Enhancing the Classification Performance of Machine Learning Techniques by Using Hjorth's and Other Statistical Parameters for Precise Tracking of Naturally Evolving Faults in Ball Bearings

Abstract

The research on identification of artificially induced faults in bearing is available in abundance in the past literature, however, the diagnosis becomes more challenging when the fault evolves naturally inside the bearing, and especially when its stages need to be precisely tracked. The conventional statistical features used commonly in the past literature do not uniquely characterize the fault status, and yield satisfactory results only in limited cases, like those for artificial faults. In this work, a new combination of fault descriptors, including three Hjorth's parameters, three statistical features and an entropy measure, is proposed and its effectiveness has been analyzed on classification performances of k-Nearest Neighbor (k-NN) and Support Vector Machine (SVM). Two datasets comprising the signals of run-to-failure tests were taken from Intelligent Maintenance Systems (IMS) and the Paderborn university repository. The data were categorized into large number of classes to closely indicate the actual fault type and size. Compared to conventional statistical features, the new combination was able to enhance the classification accuracy of k-NN and SVM, respectively, from 91.3% to 99.9%, and from 94.8% to 99.7% in the case of IMS dataset, and from 94.1% to 98.5%, and from 94.7% to 98.4% in the case of Paderborn dataset. In addition to accuracy, the performance metrics, including precision, recall, and F1-score were also improved using the proposed combinatory features.