Reliability Evaluation of Convolutional Neural Networks for Multispectral Images Segmentation on Earth Observation Microsatellites

Recent years have seen an increasing interest in the employment of Neural Networks on spacecrafts in the context of real-time data processing and decision making. However, the radiations typical of the space environment introduce potential risks of data corruption, which could compromise the reliability and usability of these systems. This thesis project aims to address this problem through a newly developed tool, able to detect faults during inference on spacecrafts. The analysis achieved with this software gives a comprehension of the reliability of the Convolutional Neural Network adopted in a specific system, under the given conditions. The case study involves the Payload Data Processor of HEO, a micro-satellite developed by Argotec for the IRIDE program. In this context, a Semantic Segmentation task with a Convolutional Neural Network has been considered. However, the tool is designed to be portable and reusable on other platforms. The tool aims to detect faults by performing on-board inference and comparing the golden outputs, obtained on the same hardware in nominal conditions, with the potentially corrupted predictions. The periodic execution of this software allows to obtain a reliability assessment of the Neural Network under analysis. The thesis opens with an overview of the Low Earth Orbit conditions and a preliminary reliability assessment of the hardware. The central part focuses on the tool logic and the adopted metrics for fault criticality evaluation. It also explores the adoption of Genetic Algorithms to create an image dataset that excites most of a given. This reflects in an increasing coverage of the fault detection mechanism. Finally, Fault Injection is used to validate the efficacy of the developed tool in detecting faults and assessing the reliability of the AI models.