Built and implemented anomaly detection and effective resource allocation for field operations for an oil and gas company.
Challenge
The client was an upstream gas producer with 500+ wells in remote areas with minimal or no operators present on ground – and had implemented sensors to monitor performance.
- Alert Fatigue: The existing reporting system generated thousands of alerts every month, leading to alarm fatigue. Most of the alerts were non-actionable.
- Identification Delay: Failures were identified only after the well had shut down leading to significant production loss.
Solution
- Tested multiple algorithms like SVM, rNN, NN, Time Series ARIMA models, etc and made a choice of final algorithms based on data availability and model quality
- Created 150+ features including trend dimensions, rate of change, second order effects and derivatives
- Created algorithms which mimic the production engineers process of identifying anomalies based on SCADA data
- Integrated with client alarm system through microservices based architecture
Impact
- Final models built after iterative tests had more than 90% of accuracy, with a median lead time of 90 minutes, with 10% of the cases having lead time in excess of 12 hours
- Classified 63% of events with three major heuristics
- Identified anomalies and patterns for optimization, reduced false alarms, improved accuracy of identifying shutdowns and reduced time gap between bad events
