Saravanan, Natarajan

An identification scheme which can be used for discrete time multi-input multi-output time invariant systems is presented. The identification scheme involves two steps; (1) The identification of a set of invariant indices (Structure identification) and (2) The estimation of the parameters of the system (Parametric identification). The technique utilizes a canonical representation of a system which is based on the notion of output injection. This canonical form is dependent on a chosen real number alpha and is therefore called the alpha-canonical form. Least square estimation technique is used for parameter estimation. The off-line version of this identification scheme is presented here. This scheme is then used to generate a linear model of the Space Shuttle Main Engine at the operating point corresponding to the 100% power level from the nonlinear dynamic engine simulation.

System identification methods are frequently used to obtain appropriate models for the purpose of control, fault detection, pattern recognition, prediction, adaptive filtering and other purposes. A number of techniques exist for the identification of linear systems. However, real-world and complex systems are often nonlinear and there exists no generic methodology for the identification of nonlinear systems with unknown structure. A recent approach makes use of highly interconnected networks of simple processing elements, which can be programmed to approximate nonlinear functions to identify nonlinear dynamic systems. This thesis takes a detailed look at identification of nonlinear systems with neural networks. Important questions in the application of neural networks for nonlinear systems are identified; concerning the excitation properties of input signals, selection of an appropriate neural network structure, estimation of the neural network weights, and the validation of the identified model. These questions are subsequently answered. This investigation leads to a systematic procedure for identification using neural networks and this procedure is clearly illustrated by modeling a complex nonlinear system; the components of the space shuttle main engine. Additionally, the neural network weights are determined by using a general purpose optimization technique known as evolutionary programming which is based on the concept of simulated evolution. The evolutionary programming algorithm is modified to include self-adapting step sizes. The effectiveness of the evolutionary programming algorithm as a general purpose optimization algorithm is illustrated on a test suite of problems including function optimization, neural network weight optimization, optimal control system synthesis and reinforcement learning control.