Enhancing classification accuracy of ball bearing faults using statistically processed features

Abstract

A new diagnostic scheme is presented for ball bearing localized faults based on pattern recognition (PR) methods, which utilize preprocessed time domain features. The features are statistically processed (FP) using their central tendency (CT) estimations, prior to the classification process. Vibration data is acquired from faulty bearings, and the features are extracted to form data set. The FP algorithm deals with outliers present in the features by suppressing them. Utilization of the smoother feature distributions reduces the unwanted impact of vibration randomness and background noise in PR based fault diagnostic procedure. This significantly enhances the classification accuracy of classifiers. The results are compared with similar work in terms of maintaining an optimum classification accuracy of a diagnostic system with minimum number of features. The proposed scheme provides 93.6% classification accuracy for four bearing faults employing three features only, and even higher using additional features.