Performance Evaluation of Machine Learning Methods for Anomaly Detection in CubeSat Solar Panels

Abstract

CubeSat requirements in terms of size, weight, and power restrict the possibility of having redundant systems. Consequently, telemetry data are the primary way to verify the status of the satellites in operation. The monitoring and interpretation of telemetry parameters relies on the operator’s experience. Therefore, telemetry data analysis is less reliable, considering the data’s complexity. This paper presents a Machine Learning (ML) approach to detecting anomalies in solar panel systems. The main challenge inherited from CubeSat is its capability to perform onboard inference of the ML model. Nowadays, several simple yet powerful ML algorithms for performing anomaly detection are available. This study investigates five ML algorithm candidates, considering classification score, execution time, model size, and power consumption in a constrained computational environment. The pre-processing stage introduces the windowed averaging technique besides standardization and principal component analysis. Furthermore, the paper features the background, bus system, and initial operational data of BIRDS-4, a constellation made of three 1U CubeSats released from the International Space Station in March 2021, with a ML model proposal for future satellite missions.