New PCA-based scheme for process fault detection and identification. Application to the Tennessee Eastman process

Abstract

This paper proposes a new principal component analysis (PCA) scheme to perform fault detection and identification (FDI) for systems affected by process faults. In this scheme, a new modeling method which maximizes the model sensitivity to a certain process fault type is proposed. This method uses normal operating or known faulty data to build the PCA model and other faulty data to fix its structure. A new structuration method is proposed to identify the process fault. This method computes the common angles between the residual subspaces of the different modes. It generates a reduced set of detection indices that are sensitive to certain process faults and insensitive to others. The proposed FDI scheme is successfully applied to the Tenessee Eastman process (TEP) supposedly affected by several process faults.