Anomaly Detection Using Deep Learning Respecting the Resources on Board a CubeSat

Abstract

We explore the feasibility of onboard anomaly detection using artificial neural networks for CubeSat systems and related spacecraft where computing resources are limited. We gather data for training and evaluation using a CubeSat in a laboratory for a scenario where a malfunctioning component affects temperature fluctuations across the control system. This data, published in an open repository, guides the selection of suitable features, neural network architecture, and metrics comprising our anomaly detection algorithm. The precision and recall of the algorithm demonstrate improvements as compared to out-of-limit methods, whereas our open-source implementation for a typical microcontroller exhibits small memory overhead, and hence may coexist with existing control software without introducing new hardware. These features make our solution feasible to deploy on board a CubeSat, and thus on other, more advanced types of satellites.