Optimization of multi-mode classification for process monitoring

Abstract

Process monitoring seeks to identify anomalous plant operating states so that operators can take the appropriate actions for recovery. Instrumental to process monitoring is the labeling of known operating states in historical data, so that departures from these states can be identified. This task can be challenging and time consuming as plant data is typically high dimensional and extensive. Moreover, automation of this procedure is not trivial since ground truth labels are often unavailable. In this contribution, this problem is approached as a multi-mode classification one, and an automatic framework for labeling using unsupervised Machine Learning (ML) methods is presented. The implementation was tested using data from the Tennessee Eastman Process and an industrial pyrolysis process. A total of 9 ML ensembles were included. Hyperparameters were optimized using a multi-objective evolutionary optimization algorithm. Unsupervised clustering metrics (silhouette score, Davies-Bouldin index, and Calinski-Harabasz Index) were investigated as candidates for objective functions in the optimization implementation. Results show that ensembles and hyperparameter selection can be aided by 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 subsequent fault detection and fault diagnosis procedures.