Vibration Signature Analysis using Rough Sets and Analogy-Based Reasoning Classification

Abstract

Various machine learning techniques along with vibration signal processing techniques are being explored for effective and accurate fault diagnosis of bearing faults. This work studies and evaluates the application of rough set based algorithms - RoughSet Classifier, LocalKnn Classifier and RseslibKnn Classifier and their combinations with 4 kinds of distance metrics - City block and Hamming (CBD and HD), City block and simple value difference (CB and VDM), Density based value difference (DBVDM), Interpolated value difference(IVDM) for bearing vibration signature analysis. The input vector fed to the classifier models is composed of statistical features extracted from pre-processed vibration signals. The different metrics used to compare the performance of the classifiers show that RseslibKnn classifier with City block and simple value difference distance metric takes considerably less time for training (9.09 sec) and classification (0.86 sec) while giving testing accuracy of 84.1442 %, thereby confirming its usefulness for real time application on large bearing fault datasets. The results obtained in this paper show the effectiveness of rough sets based algorithms, particularly KNN based classifiers, in bearing fault diagnosis.