A Fault Detection Method for Non-Gaussian Industrial Processes via Joint KICA and FVS

Abstract

The data in industrial processes have the features of non-linear and non-Gaussian. In order to enhance the accuracy of fault detection for industrial processes, and to reduce the calculation time consumption, a method is proposed to combine the joint kernel independent component analysis and the feature vector selection (FVS) to achieve fault detection in this paper. Firstly, the joint kernel function of Gaussian radial basis kernel function and polynomial kernel function is used to improve the learning and generalization ability of kernel independent component analysis (KICA) algorithm, and this can be employed to improve the accuracy of fault detection. Secondly, FVS is given to reduce the computational complexity of Joint KICA, especially in the case of large sample size. Finally, the simulation results of Tennessee Eastman (TE) process can be used to verify the effectiveness of this proposed method.