At a modern airline operations control centre, predictive maintenance systems now aspire to flag early warnings based on real-time engine health data. When a forecast indicates that an engine is likely to exceed performance limits within a defined number of days, maintenance planners can proactively adjust flight schedules, arrange for spare parts, or schedule shop visits in advance. What once triggered emergency responses or unscheduled groundings is now managed through data-driven foresight. This transformation reflects how predictive AI modelling has turned potential disruptions into planned maintenance activities, improving operational reliability and reducing cost without compromising safety.
By Sanjay Deshmukh,COO, Findability Sciences
Before AI: maintenance by pattern, gut, and schedule
Fifteen years ago, an engine showing subtle changes in exhaust gas temperature margins or vibration signatures would generate a paper trail: a technician’s note on a line maintenance card, a call into TechOps, and a conservative defensive response — ground the aircraft if in doubt, or ferry it to a maintenance base the next time it could be scheduled. Preventive maintenance relied heavily on flight-hour or cycle-based schedules, inspections driven by OEM service bulletins, and human judgment informed by experience. This approach protected safety but was inefficient: unnecessary shop visits, emergency AOG events, and expensive part shipments were common.
The limitations were structural: data lived in silos, sensor streams were high-volume but noisy, and no operational system translated complex telemetry into a simple operational decision such as “this engine needs a shop visit within X days.” That translation — days to failure — is what airlines now prize, because calendar days map directly to planning windows for crews, aircraft rotations and spare logistics.
The shift: combining data, twins and AI
The first practical step was not a model but instrumentation and integration: richer telemetry (temperature, pressure, spectral vibration), consistent logging of flight context (thrust settings, profiles), and a link to maintenance history and environmental feeds. OEMs and MROs invested in digital twins — virtual replicas of engines that run physics-informed simulations alongside live telemetry — so that patterns in sensor drift could be interpreted against expected physical behavior rather than treated as raw anomalies. Rolls-Royce and others have championed digital-twin programs that let engineers explore “what if” scenarios without turning a wrench.
Once the data plumbing was in place, airlines piloted models that forecast remaining useful life (RUL) and — crucially for operations — converted that into days to failure: calendarized RUL that factors in flight schedules and utilization. Early models were statistical or tree-based and offered explainability at the cost of flexibility. Today’s state-of-the-art prognostics blends physics constraints with deep sequence architectures — combinations of CNN-LSTM with attention or transformer-like encoders — that can capture long-range temporal patterns in noisy telemetry and fuse them with maintenance logs and environmental modes. Academic work in 2024–2025 has demonstrated improvements in RUL accuracy using CNN–LSTM–Attention hybrids and multiscale transfer learning tailored for aeroengines.
A concrete example: Air France–KLM’s operational experiment
Air France–KLM’s recent move to accelerate AI across its operations highlights how a large carrier approaches adoption. In late 2024 the group announced partnerships to deploy generative and predictive AI across its data estate — a strategic step that explicitly names predictive maintenance as a target area. For a carrier that operates hundreds of aircraft across global networks, moving from hours-long analyses to near-real-time probabilistic forecasting materially changes the cadence of maintenance decisions.
In practice, the airline’s prototype architecture mirrored the emerging industry pattern: an edge tier on aircraft performs low-latency anomaly detection and compresses high-frequency data; a cloud tier runs ensemble prognostics and digital-twin simulations; and a frontline dashboard surfaces a clear days-to-failure forecast with confidence bands and top contributing factors (e.g., rising core temperature during climb cycles, elevated particle ingestion on specific sectors). The OCC used this information to make graded decisions — limit dispatch to short sectors, pre-order parts for the next planned shop visit, or approve a non-urgent ferry to a maintenance base — turning uncertain emergencies into scheduled, auditable actions.
Best practices that made it work
Successful deployments share practical patterns:
• Align forecasts to decisions. Build the model output to answer operational questions: if the decision is “can this aircraft depart tomorrow?”, then present probability of failure within 24–72 hours, not a raw RUL number.
• Quantify uncertainty. Probability distributions, prediction intervals, and cost-weighted decision thresholds prevent overreaction to a single point estimate. Operators need to know both the most likely days-to-failure and the worst-case window.
• Human-in-the-loop workflows. Use the AI as an advisor. Maintenance planners validate and override forecasts early in rollout; their feedback then becomes training signals for model recalibration.
• Hybrid models for explainability. Blend physics-informed constraints with deep ML so that failure drivers link to physically meaningful phenomena — vital for certification and trust.
• Data governance and interoperability. Standardized data exchange (OEM, MRO, airline) and careful privacy controls make federated improvements possible without sharing raw operational data. Air France–KLM’s partnerships reflect this need to keep control of data while leveraging cloud AI capabilities.
Challenges that persist
Not all is solved. Data quality and left-censoring (engines observed only after installation) complicate life-history modeling; true failures are rare, so training sets are imbalanced; and operations evolve — flight profiles, new routings, or maintenance practices create covariate drift that must be detected and handled. Moreover, the certification and safety assurance of AI-infused prognostics remains a conservative, resource-intensive step: regulators and safety managers require rigorous demonstrations of conservatism under uncertainty and robust behavior in edge cases. Academic and industry studies stress the need for continuous validation pipelines and scenario stress tests to address these issues.
Benefits realized — and measured
When the system works, benefits cascade. Airlines report fewer aircraft on ground events, better shop utilization, and lower logistics costs from fewer urgent part shipments. Delta, Lufthansa and others have publicly discussed multi-year programs to digitize TechOps and apply ML for predictive maintenance; case studies point to measurable reductions in unscheduled removals and time-to-repair when analytics are tied to operations. For Air France–KLM, the value was not just in avoided costs but in operational confidence: planners could make marketplace commitments (e.g., guarantee service levels) with auditable risk margins.
How it changed daily work
The shift is cultural as well as technical. Line technicians still inspect physical evidence, but they now receive actionable predictions and likely causes before their walkaround. Planners orchestrate shop visits with longer lead time. Supply-chain teams stage LRUs based on probabilistic demand rather than reactive rush orders. And engineers iterate models as maintenance feedback flows back, slowly knitting a single operational truth from once-disparate systems.
Looking forward: practical recommendations
For airlines and MROs ready to scale days-to-failure forecasting, a pragmatic roadmap is clear: start with pilot fleets and clearly defined decisions; invest in data quality, feature engineering, and explainability; deploy two-tier inference (edge anomaly detection + cloud prognostics); and build continuous validation and drift detection. Federated learning across operators, larger digital-twin deployments, and standardization of uncertainty metrics will accelerate progress while preserving commercial data control. Academic advances in attention-based sequence models and multiscale transfer learning will continue to raise forecasting accuracy — but the operational win depends on integrating those models into decision workflows.
(Views expressed are personal. The writer has over 30 years of industry experience including 8 years of experience of overseeing implementation predictive AI models at leading global airlines like ANA and Jetblue.)
