Prognosis of Defect Propagation Based on Recurrent Neural Networks

Abstract

Incremental training is commonly applied to training recurrent neural networks (RNNs) for applications involving prognosis. As the number of prognostic time-step increases, the accuracy of prognosis generally decreases, as often seen in long-term prognosis. Revision of the training techniques is therefore necessary to improve the accuracy in long-term prognosis. This paper presents a competitive learning-based approach to long-term prognosis of machine health status. Specifically, vibration signals from a defect-seeded rolling bearing are preprocessed using continuous wavelet transform (CWT). Statistical parameters computed from both the raw data and the preprocessed data are then utilized as candidate inputs to an RNN. Based on the principle of competitive learning, input data were clustered for effective representation of similar stages of defect propagation of the bearing being monitored. Analysis has shown that the developed technique is more accurate in predicting bearing defect progression than the incremental training technique.