Machine Learning
Security Operations
Anomaly Detection in Industrial Control Systems
12,000+ SCADA sensors
3–7 days early warning
91% detection precision
The Challenge
12,000+ SCADA sensors generating continuous telemetry with no system to identify degradation patterns before failure. Purely reactive maintenance meant faults were only discovered after breakdown — at substantially higher intervention cost and with direct production impact.
What We Built
An ML anomaly detection pipeline processes high-frequency telemetry across all sensors continuously. Models trained on healthy operating baselines surface failure indicators 3–7 days before breakdown. Prioritised maintenance work orders are auto-raised — teams act before equipment fails.
Operational Impact
28% reduction in unplanned downtime. 91% anomaly detection precision with less than 4% false positive rate. £2.1M in avoided emergency intervention costs in Year 1.
Technology Used
Python
PyTorch
scikit-learn
Azure ML
Grafana
PostgreSQL
