Designing neural-network-based fault detection systems with D-optimum experimental conditions

Abstract

The paper deals with an application of the theory of optimum experimental design to the problem of selecting the data set for developing neural models. Another objective is to show how to design a robust fault detection scheme with neural networks and how to increase its fault sensitivity by decreasing model uncertainty. It is also shown that the optimum design is independent of the parameters that enter linearly into the neural network. The final part of this paper shows a comprehensive simulation study regarding modelling and fault detection with the proposed approach. In particular, the DAMADICS benchmark problem is utilized to verify the performance and reliability of the proposed technique.