Importance of Variables in Gearbox Diagnostics Using Random Forests and Ensemble Credits.

We consider a multivariate data matrix of size \(n \times d = 2183 \times 15\), where \(n=2183\) is the number of time segments recorded from vibration signals of two gearboxes, and \(d=15\) is the number of variables (traits) characterizing these segments. To learn about the role played by each of the 15 variables in the gearbox diagnostics, we use the Random Forest (RF) methodology with its ‘Variables Importance Plot’ (VIP) algorithm, which yields a kind of ranking of the variables with regard of their importance in the performed diagnostics. This ranking is different in various runs of the RF. We propose to use at this stage an additional module performing a specific ensemble learning yielding credits scores for each variable. It shows clearly the top most important variables.