Nadir AI transforms disposable vibration test data into a permanently learning, fleet-wide health intelligence system — ontology-grounded, privacy-preserving, and aligned with FAA and EASA safety requirements.
Current practice processes vibration data only to produce a binary verdict. Once the test concludes, raw signals are discarded. No historical record. No fleet model. No prediction. Nadir AI was built to eliminate this.
The Nadir AI framework is defined as a seven-layer reference architecture. Each layer performs a specific transformation on the input data, preserving full provenance and ontology grounding. The output of each layer is the input to the next — and every transformation is queryable, auditable, and reversible.
Follow a single bearing velocity reading from an ACES Systems test cell through Nadir AI's seven-layer pipeline. What enters as unstructured sensor output exits as a semantically governed, ontology-enriched record ready to contribute to fleet-wide learning.
Vibration signatures vary with load, speed, and flight phase. A bearing reading during cruise means something very different than the same reading on takeoff. L0 captures both the sensor stream and the full operational context that makes downstream analysis possible — assigning a unique identifier to every record and storing the raw data in an append-only buffer.
Signal capture. Raw vibration data is ingested from ACES Systems hardware or equivalent acquisition devices. Channel configuration, sampling rate (typically 10–50 kHz), and calibration factors are recorded per-channel.
Context metadata. Every record is tagged with timestamp, aircraft tail number, sensor location (station coordinates + mount type), operating regime (ground idle, taxi, takeoff, cruise, descent, landing), flight phase, environmental conditions, and recent maintenance history.
Provenance anchoring. A unique record identifier is assigned at capture time and carried through every downstream transformation. This creates the audit chain that makes L5 governance possible — any health decision can be traced back to the originating signal.
Why it matters. Including operating context at capture time is what allows the framework to normalize vibration signatures across flight conditions and to build models that are aware of the envelope in which a measurement was taken.
Traditional FFT-based vibration analysis assumes stationarity — an assumption that breaks down under transient loads, varying rotor speeds, and real-world operational noise. Nadir AI applies Ensemble Empirical Mode Decomposition (EEMD) to adaptively decompose each vibration signal into a set of intrinsic mode functions (IMFs), each representing oscillations at a different time scale. This allows fault signatures to be isolated to specific frequency bands rather than being masked in the composite signal.
Why EEMD over EMD. Standard EMD suffers from mode mixing — the same physical phenomenon can appear in multiple IMFs across different signals, making comparison unreliable. EEMD injects controlled white noise across multiple trials and averages the resulting decompositions, producing IMFs that are stable, reproducible, and physically meaningful.
Per-IMF feature extraction. From each IMF we extract statistical features (RMS energy, kurtosis, crest factor) and spectral features (peak frequency, bandwidth, spectral centroid). Each feature is stored as a first-class, typed entity in the ontology — which means downstream analytics can query the decomposed components directly.
Reproducibility. The decomposition parameters (sifting iterations, noise amplitude, ensemble trials) are recorded alongside the features, so any result can be independently re-derived from the archived raw buffer.
Operational impact. Bearing defects, gear mesh anomalies, and rotor imbalance each have characteristic signatures in different IMF bands. By making IMFs queryable, maintenance engineers can drill from "something is wrong" down to "the second-stage gear mesh is drifting" without a separate analytical pass.
| Component | RMS | Kurtosis | Peak f |
|---|---|---|---|
| IMF 1 (high) | 0.312 | 3.8 | 847 Hz |
| IMF 2 (mid) | 0.481 | 4.2 | 241 Hz |
| IMF 3 (mesh) | 0.629 | 6.9 | 52 Hz |
| Residual | 0.094 | 2.1 | — |
This is where Nadir AI's interoperability layer lives. Every feature extracted in L1 — plus every metadata tag from L0 — is mapped into a JSON-LD record structured across six orthogonal ontology dimensions. The result is a canonical health record that is queryable, comparable across aircraft and operators, and ready for federated learning.
Each canonical record links directly to established aviation and condition-monitoring standards. This is not a mapping layer bolted on top — it is embedded at the JSON-LD @context level so that every field carries its standard identity as it moves through the fleet.
Severity zones A–D, alarm thresholds, and asset class codes applied at the field level in every canonical record. Supersedes ISO 10816.
Every classified field receives an ATA chapter code (e.g. 72-00-00 for power plant). Aircraft system hierarchy preserved throughout the ontology and downstream outputs.
Data module codes referenced for any maintenance action emitted by L6. Enables direct linking of diagnostic output to task cards and work packages.
Six-layer information flow (data acquisition through advisory generation) forms the backbone reference model. Nadir AI extends this with federated learning and governance layers.
Maintenance trust requires calibrated uncertainty. A 90% confidence prediction that's actually right 60% of the time will erode maintenance engineers' trust in the system faster than any false positive. Nadir AI uses XGBoost ensembles with temperature scaling on a held-out calibration set — producing probabilities whose expected calibration error (ECE) meets safety-grade thresholds.
Ensemble architecture. For each asset class (bearing, gearbox stage, shaft), Nadir AI trains a dedicated ensemble of XGBoost classifiers for fault typing and XGBoost regressors for remaining useful life (RUL) estimation. Models consume the full ontology-mapped feature set from L2.
Temperature scaling. Raw model outputs are passed through a single-parameter temperature scaling layer fit on a validation partition. This compresses or stretches the softmax distribution until predicted confidence matches observed accuracy — without changing the ranking of predictions. ECE drops from 0.052 to 0.034 across the full pipeline.
Uncertainty + explanation. Every prediction is packaged with (a) per-class probabilities, (b) expected calibration error for the prediction, (c) a trend score — the slope of the health indicator over the last N flights, and (d) an RUL estimate with confidence interval. An explanation module surfaces the most influential features so maintenance engineers see why a given recommendation was made.
Safety posture. In aviation, false alarms trigger unnecessary inspections that waste time and erode trust. Missed faults have safety consequences. Calibrated probabilities aren't a performance metric — they're a safety requirement.
Nadir AI doesn't just analyze individual aircraft. It builds a continuously evolving, privacy-preserving AI model of fleet health — learning from every test, every flight cycle, every maintenance outcome across every platform under contract.
Current vibration monitoring produces a pass/fail verdict and then discards the underlying data. Nadir AI transforms that same data into a permanent, comparable, longitudinally trackable health record — and feeds it into a global AI model that improves with every new data point across the fleet.
The result: bearing degradation curves built from real operational data. Vibration signature drift detected across flight cycles and platforms. Cross-platform anomaly detection that surfaces failure modes before any single operator would see them. Component life prediction grounded in actual usage patterns, not theoretical schedules.
Each operator contributes to the global model — and each operator benefits from everyone else's learning. The fleet becomes smarter with every aircraft, every test, every flight hour.
Aircraft lands. ACES device connects. Instead of a pass/fail that disappears, vibration data streams into Nadir AI's pipeline — enriched, normalized, and delivered to your cloud destination within seconds. Every test becomes a permanent, queryable fleet intelligence asset.
Nadir AI's edge-deployed classification normalizes and prioritizes high-value channels for in-flight transmission. Full semantic enrichment completes on ground receipt. Maintenance crews receive actionable intelligence before the aircraft touches down — anomalies caught in the air.
All aircraft. All platforms. All test cycles. Contributing to a single federated intelligence model that predicts failures, surfaces degradation patterns, and generates maintenance intelligence that no individual operator's data could produce alone. This is the game.
The most powerful capability Nadir AI delivers — and its most critical design constraint. Every operator contributes to the global fleet model without ever exposing raw vibration data. Formal differential privacy guarantees are built in at the mathematical level.
At each training round, each operator site computes per-feature histograms of gradients and Hessians from their local dataset — clipped to a maximum L2 norm as a sensitivity bound. The central server aggregates these histograms, adds calibrated Gaussian noise for differential privacy, and selects optimal model splits using the XGBoost gain formula.
No raw vibration records ever leave the operator's environment. The only data transmitted is statistical gradient summaries — which contain no individual data points and provide formally quantifiable privacy guarantees expressed as a privacy budget ε.
The result: each operator's local diagnostic model benefits from the collective learning of every aircraft in the federated fleet, while their sensitive operational data stays within their own infrastructure. Operators who would never share raw data with competitors can still contribute to — and benefit from — shared fleet intelligence.
Validated: Federated learning achieves 95% of the performance of a fully centralized model while reducing communication bandwidth 73.3%. Privacy budget ε = 0.68 at δ = 10⁻⁵ with less than 2% utility loss versus the non-private model, accounted for using the moments accountant.
Aviation demands more than AI that works — it demands AI that is traceable, controllable, and safe. Nadir AI's governance control plane runs cross-cutting across all pipeline layers, ensuring every model update is shadow-validated before deployment and unsafe changes automatically roll back.
Governed adaptation model, shadow validation, and approval workflows aligned with FAA operational authorization requirements. Every change to diagnostic logic is traceable and reversible.
EASA 2024 rulemaking on certifiable fidelity for VHM systems. Provenance embedding and approval gates meet continuing airworthiness integration requirements.
Directly addresses the five barriers NASA identified: data availability, validation, regulatory acceptance, cost, and decision value — through governed adaptation and provenance.
IATA's framework for transitioning from reactive to predictive maintenance. Nadir AI operationalizes the "actionable, shareable intelligence" principle at the architecture level.
The final layer turns diagnostic output into maintenance action. For each aircraft under monitoring, Nadir AI produces a prioritized component list, recommended inspection type, suggested interval, confidence level, and a concise explanation — packaged for ingestion into the operator's maintenance planning system. The loop closes when outcome feedback flows back into the learning pipeline.
From probability to plan. A calibrated fault probability of 0.72 on an inner race bearing with a degrading 7-day trend and an RUL of 186 flight hours doesn't just become an alert — it becomes a specific, scoped work recommendation: borescope inspection at next scheduled check, with defer authority up to N flight hours.
Structured output. Recommendations carry the inspection type (borescope, oil analysis, vibration re-test), decision class (defer, replace, monitor), work scope tasks, interval target, and a priority score. All fields are S1000D-compatible so they can be linked directly to task cards and work packages.
Audit trail. Every recommendation carries its full provenance chain — from raw signal through L1 IMFs through the exact model version that produced it. Ops authorities can trace any maintenance action back to the originating data point.
Closed-loop learning. When a recommendation is followed, the outcome (confirmed fault, no fault found, different fault discovered) feeds back into L3 local diagnosis and L4 fleet learning. The fleet becomes smarter with every dispatch.
Validated on NASA IMS bearing run-to-failure datasets, simulated 50-aircraft × 100-cycle fleets, and controlled fault injection. Five-fold cross-validation, chronological splits, paired t-tests at p < 0.01. Effect sizes above Cohen's d = 0.8 across all primary metrics. Full methodology in Dr. Reza Nehzati's peer-reviewed paper.
Full framework vs. 87.3% raw-features baseline. 7.4pp improvement, statistically significant at p < 0.01.
ECE reduced from 0.052 to 0.034. Predicted confidences now match observed accuracy — safety-grade.
500-file corpus across three systems with inconsistent schemas, units, and naming. Field mapping recall: 94%.
Federated histogram aggregation vs. transmitting raw data. 98% of centralized model performance retained.
Request a technical briefing with the Nadir AI team. We'll map your current ACES output against the full pipeline and show you exactly what fleet health intelligence looks like on the other side.