Improved fault detection based on kernel PCA for monitoring industrial applications

The conventional Kernel Principal Component Analysis (KPCA)-based fault detection technique requires more computation time and memory storage space to analyze large-sized datasets. In this context, two techniques, Spectral Clustering (SpC) and Random Sampling (RnS), are developed to reduce the dataset size by retaining the more relevant observations while preserving the main statistical characteristics of the original dataset. These two techniques and others use the training dataset from two different industrial processes, Tennessee Eastman (TEP) and Cement Plant (CP) to be reduced and provided to build the Reduced KPCA (RKPCA) model-based fault detection scheme. The obtained results show the effectiveness of the proposed techniques in terms of some fault detection performance indices and computation costs.