Shallow Versus Deep Neural Networks in Gear Fault Diagnosis

Accurate gear defect detection in induction machine-based systems is a fundamental issue in several industrial applications. At this aim, shallow neural networks, i.e., architectures with only one hidden layer, have been used after a feature extraction step from vibration, torque, acoustic pressure and electrical signals. Their additional complexity is justified by their ability in extracting its own features and in the very high-test classification rates. These signals are here analyzed, both geometrically and topologically, in order to estimate the class manifolds and their reciprocal positioning. At this aim, the different states of the gears are studied by using linear (Pareto charts, biplots, principal angles) and nonlinear (curvilinear component analysis) techniques, while the class clusters are visualized by using the parallel coordinates. It is deduced that the class manifolds are compact and well separated. This result justifies the use of a shallow neural network, instead of a deep one, as already remarked in the literature, but with no theoretical justification. The experimental section confirms this assertion, and also compares the shallow neural network results with the other machine learning techniques used in the literature.