Gearbox condition monitoring using self-organizing feature maps

Abstract

This paper proposes a novel technique for the condition monitoring of gearboxes based on a self-organizing feature maps (SOFM) network. In order to visualize the learned SOFM results more clearly, an improved method based on the unified distance matrix (U-matrix) method is presented, in which the overall topological information condensed into the map units is considered so as to project the high-dimensional input vectors into a two-dimensional space and give a better picture of their intrinsic structure than the original U-matrix method. The feature data extracted from industrial gearbox vibration signals measured under different operating conditions are analysed using the proposed technique. The results show that trained with the SOFM network and visualized with the improved method, the feature data are mapped into a two-dimensional space and formed clustering regions, each indicative of a specific gearbox condition. Therefore, the gearbox operating condition with a fatigue crack or broken tooth compared with the normal condition is identified clearly. Furthermore, with the trajectory of the image points for the feature data in two-dimensional space, the variation of gearbox conditions is observed visually, and the development of gearbox early-stage failures is monitored in time.