Fault Diagnosis of Rolling Bearings Using Data Mining Techniques and Boosting

Ünal M., Sahin Y., Onat M., Demetgul M., Küçük H.

JOURNAL OF DYNAMIC SYSTEMS MEASUREMENT AND CONTROL-TRANSACTIONS OF THE ASME, cilt.139, sa.2, 2017 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 139 Sayı: 2
  • Basım Tarihi: 2017
  • Doi Numarası: 10.1115/1.4034604
  • Dergi Adı: JOURNAL OF DYNAMIC SYSTEMS MEASUREMENT AND CONTROL-TRANSACTIONS OF THE ASME
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus
  • Anahtar Kelimeler: rolling bearing, fault diagnosis, Hilbert transform, C5.0, boosting, SVM, power spectral density, EMPIRICAL MODE DECOMPOSITION, WEAR DEBRIS, ACOUSTIC-EMISSION, HILBERT SPECTRUM, VIBRATION SIGNAL, DECISION TREE, BALL-BEARINGS, SOUND, TRANSFORM, DEFECT
  • Marmara Üniversitesi Adresli: Evet

Özet

Rolling bearings are key components in most mechanical facilities; hence, the diagnosis of their faults is very important in predictive maintenance. Up to date, vibration analysis has been widely used for fault diagnosis in practice. However, acoustic analysis is still a novel approach. In this study, acoustic analysis with classification is used for fault diagnosis of rolling bearings. First, Hilbert transform (HT) and power spectral density (PSD) are used to extract features from the original sound signal. Then, decision tree algorithm C5.0, support vector machines (SVMs) and the ensemble method boosting are used to build models to classify the instances for three different classification tasks. Performances of the classifiers are compared w.r.t. accuracy and receiver operating characteristic (ROC) curves. Although C5.0 and SVM show comparable performances, C5.0 with boosting classifier indicates the highest performance and perfectly discriminates normal instances from the faulty ones in each task. The defect sizes to create faults used in this study are notably small compared to previous studies. Moreover, fault diagnosis is done for rolling bearings operating at different loading conditions and speeds. Furthermore, one of the classification tasks incorporates diagnosis of five states including four different faults. Thus, these models, due to their high performance in classifying multiple defect scenarios having different loading conditions and speeds, can be readily implemented and applied to real-life situations to detect and classify even incipient faults of rolling bearings of any rotating machinery.