Identifying vibration critical gearset features using a Bayesian Regularized Artificial Neural Network

In situ analyses utilizing vibration measurements is a growing field of research. Still, with few exceptions, the focus of research remains on condition monitoring of hydraulic systems and components, or the reduction of undesirable noise associated therewith. The case described here identifies and quantifies key component parameter impacts on the vibration characteristics of orbital motors, specifically those associated with the orbital gearsets. We describe a novel methodology for handling test data with a Bayesian Regularized Artificial Neural Network, where results are interpreted using a connection weight algorithm to determine the relative contribution of each input parameter on time and frequency domain vibration characteristics. Using a randomized approach, the statistical significance of different parameters’ relative importance is examined. Insights include the arduous work of specifying and creating relevant parameters for both- the independent geometric parameters and dependent vibration parameters, the challenges of interpreting the generated relationships from a neural network and how BRANN can alleviate some of these hurdles.