Process Fault Detection Using Directional Kernel Partial Least Squares

In this paper, a directional kernel partial least squares (DKPLS) monitoring method is proposed. The contributions are as follows: (1) By analysis of the relevance between the input residual and output variables, the kernel partial least squares (KPLS) residual subspace still contains output-relevant variation. (2) A new KPLS algorithm, DKPLS, is proposed to extract the output-relevant variation. Compared with the conventional algorithm, the DKPLS algorithm builds a more direct relationship between the input and output variables. (3) On the basis of the proposed DKPLS algorithm, a process monitoring method is proposed. In this monitoring method, kernel latent variables are used to explain the extracted output-relevant variation and calculate monitoring indices. Faults are detected accurately by the proposed DKPLS method. The DKPLS monitoring method is used to monitor a numerical example and the electrofused magnesium process. The experimental results show the effectiveness of the proposed DKPLS method.