(509g) Path-Sampling and Machine Learning for Rare Abnormal Events

Chemical manufacturing processes can pose serious dangers, and therefore, it is crucial to incorporate safety and reliability measures during their design. Indeed, to minimize the risk of catastrophic accidents, extensive instrumentation such as control systems, alarms, and safety interlocks, are employed routinely in chemical processes. Additionally, common reliability assessment methods such as failure mode and effect analysis (FMEA), fault-tree analysis (FTA), and the like, have proven to be effective in identifying and handling postulated abnormal events that have occurred previously or are more likely to occur, based on process historian data. However, it is difficult to predict and counteract the impact of infrequent and unforeseeable un-postulated abnormal events in real-time, which, when not considered during process design, can lead to the most serious consequences. Hence, existing reliability/safety systems, alone, might prove to be insufficient in monitoring and alerting the operator for un-postulated abnormal events.

Over the past couple of years, we developed a novel advisory system for analyzing and monitoring process reliability, consisting of multivariable alarms and reliability systems introduced using process modeling and path-sampling for un-postulated abnormal events (Sudarshan et al., 2021; Sudarshan et al., 2022). Its purpose is to augment and support existing reliability systems, suggesting actions when unanticipated reliability/quality events are approached. While path-sampling algorithms are applied routinely in molecular dynamics to analyze rare events, our work presents the first application of such methods to analyze rare abnormal events in process safety/reliability. Our analyses were demonstrated initially on an exothermic CSTR process and led to promising alarm thresholds and reliability response actions (Sudarshan et al., 2023a). We have been extending our analyses to a polystyrene CSTR exhibiting free-radical polymerization (FRP) mechanisms, known to exhibit complex, nonlinear behavior; e.g., output multiplicity, input multiplicity, isolas, nonminimum phase behavior, and the like.

Next, simple rationalization strategies were introduced, wherein the acceptability of every alarm threshold and response action was evaluated, with the alarm thresholds and/or response actions modified accordingly, based on key statistical metrics — seeking to ensure that every alarm is a quality alarm, and its response action is justified appropriately. For the exothermic CSTR, our 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 (Sudarshan et al., 2023b). Similar rationalization strategies are being applied to the advisory system developed for the polystyrene CSTR.

Additionally, the real-time performance of the rationalized alarms and reliability systems is evaluated using dynamic risk assessment, a method developed by our research group over the past two decades, in which, the risk associated is analyzed by estimating the failure probabilities of the reliability systems (Pariyani et al., 2012b; Moskowitz et al., 2015), based on multiple dynamic simulations for the process, inclusive of control, alarms and reliability systems. Expectedly, the failure probability distribution developed for the exothermic CSTR had a much lower variance as compared to one developed for a flat prior using Bayesian statistics — similar low-variance probability distributions are constructed for the rationalized alarms and safety/reliability systems for the polystyrene CSTR.

Keywords: Un-postulated Abnormal Events, Path-Sampling, Advisory System, Alarms, Dynamic Risk Assessment

References:

Sudarshan, V., Seider, W.D., Patel, A.J., Arbogast, J.E., 2021. Understanding rare safety and reliability events using forward-flux sampling. Computers and Chemical Engineering 153.

Sudarshan, V. , Seider, W. D. , Patel, A. J. , Oktem, U. G. , Arbogast, J. E. , 2022. Alarm and Safety System Design Using Forward Flux Sampling. AIChE Annual Meeting Conference Proceedings.

Sudarshan, V., Seider, W.D., Patel, A.J., Oktem, U.G., Arbogast, J.E., 2023a. Reliability Path Sampling and Dynamic Risk Analysis – Part I – Developing an Advisory System To Recognize Rare Unpostulated Reliability Events – in preparation.

Sudarshan, V., Seider, W.D., Patel, A.J., Oktem, U.G., Arbogast, J.E., 2023b. Reliability Path Sampling and Dynamic Risk Analysis – Part II – Alarm Rationalization Strategies – in preparation.

Pariyani, A., Seider, W.D., Oktem, U.G., Soroush, M., 2012b. Dynamic risk analysis using alarm databases to improve process safety and product quality: Part II-Bayesian analysis. AIChE Journal 58

Moskowitz, I.H., Seider, W.D., Soroush, M., Oktem, U.G., Arbogast, J.E., 2015. Chemical process simulation for dynamic risk analysis: A steam-methane reformer case study. Ind. Eng. Chem. Res 54, 4347–4359.