Abstract:

This study identified and classified 20 different defects in an industrial facility that contains nonlinear processes of different chemical units. The data set used is an IEEEDataPort online data set obtained from a large industrial facility. This data set, known as the Tennessee Eastman Process, includes measurements taken from 52 processing points with 20 different types of errors. Through these measurements, the Poincare drawings were obtained and the non-linear properties used frequently for each processing point were extracted. These properties are applied to the one-way ANOVA test at a statistical significance level of 5%, which indicates that there is a statistically significant difference between the error types. Both the properties and only the properties selected with ANOVA are classified using five different community learning algorithms (Boosted Trees, Bagged Trees, Subspace Discriminant, Subspace KNN and RUSBoosted Trees). The highest classification accuracy in this study was achieved by using the Subspace Discriminant algorithm push of 89.5%. A comparable level of success has been achieved with similar studies using the same data. On the other hand, ANOVA-based proprietary selection has not shown a clear advantage in the diagnosis of defects in such industrial process facilities.

Keywords:

Abstract:

This study aimed to detect and classify 20 different malfunctions in an industrial facility that involves nonlinear processes from various chemical units. The IEEEDataPort online dataset, acquired from a large industrial plant, was used in this study. It contains measures from 52 process points in Tennessee Eastman Process with 20 different fault types. We extracted two commonly used nonlinear features from Poincare Plots for each measurement point. The statistically meaningful features, which show statistically significant differences among fault types with a significance of 5%, were selected from these features. Five distinct Ensemble Learner algorithms (Boosted Trees, Bagged Trees, Subspace Discriminant, Subspace KNN, and RUSBoosted Trees) discriminated the fault types using all features and the selected features only. The maximum classifier accuracies were 89.5% for both feature sets using the Subspace Discriminant method in this study. This performance is a comprehendible result among the results achieved in similar studies. On the other hand, ANOVA-based feature selection didn't result in a clear advantage to diagnose faults in such industrial process plants.