Online Fault Prediction Based on Combined AOSVR and ARMA Models

Abstract

Accurate fault prediction can obviously reduce cost and decrease the probability of accidents so as to improve the performance of the system test and maintenance. Traditional fault prediction methods are always off-line that are not suitable for online and real-time processing. For the complicated nonlinear and non-stationary time series, it is hard to achieve exact predicting result with single models such as SVR, ANN and ARMA. Combined with the AOSVR (accurate online support vector regression) algorithm and the ARMA model, this paper presents a new online approach to forecast fault with time series prediction. The fault trend feature can be extracted by the AOSVR with global kernel for general fault modes. Moreover, its prediction residual that characterized the local high-frequency components is synchronously revised and suppled by the sliding time window ARMA model. Fault prediction with combined AOSVR and ARMA has realized better result than the single ones. Experiments on Tennessee Eastman process fault data show the new method is practical and effective.