Spacecraft fault detection, isolation and recovery without a learned model; held-out fresh scenarios 98.9% vs 29.4% for a simplified limit-to-safe-mode baseline.

Oracle7-DeepSpace 2.0 (internal)

Status: finished. The paper is written (oracle7-deepspace.pdf). The live simulation is at https://prometheus7.com/DeepSpace7/. The source release is packaged but not published. Prometheus7 Research Institute. Source SHA-256 (deepspace/*.py, recipe eval/eval_final.py:code_sha256): b936ec011372fd9860c748b39b5117c7b8b1296f85d169ccfc8a9b6f8e394408.

It is a fault-response core for deep-space spacecraft. It takes telemetry, forms confirmed facts, then fault hypotheses, then admissible responses. Each decision is either an admitted rule or a worst-case lookahead on a twin built from telemetry only, and each gets a receipt. A self-maintenance gate admits rules learned from its own verified episodes. It uses the Python standard library only, has no learned weights, and needs a fraction of one CPU.

Headline (fresh held-out seeds, full detail in README.md and the paper)

Single faults (1,700 scenarios):

  • oracle7: 98.9% [98.3, 99.3] correct, 0 harmful, 0 damage.
  • Expert fault table: 100%.
  • Simplified limit-to-safe-mode baseline: 29.4%, with 200 harmful actions and 218 damage events.

Compound faults (240): 99.2% both remedied, 0 harmful.

Self-maintenance: 9 rules admitted; all 4 planted wrong rules rejected.

Physics the twin does not model (all terms + dropout): outcome 97.7%, expert 98.0%, baseline 72.3%.

Generalisation:

  • Voyager 2, OSIRIS-APEX and MRO, added after the freeze: 1,300/1,300.
  • Parker Solar Probe: 67.1% (a hot environment is read as a stuck heater; the expert table fails identically).
  • Of 5 unknown fault types, 3 are handled without harm; a frozen sensor and a biased load sensor are not.
  • Determinism: 690/690 missions replay identically.
  • 60-day soak: 52/53 faults remedied, 0 harmful.

Real telemetry (SMAP/MSL): the detection front end is not competitive (F0.5 0.42 vs LSTM 0.71).

Result files are in results/.

Files not included

1 deployment and demo scripts for the builders' own servers are not in this repository (they name private hosts). None is part of the method or of the evaluation that produced the numbers above. SHA-256 of the sorted list of their paths: ae53c433636a66c48fd74e5b23f0248a34ee4e2bf5a3ec9ef49d2d25530e9bab.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support