Shell's Predictive Maintenance AI: Preventing Refinery Failures Before They Happen

Richard Casemore - @skarard
The Cost of Unplanned Downtime
A single day of unplanned downtime at a major refinery costs $5-10 million. Equipment failures don't just stop production—they create safety risks, environmental hazards, and cascade effects across the supply chain.
Traditional maintenance followed schedules: replace parts every X months regardless of condition. This meant either replacing healthy equipment (wasteful) or missing failures that didn't follow the schedule (dangerous).
Shell needed a better way.
Sensors Everywhere
Modern refineries are instrumented with thousands of sensors measuring:
- Temperature at critical points
- Vibration in rotating equipment
- Pressure across systems
- Flow rates and compositions
- Electrical signatures from motors
Shell's AI platform ingests data from over 30,000 sensors per facility, processing millions of data points daily.
Pattern Recognition at Scale
Equipment failures don't happen instantly—they develop over days or weeks through subtle changes invisible to human operators.
Shell's AI identifies these patterns:
- A bearing developing micro-fractures shows slightly elevated vibration at specific frequencies
- Heat exchangers fouling exhibit gradual temperature differential changes
- Pumps approaching failure display characteristic electrical signatures
The system learned these patterns from historical failure data across Shell's global operations, creating a knowledge base impossible for any individual engineer to match.
The Alert System
When the AI detects anomalies, it doesn't just raise an alarm. It provides:
Confidence levels: How certain is the prediction? Time horizons: When is failure likely? Root cause hypotheses: What's probably wrong? Recommended actions: What should maintenance teams do?
This context helps engineers prioritize. A 60% chance of compressor failure in 30 days gets scheduled maintenance. A 95% chance of turbine failure in 48 hours triggers immediate response.
Implementation Across Global Operations
Shell deployed the system across 15 refineries worldwide:
Phase 1: Data integration and historical analysis Phase 2: Model training on facility-specific patterns Phase 3: Pilot deployment with human validation Phase 4: Full operational integration
Each facility required customization—equipment varies, and failure patterns differ based on local conditions and operational history.
Results That Changed the Business
Three years into deployment, Shell reports:
- Unplanned downtime reduced by 40%
- Maintenance costs down 20%
- Equipment lifespan extended 15%
- Zero AI-predicted failures missed
- $hundreds of millions in avoided losses
The safety implications matter too. Predicted failures are controlled shutdowns, not emergency responses.
The Human Element
Shell's system augments rather than replaces maintenance engineers. The AI handles pattern detection across thousands of sensors; humans handle investigation and repair decisions.
This partnership is essential because:
- AI can detect anomalies but not explain root causes definitively
- Physical repairs require human judgment and dexterity
- Edge cases need engineering expertise to interpret
- Safety decisions require human accountability
Expanding Beyond Refineries
Success at refineries led Shell to expand predictive maintenance to:
- Offshore platforms: Where failures are especially costly and dangerous
- Pipeline networks: Detecting corrosion and stress before leaks
- LNG facilities: Protecting complex cryogenic equipment
- Retail operations: Even gas station equipment benefits from prediction
Lessons for Industrial Operations
Shell's experience offers a template for industrial AI:
-
Start with clear business value: $5-10M/day downtime costs justified significant investment
-
Data quality is foundational: Years of sensor data, properly labeled, made training possible
-
Domain expertise is essential: AI teams partnered closely with refinery engineers
-
Validate extensively: Months of parallel operation before trusting predictions
-
Scale gradually: Each facility deployment built on previous learnings
The Future of Industrial Operations
Predictive maintenance is just the beginning. Shell now explores:
- Process optimization using similar techniques
- Energy efficiency improvements
- Emissions reduction through operational adjustments
- Autonomous operations for routine procedures
The refinery of the future doesn't just predict failures—it continuously optimizes every aspect of its operation. Shell's predictive maintenance AI is the foundation for that transformation.