SO-FAR: Satellite Onboard Fault Attribution and Response

Orbit Logic is teamed with the Massachusetts Institute of Technology (MIT) Space Telecommunications, Astronomy and Radiation (STAR) Laboratory on this STTR to develop the Satellite Onboard Fault Attribution and Response (SOFAR) solution. SOFAR will attribute the faults to their probable cause. In many cases, the clues embedded in the data may be sufficiently sparse that there may be multiple possible root causes. In these situations, it is not possible to provide a definitive singular attribution source. Instead, the solution must provide hypotheses as to the root cause and the relative likelihood of each (backed by descriptive metadata explaining how the individual candidates were revealed by various trends in the supporting datasets). The Fault Attribution capability is part of a wider fault architecture this team is building that also includes fault detection and onboard autonomous fault response. Because of the modular architecture, the Fault Attribution function can simply be triggered by Fault Detection events. The “attribution report” determined by the Fault Attribution module can be used in two ways: 1) It can be delivered, via downlink, to ground operations, or 2) When sufficient institutional trust is gained in its output, it could be directly utilized by the Fault Response function of the satellite’s onboard software architecture to react in real- or near-real-time to the fault event. This team’s overall solution design will incorporate the capability associated with use case 2, but the primary focus of the Phase I research and development program will be on the representation of the Fault Attribution data, and the means by which that data is presented to the ground operations staff, who accordingly need to make decisions about how to address the situations when they arise.  The team will be developing a prototype Fault Attribution module that combines both data-driven models, such as Bayesian networks and deep learning methods, and analytical models, referred to typically as Interacting Multiple Mixture (IMM) models. We will additionally employ Bayesian inference, which provides a powerful approach to perform fault diagnosis, using conditional probabilities to model the behavior of the hosting platform’s subsystem-oriented functionality. Using this novel hybrid approach, we will explore (through testing conducted against actual mission telemetry datasets) different methods of modeling the conditional probabilities and associated loss functions, collecting metrics for attribution determination time, attribution correctness, and processing/memory resource utilization to help guide final design choices for future development and solution maturation efforts.