Hello HN community! We are a team consisting of Kamran, Raaid, Laith, and Omeed representing Constellation Space. Our website can be found at https://constellation-io.com/. Our group has developed an artificial intelligence system that can anticipate satellite connection failures before they occur. You can view a detailed demonstration in this video: https://www.youtube.com/watch?v=069V9fADAtM. Collectively, we have accumulated years of experience handling satellite operations at prominent organizations like SpaceX, Blue Origin, and NASA. At SpaceX, our responsibilities included managing the health of the Starlink constellation. While at Blue Origin, we focused on enhancing test infrastructure for the upcoming New Glenn project. During our time at NASA, we were involved in managing communications in deep space.

A recurring issue we encountered was the delay in identifying deteriorating links, often resulting in data loss by the time the problem is noticed. The primary challenge lies in the fact that satellite radio frequency links are influenced by numerous interconnected factors. When a satellite passes overhead, it becomes crucial to predict whether the link will remain stable for the subsequent few minutes. This prediction relies on factors such as orbital geometry (where elevation angles are constantly changing), tropospheric attenuation (signal loss due to humidity as per ITU-R P.676 standards), rain fade (quantified using ITU-R P.618 – rain intensity in mm/hr directly impacts signal loss at Ka-band and above frequencies), ionospheric scintillation (monitoring KP index fluctuations via magnetometer networks), and network congestion among others.

The conventional method is reactive where operators monitor dashboards and react when Signal-to-Noise Ratio drops below a specific threshold by manually redirecting traffic or switching to backup links. However, with over 10,000 satellites currently orbiting Earth and projections exceeding 70,000 by 2030, this approach is not scalable. Our system processes telemetry data from satellites, ground stations, weather radar systems, IoT humidity sensors, and space weather monitors at a rate of approximately 100,000 messages per second.

In real-time, we utilize physics-based models encompassing complete link budget equations, ITU atmospheric guidelines, orbital propagation calculations to forecast expected outcomes. Additionally, Machine Learning models trained on vast datasets sourced from actual multi-orbit operations are layered on top of these models. The Machine Learning component is particularly intriguing as we implement federated learning to respect the privacy concerns of constellation operators who are unwilling to share raw telemetry data. Each constellation trains its own local models on their proprietary data sets which are then aggregated to identify overarching patterns.

This strategy enables us to leverage knowledge across various orbit types and frequency bands; insights gained from Low Earth Orbit (LEO) Ka-band connections assist in optimizing Medium Earth Orbit (MEO) or Geostationary Orbit (GEO) operations. We achieve a prediction accuracy exceeding 90% for most link failures anticipated 3-5 minutes ahead of time which allows ample opportunity for traffic rerouting prior to any data loss occurrence. Our system is fully containerized using Docker/Kubernetes technology and can be deployed either on-premise for isolated environments or on secure cloud platforms such as GovCloud (AWS GovCloud, Azure Government) or standard commercial clouds.

Currently undergoing testing with defense and commercial partners, our dashboard offers real-time monitoring of link statuses alongside forecasts ranging from 60 seconds up to 300 seconds ahead as well as root cause analysis tools pinpointing issues like rain fade or satellite positioning below the horizon or network congestion. We provide an API interface granting access to telemetry ingestion services, predictions, network topologies snapshots, and even a natural language chat endpoint for troubleshooting purposes.

Challenges we continue to address include maintaining prediction accuracy over extended time frames (accuracy diminishes beyond 5 minutes), acquiring additional labeled failure data for rare scenarios and coordinating federated learning setups across diverse operator security protocols. We welcome feedback from professionals experienced in satellite operations, RF link modeling or large-scale time-series forecasting efforts regarding potential improvements for operational deployment effectiveness within NOC environments. Feel free to reach out with any technical inquiries you may have!