VIBRATION SIGNATURE ANALYSIS OF ROLLER BEARING USING SOFT COMPUTING TECHNIQUES

Abstract

Rolling element bearings have been extensively used in machinery to transmit the force and relative movement. REBs have such diverse applications that these are used in paper mill rollers to the space shuttle main engine turbomachinery. Roller bearings are usually much stiffer structures and have high resistance to fatigue failure compared to ball bearings of the same size. These bearings have a significant influence on the dynamic behavior of the rotating machines, and also, these bearings act as a source of vibration and noise in the system. Hence, it is critical to enhance the performance and reliability of rolling element bearings to avoid catastrophic failure of the machines. Few works are available to investigate the effect of external parameters like load change on vibration signatures. Also, defect on one rolling component of rolling elements bearing affect other component performance, compound defects are generated, and the available literature has less explored their identifications. The transient characteristics of the non-stationary vibration signal of rolling element bearings can be evaluated more extensively using empirical mode decomposition and wavelet-based methods. Extracted features from vibration signals have abundant information on bearing damage, fault initiation, and propagation; hence, this study has attempted to build more responsive features set for the bearing fault identification. A bearing test rig, Bearing Prognostics Simulator (BPS) at vibration lab, IIT Roorkee, for the analysis and data acquisition. NBC NU205E bearing has been seeded with a combination of defects on the inner race, outer race, and roller with the three different sized defects with width and depth dimensions. For the first statistical analysis, four indicators, RMS, Skewness, Crest factor, and Kurtosis, are extracted from EMD, and raw signals are analyzed. . An early level detection of fault can be done based on particular IMF and with the comparison of four significant statistical parameters. IMFs are selected based on the correlation factor of IMF and raw signal. Decomposition of signals by EMD gives high to low-frequency modes, and hence Kurtosis and Crest factor values extracted from obtained IMFs have shown better fault diagnostic information.Response Surface Methodology has been used to study the parametric effect of various operating parameters with three input factors; Load, Speed, and Defect size, and two outputs; FFT’s Vibration peak and Torque Peak-to-Peak (P2P). The Box-Behnken method has been used to get trials to plot response surfaces. Response surfaces show that change in load value doesn’t affect vibration amplitude significantly, while a slight speed variation considerably increases vibration values. On the other hand, both parameters affect torque values significantly. Defect size variation has no significant impact on the output, i.e., vibration or torque. Increasing speed sometimes decreases vibration amplitude. This is because of the “Self-peening” phenomenon of the bearing defects, where high-frequency amplitudes often decrease. Two classifiers, namely ANN and KNN, have been used for fault classification. Three sets of feature vector sets, from time-domain signals, wavelet transform signals, and empirical mode decomposed signals, have been used. Total fourteen statistical features have been extracted from each signal of the time domain and WT and EMD processed signals. An appropriate wavelet has been selected using Minimum Shannon Entropy Criterion. It has been concluded that features filtered with a supervised filter and with PCA and supervised filter give equal accuracy. This has been observed that the PCA reduces the calculation and processing time without any significant improvement in classification efficiency. Healthy bearings in both the classifier cases have nearly 100% classification efficiency, and no defect has been classified as healthy bearing and vice versa. This is a good indication of the classification process as healthy bearing shall not be replaced without any defect. More than 95% accuracy of fault detection is achieved by ANN classifier using EMD feature set. EMD has been verified as an effective feature extraction method. The adaptive characteristic of EMD has evaluated the frequency with differentiation and not with the convolution.