The traditional kernel principal component analysis is popularly used for fault detection

however

it only concentrates on the global structure of data sets and ignores the local structure when it is used to extract the nonlinear features. To solve the problem

a new method named modified kernel principal component analysis is proposed for nonlinear process fault detection and diagnosis. The idea of locality preserving is incorporated into the optimization goal of the traditional kernel principal component analysis

taking the excellence of kernel principal component analysis and manifold learning into account. The new projection space enjoys the similar global structure and the local structure

and thus

more feature information can be extracted. The modified kernel principal component analysis is used to map the data space into the feature space. Next

the feature information is classified through Fisher discriminant analysis. A monitoring statistic is established using the distance of each sample in feature space and its control limit is determined through kernel density estimation. When a fault is detected

the source of performance deterioration can be located by using a diagnosis method based on data set similarity. Finally

the results of Tennessee Eastman simulation experiment show its better effectiveness.