Enterprise AI for Predicting Asset Failures
A large vertically integrated hydrocarbon producer experienced a major unplanned shutdown due to asset failures at one of its upstream production sites and began to reevaluate its current asset maintenance systems and strategy.
The company had historically relied on traditional rules-based control systems to detect operational anomalies. However, due to the sensitivity of the control systems, a high volume of alarms was triggered daily, and operators struggled to distinguish false alarms from material deviations. The control systems could not accurately predict asset failures and operated in siloes. The company needed a better solution that could materially improve the quality of alarms, accurately predict failures before they occurred, and provide an integrated end-to-end view of assets.
To address these issues and increase the reliability of the critical systems, the oil & gas company selected BHC3™ Reliability. Within 2 weeks, a team of Baker Hughes and C3 AI experts ingested over 2 years of data for more than 2,500 assets from 27 locations to create a unified data image. The team then configured a prebuilt anomaly detection pipeline
to predict asset failures and provide root cause analyses for 2 critical upstream systems: gas compressors and water injection pumps. Using BHC3 Reliability, the team can predict asset failures with a recall of 100% up to 5 days in advance and maximize asset lifetime using time-to-failure models that provide over 2 years of visibility into asset health.
With BHC3 Reliability, the oil & gas company can generate over $10 millions of additional annual revenue and savings from increased uptime and reduced maintenance costs across the 2 systems: gas compressors and water injection pumps.
About the Company
- Vertically integrated across upstream, midstream, and downstream operations
- $100+ billion in 2019 revenue
- 50,000+ employees
- Operations in 10+ countries
Within 2 weeks, a team of Baker Hughes and C3 AI experts collaborated with project managers and subject matter experts from the oil & gas company to develop a production-ready application that could predict equipment failures for 2 types of upstream systems, gas compressors, and water injection pumps, and provide operators with early warning signals and root cause analyses to reduce unplanned downtime.
The team began by ingesting over 2 years of operational data into a unified, federated data image. The unified data image included sensor data, piping and instrumentation diagrams, and original equipment manufacturer data sheets for over 2,500 assets from 27 operating locations. The data image leveraged BHC3™ Reliability’s configurable asset templates and diagram parsing capability to model the system and subsystem hierarchy in a system-of-systems data model that is scalable and extensible to any equipment type.
The team leveraged the data model and pre-built anomaly detection machine learning pipelines to determine the optimal models for predicting asset failures. First, the team defined true anomaly events for both systems and leveraged the BHC3™ AI Suite expression engine to configure over 30 key performance indicators (KPIs). These definitions and KPIs were used as inputs for the anomaly detection pipelines.
The team utilized a semi-supervised approach and approximately 1 million data points across the 2 systems and 16 subsystems to train the ML models. Lastly, the team leveraged ML prediction feature contributions, BHC3 Reliability Failure Mode Library and recommendation engine to diagnose failure modes and recommend corrective actions. The final ML models achieved a recall of 100%, accurately predicting 100% actual failures during testing and validation up to 5 days in advance of a failure event.
Moreover, the team configured 2 time-to-failure models for each system to provide over 2 years of visibility into asset health. Visibility into necessary and major maintenance activities such as a major compressor overhaul due to fouling, allows operators to efficiently plan for long-term degradations and convert costly unplanned maintenance into lower-cost planned maintenance.
Lastly, the Baker Hughes and C3 AI developers configured BHC3 Reliability dashboards to visualize the predictive insights generated by the machine learning models. The user-friendly dashboards, configured with readily available UI components, provided a managerial view of the high-level KPIs, overall health of systems and subsystems, and AI-based prioritized risk scores that enables users to efficiently triage a high volume of alerts. Users can also utilize the application’s interpretability framework to deep dive into specific AI alerts to view the key components or tags contributing to an AI alert.
- 2 weeks from project kick-off to production-ready application
- 2 years of data from 2,500+ assets and 27 locations unified into data image
- 250+ ML features generated & evaluated
- 30+ asset key performance indicators configured
- 2 production-ready ML pipelines and 9+ ML models configured
- 3-screen BHC3 Reliability dashboard configured to visualize KPIs & ML results