Identification of component faults using automated condition monitoring methods has a huge potential to improve the prediction of machine failures. The ongoing development of the Internet of Things (IoT) will support and benefit feature selection and improve preventative maintenance decision making. However, there may be problems with the selection of features that best describe a specific fault and remain valid even when the operation mode is changing (for example different levels of load). In this study, features were extracted from vibration signals using wavelet analysis; a feature subset was selected using the random forest ensemble technique. Three different datasets were created where the load of the system was changing while the rotational speed remained the same. The tests were repeated five times by first recording the nominal condition and then introducing four faults: angular misalignment; offset misalignment; partially broken gear tooth failure; and macro-pitting of the gear. To improve previous feature selection techniques, a method is proposed where, before training a classifier, the most promising features are compared at different degrees of torsional load. The results indicate that the proposed method of using random forests to select top variables can help to choose good features that may not have been considered in manual feature selection or in individual load zones.
Validerad;2018;Nivå 2;2018-08-16 (andbra)