Fault prediction based on time series with online combined kernel SVR methods

In order to reduce the cost and decrease the probability of accidents, accurate fault prediction is a goal pursued by researchers working at system test and maintenance. Most of traditional fault forecasting methods are not suitable for online prediction and real-time processing. To solve this problem, an online data-driven fault prognosis and prediction method is presented in this paper. The operating states are forecasted with on-line time series prediction model based on the online combined kernel functions Support Vector Regression (SVR). Compared with batch SVR prediction models, online SVR has a good real-time processing performance. However, it is hard for a single kernel SVR to obtain accurate result for the complicated nonlinear and non-stationary time series. Therefore, a combined online SVR with different kernels containing global and local kernels is developed for fault prediction. For general fault modes, the fault trend feature can be extracted by global kernel. On the other hand, local kernel can reflect and revise the local changes of data characteristics in neighborhood. It has realized better result than the method of the single SVR. Experimental results for Tennessee Eastman process fault data prove its effectiveness.