Process fault detection method using time-structure KICA and OCSVM
Ni Zhang · Journal of Tsinghua University(Science and Technology) · 2012
Fault detection methods using kernel independent component analysis(KICA) can be affected by the initial values of the separating matrix and the time-structure of the information in the data.In addition,the monitoring statistics for the Mahalanobis distance may reduce the fault detection rate.These problems are eliminated by a fault detection method based on a time-structure KICA and one-class support vector machine(OCSVM).The weighted-sum of the kernel-whitened difference data time-delayed covariance matrixes is computed with the extraction of the kernel independent components converted into an eigenvalue decomposition.Then,the statistical OCSVM model is built based on the extracted independent components with a monitoring statistic to detect faults online.Simulations of independent component extraction and fault detection in the Tennessee Eastman process illustrate that this method makes full use of the time-structure of the information in the data,avoids the effect of initial values on the separating matrix,shortens the fault detection latency and improves the fault detection rate.