A statistical comparison of neuroclassifiers and feature selection methods for gearbox fault diagnosis under realistic conditions

Gearboxes are crucial devices in rotating power transmission systems with applications in a variety of industries. Gearbox faults can cause catastrophic physical consequences, long equipment downtimes, and severe production costs. Several artificial neural networks, learning algorithms, and feature selection methods have been used in the diagnosis of the gearbox healthy state. Given a specific gearbox, this study investigates how these approaches compare with each other in terms of the typical fault classification accuracy but also in terms of the area under curve (AUC), where the curve refers to the precision-recall curve otherwise known as receiver operating characteristic (ROC) curve. In particular, the comparison aims at identifying whether there are statistically significant (dis)similarities among six feature selection methods, and seven pairs of neural nets with different learning rules. Genetic algorithm based, entropy based, linear discriminants, principal components, most neighbors first, and non-negative matrix factorization are the studied feature selection methods. Feed forward perceptrons, cascade forward, probabilistic nets, and radial basis function neural nets are evaluated. Six supervised and one unsupervised learning rules are considered. Both parametric and nonparametric statistical tests are employed. A ranking process is defined to elect the best approach, when available. An experimental setup was especially prepared to ensure operating conditions as realistic as possible.