9 min

Predictive Maintenance for Aircraft

From Sensor Data to Remaining Useful Life Prediction

The Cost of Getting Maintenance Wrong

An unscheduled engine removal on a CFM56-7B (the engine powering most of IndiGo's Boeing 737 fleet) costs approximately $3-5 million and takes the aircraft out of service for 30-45 days. A scheduled removal during a planned C-check costs 40-60% less and causes zero unplanned downtime. The difference between these two outcomes is information — specifically, knowing when a component will fail before it actually does. AI-driven predictive maintenance transforms raw sensor data into Remaining Useful Life (RUL) estimates that make this possible.

Engine Health Monitoring Parameters

Modern turbofan engines generate thousands of data points per second. The parameters that matter most for health monitoring:

Parameter	Abbreviation	Normal Range (CFM56)	What Degradation Looks Like
Exhaust Gas Temperature	EGT	500-650°C cruise	Gradual rise: 1-2°C/100 cycles = normal wear. Sudden 15°C+ jump = investigate
N1 (Fan Speed)	N1	85-100% at takeoff	Increasing N1 needed for same thrust = compressor degradation
N2 (Core Speed)	N2	90-100% at takeoff	N2 creep at constant thrust indicates turbine section wear
Oil Temperature	OT	60-120°C	Sustained >140°C or rapid rise = bearing/seal issue
Oil Pressure	OP	40-100 PSI	Drop below 35 PSI = filter clog, pump wear, or leak
Vibration (broadband)	VIB	<1.0 IPS	>2.5 IPS on any axis = immediate borescope inspection
Fuel Flow	FF	Varies with thrust	Higher FF for same thrust = combustor/turbine efficiency loss
EGT Margin	EGTM	40-80°C at delivery	<15°C margin = engine approaching removal threshold

The EGT margin is the single most watched parameter. It represents the gap between the engine's actual peak EGT and the redline limit. As engine components degrade — blade tip erosion, seal wear, combustor liner cracking — the engine runs hotter to produce the same thrust. When EGT margin reaches zero, the engine must be removed regardless of other indicators.

Open data/engine-sensor-data.csv — it contains 18 months of cruise-phase engine parameter snapshots for a fleet of 12 aircraft, sampled every flight cycle. Look for the gradual EGT margin erosion and the anomalous vibration events.

Degradation Signatures

Not all failures look the same. AI models must distinguish between:

Gradual Wear (Predictable)

Bearing wear follows a bathtub curve — high infant mortality, long stable period, then accelerating degradation. The stable period can last 5000-15000 cycles. Key signatures:

Oil debris particle count increases linearly, then exponentially

Vibration spectrum shows growing amplitude at bearing characteristic frequencies (BPFO, BPFI, BSF)

Oil temperature delta (outlet minus inlet) creeps upward

Sudden Events (FOD, Bird Strike)

Foreign Object Damage causes step-change degradation:

Immediate EGT rise (5-25°C depending on blade damage)

Vibration spike — may settle to new elevated baseline or continue growing

N1/N2 imbalance if damage is asymmetric

Sometimes no cockpit indication — only detectable through trend monitoring

Intermittent Faults

The hardest to diagnose — problems that appear and disappear:

Fuel nozzle partial blockage (EGT spread between individual thermocouples varies)

Variable stator vane actuator sticking (transient N2 oscillation during accel/decel)

Bleed valve seal leakage (intermittent EGT rise only during certain operating conditions)

AI excels here because it can correlate patterns across multiple parameters and operating conditions that human analysts reviewing trend plots would miss.

Remaining Useful Life Estimation

RUL prediction is the core deliverable. Three approaches, ordered by maturity:

Physics-Based Models

Model the degradation physics (erosion rates, creep laws, fatigue crack growth) and extrapolate. Requires deep domain knowledge and accurate material property data. Works well for well-understood failure modes (e.g., turbine blade creep) but cannot handle novel degradation patterns.

Data-Driven Models

Train on historical sensor data labelled with actual failure/removal events:

Input: Last 50 flight cycles of [EGT, N1, N2, FF, VIB, OT, OP] normalized to takeoff conditions
Output: Cycles remaining until EGT margin reaches removal threshold

Architectures that work:

LSTM / GRU networks — capture temporal degradation trends, handle variable-length sequences

1D Convolutional networks — extract local degradation patterns, computationally cheaper

Transformer encoders — attention mechanism identifies which historical cycles are most informative for current prediction

Ensemble of the above — reduces prediction variance, provides uncertainty estimates

Hybrid Physics-Informed Models

The emerging best practice: embed physics constraints into neural network architectures. The network learns from data but is constrained to produce physically plausible degradation trajectories (monotonically decreasing RUL, energy conservation, etc.).

Prompt: "Given this engine's sensor history — 4200 cycles since last overhaul, current EGT margin
22°C, vibration trend +0.15 IPS over last 300 cycles, oil debris index rising — estimate
remaining useful life in cycles. Provide 10th, 50th, and 90th percentile estimates and
identify the most likely failure mode driving the prediction."

Fleet-Level vs Individual Engine Tracking

Individual engine tracking tells you when one engine needs attention. Fleet-level analysis reveals patterns that individual tracking cannot:

Batch effects — engines from the same manufacturing batch may share a defect. If engine #3 from batch 2019-Q3 fails early, check all other engines from that batch.

Operational pattern correlation — routes with frequent high-altitude takeoffs (Leh, Kullu, Bagdogra) stress engines differently than sea-level operations. AI can learn route-specific degradation rates.

Maintenance effectiveness tracking — did the last shop visit actually restore the engine to expected performance, or is it degrading faster than predicted? AI compares post-maintenance trajectories across the fleet.

Spare engine positioning — predicting which engines across the fleet will need removal in the next 60/90/120 days enables proactive spare positioning at the right maintenance base.

Maintenance Check Types

Check	Interval	Duration	Scope	Cost (Narrow-body)
A Check	500-800 FH	1-2 days	General inspection, lubrication, filter replacement	$50K-80K
B Check	6-8 months	1-3 days	More detailed inspection, component checks	$100K-150K
C Check	18-24 months	1-2 weeks	Structural inspection, system overhaul, heavy component replacement	$500K-1M
D Check	8-12 years	1-2 months	Complete strip-down, structural rebuild, repainting	$3-6M

AI optimizes the scheduling of these checks by:

Clustering component replacement needs to coincide with scheduled checks (avoid removing a panel twice)

Predicting which components will need replacement at the next C-check based on current sensor trends

Recommending check interval extensions when data shows the aircraft is in better-than-expected condition (MSG-3 philosophy)

Open data/fleet-maintenance-log.json for structured maintenance records including check types, findings, component replacements, and costs for a simulated airline fleet over 5 years.

Indian Context

HAL MRO Capabilities

Hindustan Aeronautics Limited operates India's largest MRO (Maintenance, Repair, Overhaul) facility in Bengaluru, servicing:

Tejas LCA engines (GE F404-IN20) — India's indigenous fighter programme generates unique engine health data. HAL is building AI monitoring capabilities specific to this engine's operating envelope (hot desert takeoffs from Jodhpur/Jaisalmer, high-altitude ops from Leh).

Su-30MKI engines (AL-31FP) — The IAF's workhorse fleet of 272 aircraft. These Russian engines have different degradation characteristics than Western turbofans, requiring India-specific RUL models trained on domestic operational data.

Civil MRO expansion — HAL is scaling up commercial MRO to capture India's $1.7 billion MRO market (projected to reach $4 billion by 2030 as fleet sizes grow).

Air India Engineering Services

Air India's in-house engineering arm maintains a mixed fleet (Boeing 777/787, Airbus A320neo/A350). Post-Tata acquisition, the focus is on modernizing maintenance practices — moving from interval-based maintenance to condition-based maintenance powered by AI analytics.

IndiGo's CFM56/LEAP Fleet

As India's largest airline (350+ aircraft), IndiGo operates one of the world's largest A320 fleets powered by CFM56-5B and LEAP-1A engines. The scale creates both a data advantage (massive dataset for AI training) and a logistical challenge (dozens of engine shop visits per year that must be optimized against network demand).

Open data/component-life-limits.json — it contains life-limited part data for turbofan engine components, including hard-time limits, on-condition intervals, and the parameters that drive replacement decisions.

Key Takeaways

EGT margin erosion is the primary RUL driver for turbofan engines — but single-parameter tracking misses intermittent faults and sudden events. Multi-parameter AI models catch what trend monitoring misses.

Hybrid physics-informed models outperform pure data-driven approaches — physics constraints prevent the model from predicting physically impossible degradation trajectories, improving reliability at the edges of the training distribution.

Fleet-level analysis unlocks value beyond individual engine tracking — batch effects, route-specific degradation, and maintenance effectiveness are only visible when you analyze across the fleet.

India's MRO market is ripe for AI disruption — HAL, Air India Engineering, and the growing Indian airline fleet create both the need and the data for sophisticated predictive maintenance systems.

This is chapter 3 of AI for Aerospace & Drones.

Get the full hands-on course — free during early access. Build the complete system. Your projects become your portfolio.

View course details

Ch. 2: Satellite & Remote Sensing AI

Ch. 4: Aerodynamic Simulation with AI