Executive Summary

Actuator failure on spacecraft — brushless DC motors, harmonic drives, ball screws — is difficult to predict and expensive to get wrong. Unscheduled failure risks mission loss; overly conservative maintenance schedules waste flight hardware and budget that could extend mission life.

G0 is a physics-informed AI system, built by AstraState Systems, that estimates Remaining Useful Life (RUL) for these actuator types. It is deployed as an air-gapped, on-premise container with no telemetry and no external network dependency — a deployment model built for environments where sensor data cannot leave the customer's control.

This paper reports one result: RUL prediction accuracy, validated against real bearing run-to-failure data from the IMS/Rexnord dataset, using a leave-one-out protocol across all four documented real failures. We report this number and how it was produced in full, rather than pairing it with a second, less certain claim to make the headline look broader than it is.

1. The Problem

Spacecraft actuators degrade gradually before they fail outright — bearing wear, lubricant breakdown, gear mesh degradation. The physical signal is present in motor current, vibration, temperature, and position error well before failure. The difficulty is threefold:

• Failure data is scarce. Spacecraft actuators rarely run to failure in the field — that is the point of a space program. Real run-to-failure datasets are limited to ground-test rigs and industrial analogues.
• Physics-only models are hard to calibrate across operating regimes; pure black-box ML models generalize poorly outside the exact conditions they were trained on and offer no physical interpretability.
• Any prognostics claim is only as strong as the data it was tested against, and only as honest as the paper is about what that data can and cannot support.

G0 pairs physics-informed modelling with an evaluation method designed to avoid the most common way these numbers get overstated: tuning against the same data used to report performance.

2. Methodology

2.2 Evaluation Data

The result in this paper is evaluated against the IMS/Rexnord bearing run-to-failure dataset: four documented real bearing failures, evaluated leave-one-out — the model trains on three failures and is tested on the fourth, held out entirely, repeated so each of the four failures serves as the test case once.

2.3 Avoiding Test-Set Leakage

A single, fixed model configuration was used across all four held-out splits. No per-split hyperparameter search was used to select the reported result — a per-split search was run separately during development and confirmed to produce a similar mean but is not the number reported here, because selecting the best configuration per test split would let information about the held-out failure leak into model selection. The number in this paper reflects one configuration, evaluated identically across all four real failures, with no tuning advantage given to any individual split.

3. Result: RUL Prediction Accuracy

Evaluated leave-one-out across all four documented real bearing failures in the IMS/Rexnord dataset, G0 achieved a mean RUL prediction error of 8.79 days, a 27.9% improvement over a naive baseline (mean-RUL extrapolation from the training failures).

Table 1. Per-trajectory breakdown, single fixed model configuration, no per-split tuning. Note set3_bearing3 underperforms the naive baseline on this metric alone — included here rather than omitted, consistent with reporting the full leave-one-out spread rather than a favorable subset.

This result answers the question a maintenance planner needs answered: given a real degrading actuator, how close is the predicted remaining life to the true remaining life? It does not address alarm-timing false-positive rate, which is not reported in this version of the paper.

4. Deployment Model

G0 ships as an air-gapped Docker container with a hardware-locked annual license, designed for environments — defense, aerospace, and other regulated sectors — where sensor data cannot leave customer premises under any circumstances.

• No telemetry, no phone-home, no logs leaving the deployment environment at any point in normal or emergency operation.
• Pooled-token licensing across a customer's rig fleet, with a graceful degrade-to-lockout sequence on connectivity loss rather than an abrupt hard stop, and a signed, out-of-band emergency-unlock procedure for extended outages that requires no telemetry or remote access to use.
• The shipped model artifact includes trained inference components only (context encoder, energy model, physics head); training code and raw model weights are not included in the deployed container.

5. Scope and Limitations

This paper reports RUL prediction accuracy only. It does not report an alarm false-positive rate, on either real or synthetic data. Readers evaluating G0 for a maintenance program should treat alarm-timing performance as a separate question, and should ask any prognostics vendor — including us — for the specific evaluation protocol and dataset behind any alarm-calibration number before relying on it.

The IMS/Rexnord dataset provides four real bearing failures. This is the standard public dataset for this kind of evaluation, and leave-one-out across all four is a conservative protocol, but four failures is a small sample; the per-trajectory spread in Table 1 is included specifically so readers can see that spread rather than only the mean.

6. Conclusion

G0 demonstrates real-data RUL prediction accuracy, evaluated leave-one-out against the standard public run-to-failure bearing dataset for this problem class, with a fixed model configuration and no per-split tuning advantage. We report this one result in full, including the per-trajectory spread and the naive-baseline comparison, and we do not pair it with an unverified second claim.

For a technical evaluation or deployment discussion, contact AstraState Systems.

7. Glossary

• RUL (Remaining Useful Life): the estimated time remaining before a component is expected to fail, expressed in days.
• Leave-one-out evaluation: a validation method where the model is tested on one real failure it has never seen, having been trained only on the others — repeated so every real failure gets a turn as the held-out test case.
• RUL-MAE: mean absolute error between predicted and true remaining useful life, in days.
• Naive baseline: a simple non-learned predictor (mean RUL of the training failures) used as a reference point for how much a learned model actually improves on the simplest reasonable guess.
• Held-out data: data never used during training, calibration, or model selection — used only once, at the end, purely to measure performance.
• Air-gapped: a deployment with no network connection to the outside world, used where data cannot be permitted to leave a secure facility under any circumstances.