Supervised and Unsupervised Methods for Detecting Anomalies in an Autonomous Long-Range Lunar Rover

Abstract

Autonomous space systems must navigate complex environments and accomplish concurrent tasks without continuous input and supervision from human operators. Interactions between subsystems and with the environment may lead to unintended behavior, resulting in downtime, delays, goal degradation, loss of functionality, or even catastrophic failures. The Endurance mission concept requires a lunar rover to traverse thousands of kilometers over a long period of time (multiple years) across the South Pole-Aitken Basin on the far side of the Moon. Onboard system health management is therefore required to identify anomalies and faults that pose a risk to mission objectives. To this end, we propose training both unsupervised and supervised machine learning models to detect and isolate anomalies onboard a rover over the course of a mission, supported by a pipeline for anomaly data generation that enables training and evaluation. We generate synthetic anomaly signatures using a low-fidelity mission simulator that outputs labeled datasets to enable supervised learning. We present results from a field experiment in which we deploy this supervised model as a ROS network node in a rover-ground network. In parallel, we train unsupervised models to detect anomalies in the mobility system by training on experimental field data and present results that verify the ability to detect anomalies observed by field operators as well as anomalies that were not detected by the operators. Our work demonstrates how machine learning models can detect anomalies onboard by leveraging multiple data sources, including pre-launch test data and operational data from earlier phases of the mission, and provides a pathway for improving anomaly detection as a rover's mission progresses.