Alarms prediction and classification in industrial processes using supervised machine learning techniques: A case study in an Algerian gas plant

Abstract

Alarm systems are a crucial tool designed to enhance safety levels and ensure the normal functioning of industrial plants, maintaining safe and efficient operations. During industrial process upsets, numerous conflicting and false alarms may trigger simultaneously (alarm floods), leading to confusion and creating significant challenges for operators. These alarm floods affect operators' response time making their intervention extremely difficult. In such abnormal situations, alarm classification and prioritization become crucial, significantly aiding operators by allowing them to promptly and appropriately address safety-critical alarms first, rather than dealing with false or lower-priority alarms. Meanwhile, Machine Learning (ML) is a powerful tool for information extraction that has significantly contributed to knowledge discovery and decision-making. It has been successfully applied in various fields, including fault detection and diagnosis. ML can help address the issue of process alarms by classifying and prioritizing them. This paper presents a Machine Learning-based model (Random Forest) capable of classifying and predicting alarms in industrial processes. Then, it compares its performance to well-known classifiers, including Support Vector Machines (SVM), Artificial Neural Networks (ANN), and other supervised machine learning models such as Decision Trees, K-Nearest Neighbors, and Logistic Regression. The performance of these models was rigorously evaluated based on Accuracy, Precision, Recall, F1-Score, and prediction speed. The results from our final simulations show that the RF model achieved the highest Accuracy (98.32%) and F1-Score (0.988), along with a very high Recall (0.987) and precision (0.983). While the RF model demonstrated superior predictive performance in these metrics, it had a slower prediction speed (0.3477 ms per observation) comparing to other models.