Satellite Fault Detection and Isolation using Deep Learning

While every space system is carefully designed with redundancy and protections to ensure mission success, the reality is that things can still go wrong. Thrusters can get suck, batteries deteriorate, bits get flipped, and each of these small changes can have large ramifications on mission capability and performance. Some of these faults will be obvious, but most are much more subtle and considerably harder to detect. In many cases, these are the most important faults to identify, so we can correct them before something catastrophic happens.

Identification and isolation of spacecraft faults is a particularly challenging effort. Traditionally, operators will monitor telemetry and use their expert insight to determine when things don’t look quite right. This common practice is not only costly, requiring operators to remain stationed at the mission operation center, but it is also somewhat risky as some of the faults may be inconsistent (e.g. pulse spikes) or so slowly-varying that the human eye will struggle to detect them.

To combat these costs and risks, this project investigates how to produce automated systems capable of detecting faults with dimensionality reduction and deep learning algorithms. Explicitly, we study how to use deep learning to discover latent representations of the spacecraft operational modes that can be quickly referenced on-board to detect when the spacecraft produces off-nominal behavior, so operators can be interrupted or warned accordingly.