Command Center
Live Engine
Quality is derived from residual magnitude and cross-channel consistency, not a certified sensor built-in-test. A large residual on one channel inconsistent with all physically-linked channels is flagged as a likely sensor problem rather than an engine problem.
Diagnostics
Observed residual signals are supplied to the model as input, not SHAP values or proven causal attributions. They do not represent the complete feature set (60 features total; see Model Validation → feature importance).
Channels are ranked by normalized deviation from physics model expected values, scored by rule cross-checks. For the XGBoost model's own evidence, see the fusion card above.
Not a real accelerometer recording. A short synthetic waveform is generated each tick from current RPM and active fault severity, then a small discrete Fourier transform runs on it in-browser.
Reserved page for real classifier/RUL-model evaluation results. No trained model is connected in this build, so no metric is fabricated; every field below is a placeholder until a trained model is integrated (see integration points in Engineering).
Status: —
| Metric | Value |
|---|---|
| Precision | — |
| Recall | — |
| Macro F1 | — |
| Per-class F1 | — |
| Confusion matrix | — |
| False alarm rate | — |
The fault-injection engine holds ground truth. The diagnostic layer above only ever receives resulting telemetry (never the fault label). This panel proves that separation by revealing ground truth only on request.
Prognostics
*Prototype fuel-efficiency indicator: (RPM × throttle) ⁄ fuel flow. Not a calibrated brake-specific-fuel-consumption value.
Run a mission with a fault scenario to generate an advisory.
Mission
The fault-injection engine holds ground truth (component health, onset, severity). The diagnostic layer only ever receives the resulting telemetry — never the fault label.
Replay
Run a mission in the Mission screen to record telemetry, diagnosis, and events tick by tick for time-scrub playback.
Reports
Completed simulation runs from this browser session, kept in memory only — not persisted, not real fleet history.
| Mission | Final Health | Max Anomaly | Diagnosis | RUL | Risk |
|---|---|---|---|---|---|
| No completed missions yet this session. | |||||
Engineering
Live Module StateDynamic PipelineUpdates with the simulator
0 samples / 0 s (target 20/20s, true 1 Hz spacing)
Model: — · Features: — · Sampling: 1 Hz
The main tick advances simulated mission time per frame at higher speeds while stepping the pipeline once per second internally. It uses the shared noise stream so the XGBoost 20-second window exactly matches displayed telemetry, rather than an independently re-noised realization or coarse per-tick slice.
Module DetailTechnical Spec
The simulator knows the fault. Diagnostics does not.
- The fault-injection engine holds the ground-truth degradation state (component health, onset time, severity).
- The diagnostic layer receives only the resulting telemetry stream.
- Detection is a genuine inference from sensor patterns, not a lookup of an injected label.
- Faults degrade gradually over simulated hours rather than triggering instantaneously.
- Fault injection can be scheduled before a mission starts, or triggered deterministically mid-mission with Inject Now.
Recommended telemetry columns
- timestamp, engine_id, mission_id: sequencing and identity
- altitude_ft, ambient_temp_C, throttle_pct, mission_stage: operating condition
- rpm, cht_C, egt_C, oil_pressure_bar, oil_temp_C, fuel_flow_L_h, vibration_g, injection_timing_deg: PS-mandated engine telemetry
- battery_voltage_V, alternator_current_A: electrical health
- engine_health_pct, fault_type, fault_severity: ground-truth label, used for evaluation/export only, never fed to the classifier as an input
Directional signature per fault class (7 implemented)
- Misfire: RPM drops/fluctuates, EGT unstable, vibration up, fuel flow irregular.
- Injector degradation: EGT up, fuel flow deviates, vibration up, RPM fluctuates.
- Lubrication fault: oil pressure down sharply, oil temperature up, vibration up.
- Overheating: CHT up strongly, oil temperature up, EGT up.
- Combustion instability: EGT oscillates, fuel flow irregular, injection timing deviates.
- Mechanical vibration fault: vibration dominates, other channels near nominal.
- Sensor drift: one channel diverges while physically-linked channels stay consistent with normal operation.
Scope honesty
- No physical aero piston engine is used; all telemetry is generated by the onboard simulator.
- Numeric ranges are prototype/synthetic envelopes, not certified limits for any specific aero-engine model.
- RUL and mission-risk figures are simulation-derived only (Het's ML package is a fault classifier; it does not include an RUL model).
- The primary fault diagnosis comes from Het's trained XGBoost classifier, trained and evaluated on synthetic data only (accuracy ≈ 0.596, macro F1 ≈ 0.607 on the held-out test split; see Model Validation); it is not field-validated. The rule/signature layer is a separate, transparent baseline kept as an independent cross-check.
- The vibration spectrum is a synthetic waveform, not a real accelerometer capture.
- Mission comparison history lives in page memory for the current session only.
- Role-based access control and screen gating are a prototype/browser-enforced permission model, not a server-side security boundary.
Where real components plug in
- ML model (fault classifier / RUL): replace
Service.getDiagnostics()/Service.getPrognostics()implementations: the UI reads only the returned shape, not the model internals. - Telemetry ingestion (CAN/ECU/FADEC): replace the tick-driven simulator call in
Service.getTelemetry()with a live feed; the rest of the pipeline (validation → physics → residual → dashboard) is unchanged. - Security module: see the Security screen: prototype browser-enforced permission model (role-gating and write-action locks) and reserved call sites in
Service.getSecurityStatus(). Real deployments require defense-grade server-side authentication, mTLS transport encryption, and HSM signing.
Reference Engine Parameter Basis (Rotax 914 UL/F Benchmark)
AeroTwin's thermodynamic and kinematic baselines are parameterized from public technical specifications for ~100–115 HP aero piston engines. The model is primarily based on the Rotax 914 UL/F benchmark, providing realistic operating limits (RPM, CHT, EGT, oil pressure, turbo boost response). No partnership, OEM data sharing, or official endorsement by BRP-Rotax is claimed or implied.
| Engine Parameter | Published Specification (Rotax 914 UL/F Benchmark) | AeroTwin Physics Role |
|---|---|---|
| Architecture & Layout | 4-stroke, 4-cylinder horizontally opposed boxer engine | Cylinder count & vibration harmonics |
| Cooling Architecture | Liquid-cooled cylinder heads, ram-air cooled cylinders | Dual thermal baseline (CHT vs. Oil Temp) |
| Aspiration & Boost | Turbocharged with integrated exhaust wastegate & TCU controller | Altitude compensation curve (to 15,000+ ft) |
| Fuel Delivery System | Dual constant-depression carburetors (Mechanical metering) | Fuel flow dynamics & mixture baseline |
| Displacement & Bore/Stroke | 1,211 cm³ (73.9 cu in) · 79.5 mm bore × 61.0 mm stroke | Swept volume thermodynamic modeling |
| Compression Ratio | 9.0 : 1 | Peak combustion pressure calculation |
| Power Output | 84.5 kW (115 HP) @ 5,800 RPM (Takeoff, 5 min max) / 73.5 kW (100 HP) continuous | Shaft load & throttle-power translation |
| Operating RPM Range | 5,500 RPM continuous · 5,800 RPM max · Idle ~1,400 RPM | Telemetry envelope & threshold gating |
| Fuel Delivery Note | Carbureted reference (914 UL/F). Separate electronic fuel injection (EFI) engines (e.g., Rotax 915/916 iS families) represent distinct FADEC architectures. | Fuel system degradation mapping |
The FLYGAS GAS418HA is referenced as a public turbo-normalized UAV piston engine configuration. It is evaluated for Medium-Altitude Long-Endurance (MALE) integration (source: AeroExpo directory). It serves solely as an installation benchmark for long-endurance propulsion modeling, not as a source of proprietary flight or test telemetry.
Problem statement alignment across prototype capabilities
- Section A: Digital Twin Core Framework: Virtual engine model synchronized with live data, modular architecture, real-time ingestion capability.
- Section B: Health Monitoring System: RPM, CHT, EGT, oil pressure and temperature, fuel flow, vibration, battery/alternator health, injection timing.
- Section C: Fault Detection & Predictive Analytics: Misfire, injector abnormality, lubrication issues, sensor drift, combustion instability, overheating, abnormal vibration.
- Section D: AI/ML Layer: Fault classification via Het's trained XGBoost model (primary) with the rule/signature baseline as an independent cross-check; RUL trend analysis and predictive maintenance recommendations remain the synthetic baseline pending a dedicated RUL model.
- Section E: Simulation & Replay: Historical replay, environmental simulation, high altitude, endurance, hot weather, rapid throttle transitions.
- Section F: Visualization Dashboard: Real-time health status, fault alerts, efficiency trends, maintenance advisory, mission-wise reports.
This build is a software prototype. Engine telemetry is produced by a physics-informed simulator with an operator-controlled fault-injection mode; no physical MALE UAV engine is connected. Numeric operating envelopes are prototype defaults, not manufacturer-certified limits.
RUL, anomaly-score, fault-probability and mission-risk outputs are simulation-derived against synthetic degradation trajectories only. Field-grade validation would require a named reference engine, its performance maps, and real test-rig or flight telemetry (none of which are claimed here).
Module StatusBuild AuditCorrected against actual build state
- XGBoost diagnostic classifier (primary)
- Rule / signature diagnostic baseline (cross-check)
- Diagnostic cross-check display (not probability-fused)
- Model validation report rendering (Het's real eval data)
- 1 Hz ML sampling layer, sharing one authoritative telemetry trajectory with the display (no separate noise realization)
- Backend ML inference (swap client-side runtime for a server)
- Live telemetry adapter (Service.getTelemetry() call site)
- Edge deployment (interfaces defined, not executed)
- Real engine telemetry
- CAN / SocketCAN connection
- Cybersecurity implementation
- Edge deployment (physical execution)
- Real target-engine calibration
- Dedicated RUL ML model (Het's package is diagnosis-only)
Deliverables ChecklistSpecification Audit
- Functional software demonstrator
- Digital twin architecture design
- Engine simulation model with mission profile
- Fault-injection engine (scheduled + live)
- XGBoost diagnostic classifier (integrated, primary)
- Rule/signature cross-check baseline
- RUL estimation module (synthetic baseline)
- Mission-risk scoring
- Sensor data-quality module
- Event timeline
- Mission replay capability
- JSON/CSV mission export
- Application shell, 3 role-based workspaces
- Model validation report (real eval data)
Build PhasingEngineering Roadmap
| Phase | Scope | Track |
|---|---|---|
| 1–3 | Simulator, mission profile, telemetry schema, application shell | core |
| 4–5 | Physics baseline, residual analysis, fault injection | core |
| 6–8 | Anomaly detection (rule), synthetic RUL, mission risk | core |
| 9–10 | Event timeline, replay, reports | core |
| 11 | XGBoost diagnostic classifier integration (1 Hz sampling, cross-check, validation report) | core |
| 12 | CAN / SocketCAN ingestion path, backend ML inference, edge deployment | integration-ready / planned |
| 13 | Security implementation, real target-engine calibration, dedicated RUL model | planned |
Security
Multi-Device Threat Monitor SECURITY DEMO STATE
Real-time detection of coordinated unauthorized access (client-side demo state)
| Device | Type | Status | Requests | Action |
|---|
No coordinated attack detected.
AeroShield Security
Emergency protection for unauthorized multi-device access and telemetry integrity.
Module StatusBuild AuditCorrected against actual build state
- role-gated screen/action visibility
- session-local audit log
- client-side telemetry range/delta checks
- passkey-protected lockdown workflow
- packet-level cryptographic authentication
- TLS/mTLS channel encryption
- real device attestation/fingerprinting
- HSM-backed model-weight signing
Architecture Integration Modules
Reserved integration points for zero-trust cryptographic verification, role-based access control, and telemetry anti-tamper interfaces for edge and ground telemetry streams.
User/role identity verification for Operator, Maintenance and Engineer access.
Service.getSecurityStatus().identityPrototype browser-enforced permission model gating screens and write actions per role. Not a certified defence-grade or server-side security boundary.
AeroTwin.permissions.canAccess(role, screen, action)Real-time range validation, sudden delta detection, and thermodynamic cross-sensor consistency monitoring.
checkTelemetryIntegrity()Cryptographic packet authentication, sequence replay protection, and UAV/Ground/ECU channel validation.
AeroTransportSecurity.check()Immutable log of mission actions, security alerts, fault injections, and administrative overrides.
auditSecurity()Cryptographic signature and hash verification for physics models, ML XGBoost weights, and fault thresholds.
AeroModelIntegrity.verify()