Enterprise AI for Preventing Downtime of End Flash Gas Compressors

Challenge

A major player in the global liquified natural gas (LNG) industry configured the BHC3™ Reliability application to help operators accurately predict compressor failures, reduce false alarms, and better leverage the high volume of control systems data.

End flash gas compressor systems, which pressurize the vapor produced during the LNG refrigeration process, are managed by direct control systems (DCS). The existing control systems lack the ability to accurately predict compressor failures and do not provide the root cause analysis required for troubleshooting. The resulting unplanned shutdowns cause significant delays and increase maintenance costs. To address these issues and increase the reliability of its compressors, the LNG company selected BHC3 Reliability.

Approach

The team ingested 1.4 billion data points from telemetry data, event failure history and asset hierarchy extracted from piping and instrumentation diagrams to create a unified and federated data image.

The team then leveraged this data image and a prebuilt anomaly detection machine learning pipeline as the foundation to determine the optimal model for predicting end flash gas compressor failures. After further testing, the final model detected 89% of unplanned shutdowns with a five-day proactive notice and reduced the number of false alarms by 55%. Lastly, the developers configured BHC3 Reliability dashboards to visualize the predictive insights providing a managerial view to maintain overall asset health of end flash gas compressors systems and subsystems.

Project Objectives

• Create a unified, federated data model integrating disparate data sources (e.g., asset hierarchy, sensor tags, failure events, P&ID)
• Deliver a user-friendly application that uses machine learning to identify high-risk end flash gas compressor units and predict unplanned shutdowns in advance
• Reduce the number of false alarms created by rules-based systems (DCS)

About the Company

• $2B+ in annual revenue
• 1,000+ employees
• 20+ million tons per annum of LNG production capacity

Project Highlights

• 8 weeks from kick-off to production-ready application
• 1.4 billion data points ingested from 4 disparate data sources and 3 years of data
• Created unified object model to represent asset hierarchy, telemetry data, and shutdown and events logs
• 1,000+ machine learning features tested
• 50+ model configurations built and tested
• 5 BHC3 Reliability application user interfaces configured