Optimisation of SVM Methodology for Multiple Fault Taxonomy of Rolling Bearings from Acceleration Records

Abstract

Health monitoring of bearings is very critical for satisfactory working of complex machinery. Thus, the ability to detect bearing faults and classify them based on their nature becomes very important aspect of health monitoring of machines. In the machine learning methodology for the fault taxonomy, the support vector machine (SVM) is well recognized for its generalization capabilities. In this work, the taxonomy of rolling element bearing faults has been discussed. Acceleration signatures are classified by the support vector machine (SVM) learning algorithm. The tuning of the SVM and kernel parameters is necessary for better taxonomy. The novelty of the paper is in comparing the ability to classify a set of faults by the tuned SVM and kernel parameters with the help of grid-search method (GSM), genetic algorithm (GA) and artificial bee colony algorithm (ABCA). Four fault settings along with no-fault condition were considered. Three statistical features were obtained from acceleration signatures. The fault taxonomy was performed at the identical rotational speed at which signals were captured. The taxonomy capability is observed and it depicted a very good prediction performance especially at higher speeds.