Every time a ground crew discovers a fault at the gate — hydraulic pressure reading wrong, engine vibration outside spec — a clock starts ticking. The aircraft goes on ground (AOG). Gate neighbors shuffle. Passengers get rebooked. The airline pays anywhere between $10,000 and $150,000 per hour in lost revenue, crew repositioning, and emergency logistics, according to IATA benchmarks cited in OxMaint’s March 2026 MRO analysis.
Over 60% of those AOG events involve failures that predictive AI systems can detect 15 to 30 days in advance. The technology to prevent most of them already exists. The gap is in how — and whether — it gets deployed.
The aviation IoT market grew to $11.03 billion in 2026, up from $9.13B the year before, a 20.8% jump driven largely by predictive maintenance deployments and real-time aircraft monitoring. That growth is happening in actual systems, not in proof-of-concepts.
The problem with maintenance by calendar
Aviation has operated on one core logic for decades: replace components on a fixed schedule based on flight hours or calendar dates. Time-based maintenance is safe in the sense that it prevents operating parts past their design life. It’s wasteful in ways the industry has quietly accepted as normal.
30 to 40% of components replaced under fixed-interval schedules still have useful life left when they come out. Meanwhile, a meaningful share of unplanned failures happen between scheduled checks — components that degraded faster than the interval assumed. The interval was designed for the average case, and the average case is not the problem.
The global MRO market stands at $85 billion, with roughly 40% absorbed by reactive, unplanned repairs. The emergency repair premium alone is 4.8 times the cost of a planned maintenance event, according to ATA MSG-3 industry cost analysis.
Condition-based maintenance asks a different question. Not “when did we last replace this?” but “what does the sensor data say about this component right now?” The first question is answerable with a spreadsheet. The second one requires embedded sensors and real-time analytics.
What the sensor layer actually does
A modern commercial aircraft generates over 1TB of sensor data per flight. Engines, APUs, landing gear, hydraulics, avionics — every major system can carry sensors tracking vibration frequency, temperature profiles, pressure, and operating hours.
Raw telemetry by itself is noise. The value comes from ML models trained on OEM baseline profiles and historical failure databases, detecting micro-anomalies weeks before they show up on a flight deck indicator. Boeing’s AnalytX fleet data puts the average advance warning at 21 days — long enough to schedule a repair at base, order parts at standard rates, and send a crew that arrives knowing exactly what they’re fixing, rather than diagnosing a mystery failure at 2 AM at an outstations airport.
Cloud-dependent architectures have a real problem in this environment, though. Satellite connectivity at cruising altitude is still intermittent. More critically, some monitoring tasks need sub-millisecond response times that a cloud roundtrip cannot deliver. Edge AI becomes less a preference and more a hard requirement.
What edge AI means on an aircraft
Edge AI means inference runs on the device, not in a remote data center. A neural processing unit (NPU) embedded in the system-on-chip handles anomaly detection locally, flags deviations in real time, and sends only relevant event data upstream — the sensor feed does not need to stream continuously to the cloud.
The hardware ecosystem has caught up. New SoCs from NXP, MediaTek, and STM32 now ship with dedicated NPU cores built for edge inference workloads. Zephyr RTOS has been gaining real traction for secure, low-power connected devices — Embedded World 2026 made clear it has become the default choice for safety-aware embedded development. For aviation environments where power budgets are constrained and failure modes matter, hardware choices compound fast.
The architecture that works in production divides responsibility deliberately: sensors capture telemetry at the component level, on-device inference runs anomaly detection and flags deviations from OEM baselines, an edge gateway aggregates events across systems, and the cloud handles model retraining and fleet-wide analytics. Edge handles speed and resilience. Cloud handles learning and scale. The boundary between them is a design decision, not a default setting.
The certification problem nobody explains well
DO-178C governs software in airborne systems. DO-254 covers hardware. Neither was written with adaptive machine learning in mind. EASA has been developing AI-specific guidance — its framework for Level 1 and Level 2 AI systems was on track for finalization through 2026 — but the certification process for any specific product still involves significant documentation overhead, traceability evidence, and time.
The practical path through this is scope definition. Predictive maintenance systems that sit in the monitoring and analytics layer — anomaly detection, alert generation, automated work order creation — rather than directly in flight-critical control paths face a much cleaner regulatory situation. They operate outside the certified avionics environment. Getting the boundary wrong early adds years to deployment timelines.
Teams that scope this correctly see real numbers. Operators running mature AI predictive programs are reporting 35% fewer unscheduled AOG events within 12 months and 18-25% lower total MRO costs compared to time-based preventive maintenance. Those results make the case internally for expanding scope over time, eventually including deeper integration with certified systems.
When embedded AI moves into the cockpit
The predictive maintenance conversation mostly stays in the MRO world: ground teams, data platforms, work order automation. But there’s a parallel track where embedded AI is taking on tasks directly alongside pilots.
The ReadU6 project is a useful illustration of what this engineering actually looks like. Built with a UK aviation client, it’s an AI-powered communication device that processes ATC instructions in real time, filters cockpit noise, and surfaces structured command text for pilots — all running on a custom embedded system built on a high-performance single-board computer, physically designed for cockpit constraints: compact, low-power, anti-glare.
The AI side ran NLP-based speech recognition and translation models trained over eight months on real ATC communication data. Getting reliable performance under noise, strict latency requirements, and safety constraints meant the hardware-software boundary was the actual engineering problem. Not the model architecture.
That dynamic is increasingly true in predictive maintenance too. The intelligence moving closer to the aircraft changes what the work is: less about which algorithm to run, more about what runs where, how it fails safely, and what the certification scope covers.
Where deployments actually stall
Sensor coverage gaps are the most common silent failure mode. A model trained on partial sensor data produces partial predictions, and partial predictions create a false sense of coverage that is arguably worse than having no predictive system at all. Mapping the actual instrumented surface area before claiming predictive capability is unglamorous work that most teams skip.
The edge-cloud boundary also needs to be an explicit architecture decision. What runs on-device, what runs on the ground gateway, what goes to the cloud — this needs to be decided deliberately, because changing it mid-project means redesigning data contracts and revalidating the inference pipeline.
And maintenance workflow integration is where ROI actually materializes. A predictive alert that generates a push notification but not a work order usually gets ignored by the time the ground team checks their inbox. Closing the full loop from anomaly detection to parts pre-order to crew assignment is where time-to-repair cuts of up to 40% show up. The analytics layer is the easy part. The workflow integration is where the money is.
If you’re mapping where to start, Allmatics works with aerospace and aviation teams on embedded IoT development and AI system integration — from product discovery through deployment. The scoping process is usually where the clearest picture of what to build first, and what to leave for later, emerges.