Smart Fault Detection in Satellites: Review of Lightweight ML Models for Real-Time Anomaly Detection

Fault detection in satellites is critical for ensuring mission reliability, longevity, and operational safety. Traditional fault detection methods, such as threshold-based monitoring, are limited in adaptability to dynamic and complex system behaviors, often leading to missed or false anomaly detections. On the other hand, deep learning-based approaches, while powerful, are computationally expensive and unsuitable for real-time onboard processing due to resource constraints. This study explores the use of lightweight machine learning models, including Decision Trees, Random Forest, and Feature-Weighted Random Forest to efficiently detect anomalies in satellite telemetry data. Using publicly available datasets, we evaluate these models based on accuracy, computational efficiency, and feasibility for real-time implementation in resource-constrained satellite systems. Our results demonstrate that Decision Trees achieve an optimal balance between accuracy and low computational overhead, making them highly suitable for onboard fault detection. While Random Forest offers improved accuracy, it comes at the cost of increased computational demand, making it less ideal for real-time operations. These findings underscore the potential of lightweight AI-driven fault detection for enhancing satellite autonomy and mission resilience without requiring extensive onboard computational resources. Future work will focus on deploying these models in real-time satellite environments, optimizing hybrid AI techniques for improved fault detection accuracy, and exploring adaptive learning approaches to enhance model robustness in long-duration space missions.