Aircraft predictive maintenance that turns the night's warnings into planned work before they ground an aircraft
Warnings become planned work before they ground an aircraft.
See one case, screen by screen ↓Aircraft predictive maintenance, sometimes called prognostics and health management, turns aircraft health-data warnings into planned work for maintenance control and engineering. Eight AI agents read the night's ACARS fault messages, condition-monitoring downloads and engine and APU trends, account for every warning, compare each trend with the fleet and with past failures, argue against it, and fit the work into a ground time with the part and the tooling. The maintenance controller asks the agents questions in plain words, sees the cited evidence, and accepts, defers or dismisses each task with a reason code and an e-signature.
How to choose AI for airlines: a buyer's guide to agentic AI solutions →
Updated
Every night a network fleet sends on the order of a thousand fault and condition messages by datalink (ACARS) and post-flight download. Most repeat, many are harmless — a power transfer at the gate, a maintenance test, a door cycle — and some are already covered by an open MEL deferral or work order. The aircraft health monitoring engineer and maintenance control read that stream by hand, across the airframer's portal, the engine and APU condition reports and the M&E system, and on a busy night it is triaged, not read.
The failures that matter are the slow ones: an IDG oil outlet temperature creeping up over eleven flights, an EGT margin eroding faster than the fleet. No single message is alarming. When a trend is caught, it reaches maintenance control as an email that knows nothing about where the aircraft will be, the hangar, the qualified technicians, the spare or the tooling. So it is watched until the part fails away from base — an unscheduled removal, an aircraft on ground, and on an ETOPS twin a lost generator can mean a diversion.
Estimated split of one night's review for maintenance control and health monitoring at a network carrier, by hand and with the solution. Hours, not days.
Eight agents condense the night's warnings, argue against every trend, fit the work into the plan and learn from dismissals; maintenance control decides.
Corvane Airways flies 300 aircraft from its hub at Denver. On the night of October 14, 268 of them flew and sent 1,240 warnings. One of them, a 777-300ER due to fly to London tomorrow, has a left IDG on the curve of two earlier failures. Here is how the maintenance controller works it with the agents — reads their briefing, asks them questions, checks their evidence and signs — screen by screen, in the working solution.
Rosa leads the night shift at MCC Denver. Her screen opens with the planner agent's briefing: 1,240 warnings condensed to 7 proposed tasks and 24 patterns to watch, two tasks waiting on an engineer, and one feed late — Halvard Air's 22 aircraft were not read after 02:10. Each task carries its decision deadline, N841CV's left IDG first, decide by 20:05.
“One matters most: N841CV's left IDG is running on the curve of two earlier failures and flies to LHR tomorrow; it can be changed at DEN tonight.”
The overnight stream runs down the left; the agents condense it in the middle. The intake agent folded 612 repeats — same fault, same tail, same flight phase. The triage agent set aside 388 known harmless messages by rules R1–R14 and linked 142 to an open deferral or work order. The trend agent grouped the 98 signals left into 31 patterns: 7 proposed tasks, 24 on the watch list.
612 + 388 + 142 + 98 = 1,240. Nothing is dropped without a rule or a record.
Rosa opens “388 known harmless”. Every rule is listed with what it sets aside and how many it caught tonight: R1, power transfer at the gate, 96; R2, maintenance test in progress, 71; R4, entry door cycles, 52. Any line can be pulled back by a person and becomes a signal again — the nuisance alerts are filtered, never hidden.
“Every line below can be pulled back by a person.”
N841CV's left IDG oil outlet temperature rise has climbed every flight for 11 flights, from 24 °C to 33 °C, against a fleet normal of 21–26 °C, while the right IDG stays normal. It matches N836CV (March 2025, similarity 0.89), which disconnected over the Atlantic and sat 31 hours at LHR, and N847CV (January 2026, 0.84), 19 hours at ORD. On both, the failure came 3 to 4 flights after today's value.
Both teardowns found generator bearing wear with a partly blocked scavenge filter.
The check agent tried to explain the trend away: a temperature sensor (the right IDG reads normal on the same flights, and the left sensor was replaced in June), hot-weather ground running (the rise holds on cold-day flights), and the oil service on Oct 6 (the trend began Oct 2). None fits, confidence 0.91. The planner agent found the slot: DEN tonight, Hangar 2, 23:00–04:30, in an 18 h 35 min ground time before CV 908 to LHR, a 9 h 20 min ETOPS flight.
“If it is not done: an inoperative IDG does not allow an ETOPS departure.”
Rosa asks in plain words. The planner agent answers with its sources: CV 908 is the second flight from now, and the earlier failures came 3 to 4 flights after today's value. If the IDG fails at LHR the return may not dispatch, LHR holds no spare, and N836CV's version of that cost 31 hours on the ground.
“I would not wait.”
Rosa chooses A1 · Trend confirmed, work fits the plan, and signs as maintenance controller. Accepting releases the planner agent's draft WO-26-81244 to the DEN line and tells the line maintenance manager. Until that moment nothing was booked, released or moved.
“I accept this task as Rosa Delgado, Maintenance Controller — e-signature.”
On the station plan N841CV lands at 22:05 as CV 1417 from SFO, and the IDG change sits inside its ground time beside the night's A-check and daily checks. WO-26-81244 carries the ground time, Hangar 2 free 23:00–07:00, two technicians qualified on the 777 plus an inspector, IDG S/N CV-5102 in DEN stores and the IDG sling in the tool crib. The parts agent proposes moving one IDG from ORD to DEN to restore the hub's stock.
The parts agent reserved nothing: it never orders or moves a part.
In 30 days Rosa and Kwame dismissed FWD CARGO DOOR at the gate 63 times as a known harmless message. The feedback agent drafted R15 — door messages within 3 minutes of a cargo-door cycle on the ground, with no message in flight — and replayed it: 1,412 messages it would have set aside in 90 days, and 0 of 19 confirmed failures hidden. The reliability engineer approves; the Director countersigns.
“No rule changes itself.”
The shop's teardown of IDG S/N CV-4418 finds generator bearing wear and a 60% blocked scavenge filter. The outcome agent records the prediction as confirmed and writes the avoided AOG record: 31 hours on the ground at LHR and a canceled return avoided, an estimated $412,000, with its working from Corvane's own N836CV event.
No avoided AOG is claimed without a finding; dollars are always marked estimated.
The same night of health data, seen by the five people who carry it — what their morning looked like, and what it looks like now.
Eight agents condense the night's warnings, argue against every trend, fit the work into the plan and learn from dismissals; maintenance control decides.
Reads every fault and condition message from datalink, post-flight downloads, the engine and APU condition feeds and the tech log, and folds repeats and duplicates — same fault, same tail, same flight phase.
Sets aside known harmless messages by the airline's signed rules and links messages to the open MEL deferral or work order that already owns them.
Groups the signals into patterns, computes each trend against the fleet, and finds the earlier failures it resembles in the airline's own removal and teardown history.
Tries to explain each pattern away — a sensor, the season, recent maintenance, a software change — on its own data, and can block a task.
Finds where the aircraft will be, the ground time, the hangar, the qualified technicians and the next flight that matters, and drafts the work order.
Finds the spare and the tooling at the station, and proposes transfers that keep the stations stocked.
Reads every dismissal and its reason code, finds the patterns, drafts rules and replays them on the last 90 days before anyone approves them.
Reads the shop's teardown report for every predicted removal, records confirmed or no fault found, and writes the avoided AOG record and the predictions and outcomes report.
Decides every task; the reliability engineer reviews below 0.80 and approves rules; the Director countersigns. The agents propose; a named person decides.
Every screen carries the same conversation: ask why, ask what if, or tell the agents to replan. Answers cite their source, and nothing is released, booked or moved until a person says so.
Why is N841CV first?
Its left IDG's oil outlet temperature rise climbed from 24 °C to 33 °C over 11 flights, while the right IDG stayed normal. Two earlier IDGs failed on the same curve: N836CV disconnected over the Atlantic and sat 31 hours at LHR; N847CV sat 19 hours at ORD. Its second flight from now is CV 908 to LHR, an ETOPS flight it may not fly with an IDG out. It lands at DEN at 22:05 with a spare in stores.
Find a slot for T2 at DEN instead
N722CV overnights at DEN on Fri Oct 16, 22:40–05:50; Hangar 4 bay 2 is free 23:00–03:00 and two A321neo technicians are on shift. The valve has to come from ORD on CV 553 at 14:10. That is a day sooner than the ORD slot on Saturday night and keeps the job at the hub.
Which aircraft are at risk of an AOG this week?
Two, if nothing is done: N841CV (left IDG, an ETOPS departure to LHR on Friday) and N877CV (outflow valve; a second motor fault would ground it at IAH before its GRU flight). Both have a slot proposed.
How good were your predictions last quarter?
Of 22 predicted removals torn down since Jul 17, 19 were confirmed and 3 were no fault found — 86%. That is an estimated 214 AOG hours avoided; every number links to its teardown report.
13 screens from the working solution, on its sample data. Pick one to see it large.
The thresholds, rules, reason codes and clocks every night is run against are settings, changed only with a reason and a countersign.
| Setting | Default | Choose from |
|---|---|---|
| Propose to maintenance control at or above | 0.80 confidence | Set by the reliability engineer, countersigned by the Director |
| Below this, the watch list | 0.55 confidence | Between the two, an engineer reviews first |
| Count a prior failure as similar at or above | 0.75 | Any change is replayed on 90 days first |
| Harmless-message rules | R1–R14, each versioned | Each rule on or off; new rules proposed from dismissals |
| Reason codes | Accept A1–A3 · Defer D1–D3 · Dismiss X1–X4 | Plus engineer E1–E2 and rule L1–L2 |
| Decide a task for tonight | 2 hours before the aircraft lands | Then escalated to the Duty Manager, Maintenance Control |
| Engineer review | within 4 hours | Then escalated to the reliability lead |
| Morning briefing | 06:30 every day | In-app and email, to maintenance controllers and health monitoring |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample airline data (the $412,000 is the solution's own estimate: 31 AOG hours at $9,800 an hour plus $108,200 for the canceled return and re-accommodation, from the sample airline's earlier event) · “target” = the design goal, measured in the live solution · “estimated” = our estimate (the split of a night's review hours) · “typical” = published figures (an airframer's published estimate of $10,000 to $150,000 an hour for an aircraft on ground; AviationPros, 2019 — 700–1,000 MB of data per flight on newer types). Regulations named as context: 14 CFR 121.373 (CASS), 121.374 (ETOPS), 121.380 (records), FAA AC 120-17B (reliability programs). Corvane Airways, Halvard Air and the people, tails, flights, work orders and serial numbers named on this page are fictional — characters and sample data in the working solution.
Aircraft predictive maintenance, sometimes called prognostics and health management, turns aircraft health-data warnings into planned work for maintenance control and engineering. Eight AI agents read the night's ACARS fault messages, condition-monitoring downloads and engine and APU trends, account for every warning, compare each trend with the fleet and with past failures, argue against it, and fit the work into a ground time with the part and the tooling. The maintenance controller asks the agents questions in plain words, sees the cited evidence, and accepts, defers or dismisses each task with a reason code and an e-signature.
AOG prevention starts with catching the slow trend — an IDG temperature creeping up over eleven flights, an EGT margin eroding faster than the fleet — and turns it into a task that arrives planned: the station, the ground time, the hangar, qualified technicians, the spare and the tooling, before the next flight that matters. In the demo a left IDG is changed at the hub before a 9 h 20 min ETOPS flight instead of failing at an outstation.
The intake agent folds repeats of the same fault, tail and flight phase, and the triage agent sets aside known harmless messages by the airline's signed rules and links others to the open deferral or work order that owns them. Every message ends with an outcome and its reason, and any one can be pulled back. On the sample night 1,240 warnings became 7 proposed tasks.
The outcome agent reads the shop's teardown report for every predicted removal and records it as confirmed or no fault found. On the sample airline's last 90 days, 19 of 22 removals torn down were confirmed — 86% precision — broken down by the confidence each was proposed at, so the thresholds are tuned on the airline's own outcomes.
It feeds them. Every prediction, decision, rule change and outcome is on a hash-chained audit trail, an evidence pack per task collects the sources, and a quarterly predictions and outcomes report goes to the Director to sign. It supports a reliability program run along the lines of FAA AC 120-17B and CASS under 14 CFR 121.373, and maintenance records (14 CFR 121.380) stay in your M&E system.
Yes. The trends screen shows EGT margin and loss per 100 cycles for every engine and start EGT for every APU against the fleet band, corrected to standard day, and ETOPS engines also carry the engine condition and oil consumption monitoring of the ETOPS maintenance program (14 CFR 121.374). An outlier becomes a proposed task, such as a borescope, like any other trend.
Yes. The agents propose and draft; they never book a slot, release a work order, order or move a part or change a rule. The maintenance controller decides every task with a reason code and an e-signature, the reliability engineer reviews tasks below 0.80 confidence and approves rules, the line maintenance manager approves part transfers, and the Director countersigns.
The Agentic Solution Engine builds and deploys this aircraft predictive maintenance software from your requirements and documents — your reliability program, MEL, harmless-message rules and a history of past removals and teardowns — and it goes live once every quality gate has passed. Before going live it replays a past night end to end, and we will walk you through it on your own nights first.
No tail flies past an AD, and every line of the status report is proven.
1-min demoService Bulletin EvaluationEvery bulletin decided in days on the fleet’s own evidence, every tail and dollar cited.
1-min demoDeferred Defect ManagementNo aircraft flies past its MEL limit, and every fix lands on a planned overnight.
1-min demoWe'll run Aircraft Predictive Maintenance on a replay of your own past nights of health data, deferrals and removals.