Identification of optimal wavelet packets for machinery fault diagnosis using Bayesian neural networks

Abstract

This paper presents an approach to automatically extract optimal features using Bayesian neural networks. Hyperparameters are used in Bayesian neural networks to reflect the significance of inputs to the underlying model. The relevant input variables can be determined by judging the magnitudes of these hyperparameters. This automatic relevance determination technique is applied on the signals’ wavelet packet transform to select optimal feature combination for fault diagnosis, with each feature element extracted from a wavelet packet. The proposed method integrates feature selection and pattern recognition, leading to a compact and automatic pattern identification procedure. A case study on rolling element bearing fault diagnosis showed that the selected features with reduced dimension derived accurate diagnosis results when compared with the original features, demonstrating that the proposed approach provides automatic fault diagnosis.