Kernel-based fault diagnosis on mineral processing plants

Abstract

Safe operation, environmental issues, as well as economic considerations all form part of the wide range of driving forces for the development of better fault diagnostic systems on process plants. The continuous search for novel methods for fault detection and identification resulting from these incentives has recently drawn attention to support vector machines as a means towards improved fault diagnosis. These kernel-based methods are in theory capable of better generalization, particularly as far as large systems are concerned, since their performance is not dependent on the number of variables under consideration and recent studies underlined their promising role in diagnostic systems. However, integration of these methods into the classical multivariate statistical process control framework is complicated by difficulties in the identification of the original variables associated with detected faults. In this paper, a general strategy for process fault diagnosis is proposed. First, kernel methods are used to remove nonlinear structure from the data, if present, after which the residuals from the data are used to monitor the process. A novel element of the strategy is the use of one-class support vector machines to estimate nonparametric confidence limits for these residuals. Using these limits in conjunction with Gower and Hand biplots and standard statistics collectively constitute a powerful approach to monitoring process systems, as demonstrated by several case studies on mineral processing systems.