Process Fault Detection Employing Feature Selection and Linear Discriminant Analysis

Classification methods such as linear discriminant analysis (LDA) have been widely applied to fault detection in industrial processes. In this case, the problem consists of classifying the operation as normal or faulty on the basis of monitored variables. If the number of such variables is large, selection techniques may be used to choose an informative subset of features in order to obtain a classifier with better generalization properties. In fact, the use of too many features may cause overfitting problems. In LDA, the presence of multicollinearity among the features may also lead to poorconditioning issues, which are a known cause of generalization problems for the resulting classifier. This paper presents a fault detection approach that employs the Successive Projections Algorithm (SPA) as a feature selection technique for use with LDA. SPA is a recently proposed technique, which was specifically designed to minimize multicollinearity among the classifier inputs. The joint use of SPA and LDA has provided good results in several pattern recognition problems. However, their application in process fault detection has not been previously reported. The performance of the proposed approach was assessed in a simulated case study involving the Tennessee Eastman process, which is a reactor-separatorrecycle system widely used as a benchmark in fault detection studies. The simulations involved 22 measured variables under normal operating conditions, as well as eight different types of faults. The LDA classifier inputs comprised present and time-lagged values of the measurements. SPA was thus applied to select not only the measurements to be considered for fault detection, but also the time lags to be employed. The results were evaluated in terms of overall classification performance, as well as sensitivity and false alarm rate. These metrics were obtained for a test set, which was not employed for feature selection or classifier training. As a result, a classification accuracy of 100% was obtained for six of the eight fault types. The accuracy for the remaining faults ranged from 75% to 85%. For comparison, two other classification techniques were also employed, namely κ-Nearest Neighbours (KNN) and LDA with feature selection by a Genetic Algorithm (GA-LDA). As a result, SPA-LDA was found to be superior to KNN and comparable to GA-LDA. In addition, SPA-LDA provided more parsimonious classification models as compared to GA-LDA.