Remaining useful life estimation approach of rolling bearing based on the combined support vector classification and regression algorithms

To overcome the insensitivity of health indicators to faults and the difficulty in balancing model complexity with training efficiency, a remaining useful life prediction method of rolling bearing based on the combined support vector classification and regression algorithms is proposed. The time-domain features characterizing all fault information across various degradation stages are extracted from the raw vibration data of rolling bearing, and then the health indicator is constructed by using of principal component analysis method in the negative time series. The operating states of rolling bearing are divided by the support vector classification method into three stages, namely the stable period, the degradation period, and the failure period. The support vector regression model of different stages is built by the radial basis kernel function, and both its complexity and prediction performance are bettered by automatic optimization of RBF kernel parameters and penalty coefficients through grid search. The experimental results based on the PRONOSTIA dataset show that the accuracy rate of stage recognition reaches 98.69% and the error in lifespan prediction is only approximately 4.36%. The sensitivity and engineering applicability of this method for early degradation of rolling bearing have been verified.