Mixture Probabilistic Linear Discriminant Analyzer for Process Fault Classification

Fault misclassification results in misleading locations of root faults, which brings economic loss and safety concerns. In order to enhance the performance of fault classification, this paper proposes a novel mixture fault classifier based on probabilistic linear discriminant analyzer (PLDA). Unlike the probabilistic model with a single latent variable, more useful information can be extracted by PLDA using within- and between-class latent variables. With the introduced state and mixture variables, the proposed mixture PLDA (MPLDA) classifier assigns a test sample to different components for each class in probability. For the class label decision, a robust state inference strategy is developed, which includes investigating the effect of the shared between-class variable on conditional probability and adopting the voting scheme to collect evidences from training samples to identify the class source of the test sample. After classifier construction, the model parameters are obtained by the expectation-maximization (EM) algorithm. The validity of MPLDA classifier in fault classification is then illustrated by the application of the Tennessee Eastman (TE) process.