Advances in data mining for dissolved gas analysis

This paper reports NGC's continued application and refinement of a data mining technique based on the Kohonen neural network. The technique has been applied to NGC's database of transformer dissolved gas-in-oil analysis (DGA) measurements for high voltage transformers. The technique has proven able to highlight bad data and 'blind test' data, and has been optimized to reveal the early stages of potential plant problems. A number of key types of transformer condition have been distinguished by it, including for example three kinds of partial discharge. The Kohonen technique has been successfully applied to transmission, distribution and generator transformers. In addition a practical tool for DGA interpretation is being developed. We are now looking to expand the use of the technique to other monitored parameters.