Why Predictive Maintenance Doesn't Work for Satellite Fault Detection

In industrial vibration analysis, mechanical failure has well-defined physics. A defective bearing generates specific, known vibration frequencies. A few well-placed accelerometers and a spectral analysis over the FFT can diagnose the problem in seconds. The domain knowledge is so strong that you rarely need a complex model: a statistical one over the right signal is, most of the time, optimal.

Satellite telemetry doesn't have that luxury.

I spent several years working at one of Europe's leaders in predictive maintenance, and now I build software around distributed systems that ingest telemetry from physical environments and turn it into intelligence. New space is one of the sectors I'm focused on and it's a sector placing growing importance on fault detection. Naturally, I wanted to know how what I'd learned in industrial predictive maintenance would translate.

So when I went looking into satellite telemetry, my starting question wasn't "how do I detect an anomaly?" It was: what does this domain break that I had assumed was solid ground? The answer turned out to be less about model complexity and more about the structure of the problem itself.

The benchmark that maps the problem

The most rigorous map of this structure I've found is the ESA Anomaly Detection Benchmark (ESA-ADB). A joint effort by operations engineers and ML experts, it trained and compared a wide range of anomaly detection algorithms on real satellite telemetry, from LSTMs for forecasting to autoencoders to plain statistical methods.

The 2025 paper's result is telling: neither the simple statistical methods nor the more sophisticated deep learning methods solve the problem satisfactorily. There's a promising recent development, though. A 2026 paper testing temporal-enhanced transformer architectures reports strong results for anomaly detection on irregular-interval spacecraft telemetry, tested (among other datasets) on OPSSAT-AD, another ESA benchmark built from OPS-SAT CubeSat telemetry. I'm still looking into whether these transformer-based approaches have been tested against the ESA-ADB dataset specifically.

Five reasons satellite telemetry breaks the standard playbook

1. Dimensionality and volume exceed manageable limits. A modern satellite can carry thousands of telemetry channels. The ESA-ADB dataset alone spans roughly 17.5 years of real-world telemetry across multiple missions, comprising over a billion data points. At that scale, even trivial decisions, like which evaluation metric to use, become engineering challenges. The benchmark's authors had to rule out theoretically superior metrics simply because their computational cost was infeasible at this volume.
2. Dependencies form a network too large to map by hand. With vibration data, you know exactly what to look for. In a satellite, a genuine anomaly might not show up in any single channel. It only emerges from the relationship between several of them: a temperature rise paired with power consumption that fails to drop as expected, for instance, a pattern across dozens of sensors that looks perfectly normal channel by channel.

   This isn't a theoretical complication. It's the reason most unsupervised algorithms fail at something operationally critical: when they flag a moment as anomalous, they often can't tell you which channel is responsible. Most collapse data from all channels into a single aggregate score, so they raise the alarm but offer no diagnosis.

3. The data behaves erratically by design. This is, to me, the most underestimated problem. Real satellite telemetry comes with:
   • Irregular, asynchronous sampling rates. One channel might report every second, another every few minutes, neither on a perfectly periodic schedule.
   • Data gaps, caused by the satellite entering a quiescent state or by ground-station communication issues.
   • Trends that reflect expected physical degradation of a component, not an anomaly.
   • Operational concept drift. The satellite's expected behavior genuinely changes depending on its operating mode or mission phase.

   Most anomaly detection literature assumes time series that are uniformly sampled and stationary. Neither holds here. The benchmark itself requires a resampling preprocessing step just to run most of the available algorithms.

4. Different channel types can't be treated as equal. Not every channel is a classic continuous signal oscillating around a mean. You get monotonic channels (cumulative counters like total operating hours, where the change rate is what matters), near-constant channels (where standard-deviation-based methods break down), categorical channels, binary channels, and telecommands. Telecommands are instructions sent from Earth that change the satellite's execution mode, and with it, what "expected behavior" even means from that point on. Applying a single normalization pipeline and a single model to all of these, as many algorithms do, is itself a source of error.
5. Noise from the space environment has no earthbound equivalent. Radiation, extreme temperature swings, electromagnetic interference: the space environment injects measurement noise unlike anything you deal with on the ground. One of the benchmark's more revealing findings is that a DC-VAE, a fairly sophisticated deep learning method, performed no better than a plain statistical baseline. The authors trace this to noise combined with high sample rates generating a flood of false positives, a problem no amount of architectural complexity managed to solve.

Detection isn't enough

As that last point suggests, most tested models raised a lot of false alarms. This matters more than it might in other domains, and it connects to something I keep coming back to when I think about technical problems: they have to solve business problems. In satellite operations, a flood of false alarms defeats the purpose. Operators need to stay focused on real operational issues, and that need only grows as missions and fleets scale up.

There's a further nuance the benchmark makes explicit that's rarely discussed outside this domain: detection alone isn't the goal. When an anomaly is detected matters, not too early, not too late, and how the detection aligns in time with the actual event matters more than a simple binary overlap. That's why the benchmark doesn't rely on standard metrics like F1-score. It uses metrics purpose-built to evaluate detection timing and temporal localization quality. Standard academic metrics were designed for cleaner problems and don't capture what operators actually need.

What this means

The honest takeaway from the paper is that even its best-performing model, the Telemanom-ESA, an ESA version of the NASA Telamanom, which is basically an LSTM heavily hand-tuned for this specific task, works because it's highly parameterized and calibrated to this particular data, with no guarantee it transfers to other missions.

This is the thesis I wanted this first article to land on: the idea that the domain constrains the model isn't new to anyone who has deployed ML in production. What's specific to this problem is how far that constraint goes. Usually the domain adds restrictions, but standard tools still get you most of the way there. Here, that's not true. Irregular sampling breaks the seasonality assumptions most models take for granted. The absence of well-characterized failure modes leaves models without a useful prior. And the sheer heterogeneity of channel types rules out a single, unified processing pipeline. In satellite telemetry the domain invalidates a good part of the default toolkit you bring.

This is the first piece in a series where I'm working through the FDIR problem for the new space sector. Next, I want to look more closely at how transformer-based approaches address the irregular-sampling problem directly, and whether that generalizes better across missions than the hand-tuned LSTM baselines do, which is the gap this piece leaves open. If you're working on telemetry and ground segment systems in the space sector, I'd like to hear from you.

References

• Kotowski, K. et al. European Space Agency Benchmark for Anomaly Detection in Satellite Telemetry. arXiv:2406.17826.
• Transformer-based anomaly detection for satellite telemetry data. Acta Astronautica.