Failure prognosis of rolling bearings using maximum variance wavelet subband selection and support vector regression

Machinery failure prognosis is one of the major tasks of condition-based maintenance in which the current issues of the machines are diagnosed, and the remaining useful life (RUL) is estimated by monitoring its present condition. This paper proposes a data driven approach for RUL estimation of rolling bearings. Health indicators (HI) of training bearing and test bearings are constructed by fusion of three features (semivariance, rms and variance) of the wavelet subband that has the maximum variance. Selection of the maximum variance subband helps in clear visualization of the fault progression with time. The dimensions of the selected features are reduced using principal component analysis. The dimensionally reduced features are then subjected to moving averaging over fixed window size and normalizing the moving average. By observing the run to failure degradation profile of training bearings, their respective failure thresholds are estimated. To estimate the failure thresholds of test bearing, HI of the test bearings are matched with the HIs of all the training bearings using bicubic interpolation and goodness of fit function. Support vector regression (SVR) models are used to predict the future degradation profile of test bearings and are constructed using HIs of the bearings. The failure threshold and SVR model finally predict the RUL of the bearings. PRONOSTIA data set is used for validation of the proposed algorithm. The prediction results are compared with some similar works on aforesaid dataset and have proved to be better in terms of prognosis evaluation parameters, viz. mean error, score and standard deviation error.

Automated summary

This paper proposes a data-driven approach for estimating the remaining useful life (RUL) of rolling bearings. By constructing health indicators (HI) using the maximum variance subband of the wavelet features, reducing feature dimensions using principal component analysis, and using support vector regression (SVR) models to predict the future degradation profile, the RUL of the bearings can be accurately predicted.