An EMD-based long-term LSSVR for machine condition prognostics

Machine condition prognostics is very important for system safety and condition-based maintenance. Only one-step ahead forecasting of machine condition is considered because of the poor performance of multi-step ahead forecasting. To find effective method for non-stationary long-term machine condition prognostics, this paper firstly reviews multi-step ahead forecasting strategies followed with a comparison of different multi-step ahead forecasting strategies using support vector regression (SVR) and least squares support vector regression (LSSVR) in two datasets, and then develops an recursive multi-step LSSVR (MSLSSVR) model by introducing empirical mode decomposition (EMD) to realize machine condition prognostics. EMD is used to get more stationary signals instead of the non-stationary original signal, and MSLSSVR is constructed to make multi-step prediction with decomposed signals individually. All predicted values are combined eventually to get the future trajectory of condition indicator. Moreover, the proposed algorithm is validated in the TE process. The experimental result shows that the proposed EMD-MSLSSVR model can forecast the failure status in advance, and the performance is satisfied.