Gains:
- Ability to distinguish trend, anomaly and prognostic (RUL) analysis and use artificial intelligence as a pattern marker in sensor data
- Ability to distinguish with physical confirmation whether an anomaly is caused by the measurement chain (sensor/cable/calibration) or component
- Ability to understand that prognostic prediction is not a definitive date and that only the approved manufacturer's program determines maintenance intervals.
Aircraft maintenance oscillates between three scheduling philosophies. Corrective maintenance repairs a fault after it occurs. Preventive/periodic maintenance replaces parts at a certain flight hour or cycle. Predictive maintenance - monitoring the actual condition of the component and intervening "just in time" before it fails - is the most modern approach and is the area where artificial intelligence (AI) adds the most value. In this unit, we will cover trend reading from sensor data, prognostics and its limits.
Where does data come from?
Modern aircraft constantly produce data. Major sources:
- ACARS/QAR/FDR (Aircraft Communications Addressing and Reporting System / Quick Access Recorder / Flight Data Recorder — systems that record/transmit flight parameters): engine parameters, system states, flight data.
- HUMS (Health and Usage Monitoring System monitors vibration and gearbox health, especially in helicopters): vibration, temperature, cycle numbers.
- Engine trend monitoring: Flight-to-flight monitoring of parameters such as EGT, N1/N2 (engine rotor speeds), fuel flow, oil pressure/temperature.
- CMS/CMC maintenance messages and BITE records.
This data is voluminous and contains slow trends (the insidious climb of a value over flights) that the human eye cannot detect. This is where AI and statistical models come into play.
Trend, anomaly and prognostics: three separate jobs
There are three concepts that should not be confused:
- Trend analysis: The trend of a parameter over time. For example, EGT margin loss is a classic indicator of engine performance degradation.
- Anomaly detection: Sudden/statistical deviation from expected behavior. AI models learn the “normal” pattern and flag the outlier.
- Prognostic (RUL): Remaining Useful Life — estimate of remaining useful life. "This component reaches its limit after approximately X cycles." This is the most ambitious and the most ambiguous.
AI helps with all three, but uncertainty increases with all three: trend is the most reliable, prognostic is the most speculative. A RUL estimate is never an exact date; It is a probability range and is evaluated within the framework of engineer's interpretation and manufacturer limits (hard time, on-condition maintenance programs).
Caution: A prognostic estimate is a good trigger to “look sooner”; "Remove before the limit" or "extend the limit" cannot be justified. Maintenance intervals are determined only by the approved maintenance program (MPD/MRB) and manufacturer limits. AI prediction doesn't change these, it just draws attention.
Two error types: missed fault and false alarm
Predictive systems have two-way risk. False negative (missed failure): model does not flag a problem, component fails unexpectedly — safety risk. False positive (false alarm): flags a non-pattern problem, unnecessary disassembly, unnecessary cost and — importantly — risk of introducing new faults during unnecessary disassembly (maintenance-induced fault). A good program balances the two; It is wrong to blindly replace parts at every alarm and ignore every alarm. AI marks the pattern; The engineer determines the threshold and intervention.
Tip: Start an alarm with "why now?" ask. Is it a real trend or sensor drift/data error? Part of the anomaly is not in the component, but in the measurement chain (sensor, cable, calibration). Don't say "malfunction" without physical confirmation.
three mini cases
Case 1 — EGT margin warned the trend early. One engine's EGT margin has slowly decreased by 6°C over the last 60 flights. It was not noticeable when looking at individual flights; The AI trend chart made the trend clear. The engineer scheduled a boroscopic inspection according to the manufacturer's trend guide and found premature wear on the turbine blade. Planned intervention was provided instead of unexpected failure; gain: preventing an AOG (Aircraft on Ground) incident.
Case 2 — False alarm hunted. HUMS triggered a vibration alarm for a gearbox. Rather than immediately disassembling, the team first checked the sensor mount and cable: a vibration sensor's connection had become loose. The signal improved, the alarm went off. An unnecessary gearbox dismantling (high cost + risk of new errors) is avoided.
Case 3 — Prognostic misused, corrected. A planner wanted to postpone the replacement of a component based on AI's estimate of "RUL 300 cycles". The senior engineer reminded: the component was at the hard time limit, and the limit was fixed in the approved program. Prognostic prediction was a recommendation, hard time was a necessity. The change was made on schedule; A possible compliance violation has been prevented.
Four copyable templates
Role: Engine/parameter trend reading assistant.Task: Describe whether there is any trend or deviation in the [parameter] values of the last [N] flights below.Rules:- Just describe the pattern; DO NOT diagnose the fault.- If there is a significant climb/decline, indicate in which flight range it is.- Remind that the evaluation is a pre-qualification and the limits are defined in the manufacturer's program. Data: [CSV/paste table]
Role: Anomaly prioritization assistant. Task: Group the following alarm/anomaly list into "requires physical confirmation first" and "possible sensor/data error". Rules: Suggest the first thing to check for each (sensor/cable/calibration etccomponent); The final decision is with the engineer. List: [anomaly records]
Role: Sensor vs component separation advisor. Task: For the following anomaly, whether the measurement chain (sensor-cable-calibration) or the component is the primary suspect, list with reasons. Rules: Check the measurement chain before recommending part replacement. Anomaly: [description + data]
Role: Trend report draft assistant.Task: Write a short draft report to be presented to the engineer from the following trend finding.Rules:- Finding | possible meaning (cautious) | recommended check | use relevant manufacturer's guide headings. - Exact RUL/dating; Add "with manufacturer's limit and engineer approval". Finding: [trend summary]
Weak prompt / Strong prompt
Weak: "Will this engine fail? How many flight lives is left?"
It forces the AI to deliver a definitive prophecy; The "remaining life" number produced is unfounded and dangerous.
Strong: "Below are the EGT margin and N2 values for the last 60 flights. Just describe the trend and deviation; indicate in which flight interval it changes; give fault diagnosis and definitive remaining life; state that this is a pre-qualification, the limit is defined in the manufacturer's trend guide and the decision is up to the engineer."
This prompt positions the uncertainty correctly and leaves the decision to the human.
Table: Three analysis types and confidence levels
Analysis
what does it say
trust
Appropriate role of AI
trend
trend direction
high
Making slow change visible
Anomaly
sudden deviation
medium
Marking the outlier, prioritizing
Prognostic (RUL)
Remaining life estimate
Low/unclear
Trigger "Look sooner"
decision
Remove/replace/defer
—
Human + approved program
Common mistakes
- Mistaking prognostication for definitive date. RUL is a probability range, it does not override hard time.
- Changing parts at every alarm. False alarms generate new errors due to disassembly.
- Mistaking a sensor/data error for a component failure. Eliminate the measuring chain first.
- Looking at a single flight and missing the trend. Slow trends appear only in the time series.
- Trying to change the maintenance interval with AI. Only the approved program determines the intervals.
In summary
Predictive maintenance is the art of extracting meaningful signals from voluminous sensor data, and AI is a powerful assistant here: it makes slow trends visible, flags anomalies, draws attention to the right place. But the level of confidence decreases from trend to prognostic; RUL is not an exact date. With every alarm "sensor or component?" ask; Only the approved manufacturer's program determines maintenance intervals; The intervention decision and signature belong to the engineer.
Application task
Give your (anonymized) parameter series to the AI with the first template and get the trend recipe. Then, distinguish between "sensor or component" with the third template. Compare with the actual manufacturer trend guide: Is the trend marked by the AI real, is the threshold correct? Write a one-page assessment and answer the question "what physical check is required?"
checklist
- [ ] I considered trend, anomaly and prognostics as separate concepts.
- [ ] I asked the AI for a pattern description, not a diagnosis.
- [ ] For each anomaly, I first eliminated the measurement chain (sensor/cable/calibration).
- [ ] I did not use the RUL estimate as an exact date.
- [ ] I determined the maintenance interval according to the approved manufacturer's program.
- [ ] I left the intervention decision and signature with the engineer.