Advancing Machine Learning in Industry 4.0: Benchmark Framework for Rare-event Prediction in Chemical Processes
This work provides a domain-specific framework for industrial operators to enhance safety and reliability in chemical plants through better rare-event prediction.
The authors tackled the problem of predicting rare abnormal events in chemical processes by developing a benchmark framework to compare various machine learning algorithms, finding that sophisticated models like XGBoost and TabNet achieved optimal performance with reduced RMSE and improved alarm efficiency.
Previously, using forward-flux sampling (FFS) and machine learning (ML), we developed multivariate alarm systems to counter rare un-postulated abnormal events. Our alarm systems utilized ML-based predictive models to quantify committer probabilities as functions of key process variables (e.g., temperature, concentrations, and the like), with these data obtained in FFS simulations. Herein, we introduce a novel and comprehensive benchmark framework for rare-event prediction, comparing ML algorithms of varying complexity, including Linear Support-Vector Regressor and k-Nearest Neighbors, to more sophisticated algorithms, such as Random Forests, XGBoost, LightGBM, CatBoost, Dense Neural Networks, and TabNet. This evaluation uses comprehensive performance metrics, such as: $\textit{RMSE}$, model training, testing, hyperparameter tuning and deployment times, and number and efficiency of alarms. These balance model accuracy, computational efficiency, and alarm-system efficiency, identifying optimal ML strategies for predicting abnormal rare events, enabling operators to obtain safer and more reliable plant operations.