Path-Sampling and Machine Learning for Rare Abnormal Events: Application to Polymerization CSTRs

Abstract

Previously, we developed novel, unidirectional, static, multivariate, alarm systems for rare un-postulated abnormal events, demonstrated successfully for an exothermic CSTR. Herein, our techniques are improved significantly for a more-complex polystyrene CSTR, operating in its unstable region, capable of abnormal shifts to two undesirable regions; i.e., unsafe and unreliable regions. BG-FFS, a path-sampling algorithm, is utilized to simulate efficiently multiple rare abnormal trajectories; then, the XGBoost machine learning algorithm is utilized to develop accurate predictive models for committer probabilities; i.e., pB as a function of key process variables – such models, when deployed in real-time, result in improved bidirectional dynamic multivariate alarm systems, capable of response actions using real-time pB predictions. Then, using our rationalization strategies, the initial alarm systems are evaluated and modified, followed by DRAn (Dynamic Risk Analysis) studies and sensitivity analyses to investigate the effects of varying other process parameters to achieve more effective response actions.