Incremental Cluster Learning Neural Network Application to GIS Internal Diagnostics

Abstract

The power equipment such as GIS (Gas Insulated Switchgear) plays an important role for stableness and reliance of electric power supply. Although the advantages of GIS are maintenance free and long-term stability, there exists the necessity to develop predictive maintenance system to avoid fatal mal-functions in case. The objective of predictive maintenance is detecting small symptoms of partial discharge within GIS in service. The NN (Neural Network) characteristics of learning and self-organization are expected to show better performance for realization of predictive maintenance system. This paper presents the NN architecture of ICLNN (Incremental Cluster Learning Neural Network), the architecture of diagnostic system, and experimental results using simulated abnormalities. The purpose of development of ICLNN is real time learning of normal status at the actual site. This is required because NN should be trained for the background noise to define normal status. The normal status is easy to define and lots of data for this status are available, though complete sets of abnormal status are difficult to define and need many experiments to obtain training data. The experiments are conducted using simulated abnormalities of various lengths of metal particles. The results exhibit the advantages of ICLNN diagnostic system over conventional amplitude threshold system for some cases. They also show the limitation of the performance of both systems.