Bearing fault diagnosis using discrete Wavelet Transform and Artificial Neural Network

Abstract

Rotating machinery has vast industrial applications in fields of petroleum, automotive, HVAC and food processing. Rotating machineries use bearings to perform rotational or linear movement of various subcomponents while reducing friction and stress. Compared other types of bearing, REBs offer a good balance of key attributes like friction, lifetime, stiffness, speed and cost. Hence, real-time monitoring and diagnosis of bearings is crucial to prevent failures, improve safety, avoid unforeseen downtime of production assembly lines and lower cost. We propose an approach based on Wavelet Transform and ANN for analysis of vibration signals from a rolling element bearing to identify and multi-classify its component defects. The vibration signals from the REB being analyzed are passed over to the software setup consisting of Wavelet Transform and ANN. To remove noise and extract the relevant features from this signal, we pass the vibration signal through a Wavelet transform. These features are retrieved using time domain parameters like Skewness, Kurtosis, RMS and Crest Factor and they are used as an input for ANN classifier. The role of the ANN is to classify the bearing fault features produced by the Wavelet Transform and identify bearing faults, if any. To this end, we have designed a feedforward topology ANN using the sigmoid transfer function. The ANN training methodology uses three learning paradigms - namely, Levenberg-Marquardt, Resilient Back-propogation and Scaled Conjugate method. The learning models generated by each algorithm are tested to find the one which gives better accuracy. The outcome of this experiment indicates that DWT and ANN can together achieve good accuracy and reliability in detection and classification of bearing faults.