Kohonen Neural Networks for Machine and Process Condition Monitoring

Abstract

Kohonen feature map neural networks have been used for a variety of successful applications since their introduction in the late eighties. They have one major advantage over their more common peers in that they are capable of unsupervised learning. This property makes them ideal for machine health monitoring situations.

Unsupervised learning allows the network to represent single or multiple classes according to distribution and density. Novelty detection is then possible on-line or classes can be labelled to give diagnosis. This presentation explains the special nature of machine monitoring applications in data availability and desired diagnosis information and provides examples of such systems working in different environments.