PREDICTIVE MAINTENANCE · CASE STUDY
A field radio that tells you it's failing two weeks early.
Below is the story of RDO-MTN — a 902 MHz repeater at Sierra Tower, west of CS-038. Over fourteen days, its signal quality, packet retry rate, and chassis temperature drifted in ways too gentle to trip a threshold alarm — but loud enough for a behavioral model to flag with high confidence. Scrub the timeline. Watch the four traces diverge. Watch the AI catch it.
02 / TIMELINE
14 Days of Quiet Drift
Hit Play to walk forward through the case study. Each panel shows one signal the radio publishes; the bottom row is the AI’s confidence that it’s degrading. Pay attention to when the bar turns amber.
Retry rate ticks above baseline. Statistically not significant — but the model notices.
Joint behavior across RSSI / temp / retries fits a known degradation signature. Confidence climbing.
AI confidence crosses 75%. Service window recommended. Estimated time to fault: 4 days.
Traditional SCADA would just now notice — packet loss crosses 5%. Aevus had this 48 hours ago.
End of horizon. Without intervention, the link drops here.
03 / OUTCOME
Two Possible Day 14s
Same radio, same drift, same physical failure mode. The only variable is whether anyone was watching for it.
- D 0–11
Telemetry stays "inside the box." Threshold alarms remain silent.
- D 12
Packet retry crosses 5%. SCADA fires a HIGH alarm at 02:14.
- D 13
Crew dispatched. Replacement radio not in regional stock; truck rolls from Austin.
- D 14
Link drops mid-day. 6 wellsites lose telemetry. Production deferred.
- D 4
Behavioral model flags subtle retry drift. Logged, not surfaced to the operator.
- D 7
Multi-signal pattern recognized. Predictive watchlist auto-includes RDO-MTN.
- D 10
Confidence > 75. NOC receives a single ticket: "Service RDO-MTN within 96h."
- D 11
Spare radio dispatched alongside next scheduled site visit. Cutover in 22 min.
Related Articles
Where AI earns its keep in the control room.
The warning was there for 11 days before the outage.
Time-series databases, deadband compression, and query patterns.
Ready to see how this applies to your operation? Start a pilot conversation — no commitment, no field changes.