Fault detection using Linear Discriminant Analysis with selection of process variables and time lags

This paper is concerned with the selection of process variables and time lags for fault detection. For this purpose, a feature selection technique known as Successive Projections Algorithm (SPA) is employed with Linear Discriminant Analysis (LDA) classifiers to discriminate between normal operating conditions and faults. SPA was originally designed to minimize multicollinearity among the selected features, which is a known cause of generalization problems for LDA. In the present work, a modification to the basic SPA formulation is proposed to place larger emphasis on the selection of relevant features for the classification task. The proposed SPA-LDA methodology is illustrated in a case study involving the Tennessee Eastman benchmark process. For comparison, a genetic algorithm (GA) for feature selection is also employed. In this study, the pre-selection of process variables was found to improve the accuracy of the resulting classifiers. In practice, such a pre-selection would have the extra advantage of reducing the number of sensors required to detect a given fault. Moreover, the proposed modification in SPA-LDA resulted in an improvement of average classification accuracy from 88% to 96%. This result was similar to that obtained by GA-LDA (97%). However, the SPA-LDA classifiers were found to be less sensitive to measurement noise.