Batch process monitoring based on support vector data description method

Abstract

Process monitoring can be considered as a one-class classification problem, the aim of which is to differentiate the normal data samples from the faulty ones. This paper introduces an efficient one-class classification method for batch process monitoring, which is called support vector data description (SVDD). Different from the traditional data description method such as principal component analysis (PCA) and partial least squares (PLS), SVDD has no Gaussian assumption of the process data, and is also effective for nonlinear process modeling. Furthermore, SVDD only incorporates a quadratic optimization step, which makes it easy for practical implementation. Based on the basic SVDD batch process monitoring approach, the method is further extended to multiphase and multimode batch processes. Two case studies are provided to evaluate the monitoring performance of the proposed methods.