Data description and noise filtering based detection with its application and performance comparison

Abstract

Abnormalities should be detected reliably so that corrective actions can be taken to maintain a high quality level of products. Fault or outlier detection provides early warning for a fault and identification of its assignable cause. The availability of large real-time datasets has motivated the study of data-driven approaches to fault detection. Recently, many powerful kernel-based nonlinear learning techniques have been developed and shown to be very effective tools. As one of one-classification methods, SVDD is able to define a boundary around samples with a volume as small as possible. This paper proposes a data description-based detection method combined with orthogonal filtering for enhanced detection abilities. The orthogonal filtering as a preprocessing step is executed before SVDD modeling to remove unwanted variation of data. The performance of the proposed method was demonstrated using data of two test processes. The case study has shown that the proposed method produces reliable detection results. In addition, the use of SVDD combined with orthogonal filtering step outperformed other one-class classification-based detection methods.