A New Kernel-Based Classification Algorithm for Systems Monitoring: Comparison with Statistical Process Control Methods

Abstract

The paper presents a new Kernel-based monitoring algorithm compared with statistical process control methods, such as DISSIM and MS-PCA and some other methods widely used in process control applications. The proposed algorithm is a modified version of the well known support vector domain description (SVDD). The last one is commonly used for one-classification problems, named also novelty detection. In this paper, we have used a modified SVDD endowed with useful tools to manage multi-classification problems. The proposed classifier is also able to deal with stationary as well as non-stationary data. The principle is based on the dynamic update of the training set through a recursive deletion/insertion procedure according to adequate rules. In order to reduce the computational complexity and improve the rapidity of convergence, the algorithm considers in each run a limited frame of samples for the training process. To prove its effectiveness, the approach is assessed at first on artificially generated data. Then, we have performed a case study applied on chemical process.