An Automated Workflow for Condition Monitoring of Centrifugal Compressors Using a Combined Data-Driven and Physics-Based Approach

Abstract Equipment degradation is inevitable during operation. The breakdown of equipment is costly due to replacement expenses and may cause a facility to shut down, leading to higher costs and production losses. Hence, equipment maintenance is essential to ensure a satisfactory level of reliability during its lifecycle. Maintenance strategies may adopt a reactive approach, in which maintenance happens after failure and should be avoided, or a preventive approach, in which maintenance happens regardless of the state of the system. This work aims at addressing the challenges posed by system deterioration and costly downtime through the development and implementation of a condition monitoring solution for a centrifugal compressor situated in an offshore platform. To achieve this objective, a comprehensive methodology was developed containing two modules: fault detection and prognostics. In the fault detection module, a hybrid approach was adopted by combining measured and simulated data generated using physics-based models, where a machine learning model is trained to detect anomalies and generate alarms. In the prognostics module, a degradation state variable is introduced to predict the time to failure of the centrifugal compressor. By leveraging process simulation to construct a physics-based model, the system can track changes in design polytropic efficiency using real-time sensor data to make predictions of time to failure using a Kalman filter algorithm. We demonstrate the methodology through examples using actual centrifugal compressor measurements collected from Valhall public industrial data. The fully automated solution provides real-time insights on the operational state of the centrifugal compressor, enabling proactive maintenance strategies. The analysis of centrifugal compressor sensor data is streamlined to detect anomalies and trigger alarms for different sensors, enabling prediction of time to failure within a few days before the compressor shutdown happens. The approach presented in this work integrates the strengths of both physics-based and data-driven methods. The methodology exploits the alarm generation and complements it with a time-to-failure prediction using a Kalman filter algorithm. The presented solution contains a real-world application of machine learning in condition monitoring and has the potential to enhance predictive maintenance planning and prevent costs associated with downtime.