Recurrent neural networks for long-term prediction in machine condition monitoring

Abstract

A new approach to multi-step prediction of machine health status using recurrent neural networks (RNNs) was developed. Based on the principle of competitive learning, input data to the networks were preprocessed and clustered for more effective representation of similar stages of the process being monitored. The developed technique has shown to provide more accurate failure progression prediction than the commonly used recurrent network techniques, as demonstrated by experiments using defect-seeded rolling bearings.