Modular AI for Faults: Local Watch and Efficient Response

As NASA continues to develop and launch satellites, rovers, landers, and other spacecraft, there is a growing need to increase the level of autonomy of each system. As the number of spacecraft increases, it becomes increasingly difficult to actively monitor each of them for off-nominal behavior, even with the help of a large staff. Furthermore, time-sensitive faults might require a faster response time than what human operators or mission-control can provide. Therefore, one of the key capabilities for an autonomous system is the ability to manage faults and off-nominal behavior. The paper will describe a new framework called MAIFLOWER (Modular AI for Faults: Local Watch and Efficient Response) designed to provide generalizable, automatic anomaly detection, diagnosis and recovery capabilities. MAIFLOWER offers several key benefits. First, MAIFLOWER provides a robust fault detection interface to quickly, accurately, and autonomously identify anomalous behavior from onboard telemetry data. Second, MAIFLOWER integrates several anomaly detection methods from knowledge-based approaches such as model-based reasoning (MBR) to data-driven methods like machine learning (ML) and thermodynamic analysis into a single package, leveraging strengths of each method while mitigating their weaknesses. Third, the fault management modules of MAIFLOWER are designed to cue adaptive execution modules as well as trigger replanning and rescheduling procedures to appropriately respond to faults, creating a closed loop execution environment for continuous anomaly detection and fault management. The paper also describes the application of MAIFLOWER to Astrobotic's Vertical Solar Array Technology Optimized for Lunar Traverse (VOLT), a mobile power generation and transmission platform being developed as part of Astrobotic's lunar power grid architecture.