Missing measurement estimation of power transformers using a GRNN

Abstract

Many industrial devices are monitored by measuring several attributes at a time. For electrical power transformers their condition can be monitored by measuring electrical characteristics such as frequency response and dissolved gas concentrations in insulating oil. These vectors can be processed to indicate the health of a transformer and predict its probability of failure. One weakness of this approach is that missing measurements render the vector incomplete and unusable. A solution is to estimate missing measurements using a General Regression Neural Network on the assumption that they are correlated with other measurements. If these missing values are completed, the entire vector of measurements can be used as an input to a pattern classifier. To test this approach, known values were deliberately omitted allowing an estimate to be compared with actual values. Tests show the method is able to accurately estimate missing values based on a finite set of complete observations.