Neural networks for the identification of gas cylinder faults

Abstract

The safety of gas cylinders in domestic applications is of utmost importance. A crucial stage in the appraisal of the risk for failures and possible faults is occurring during the filling process. In Cyprus, this task is currently done by specialized workers who monitor the cylinders during filling. In order to explore the possibility for an automated risk appraisal and consequent screening, a system of fault identification using vibrational time series and neural network classification has been used. Two systems have been attempted. One using a multi-slab feedforward neural structure employing backpropagation-type learning, and a Kohonen self-organizing map. The results were also compared with different simple statistical methods. The feedforward net, proved to be slightly better responding than the Kohonen map for this particular problem.