Fault Diagnosis and Detection Based On Combination with Gaussian Mixture Models and Variable Reconstruction

This paper presents a fault diagnosis approach that is the combination with Gaussian mixture models and variable reconstruction. Usually, the traditional multivariate process monitoring techniques has the fundamental assumption that the operating data should follow a unimodal Gaussian distribution, but it often becomes invalid due to the practice different operating conditions. The Gaussian mixture models method can overcome above problems and make the fault diagnosis to be more accurate than before. And fault diagnosis based on principal component analysis is to use contribution plot to locate the fault sources, but it often results in indistinct or incorrect diagnosis. Thus the variable reconstruction approach is introduced to resolve the problem. As a result, a novel multimode process monitoring approach based on the combination with Gaussian Mixture Model and variable reconstruction is proposed. The combination method is illustrated for a simulated Tennessee-- Eastman process (TE) which is tested for fault diagnosis and detection.