Application of adaptive learning networks for the detection of failing power system components

Abstract

A system capable of monitoring mine electrical power systems and detecting component failure in early stages could significantly improve power system safety and availability. Such monitoring would require a method of evaluating electrical features, calculated from terminal values, for indications of component deterioration. Research is being conducted by the US Bureau of Mines to examine the use of mathematical models to aid in this evaluation by creating polynomial networks called adaptive learning networks that can indicate deteriorated conditions in cable-connected motor systems. This process uses laboratory training data to select the electrical features most significant for accurately modeling cable-motor system conditions and forms mathematical expressions relating these features to the presence and severity of deterioration. Particular attention is given to PNETTR-4X, a modelling software package that forms adaptive learning networks from input training data using an unsupervised learning process. Models developed thus far can process readily measured terminal information and quantify deterioration power and current to within 3% of motor full load values. >