Project Overview
Onboard and ground-based fault classification using supervised and self-supervised learning to detect anomalies across critical spacecraft components before they lead to failure.
The Problem
Spacecraft operate in extreme environments where sensor data is noisy and traditional threshold-based alarms trigger massive amounts of false positives. Ground teams waste hundreds of hours investigating non-issues while missing subtle, complex faults.
Our Solution
We developed a hybrid AI architecture utilizing Self-Supervised Learning (SSL) to pre-train on vast amounts of unlabelled historical telemetry, fine-tuned with supervised classification for known fault signatures. The model can accurately classify 40+ distinct spacecraft faults with extreme precision.
System Architecture
The system processes multi-modal sensor data (thermal, power, vibration) through a series of Convolutional Neural Networks (1D CNNs) and Vision Transformers (ViTs) applied to spectrograms, deployed on a high-throughput edge inference cluster.
Impact & Results
Reduced false alarm rates by 87% compared to legacy threshold systems. Successfully detected a reaction wheel degradation signature 4 weeks before the manufacturer's predicted failure time.