Bearing fault detection and classification by wavelet-artificial neural network and wavelet-energy singular value ratio

Abstract

In this dissertation, Wavelet-ANN (Artificial Neural Network) and Wavelet-ESVR (Energy Singular Value Ratio) are proposed for detection and classification of faults in systems with periodical characteristics. Bearings in particular display such periodical characteristics. Bearings are one of the most critical components in rotary machinery and the majority of failures arise from defective bearings. Early warning in bearing deterioration is essential and it can prevent substantial machine downtime. In the first method an ANN is trained to diagnose bearing health by extracting information from discrete Wavelet coefficients. In the second approach, ESVRs are used to detect variations in the continuous Wavelets coefficients as symptoms of bearing health. Computer-simulated data and real bearing vibration data were then applied to perform initial testing and validation of these approaches. The test results show that the proposed methods are effectively detecting different bearing faults.