Dynamic process fault detection based on D-SVDD

Abstract

There are dynamic and non-linear characteristics in industrial processes. These features make the faults difficult to be detected. To improve the situation, the dynamic support vector domain description (D-SVDD) is proposed. This method is based on dynamic expanding matrix and the support vector domain description (SVDD). Compared with the traditional support vector domain description method, the new method takes the sequence correlation between variables into account. So it is more suitable for dynamic process data than traditional method. And it builds a more accurate model. First of all, the appropriate number of dynamic steps are chosen and the dynamic augmented matrices are constructed. Then the dynamic augmented matrices are used to calculate the boundary of the minimum hypersphere by means of SVDD. In this way, the distribution range of the data is described. In the end, the monitoring statistics of new samples are calculated. And the monitoring statistics are compared with the control limit obtained in the previous step to determine whether there are faults or not. The proposed method is applied to the Tennessee Eastman Process. The results show that the fault detection effect of D-SVDD has been significantly improved.