The generation of the sensitivity functions which are of central importance in this approach is described in this thesis, involving use of a signal convolution approach and a two-stage method both in the time-domain and frequency-domain. Three different forms of input signal, which involve the step input, the impulse input and the extended pseudorandom binary sequence (PRBS) input signal are selected in the calculation of the sensitivity functions. The two-stage approach to generating the sensitivity functions of the controller parameters in systems with unknown plants has been investigated for the first time in this research work. The advantage of this novel approach is that there is no limitation on the form of the test input signal. The sensitivity functions can be obtained from measurements directly without any calculations. No problems of implementation arise with the sensitivity filters required. In the controller tuning process, the least squared approach is used to provide the figure of merit for each projected system response. The changes of the controller parameters are altered to minimise the difference between the response of the actual system and the desired response. The details of an application of the tuning procedure using the signal convolution approach for generating the sensitivity functions for a two-tank system with two inputs two outputs both in the time-domain and the frequency-domain are given. Special consideration is given to the accurate modelling of the two-tank system upon which this work is based. Questions of plant nonlinearity and measurement noise and their effects on the tuning process are given careful consideration but no significant problems were encountered. In order to prove that the technique is suitable for more complex problems, the technique has also been applied successfully to helicopter flight control system design optimisation. This is of potential interest as a means of reducing the period of time for test flying and design modification for practical helicopter flight control systems. The tuning process is a very "visible" one and likely to be attractive for applications of this kind. From the results of these two applications of the tuning technique it can be seen that the tuning process is very effective although the initial responses of the system may be far from the desired responses. In fact, the adjustment procedure provides fast is convergence in the cases considered. Significant progress is made at each adjustment without any oscillations in parameter values. The number of experiments needed to generate the sensitivity information needed for controller tuning is, in general, significantly smaller than that required for a traditional parameter perturbation method for sensitivity function generation