Design of an auto-associative neural network by using design of experiments approach

Data driven Computational intelligence methods have become popular in Fault detection and isolation (FDI) due to relatively quick design and not so difficult implementation on real systems. In this paper a research work on a Taguchi DoE approach for training the auto-associative neural network to extract non-linear principal components of a system, is presented. Design of such network was first proposed by Kramer however for achieving robustness to unspecified parameters such as noise level and disturbances, a design of experiments methodology can be used to optimally define network structure and parameters.