Snapshot
Who: Regional gathering & transmission operator (anonymized) — mixed vegetation, seasonal crop interference.
Problem: Annual flyovers missed fast-moving temperature deltas; spreadsheets couldn’t fuse IR with SCADA transient hints.
What we shipped: Nightly ortho differencing, calibrated IR normalization, and a Bayesian layer that down-weights anomalies correlated with heat waves or controlled burns.
Challenge
False positives eroded trust with field supervisors. Every alert needed imagery, map location, and a one-sentence hypothesis.
Approach
We replayed five years of maintenance tickets to train the ranker on what actually got validated in the field vs. closed as noise.
Solution components
Rollout timeline
Months 1–2 · Historical replay
Matched known incidents to model scores; tuned vegetation masks by county.
Months 3–4 · Shadow alerts
Integrity managers received ranked list; field response unchanged.
Months 5–6 · Live queue
Integrated with CMMS; SLA timers per severity band.
Quarter 3 · Expansion
Added third imagery vendor without breaking normalization pipeline.
“We stopped debating whether a pixel was ‘orange enough.’ The queue told us where to drive first—and the photos held up in the morning meeting.”
— Director of Pipeline Integrity, anonymized operator
Results & controls
- Four validated issues surfaced 3–6 weeks earlier than the historical foot-patrol cycle would have caught them.
- Estimated $1.7M in avoided product loss, environmental response, and regulatory exposure (internal ERM model).
- Emergency crew mobilizations −22% after de-prioritizing correlated false clusters.
- Each alert packet includes imagery hash, model version, and weather station IDs.
Representative scenario for marketing purposes. Regulatory reporting obligations differ by region; figures are internal estimates unless externally audited.