Experimental design of a diagnostic system based on unsupervised and supervised learning methods

Abstract

This paper presents results of a research conducted in order to apply artificial neural networks for building a technical diagnostic system. In a developed diagnostic process, self-organized and feedforward neural networks are used. A self-organized neural network is introduced for reduction of extensive input data set and a new unsupervised learning algorithm is developed for that purpose. Next, the trained feedforward neural network is applied to map the current system state into corresponding diagnostic indicators. Several factors which have the greatest impact on the accuracy of the neural network based on the diagnostic system are considered. These factors are: the level of reduction of the input data set, selection of the most informative features in an input data samples, and the configuration variation of the feedforward neural network. The developed diagnostic system is tested on the problem of early detection and classification of gas leaks in a pipeline network for the natural gas transmission.