Application of DFAR model to the trend prediction of fault in rotary machines

Aiming at the feature values of vibration signals of rotary machines, which sometimes have non-stationary and nonlinear development characteristics, DFAR (differential functional-coefficient autoregressive) model was presented. DFAR established nonparametric prediction model by using functional-coefficient autoregressive model. It automatically selected fractional or integral difference to process the original data and estimated the optimal model parameters according to the improved cross-validation criterion. The functional-coefficient autoregressive model can make the model parameters change with the model dependence variable. The fractional difference can extract the stable trend information from time series and it won't lose low frequency because it has no defect of over-difference. So, DFAR can approximate the nonlinear time series better. As shown in the experiments, DFAR can improve the precision of predicting development trend of the non-stationary and nonlinear fault feature values.