1. Abstract
Modern satellite missions generate continuous, high-dimensional telemetry that exceeds the practical capacity of human operators to monitor in real time. This paper proposes a software-first reference approach for telemetry anomaly detection and operator decision support in mission operations centres. The approach is intentionally hardware-agnostic, open to integration with existing ground systems, and aligned with the public priorities of the UK Space Agency, the European Space Agency, and NASA.
We describe a layered architecture combining streaming ingest, schema-aware feature extraction, hybrid statistical / ML detection, an operator-facing decision-support layer, and a security envelope. We close with a deployment posture, an evaluation methodology, an explicit ethics and safety stance, and a 24-month implementation roadmap.
2. Context & Motivation
The global space economy reached an estimated $613B in 2024 and is widely projected to grow toward $1.8T by 2035. A growing share of this value sits not in launch but in downstream services: data products, analytics, resilience tooling, and intelligent operations. The UK Space Agency's published priorities, ESA's investments in intelligent systems for space safety and operations, and NASA's SBIR pipeline all explicitly identify technology-supported mission operations as a strategic capability gap.
Current operations teams typically rely on threshold-based alarms, manual playbooks, and post-incident review. As constellations grow from single satellites to fleets of dozens or hundreds, this approach scales poorly. Operators face alert fatigue, slow anomaly localisation, and limited cross-mission learning.
A practical, software-first SI layer — one that augments rather than replaces operators — directly addresses this gap without requiring new spacecraft, new launch capability, or new physical infrastructure. This is the lane UTZYx proposes to occupy.
3. Scope & Non-Goals
In Scope
- ▸ Ground-segment software for telemetry analysis
- ▸ Anomaly detection on housekeeping & payload streams
- ▸ Operator decision-support and assistive interfaces
- ▸ Simulation harness for offline development
- ▸ Integration patterns for existing MOC tooling
Out of Scope
- ○ Spacecraft autonomy and on-board SI
- ○ Launch vehicle systems
- ○ Radio-frequency hardware design
- ○ Closed-loop command authority (operator remains in the loop)
- ○ Mission-critical certification (treated as future work)
4. Reference Architecture
The reference architecture is a five-layer pipeline. Each layer has a clear contract and can be implemented or replaced independently, which is critical for integration with existing mission operations centres (MOCs).
┌───────────────────────────────────────────────────────────────┐ │ L5 · Operator Decision Support (UI + LLM-assisted runbooks) │ ├───────────────────────────────────────────────────────────────┤ │ L4 · Detection & Reasoning (statistical + ML + rules) │ ├───────────────────────────────────────────────────────────────┤ │ L3 · Feature Store (schema-aware, time-aligned) │ ├───────────────────────────────────────────────────────────────┤ │ L2 · Stream Ingest & Quality (Kafka / Pulsar + validation) │ ├───────────────────────────────────────────────────────────────┤ │ L1 · Ground Segment Adapter (CCSDS / SLE / proprietary) │ └───────────────────────────────────────────────────────────────┘
L1 · Ground Segment Adapter
Reads telemetry from existing ground systems via standard interfaces (CCSDS, SLE) or vendor-specific bridges. Operates read-only by default.
L2 · Stream Ingest & Quality
Durable streaming (Kafka or Pulsar) with schema validation, late-arrival handling, and quality flags surfaced as first-class metadata.
L3 · Feature Store
Time-aligned, schema-aware feature extraction. Windowed aggregations, contextual features (orbit phase, eclipse, ground-station handoff).
L4 · Detection & Reasoning
Hybrid stack: residual analysis, density estimators, and supervised classifiers where labels exist. Rule layer encodes operator playbooks.
L5 · Operator Decision Support
Human-first UI. Anomaly cards summarise context, suggest playbook steps, and link to historical similar events. Operator stays in command authority.
Cross-cutting · Security Envelope
Zero-trust between layers, signed artefacts, full audit trail. Detailed in Section 9.
5. Data Layer
Telemetry from a typical small-satellite housekeeping subsystem yields hundreds to thousands of channels sampled at variable rates. The data layer makes three commitments:
- Schema as truth. Every channel carries a versioned schema with units, expected range, and physical meaning. Detectors operate on structured features, not raw bytes.
- Time as a first-class dimension. All features are time-aligned to mission time, with explicit handling of ground-station gaps and signal-loss periods.
- Context as metadata. Orbit phase, eclipse status, manoeuvre windows, and operator notes are joined into the feature store and made available to downstream models.
6. Modelling Approach
We deliberately avoid a monolithic deep model. Mission operations data is heterogeneous, partially labelled, and subject to long tails of rare events. The proposed stack is intentionally hybrid:
Tier A
Statistical baselines
EWMA, robust z-scores, seasonal-trend decomposition. Cheap, explainable, hard to argue with.
Tier B
Unsupervised structure
Isolation Forest, density estimators, autoencoder reconstruction error on multivariate windows.
Tier C
Supervised + LLM assist
Gradient-boosted classifiers where incident labels exist. LLM layer summarises context, never issues commands.
No commands, ever, from the model. The output of L4 is always an annotated event: severity, evidence, similar historical incidents, and a suggested playbook. The operator decides.
7. Evaluation Methodology
Evaluation is held to four metric families:
- Detection quality: precision, recall, and F1 measured against a curated incident catalogue, with per-severity breakdown.
- Operator burden: alerts per operator-hour, time-to-acknowledge, and false-alert rate. SI is only useful if it reduces, not increases, cognitive load.
- Time-to-localise: from alert raised to root subsystem identified, compared against the legacy baseline.
- Drift & calibration: rolling PSI / KS tests on input features and model outputs; calibration plots reviewed monthly.
We commit to publishing reproducible benchmarks against open datasets such as NASA's public SMAP / MSL telemetry releases and any synthetic harness we ship. Unrepeatable claims are not made.
8. Deployment & Operations
The reference deployment is containerised, runs on commodity infrastructure, and is designed to be placed alongside — not inside — the mission-critical ground system. Three deployment postures are supported:
Sidecar
Read-only stream from existing MOC; UTZYx layer runs in parallel, no command authority.
Sovereign
Self-hosted in customer / national infrastructure. Air-gappable. No external telemetry.
Managed
Hosted by UTZYx for downstream-data customers who do not operate their own MOC.
9. Security & Resilience
Space ground systems are classified as critical infrastructure in multiple jurisdictions. The security posture is non-negotiable:
- ▸ Zero-trust between layers; mTLS for all internal traffic.
- ▸ Signed, immutable model and configuration artefacts.
- ▸ Full append-only audit trail of every detector firing and operator action.
- ▸ Supply-chain SBOM for every release; dependencies pinned and scanned.
- ▸ Defined incident-response runbooks tested quarterly.
- ▸ No outbound telemetry to UTZYx in sovereign deployments.
10. Ethics, Safety & Discipline
This work is positioned with the same discipline as the rest of the UTZYx Space Initiative. We make no claim of operational satellites, launch capacity, certification, or agency endorsement. The system described here is a reference approach — a basis for prototypes, pilots, and grant-funded research.
- ▸ Operator-in-the-loop is mandatory; the system never sends commands.
- ▸ Dual-use considerations are tracked per deployment.
- ▸ No use of customer telemetry to train shared models without explicit consent.
- ▸ Vulnerability disclosure follows responsible-disclosure conventions.
11. Implementation Roadmap
Open reference implementation
Public reference of L1–L4 against open NASA SMAP/MSL datasets. Reproducible benchmarks. White-paper iteration based on community feedback.
First paid pilot
Sidecar deployment with one downstream EO operator. Targets: ≥30% reduction in time-to-localise on the agreed incident catalogue.
Grant-funded resilience extension
UKSA / ESA Business Applications proposal extending the security envelope and adding the sovereign-deployment posture.
Hosted-payload software contribution
Subject to partner availability: ground-side SI software contribution to a third-party CubeSat or hosted-payload mission.
12. References & Further Reading
- Space Foundation, The Space Report 2024. Industry estimate of the global space economy.
- World Economic Forum & McKinsey, Space: The $1.8 trillion opportunity for global economic growth, 2024.
- GOV.UK / UK Space Agency, public funding announcements and the National Space Innovation Programme (NSIP) call documentation.
- ESA Space Solutions, Business Applications & Space Solutions programme documentation, business.esa.int.
- NASA Small Business Innovation Research (SBIR) / STTR programme, sbir.nasa.gov.
- Hundman, K. et al., Detecting Spacecraft Anomalies Using LSTMs and Nonparametric Dynamic Thresholding, KDD 2018 (NASA SMAP/MSL dataset).
- CCSDS Recommended Standards for Space Link Extension and packet telemetry, public.ccsds.org.
All cited figures are publicly reported industry estimates or documented programme parameters. UTZYx is not a recipient of these grants and is not endorsed by NASA, ESA, the UK Space Agency, or any other agency.