Abstract:

For several complex industry processes，the original fault sources are difficult to identify by using kernel principal component analysis（kernel PCA）methods. And during the modeling and online dynamic monitoring process，the calculation of the kernel matrix K is a bottleneck problem for a large data set. An integrated fault diagnosis method based on feature sample extracting and kernel PCA was developed. Firstly，a feature extraction method was adopted to pre-process the modeling data set for solving the calculation problem of the kernel matrix K. Secondly，Hotelling statistics，T² and SPE of kernel PCA were adopted to detect system fault. Once fault was detected，the gradient algorithm of kernel function was used to define two new statistics，C_T² and C_SPE，which represented the contribution of each variable to Hotelling T² and SPE respectively. According to the degree of contribution，the fault variables might be identified from these correlative variables. To demonstrate the performance，the proposed method was applied to the Tennessee Eastman（TE）process. The simulation results showed that the proposed method could effectively identify various types of fault sources.