AUTHOR=Webb Z. T. , Nnadili M. , Seghers E. E. , Briceno-Mena L. A. , Romagnoli J. A. 

TITLE=Optimization of multi-mode classification for process monitoring

JOURNAL=Frontiers in Chemical Engineering

VOLUME=Volume 4 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/chemical-engineering/articles/10.3389/fceng.2022.900083

DOI=10.3389/fceng.2022.900083

ISSN=2673-2718

ABSTRACT=Data mining and knowledge discovery (DMKD) focuses on extracting useful information from data. A particular application of DMKD, fault detection and diagnosis (FDD), seeks to identify anomalous states in the process by using dimensionality reduction (DR), clustering (CL), or a combination of both coupled with a fault classification application. However, the selection of the appropriate methods for a particular application can be difficult. Moreover, the selection of appropriate hyperparameters for each method is also a challenging task. In this contribution, several ensembles for FDD were studied using the Tennessee Eastman Process and an industrial pyrolysis process. For each ensemble, the hyperparameters were selected using automatic Machine Learning based on a multi-objective optimization. The evolution of unsupervised clustering metrics (silhouette score, Davies-Bouldin index, and Calinski-Harabasz Index) with accuracy during optimization was tracked to assess the usefulness of these metrics to optimize the performance of the ensembles in the absence of labels. Results show that ensembles and hyperparameter selection can be enhanced by using multi-objective optimization. It was found that Silhouette score and Davies-Bouldin index are strong predictions of the ensemble’s performance and can then be used to obtain good initial results for fault diagnosis.