Using a novel combination of process modeling and molecular dynamics-based path-sampling algorithms, rare abnormal un-postulated unsafe/unreliable events were simulated efficiently, with a robust, predictive model developed using supervised machinelearning — this model was then utilized to develop novel, improved multivariate alarm systems

The multivariable alarm systems also includes automated/operator-assisted response actions that monitor the process in real-time to prevent abnormal 'crashing' to the undesirable operating region, supporting existing systems

Methods demonstrated initially on a simple exothermic CSTR led to promising multivariate alarm systems, with current focus on extending our analyses to a complex styrene polymerization CSTR with multiple potential undesirable operating regions

Tree-based ensemble machine learning algorithms (XGBoost, LightGBM) leveraged for developing highly accurate predictive models, capable of providing real-time tracking of the committer probabilities, resulting in improved, dynamic multivariate alarm systems

Rationalization strategies introduced to evaluate the quality and effectiveness of the multivariate alarm systems, with appropriate modifications made based on key statistical metrics

Strategies resulted in a significant reduction in the number of nuisance alarms, focusing on only quality alarms, which, if ignored,were more likely to result in an abnormal shift in operation to the undesirable regions

Finally, the real-time risk of the rationalization systems were quantified using dynamic risk analysis and Bayesian Analyses, resulting in low-variance posterior distributions, updated easily online as more data are received