A fault diagnosis method of kernel Fisher projection was proposed to solve issues of nonlinear data

complex classes

and difficult fault-diagnosing in chemical processes. The proposed method provided a uniform solution for partial sample mix-up induced by a large difference in category distances and nebulous classification of different category's boundary data after projection of the original data sample. First

an improved category distance was used to change sample distribution in the projection space so that the sample had a good projection. Then

boundary data was screened out by a defined threshold parameter and classified by improved K-Nearest Neighbor (K-NN) algorithm

which none-boundary data was classified by Mahalanobis distance. Simulation in a TE process showed that training accuracy was increased by 2.25% and testing accuracy was increased by 2.41% in the first experiment

whereas training accuracy was increased by 4.75% and testing accuracy was increased by 6.75% in the second experiment. Therefore

the method improved both fault diagnosis time and accuracy.