Diagnosing batch processes with insufficient fault data: generation of pseudo batches

Abstract

To ensure the safety of a batch process and the quality of its final product, one needs to quickly identify an assignable cause of a fault. Cho and Kim (2003) recently proposed a diagnosis method for batch processes using Fisher's Discriminant Analysis (FDA), which showed a satisfactory performance on industrial batch processes. However, their method (or any other method based on empirical models) has a major limitation when the fault batches available for building an empirical diagnosis model are insufficient. This is a highly critical issue in practice because sufficient fault batches are likely to be unavailable. In this work, we propose a method to handle the insufficiency of the fault data in diagnosing batch processes. The basic idea is to generate so-called pseudo batches from known fault batches and utilise them as part of the diagnosis model data. The performance of the proposed method is demonstrated using a real data set from a PVC batch process. The proposed method is shown to be capable of handling the data insufficiency problem successfully, and yields a reliable diagnosis performance.