Outliers analysis and one class classification approach for planetary gearbox diagnosis

An application of outlier analysis for diagnosis of gearboxes working under non-stationary operating conditions is considered. The analysis is performed using quite large data characterizing (by 15 power spectrum amplitudes) two gearboxes, one in bad and the other in good condition. Analysis of Mahalanobis Distances in the two data sets shows that both of them are highly heterogeneous and non-Gaussian. A mapping of the data from the set 'good' to two dimensions by the non-linear Neuroscale method permits to visualize the data in a plane. Using the derived mapping, some decision boundaries (DBs) - permitting to identify a given fraction alpha = 0.05 of low density outliers - are constructed for the 'good' set. This is done using three methods: 1. Parzen kernel density, 2. mixture of Gaussians and 3. Suport Vector Data Description (SVDD). Each of the obtained DBs was tested using the bad data. For all 3 methods more than 98% of the bad data points were found outside the DBs constructed for the 'good' set.