A Hitchhikers Guide to Anomaly Detection: Machine Learning-based approaches to anomaly detection and Failure diagnosis in spacecraft avionics systems using low power devices & flight ready devices

The rapid growth of satellite constellations, alongside the increasing complexity of satellite systems, has highlighted the need for autonomous satellite health monitoring systems. Traditional human-based telemetry analysis is becoming increasingly inadequate to manage the vast amount of data generated by these systems. As a result, the development of automated anomaly detection frameworks is crucial to ensuring the operational reliability of space-based assets. This thesis investigates machine learning (ML)-based approaches for onboard anomaly detection in small satellites, with a focus on real-time, in-orbit, edge deployment. The initial aspect of this research is the exploration of supervised versus unsupervised learning methods for anomaly detection. Supervised learning techniques, which require labeled data, are commonly used in controlled environments where anomaly types are well defined. However, for space-based applications, where labeled datasets are often sparse or unavailable, unsupervised methods become crucial. Unsupervised anomaly detection models, such as autoencoders, are capable of identifying outliers in telemetry data without the need for explicit labels. This approach provides a flexible solution for detecting novel or unknown anomalies in satellite operations. Therefore, we focused on developing a catalog of univariate anomaly detection models using unsupervised methods. These models were then deployed on an array of Commercial Off-The-Shelf (COTS) edge AI devices to evaluate their performance in space-based environments. The COTS devices, known for their cost-effectiveness and adaptability, presented an ideal platform for testing the feasibility of real-time anomaly detection in constrained environments. Next, we turned to the development of the EIRSAT-1 dataset. With the lack of publicly available data, we collaborated with the EIRSAT-1 team to develop a dataset for multivariate anomaly detection in satellite telemetry. The creation of this dataset was crucial due to the lack of publicly available, high-quality satellite telemetry data that could be used for model development. With the EIRSAT-1 dataset, we trained and evaluated advanced, multivariate models such as Long Short-Term Memory (LSTM) autoencoders, which are well-suited for detecting complex temporal patterns in multivariate telemetry data. To enhance the robustness of the anomaly detection system, we also developed a multiclass classifier to address labeling issues, ensuring that detected anomalies could be accurately categorized. This classifier was integrated into the overall framework to refine the detection process and improve the model's ability to distinguish between different types of anomalies in real-time, leveraging the benefits of both supervised and unsupervised architectures without the drawbacks of either. Finally, a significant challenge in deploying AI-based anomaly detection systems in space is ensuring their trustworthiness for satellite operators. To address this, we incorporated Explainable AI (XAI) techniques, specifically Local Interpretable Model-agnostic Explanations (LIME), into our model. These methods provide transparency into the decision-making process of the AI, allowing satellite operators to understand the reasoning behind the model's anomaly detection outputs. This transparency is critical to gaining the trust of satellite operators and ensuring that the AI system can be effectively integrated into mission-critical operations. Through this work, we present a comprehensive framework for onboard anomaly detection in small satellites, leveraging unsupervised learning, edge AI deployment, multivariate model development, multiclass classification, and explainable AI techniques. The findings of this thesis contribute directly to the advancement of autonomous satellite health monitoring systems, enabling more efficient, reliable, and transparent anomaly detection in space-based applications.