SYSTEMS
HomeProductWhy usComplianceValidationEvaluateFAQContact
Prognostics for spacecraft actuators

Find out whether prognostics works on your actuators, before you commit.

G0 turns rig telemetry into remaining-life estimates with uncertainty, running air-gapped on your own hardware. We test it first on your own run-to-failure data, with predictions committed in advance. It is designed for brushless DC motors, harmonic drives and ball screws, and it is early-stage: the limits are stated on the Validation page.

Request a pilotWhat we have and have not tested
Recorded G0 run · simulator data
NOMINAL
NODE_ISOLATED
Health
0.00
Remaining life
–
Run
0%
Open this run in Evaluate
The pilot

Test it on your own run-to-failure data first.

1 · YOU SEND DATA

Historical run-to-failure data from your rig, with the outcome hidden.

2 · WE COMMIT

We commit hash-checked predictions before we see the outcome, so they cannot be altered afterwards.

3 · YOU COMPARE

You compare our predictions with what actually happened.

4 · WE REPORT PLAINLY

We report G0 next to simple baselines, and tell you plainly if it does not beat them.

We are looking for 2–3 design partners. Duration, cost and any case-study arrangement are agreed up front, before you share any data.

Deployment

Deploys inside your facility, and your security team can check it.

Each of these is something your own team can inspect, without taking our word for it.

OFFLINE INSTALL

Delivered as a signed container image and loaded without internet or registry access.

ZERO EGRESS

No outbound connections. A local console for your team, plus a separate authenticated, read-only admin view.

NO TELEMETRY

No usage data, sensor data, logs or results leave the deployment. We receive none of it and have no route to.

SIGNED MODEL FILES · CPU ONLY

Model files are cryptographically signed so you can confirm what you are running. CPU only, no GPU needed.

YOUR LICENCE SERVICE

The licence service ships in the bundle and runs on your hardware, pooled across your rigs.

PER-RIG HISTORY

Telemetry, predictions, faults and state changes are logged per rig to storage you mount.

Full deployment specification · Data sovereignty

The problem

Wear-out is gradual, invisible until late, and expensive to find the hard way.

Life tests run for weeks on hardware that is slow to replace. Teams stop early and learn little, or run to failure and lose the unit and the schedule slot.

Prognostics and health management (PHM) estimates how much useful life a component has left. G0 is designed to apply it to rotary actuators on test, such as brushless DC (BLDC) motors, harmonic drives and ball screws, using the signals your rig already records, so a test lead can decide whether to extend, inspect or stop.

From rig data to a decision in three steps.

1 · CONNECT

Ingest rig data as CSV: motor current, temperature, vibration, position error and torque. No new instrumentation.

2 · MODEL

A physics-informed model tracks each actuator's degradation continuously, on your hardware, with no network access.

3 · ACT

Health state, remaining life and alarms update with every batch of data, a basis for extend, inspect or stop decisions.

Outputs

What your test team gets.

RUL WITH UNCERTAINTY

Remaining useful life in days with a stated interval, not just a single number. Interval calibration is not yet reported.

HEALTH STATE AND ALARMS

Healthy to failure imminent at a glance, plus degradation and sudden-change alarms with thermal and electrical false-alarm suppression.

LOCAL CONSOLE

A per-rig console and history, entirely inside your network.

Evidence

Early-stage, with the limits stated.

Early-stage: tested on 4 real bearing failures, ~16% lower error than a naive baseline, with limits stated; pilot partners wanted.

What we have and have not tested

Next step

Find out whether G0 works on your data.

Request a pilotOr watch a recorded run →
Product

G0 is designed to predict the remaining useful life of spacecraft actuators from signals you already record.

What G0 is designed to track, what it needs from your rig, what it hands your test team, and how it deploys.

What G0 is designed to track

G0 is designed to follow the mechanisms that end rotary-actuator life: bearing fatigue, gear crack growth, thermal stress and, in vacuum and TVAC tests, the condition of solid lubricant. It infers them indirectly from signals the rig already produces, so there is no sensor on the failing component to add or to fail.

Tested so far: bearing degradation only, from vibration features, on public run-to-failure data. Gear crack growth, thermal stress and lubricant condition have not been validated on real failures. What we have and have not tested.

Designed for

Brushless DC (BLDC) motors, harmonic drives and ball screws, through bearing, gear and lubricant wear. G0 has not yet been validated on any of these actuator types.

Physics-informed approach

G0 is physics-informed: degradation physics constrains what the model may predict, and learning from data covers what physics cannot capture. The aim is a remaining-useful-life (RUL) estimate that stays physically plausible with little failure data. Why this approach.

What it needs from your rig

  • Five signals: motor current, temperature, vibration RMS, position error and torque. Validation to date used vibration features only; the other four signals have not been tested on real failures.
  • One optional signal: vacuum pressure, which enables the vacuum end-of-life alert.
  • CSV ingestion, processed at 10 Hz. Most existing data-acquisition exports work as they are.

What it hands your test team

  • Remaining useful life in days with an uncertainty interval.
  • Health state: healthy, minor wear, degraded, failure imminent.
  • Degradation alarm when the trend crosses a calibrated level.
  • Sudden-change flag for step events such as debris or shock.
  • False-alarm suppression for thermal-expansion transients and electrical events such as radiation-induced current spikes.
  • Per-rig history of telemetry, predictions, state changes and faults.

Reading the uncertainty

A remaining-life figure without error bars cannot support a decision. G0 reports an interval and separates two sources of uncertainty: noise in the measurement, and distance from what the model has seen before. When the second rises, the actuator is behaving in a way G0 has not met, and the estimate should be read accordingly.

Interval calibration, meaning how often the interval contains the true remaining life, is not reported yet. Until it has been measured on real failures, read the interval as indicative rather than as a stated confidence level.

Where G0 fits

G0 is designed for ground-test and life-test campaigns on rotary actuators: thermal-vacuum (TVAC) testing, qualification and acceptance runs, and mechanism development rigs. It adds a remaining-life estimate with uncertainty alongside your existing test procedures; it does not replace qualification testing. It has not yet been validated on spacecraft actuators, so the first step is a pilot on your own run-to-failure data.

Deployment

G0 is delivered as a single Docker container that you run on your own hardware, inside your own network. This is what your IT and security teams will want to see.

PackagingOne Docker container image. Compiled, with no source code in the image.
InstallationLoaded offline from a signed bundle. No internet or registry access needed.
RuntimeRuns as a non-root user. One container serves many rigs, each with its own slot.
NetworkNo outbound connections. The console is served on one local port; the admin view is on a separate port, authenticated and read-only.
StorageMounted volumes for signed model files, licence, input CSVs and per-rig history.
ComputeCPU only. No GPU required.
LicensingHardware-locked, annual. The licence service ships in the bundle and runs on your hardware, pooled across your rigs. Staged degradation in an outage; signed emergency unlock.
IntegrityCryptographically signed model files.

Your bundle ships with its own deployment instructions and compose file.

Vacuum end-of-life alert

Lubricant-depletion alert for vacuum and TVAC testing. When a vacuum pressure signal is provided, G0 flags end-of-life lubricant conditions at high torque.

It is raised at end-of-life conditions. It is not an early warning, and it cannot yet tell true depletion from ordinary wear-out. It stays silent at ambient pressure, which you can see on the Evaluate page.

Evidence

Early-stage: tested on 4 real bearing failures, ~16% lower error than a naive baseline, with limits stated; pilot partners wanted. What we have and have not tested. Security and data handling are on the Compliance page.

Request a pilot
Why AstraState

Prognostics built for scarce failure data.

Hybrid prognosis for scarce failure data, evaluated the way a reviewer would.

Hybrid prognosis, by design

Prognostics literature distinguishes physics-based, data-driven and hybrid approaches. Physics alone is brittle across operating regimes, and data alone starves when failures are rare. G0 is physics-informed: degradation physics constrains what the model may predict, and learning from data covers what physics cannot capture.

The aim is an estimate that stays physically plausible with little failure data, and an explicit signal when conditions fall outside what the model has experienced. See what G0 produces on the Product page.

Evaluated the way a reviewer would

We test on real failures the model has never seen (leave-one-out), with one fixed configuration. Choosing the best setting per test failure would leak the answer into the result, so we do not. Preprocessing uses only data available at prediction time. An earlier version of our analysis normalised each failure using statistics from its whole life, which would not be available in real operation; we corrected it on 24 October 2026. We compare against two simple baselines, a naive one and an RMS-only one, and we say where G0 did not beat them: on the corrected protocol it beat the naive baseline on one of four failures and was slightly worse on the other three. Full validation results.

What we do not claim

  • No alarm false-positive rate is reported on real or synthetic data in our white paper. Ask any prognostics vendor, us included, for the protocol behind such a number.
  • The vacuum alert is an end-of-life indicator, not an early warning.
  • Four real failures is a small sample, and the average gain comes mainly from one of them.
  • Performance on spacecraft actuators, motors, harmonic drives or ball screws: only bearing data has been tested.
  • Calibrated uncertainty intervals: how often they contain the true value is not reported yet.

Built for closed environments

Most prognostics products assume a cloud connection. G0 assumes the opposite: air-gapped, on your hardware, no telemetry. Data sovereignty is the starting constraint of the design, not an add-on. Deployment and security details.

Why existing approaches stall

Prognostics research has long noted that actuator fault-progression data is scarce, because few units are ever run to failure. Physics-only models are hard to calibrate across operating regimes. Purely data-driven models need failure histories that do not exist, and extrapolate poorly outside their training conditions.

Who it is for

Programme and test leads, and the reliability and mechanisms engineers who answer to them, running life, thermal-vacuum and qualification campaigns on rotary actuators: at spacecraft primes, subsystem suppliers and test laboratories where data cannot leave the building.

Why now

Test campaigns are getting longer and costlier per article while failure data stays scarce. Practical on-premise deployment of modern models changes what a test team can know during a run, not only after it.

Compliance

Your sensor data never leaves your facility.

Air-gapped by design, with no telemetry and no outbound connections.

Air-gapped by design

G0 runs entirely inside your network. It needs no internet connection and makes no outbound calls, in normal operation or in emergencies. Your security team can verify that by inspecting the deployment's network behaviour. Deployment specification

No telemetry

No usage data, sensor data, logs or results leave the deployment. We do not receive them and have no route to.

Sensitive environments

G0 is designed with ITAR-sensitive environments in mind. That describes a design principle. It is not a certification or compliance claim, and your own export-control and security review still applies.

Licensing without connectivity

Licences are pooled across your rigs and checked inside your network. If the licence service becomes unreachable, G0 degrades in stages toward lockout instead of stopping abruptly. For extended outages, a signed emergency-unlock procedure works out of band, with no telemetry or remote access.

What is in the container

Trained inference components only. Training code and raw model weights are not shipped. Model files are cryptographically signed so you can confirm what you are running.

What G0 does and does not claim

  • Does: run inside your network, make no outbound connections, send no telemetry, and ship signed model files.
  • Does not: claim any certification, including ITAR. G0 is designed with ITAR-sensitive environments in mind, which is a design principle, not a compliance claim.
  • Does not: replace your own export-control and security review.

See the deployment specification or contact us about your security review.

This website

The Evaluate page replays recorded simulator runs. It uses no customer data and does not run G0 on anything you enter.

Validation

Validation: what we’ve tested and what we haven’t

Early-stage, on public bearing data. The limits come first.

Correction · updated 24 Oct 2026

v1 of this page and of the white paper normalised each failure by its own full-life statistics, which a deployed system cannot do. We corrected this and added baselines; the headline moved from 8.79 days (27.9%) to 10.1 days (about 16%).

Status

G0 is in pre-commercial validation. Few spacecraft-actuator failure records exist, so we test on public run-to-failure data and are looking for partners to validate on real actuator data.

Result: public bearing data (IMS/Rexnord)

Evaluated leave-one-out on the four documented IMS/Rexnord failures, G0 achieved a mean remaining-life error of 10.1 days, about 16% lower than a naive baseline (12.1 days) and 14% lower than an RMS-only baseline (11.8 days). G0 beat the naive baseline on one of four failures (set2_bearing1) and was slightly worse on the other three, so the average gain comes mainly from one failure. It beat the RMS-only baseline on two of four.

Held-out failureG0 (days)Naive (days)RMS-only (days)G0 vs naive
set1_bearing310.289.749.73−5.5%
set1_bearing410.049.748.47−3.1%
set2_bearing15.8215.128.78+61.5%
set3_bearing314.3713.8420.18−3.8%
Mean10.1312.1111.79+16%

Mean of three training seeds (42, 43, 44), one fixed configuration, no per-failure tuning. Features are normalised using only each trajectory’s first 20 snapshots and training-fold statistics, so nothing in preprocessing uses information from the held-out failure’s future. The naive baseline is the mean training RUL capped at 40 days, matching the model’s training target. Positive percentages mean G0’s error is lower. These runs used vibration features only.

IMS bearing data: Center for Intelligent Maintenance Systems, University of Cincinnati, with Rexnord; hosted in NASA’s Prognostics Data Repository. Not a NASA test and not NASA-endorsed.

Other benchmarks

We also tested two further public bearing datasets (University of Ferrara, 6 failures; FEMTO/PRONOSTIA, 17 failures). On these, G0 showed no reliable advantage over simple baselines. Protocols and per-failure results are available on request.

What this does not show

  • Performance on spacecraft actuators, motors, harmonic drives or ball screws.
  • The current, temperature, torque and position-error inputs on real failures. The validation used vibration features only.
  • Interval calibration: how often the uncertainty intervals contain the true value is not reported yet.
  • Alarm false-alarm rates on real failures.

Real versus synthetic

The bearing result uses real data only. The runs replayed on the Evaluate page are simulator output, with thermal and shock events injected where labelled. They show how G0 behaves. They are not evidence of accuracy on real hardware.

The vacuum end-of-life alert has been tested on simulator data only. It fires at end-of-life conditions and cannot yet separate true lubricant depletion from ordinary wear-out. We quote no lead time.

How we close the gap

Blind pilot validation on your run-to-failure data: we commit predictions in advance and hash-check them, and you compare against what happened. Find out more on the Evaluate page, or go straight to a pilot.

Request a pilot

Technical white paper · v2 · 24 October 2026

v2 supersedes the September 2026 version. v1 normalised each failure by its own full-life statistics, which a deployed system cannot do; we corrected this and added baselines, and the headline moved from 8.79 days (27.9%) to 10.1 days (about 16%).

Download the white paper (PDF, 4 pages)

Read the white paper

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 is designed to estimate 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 primary result: RUL prediction error on real bearing run-to-failure data from the IMS/Rexnord dataset, using a leave-one-out protocol across all four documented real failures, with causal preprocessing and compared against a naive baseline and an RMS-only baseline. It also summarises two further public bearing benchmarks on which G0 showed no reliable advantage over simple baselines. Everything in this paper concerns bearing vibration data; nothing here validates G0 on spacecraft actuators.

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

2.1 Preprocessing and baselines

Features are normalised using only each trajectory’s first 20 snapshots and training-fold statistics, so nothing in preprocessing uses information from the held-out failure’s future (causal preprocessing). The September 2026 version normalised each failure using statistics from its full life, which a deployed system cannot do; that has been corrected and the results restated. Two baselines are reported: a naive baseline (the mean training RUL, capped at 40 days to match the model’s training target) and an RMS-only baseline that uses vibration RMS alone.

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. Only vibration features were used in this evaluation, although the product takes five input signals. Data provenance: IMS bearing data from the Center for Intelligent Maintenance Systems, University of Cincinnati, with Rexnord, hosted in NASA’s Prognostics Data Repository. This is not a NASA test and not NASA-endorsed.

2.3 Avoiding Test-Set Leakage

A single, fixed model configuration was used across all four held-out splits, and the reported results are the mean of three training seeds (42, 43, 44). No per-split hyperparameter search was used to select the reported result, 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 on the four documented IMS/Rexnord failures, G0 achieved a mean remaining-life error of 10.1 days, about 16% lower than a naive baseline (12.1 days) and 14% lower than an RMS-only baseline (11.8 days). G0 beat the naive baseline on one of four failures (set2_bearing1) and was slightly worse on the other three, so the average gain comes mainly from one failure. It beat the RMS-only baseline on two of four.

Held-out failureG0 (days)Naive (days)RMS-only (days)G0 vs naive
set1_bearing310.289.749.73−5.5%
set1_bearing410.049.748.47−3.1%
set2_bearing15.8215.128.78+61.5%
set3_bearing314.3713.8420.18−3.8%
Mean10.1312.1111.79+16%

Table 1. Mean of three training seeds (42, 43, 44), one fixed configuration, no per-failure tuning. Features are normalised using only each trajectory’s first 20 snapshots and training-fold statistics, so nothing in preprocessing uses information from the held-out failure’s future. The naive baseline is the mean training RUL capped at 40 days, matching the model’s training target. Positive percentages mean G0’s error is lower.

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.

3.1 Other public bearing benchmarks

We also tested two further public bearing datasets: University of Ferrara (6 failures) and FEMTO/PRONOSTIA (17 failures). On these, G0 showed no reliable advantage over simple baselines. Protocols and per-failure results are available on request.

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 on bearing vibration data only. It does not report an alarm false-positive rate, on either real or synthetic data, nor interval calibration (how often the uncertainty intervals contain the true value), nor results using signals other than vibration features, nor results on spacecraft actuators, motors, harmonic drives or ball screws. 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, and the average gain comes mainly from one of them. Table 1 shows the per-failure spread rather than only the mean.

6. Conclusion

On public bearing data, G0 showed a modest reduction in mean RUL error against a naive baseline (about 10.1 against 12.1 days), driven mainly by one of four failures, and no reliable advantage over simple baselines on two further public benchmarks. This is early-stage evidence on bearing vibration only. It does not establish performance on spacecraft actuators; for that we propose blind pilot validation on partners’ own run-to-failure data.

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, capped at 40 days to match the model’s training target) used as a reference point for how much a learned model actually improves on the simplest reasonable guess.
  • RMS-only baseline: a simple predictor that uses vibration RMS alone, reported alongside the naive baseline.
  • Causal preprocessing: normalisation that uses only data available up to the time of each prediction, with no statistics from the future or from the failure’s whole life.
  • Interval calibration: how often a stated uncertainty interval actually contains the true value. Not yet reported for G0.
  • 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.
FAQ

The questions test and programme leads ask.

Straight answers on data, validation, limits and fit.

About G0

What is G0?

G0 is AstraState Systems' physics-informed prognostics software. It is designed to turn telemetry your test rig already records into a remaining-useful-life estimate with an uncertainty interval for spacecraft actuators such as brushless DC motors, harmonic drives and ball screws. It is early-stage: so far it has been tested only on public bearing data. It ships as a Docker container and runs air-gapped on your own hardware. See the Product page.

What is RUL?

Remaining useful life (RUL) is the time an actuator is expected to keep operating before it reaches end of life. G0 reports it in days with an uncertainty interval, so a test lead can decide to extend, inspect or stop.

What is prognostics and health management (PHM) for actuators?

Prognostics and health management (PHM) is the practice of tracking the condition of a component, estimating how much useful life it has left, and acting before it fails. For rotary actuators such as BLDC motors, harmonic drives and ball screws, that means turning signals a test rig already records (current, temperature, vibration, position error and torque) into a remaining-useful-life estimate and alarms. G0 is designed for that job; so far it has been tested only on public bearing data.

What actuator types is G0 designed for?

Brushless DC motors, harmonic drives and ball screws, through bearing, gear and lubricant wear. It has not yet been validated on any of these actuator types; the only real-failure testing so far is on bearings.

How is G0 different from generic machine learning?

G0 builds in the physics of degradation and reports uncertainty, so it is designed to stay plausible with little failure data and to tell you when it is outside its experience.

Does G0 replace qualification testing?

No. It informs decisions during testing. Qualification criteria and sign-off remain with your own process and standards.

Data, security and ownership

Does telemetry or sensor data leave my environment?

No. G0 runs air-gapped with no telemetry and no outbound calls. We never receive your sensor data, predictions or logs.

Who owns the data and the results?

You. G0 runs on your hardware and everything it produces stays there.

Can G0 run air-gapped?

Yes. Installation and operation are fully offline: G0 is loaded from a signed bundle, needs no internet or registry access and makes no outbound connections. Details on the Compliance page.

Is G0 ITAR certified?

G0 makes no ITAR certification or compliance claim. It is designed with ITAR-sensitive environments in mind: it runs on your hardware, inside your network, with no telemetry and no outbound connections. Your own export-control and security review still applies. See the Compliance page.

What if the licence service is unreachable?

G0 degrades in stages toward lockout. For long outages, a signed emergency-unlock procedure works without any remote access.

Deployment and integration

What data does G0 require?

Motor current, temperature, vibration RMS, position error and torque from your rig, as CSV. Vacuum pressure is optional and enables the vacuum end-of-life alert. For validation, at least one run to failure helps most. Our own validation to date used vibration features only.

What is a run-to-failure dataset, and why does it matter?

A run-to-failure dataset records a component from healthy operation through to failure. It is the only way to check a remaining-life prediction against what really happened, which is why validation depends on it and why it is scarce for spacecraft actuators. It is also what we ask for in a pilot: historical run-to-failure data from your rig, with the outcome hidden.

What does integration involve?

Pointing G0 at your existing data-acquisition exports. There is no new instrumentation and no change to your test procedure.

How is G0 deployed?

As a single Docker container, loaded offline from a signed bundle and run on your own hardware. It serves many rigs, mounts your storage for data and history, and makes no outbound connections.

What hardware does it need?

CPU only; no GPU is required. We will confirm sizing for your number of rigs and data rates during the briefing.

Validation, accuracy and limits

How is G0 validated?

On public bearing data. On the IMS/Rexnord run-to-failure dataset (four real failures, leave-one-out, one fixed configuration, no look-ahead in preprocessing), mean remaining-life error was about 10.1 days against 12.1 for a naive baseline and 11.8 for an RMS-only baseline. G0 beat the naive baseline on one of the four failures and was slightly worse on the other three; it beat the RMS-only baseline on two of four. On two further public bearing datasets it showed no reliable advantage over simple baselines. These are vibration-only bearing results and should not be read as a validation on spacecraft actuators. What we have and have not tested.

How accurate is it?

On four real bearing failures, leave-one-out, mean remaining-life error was about 10.1 days against 12.1 for a naive baseline, roughly 16% lower, with the gain coming mainly from one failure. We do not yet report accuracy on spacecraft actuators, interval calibration or an alarm false-alarm rate.

Are the uncertainty intervals calibrated?

Not yet reported. We have not yet measured how often G0’s intervals contain the true remaining life on real failures, so read them as indicative rather than as a stated confidence level.

Is the vacuum alert an early warning?

No. It fires at end-of-life lubricant conditions in vacuum or TVAC tests at high torque, and cannot yet separate true depletion from ordinary wear-out.

Can I trust the numbers on the Evaluate page?

They are real G0 output on recorded simulator runs, with run ID and model version shown for each. They show behaviour, not accuracy on your hardware.

Pilots and next steps

What is blind validation?

You give us run-to-failure data without revealing the outcome. We produce predictions from the data alone and hash-check them so they cannot be altered afterwards. You then compare them with what actually happened. Blind validation is currently offered as a pilot, so the process may evolve with early partners.

What happens after I request a pilot?

We reply within 3 working days to discuss your setup and data. If it is a fit, we can run a pilot blind validation on your data and walk you through the results before you decide anything. We are looking for 2–3 design partners; duration, cost and terms are agreed up front, before you share any data.

How long does a pilot take, and what does it cost?

Duration, cost and any case-study arrangement are agreed up front, before you share any data, and depend on your data and setup. We are looking for 2–3 design partners. Request a pilot and we will reply within 3 working days.

Contact

Bring a test setup. We will tell you if G0 fits.

A working conversation, with a reply within 3 working days.

What to expect

Requesting a pilot starts a working conversation, not a sales call. Bring a test setup, and we will tell you honestly whether G0 fits, what data you would need, and whether a pilot blind validation on your own run-to-failure data makes sense. We reply within 3 working days. Prefer email? Write to hello@astrastatesystems.com.

From first look to decision

Explore recorded runs and size your own cost of failure on the Evaluate page, with no data exchanged. What G0 needs from your rig is on the Product page. If it fits, we can run a pilot blind validation on your run-to-failure data. This is an early-stage protocol we are still refining with first partners. You hold the outcome, we commit our predictions in advance and hash-check them, and only then do you compare. You review the results, then decide. We are looking for 2–3 design partners; duration, cost and any case-study arrangement are agreed up front, before you share any data.

Thanks — we'll follow up within 3 working days.

Evaluate

Watch G0 work through a full run to failure.

Replay recorded runs and inspect what G0 reports at every point: health, remaining life with its uncertainty, and the alarms with their timing. Or size the cost of unplanned failure on your own programme. Each view has its own share link, so a test engineer and a budget holder can send each other exactly what they saw.

Failure modes (choose up to two)

Test condition

Vacuum end-of-life alert: for vacuum and TVAC testing. When a vacuum pressure signal is provided, G0 flags end-of-life lubricant conditions at high torque. It is raised at end of life and is not an early warning.

REAL PIPELINE OUTPUTSpeed 2×

Predict, then reveal

At what point in the run (% of life) will the degradation alarm first fire for the first selected mode?

50%

Events log

Provenance

About these demonstrations

The runs shown are real G0 pipeline output on recorded simulator runs, with thermal and shock events injected where labelled. They show how G0 behaves, not accuracy on your hardware. The cost-of-failure view is a calculator driven by your own inputs; any figures shown are illustrative, not industry statistics.

What G0 produces · Validation on real failures

Request a pilot

What does an unplanned failure cost you?

Enter your own figures. Nothing is pre-filled from an industry average, because no reliable public failure rate exists for actuators in ground test.

Reference points for cost of one failure (US dollars, selecting one switches the currency), from real NASA project status reports (2004 dollars, small instrument programme): late thermal-vacuum failure with rework and retest about $30k; 35 days of schedule contingency about $50k+; launch delay about $226k. Source. Large programmes will be far higher.

The 30% default is the low end of published industrial benchmarks for condition-based versus reactive maintenance (roughly 30–40% lower maintenance cost, from a US DOE guide that traces to a 2001 pump-industry article). It comes from industrial equipment, not spacecraft actuators. Replace it with your own view.

Estimated annual unplanned-failure exposure

Fleet × failure rate × cost × criticality. Arithmetic on your inputs, not a forecast or a quote.

Request a pilot on your data

If this estimate is worth acting on, tell us about your setup and we will reply within 3 working days.

Thanks — we'll follow up within 3 working days.

ProductWhy usComplianceValidationEvaluateFAQRequest a pilotWhite paper (PDF)
© 2026 AstraState Systems