Fault Detection on Bearings and Rotating Machines based on Vibration Sensors Data

In this work a comparative study among the known fault detection techniques Local Outlier Factor and Isolation Forest as well as a proposed methodology called Standardised Mahalanobis Distance is presented. The study is focusing on the challenging problem of fault detection on bearings and rotating machines using vibration sensors’ data. During the first phase of the experiments, all models are applied and evaluated using cross-validation on a dataset created in lab by obtaining vibration signals of a rotating machine. In the second phase, the outlier detection techniques including the proposed one, are applied and evaluated on a popular, public dataset. In both phases, various parameters’ combinations are tested in order to find the most efficient set for each technique. As can been derived by the evaluation results, the Standardised Mahalanobis Distance methodology outperforms Local Outlier Factor and Isolation Forest on fault detection on voltage drop down of rotating machines in the case the voltage value of the abnormal condition is not close to the nominal. In addition, the evaluation results from the public dataset indicate that Standardised Mahalanobis Distance is able to identify outliers before an outer race fault on a bearing occurs, in a more efficient and solid way than Local Outlier Factor and Isolation Forest models. The proposed approach is applied also on a real world scenario in the premises of major lift manufacturer, using custom vibration sensors and it is currently under further evaluation.