Skip to content
CubeSTEM MissionLabCubeSTEMMissionLab
Engineering workbench — a technical surface, not a school lesson. Go to the lessons

MissionLab · AI-Alpha-1

Spacecraft AI Playground

EPS Spacecraft Health — Synthetic Telemetry

Simulation-trained / analysis onlyModel Lab unavailable

Source: canonical Digital Twin evidence. Alpha-1 uses simulated Digital Twin evidence only. No hardware-in-the-loop, measured-physical, flight, or qualification evidence is claimed.

Re-derives the accepted model and dataset on the server, evaluates that exact model under declared simulated shifts against a transparent baseline, and estimates static fit against a declared development-proxy target. Analysis only — no target measurement.

Model Lab is not configured or could not be reached.

Engineering source

Canonical Digital Twin evidence

The Playground consumes accepted Core/G5 evidence through the Dataset Factory. It does not run spacecraft physics in the browser.

Source authority
digital_twin_core
Source class
simulated
Mission reference
Alpha-1 reference mission context (when Dataset Factory supplies it)
Scenario / case count
7 accepted EPS scenario IDs

Scenario / case list

  • nominal-no-fault → intended nominal
  • solar-source-derate → intended anomaly
  • battery-discharge-inhibit-eclipse → intended anomaly
  • rail-forced-open → intended anomaly
  • sustained-overload-trip-recovery → intended anomaly
  • persistent-overload-retrip → intended anomaly
  • combined-fault-protection → intended anomaly

Selected feature channels (IDs + units)

  • Battery state of charge · percent · cubestem.twin.channel.truth.battery_soc.v1
  • Solar generation power · W · cubestem.twin.channel.truth.solar_generation_power.v1
  • Served load power · W · cubestem.twin.channel.truth.served_load_power.v1
  • Unserved power · W · cubestem.twin.channel.derived.unserved_power.v1
  • Battery charge/discharge power · W · cubestem.twin.channel.truth.battery_charge_discharge_power.v1

Excluded label-leaking channels

  • EPS fault active (truth flag) · flag · cubestem.twin.channel.derived.eps_fault_active.v1
  • EPS protection active (truth flag) · flag · cubestem.twin.channel.derived.eps_protection_active.v1
  • EPS rail trip active (truth flag) · flag · cubestem.twin.channel.derived.eps_rail_trip_active.v1

Fault, protection and rail-trip truth flags are excluded from training features so the classifier cannot read the answer directly.

Learning

Error interpretation

Explanatory guidance only — not engineering authority and not a flight-performance claim.

  • False positive: the model predicted anomaly when the held-out label was nominal. In operations this can waste operator attention; here it is a teaching signal about over-sensitivity.
  • False negative: the model predicted nominal when the held-out label was anomaly. In operations this can hide a real off-nominal condition; here it shows under-sensitivity on held-out cases.
  • Why models can perform poorly: small held-out sets, class imbalance, distribution shift across fault families, or features that do not separate the intended label can all produce low metrics. That outcome is still a valid experiment.
  • Simulated performance is not HIL/flight performance: these metrics come from simulated Digital Twin evidence only. They do not prove hardware-in-the-loop behaviour, measured physical truth, onboard resource fit, or flight readiness.

Roadmap

What comes next

Truthful progression only. Future Alpha capabilities are planned and unavailable here.

  1. AI-Alpha-1

    Synthetic telemetry + real ML

    Canonical simulated evidence → dataset → Model Lab → analysis-only Passport.

    CURRENT

  2. AI-Alpha-2

    Robustness + TinyML / resource fit

    Not available in this release.

    PLANNED / UNAVAILABLE

  3. AI-Alpha-3

    Actual edge / PIL

    Not available in this release.

    PLANNED / UNAVAILABLE

  4. AI-Alpha-4

    HIL / synthetic-to-HIL transfer

    Not available in this release.

    PLANNED / UNAVAILABLE