Kernel Principal Component Analysis Improvement based on Data-Reduction via Class Interval

Abstract

Kernel Principal Component Analysis (KPCA) is an effective nonlinear extension of the Principal Component Analysis for fault detection. For large-sized data, KPCA may drop its detection performance, occupy more storage space for the monitoring model, and take more execution time in the online part. Reduced KPCA pre-processes the training data before applying the KPCA method, the proposed approach selects samples based on class interval to reduce the number of observations in the training data set while maintaining decent detection performance. This approach is applied to the Tennessee Eastman Process and then compared to some of the existing approaches.