Gas Discrimination Method for Detecting Transformer Faults by Neural Network

Abstract

A method available for early detection of abnormality in an oil-filled transformer is discribed, in which four gas sensors having different characteristics and neural network are used to identify gas species (H₂, CH₄, C₂H₄, C₂H₂ and mixture of two species). In order to improve the selectivity of gas sensors, the time response patterns induced by changing sensor temperature, and the stationary sensor output is identified by neural network. Furthermore, the mixture ratio of gases is derived by using the stationary sensor output in response to the changing sensor temperature.
Gas species are well discriminated, and the mixture ratio derived from sensor output agrees well with the measurement by gas chromatography. Therefore, it is confirmed that our method is applicable to the transformer diagnostic technology.