Fault Prediction of Nonlinear System Using Time Series Novelty Estimation

Many complex nonlinear systems have the characteristics such as difficult modeling, dangerous testing, and high cost to experiment with fault. Aim at these points, a new kind of novelty estimation method based on an on-line Least Square Support Vector Regression (LS-SVR) algorithm is presented for fault prediction of nonlinear system in this paper. Firstly, a set of characteristic token values, which represent normal state of time series at the beginning, are selected as the training data to establish the on-line LS-SVR model. Then, the output values of the on-line model are analyzed to estimate novel pattern. Once the novel pattern is affirmed, the fault of system is forecasted immediately. The proposed method needs neither the math model of system nor the fault training data and primary knowledge, and it is believed with good real-time capability. The results of simulation on Continuous Stirred Tank Reactor (CSTR) show the efficiency.