Abstract: A new method, distance mapping, is presented in order to visualize the trained results by self-organizing maps (SOM) apparently. By means of similarities evaluated based on Euclidean distances between input vectors and output neurons weights combining with the distribution of fixed lattices in the network, high-dimensional input vectors are projected into a two-dimensional space. SOM is employed in fault recognition and condition monitoring of gearbox combining with the proposed visualizing technique. It is proved that feature points under gear normal, tooth cracked and tooth broken conditions are mapped into different areas on two-dimensional space more clearly by distance mapping than U-matrix method, which helps distinguish gearbox conditions correctly. With the trace of the image points for gear feature data on the plane, the variation of gearbox conditions is observed visually, and furthermore, early gear failures occurrence and its varying trend is monitored in time.

Key words: Condition monitoring, Distance mapping, Self-organizing maps, U-matrix, Visualization