A comparative analysis of classical and one class SVM classifiers for machine fault detection using vibration signals

Abstract

Early and efficient fault detection is very important in today's complex and sophisticated automated industry. For fault detection, many techniques have been employed among which the support vector machines (SVM) is a popular one owing to its many attractive features like fast classification, good handling capability of non-linear behavior of the data, and providing a global optimum for classification. This article presents the use of SVM and one of its variants i.e. one class SVM for fault detection in a rotation based machinery. The rotating machines give vibrational signals that can be analyzed to monitor the machines' health. The fundamental idea and implementation technique of classical SVM and one-class SVM are discussed. The vibration signals are obtained followed by feature extraction in time and frequency domain and on this basis, fault classification is performed. The performance of the said classifiers is compared for the Intelligence Maintenance Systems (IMS) bearing vibration data with the introduction of step and incipient faults respectively. Presence of incipient fault makes the classification very difficult. Afterwards, the classifier failure condition is calculated and the decision value plots are explicated. Classification results obtained using one class SVM are superior than classical SVM as advocated by our simulations.