Pattern classification of bearing faults in PMSM based on time domain feature ensembles

Abstract

This paper aims to identify an effective pattern classification method
that can be employed using vibration and current data to identify bearing conditions.
The authors attempted non-conventional time-domain features to detect the bearing
conditions in permanent magnet synchronous motors (PMSM). This study uses two
case studies with eight datasets from Paderborn University to identify the bearing
conditions of 3 and 12 classes. Support vector machine, k-nearest neighbor, random
forest, decision tree, and naive Bayes classifiers are attempted with 10% holdout
validation for 4 data sets with 31 feature ensembles. Also, this paper investigates the
Henry Gas Solubility Optimization (HGSO) feature selection approach for identifying
the most discriminant features. The effectiveness of these discriminant features is
verified with three bearing conditions diagnosis. Results have shown, that four feature
ensembles with 2 to 10 features outperformed support vector machine, k-nearest
neighbor, and random forest classifiers. In contrast to previous relevant studies, the
proposed features are useful in identifying PMSM-bearing conditions with excellent
accuracy in vibration and combined current signals under a wide range of operating
conditions.