Novel fault subspace extraction methods for the reconstruction-based fault diagnosis

Abstract In fault diagnosis, partial least squares (PLS) is a popular data-driven method to identify abnormal key performance indicators (KPI). However, there are two problems in fault diagnosis when using PLS, including inaccurate fault subspace extraction and unidentified false alarms. In the first problem, the improved PLS (IPLS) model is adopted to obtain a precise subspace through orthogonal decomposition. In addition, to eliminate the normal value in fault data, a quality-related fault subspace (QRFS) extraction method is proposed, which can extract a purer quality-related fault subspace. In the second problem, to provide feedback for false alarms, a modified IPLS (M-IPLS) model is proposed to extract the quality-unrelated fault subspace. Based on the proposed fault subspace extraction methods, the fault can be reconstructed by a lower dimensional fault subspace and false alarms with feedback can improve the efficiency of diagnosis. Finally, two examples, including a numerical simulation and the Tennessee Eastman process (TEP), are used to show the effectiveness of the proposed method.