Autoassociative neural networks with an application to fault diagnosis of a gas turbine engine

With the growth in accessibility of computer technology, the acquisition of data has become tremendously simplified. While this has been a boon to scientists and engineers studying systems, the challenge has become to try to extract knowledge from increasingly larger and higher dimensional databases. This thesis approaches the problem by establishing the following hypotheses: (a) Prior knowledge of the data e.g. a first principles model is not required to make a valid assessment of the data; (b) High dimensional data can be reduced to a lower dimension without a significant loss of information; and (c) Autoassociative neural networks can provide an equivalent or more flexible means of achieving dimension reduction when compared to standard statistical methods.; From these three hypotheses, this work demonstrates the use of autoassociative neural networks in reducing the dimensionality of high dimensional data to two or three. Such a reduction then allows the direct visualisation of the data by a human observer for the extraction of relevant features.; The use of autoassociative neural networks is explored in detail from a theoretical and practical point of view. In order to provide a benchmark, the results of the methodology using an autoassociative neural network are compared to classical statistical techniques. The comparisons are made for a series of toy problems and in a case study, namely the fault diagnosis of a gas turbine engine. (Abstract shortened by UMI.)...