An Approach to Rolling Bearing Fault Diagnosis using Fractal Descriptors and Regularized Least Squares

Rolling element bearings (REBs) are an inevitable part of every rotating component in machinery and are subjected to the maximum wear and tear. The early detection and diagnosis of faults greatly reduces cost of maintenance concurred during a machine downtime. The present work aims at characterizing vibration signal by using fractal descriptors obtained from the Variational Mode Decomposition (VMD) domain and its application in remaining useful life (RUL) prediction. The time-domain vibration signals from rolling element bearings are obtained from a run-to-failure experiment. Both linear and non-linear fractal descriptors are estimated from the band-limited Intrinsic Mode Functions (BLIMFs) using Higuchi's Fractal Dimension (HFD), Katz Fractal Dimension (KFD) and kurtosis. The extracted features in multi-resolution domains are then used to characterize faults using Regularized Least Squares (RLS) and three different kernels viz., linear, random and gaussian with two different cross validation techniques - Leave One Out (LO) and Hold Out (HO). It is found that Radial Basis function (RBF) kernel with hold out option gives maximum classification accuracy.