# Multi-Modal Anomaly Detection
The defensible AI play in F&G safety. Combines multiple data streams to eliminate false alarms and predict failures.
## Architecture
1. **Computer vision** on existing CCTV infrastructure
2. **IoT sensors**: heat, gas, vibration, pressure, flow
3. **Time-series models** trained on facility-specific baselines
4. **BMS/SCADA integration** for automated response
## Why It's Defensible
Facility-specific learning creates [[Switching Costs]]:
- Each plant has unique "normal" signatures
- Models improve with operational history
- Retraining from scratch = months of baseline collection
This is [[Embedding]] into operations. Hard to rip out once calibrated.
## What It Solves
- Steam vs smoke discrimination
- Flaring vs fire detection
- Equipment heat vs thermal anomaly
- Sensor drift compensation
- Predictive failure detection (2-4 weeks ahead)
## Revenue Model
SaaS per facility + % of downtime prevented. Outcome-based pricing aligns incentives: customer pays when you save them from the $84M/year [[False Alarm Problem in F&G]].
## Links
- [[F&G Safety Opportunity MOC]]
- [[Defensibility Principles MOC]]
- [[Predictive Maintenance in O&G]]
- [[False Alarm Problem in F&G]]
---
#ai #technology #defensibility