Data-driven forecasting for anomaly detection in a compressor unit

Abstract

Equipment reliability is crucial for refineries and timely anomaly (equipment failures, sensor faults, wear and tear, unexpected inputs etc.) detection is essential to keep equipment running safely, improve performance, and have an effective maintenance strategy. Modern refineries generate large amounts of data. Combined with machine learning, models that can monitor the operation of complex processes and equipment in real-time can be learned. These models can guide operators, and engineers in identifying faults. This study proposes a data-driven approach to detect anomalies of a reciprocating compressor in a petrochemical refinery. The idea is to capture the regular operating behavior of the compressor with a learned model and compare its predictions with measurements. As such, a model that forecasts future sensor outputs given past measurements is trained from real-world historical data. Deep neural networks with recurrent layers are utilized. After training, the forecasted measurements are compared with the observed measurements and any large deviations are flagged as potential anomalies. The approach is evaluated both on historical and real-time data. The results demonstrate that the approach can be used as an anomaly detection decision-aid for operators and engineers. The approach has the potential to facilitate rapid actions, to help avoid major faults, and for reducing operator fatigue and cognitive load, letting them focus on higher level tasks such as monitoring entire processes versus single equipment.