5.4.2 Feature-based analysis of reproducible bearing damages based on a neural network
- Event
- 20. GMA/ITG-Fachtagung Sensoren und Messsysteme 2019
2019-06-25 - 2019-06-26
Nürnberg, Germany - Chapter
- 5.4 Zustandsüberwachung
- Author(s)
- A. Beering, J. Döring, K. Krieger - ITEM Universität Bremen (Deutschland)
- Pages
- 451 - 456
- DOI
- 10.5162/sensoren2019/5.4.2
- ISBN
- 978-3-9819376-0-2
- Price
- free
Abstract
In the interest of detecting damage to tapered roller bearings at an early stage and avoiding further consequential damage in future, an investigation of reproducible damage to bearing inner races, outer races and rolling elements is carried out. The vibration signals generated by the contact of the damaged surfaces during bearing runtime are recorded via a piezoelectric vibration transducer. Different scenarios with regard to the rotational speed and the size of the damage are investigated. Based on a calculation of features, a feed-forward neural network is trained and used to classify the damage. To assess the quality of the neural network, the receiver operator characteristic (ROC) and the area under curve (AUC) are compared for the neural network as well as for other popular classification algorithms such as support vector machine (SVM), decision tree and k-Nearest-Neighbor (KNN). Based on the AUC values, this approach showed that the neural network used has the best performance for the classification of bearing damage with an AUC of 0.994 and an overall classification accuracy of 93.5 %.