Abstract
A classical method of determining the condition of a gearbox has been the analysis of vibration signals. The characteristics of a vibration signal captured from a gearbox are heavily dependent on systems parameters such as the speed profile and meshing harmonics as well as variations from a variety of sources. One means by which the influence of these parameters can be minimized is through synchronous averaging techniques. However, different synchronous averaging techniques return different signal outputs, and the ability to extract useful information regarding the condition of a gearbox thus varies as well. This research seeks to objectively compare the strengths and weaknesses of two synchronous averaging techniques with respect to gearbox condition classification performance. By using a numerical model, vibration signals from an idealized gearbox with known defects and uncertainties are generated with both a constant speed and varying speed. These signals are then processed using time-synchronous averaging and angle-frequency domain synchronous averaging, respectively, to return feature enriched signal data. Both sets of synchronously averaged data are then processed through a simple, general benchmark convolutional neural network to predict the condition of the gearbox. The classification performance and consistency returned highlights the strengths of synchronous averaging as a preprocessing step for gearbox classification.
