A pruned support vector data description-based outlier detection method: Applied to robust process monitoring

Abstract

Support vector data description (SVDD) is a boundary-based one-class classifier that has been widely used for process monitoring during recent years. However, in some applications where databases are often contaminated by outliers, the performance of SVDD would become deteriorated, leading to low detection rate. To this end, this paper proposes a pruned SVDD model in order to improve its robustness. In contrast to other robust SVDD models that are developed from the algorithmic level, we prune the basic SVDD from a data level. The rationale is to exclude outlier examples from the final training set as many as possible. Specifically, three different SVDD models are constructed successively with different training sets. The first model is used to extract target points by means of rejecting more suspect outlier examples. The second model is constructed using those extracted target points, and is used to recover some false outlier examples labeled by the first model. We build the third (final) model with the final training set consisting of target examples by the first model and false outlier examples by the second model. We validate our proposed method on 20 benchmark data sets and TE data set. Comparative results show that our pruned model could improve the robustness of SVDD more efficiently.