Comparison of Performance of Machine Learning Methods for Bearing Faults Classification Using Time-Domain Features

This article deals with comparison of a classification success rate of machine learning methods used for vibration signals captured on damaged bearing. The most significant and the most used methods suitable for this vibrodiagnostic field have been selected, namely support vector machine, k-nearest neighbor, Mahalanobis-Taguchi system and principal component analysis for reduction of a dimensionality of a features space. All the methods have been implemented in MATLAB and their performance has been analyzed using same input datasets. As the input for all the aforementioned methods, data from bearing shortened life-cycle test, done by Bearing Data Center at Case Western Reserve University, have been used. Intentionally, computationally intensive pre-processing procedures have been excluded from the computational chain, as only time domain features have been used for the classification process. Resulting from this study, a classification error rate, an execution time, an implementation difficulty and a robustness of the algorithm strongly differ among all the methods while using the same input data and the same number and type of the input extracted features.