System identification and its applications, with emphasis on direction-dependent processes

In the first sub-section of the thesis, signal design for both linear and nonlinear system identification is considered. To identify a linear system using a perturbation test, a binary signal is sufficient and has the advantage of maximising the power available within a specified peak-to-peak amplitude. For this purpose, a program was written to generate five classes of binary and near-binary signal. However, to identify a nonlinear system with a Hammerstein structure, a multi-level signal is required, and methods to optimise such a signal are proposed. In the second sub-section, the detection of the departure from linearity for direction dependent processes is considered. It was found that only signals based on maximum length sequences allow the detection of such characteristics due to the coherent patterns formed in the cross correlation function. The 'combined' linear dynamics of the system are identified. The modelling of such processes using Wiener and neural network models is investigated. Practical results from an electronic nose are presented. The control of direction-dependent processes using the PID controller is then examined, with the design rules set according to the identified 'combined' dynamics. The thesis then moves on to the topic of autotuning. The autotuning of Smith predictors for processes with significant dead time is considered. The frequency response of the process is identified in closed-loop using a multi sine signal. Tuning rules for robust control are suggested which relate the controller parameters to the process parameters. A real application using a hot-air flow device is illustrated. The final sub-section of the thesis looks at the identification of Wiener-Hammerstein models. A new technique using linear interpolation is proposed which is based on the symmetry properties of the Volterra kernel. This method has the advantages that only a single experiment is needed, and it is simple to use since no optimisation or recursive computations are required. Simulation examples are provided to illustrate the effectiveness of the technique, and its robustness in the presence of noise and input signal distortion.