Machine Condition Monitoring and Fault Diagnostics with Imbalanced Data Sets based on the KDD Process

Abstract

Abstract: Machine condition monitoring is a maintenance strategy, which enables real-time diagnostics and prognostics of machine conditions. One major problem of machine fault diagnostics is that data presenting faulty behavior is significantly underrepresented resulting in poor classifier accuracy for the faulty classes. Since misclassification of faulty behavior results in unplanned machine breakdowns and thus economic loss, improved fault diagnostics classifiers handling data imbalance are of importance. This paper addresses the problem of binary and multi-class imbalanced data sets. Based on the structural KDD process for data mining, required steps for imbalanced data sets are defined. The KDD process is evaluated through a real-world industrial case including data sampling and the development of a support vector machine classifier.