On-line fault detection method based on modified SVDD for industrial process system

To tackle problems of on-line fault detection in industrial process system, a method based on modified SVDD (support vector data description) is presented. Since the performance of SVDD is strongly influenced by kernel parameter selection and hyper-sphere class-boundary. A new criterion is presented to optimize the kernel parameter, which is performed by measuring the non-Gaussian value of kernel sample vector. Moreover, when the kernel sample vector is with uneven distribution, there are certain risks in using hype-sphere as class-boundary in comparison with hype-ellipse class-boundary. KPCA (kernel PCA) is employed to adjust the hyper-sphere to acquire more reasonable class-boundary, in which the length of ellipse¿s major axis in each principal component direction is computed and then equalize each major axis by scaling method. The feasibility and effectiveness of proposed approach is illustrated by the application of standard data and fault diagnosis simulation experiment.