Efficient Selection of Data Samples for Fault Classification by the Clustering of the SOM

In this paper we propose a sample selection procedure for improving accuracy of supervised classifiers in fault classification tasks. To generate faulty samples, a laboratory testbed is constructed and to avoid loss of a 3-phase AC induction motor (due to high short-circuit currents) resistors are used to limit current levels. This gives rise to short-circuit faults of different impedance levels, which may generate data samples difficult to classify as normal or faulty ones, specially if the faults are of high impedance (easily misinterpreted as non-faulty samples). Aiming at reducing misclassification, we use the clustering of the SOM approach [1] with modified information criteria for cluster validation. By means of comprehensive computer simulations, we show that the proposed approach is able to cluster successfully the different types of short-circuit faults and can be used for the purpose of sample selection.