Explainable Deep Learning for Satellite Telemetry Anomaly Detection: A Systematic Literature Review

Satellite and spacecraft telemetry systems produce vast amounts of multivariate time-series data, which are critical for spacecraft health monitoring, fault diagnosis, and mission assurance. Modern developments in artificial intelligence and deep learning have led to dramatic improvements in anomaly detection performance. Yet, many approaches still struggle with explainability, causal reasoning, the lack of standardized benchmarks, high computational cost, and limited on-board deployment feasibility. This paper presents a systematic literature review of AI-based telemetry anomaly detection, focusing on deep learning, explainable AI, transformer learning, graph neural networks and emerging language-model-inspired techniques. We conducted a systematic review of a curated corpus of 47 primary studies, accessed through Scopus, IEEE Xplore, ScienceDirect, Springer Nature Link and Taylor & Francis Online. The review followed PRISMA guidelines and incorporated thematic synthesis, taxonomy development, quality assessment, dataset mapping, and research-gap analysis. The review shows a clear trend from traditional machine learning to recurrent neural networks, autoencoder-based methods, transformer-based temporal learning and hybrid frameworks. It also stresses the increasing importance of explainability and context reasoning for autonomous spacecraft operations. We propose a conceptual framework to address key gaps and limitations across transformer learning, graph causal reasoning, explainable AI components, retrieval-augmented language models, and uncertainty-aware intelligence. This framework can support the development of more interpretable, autonomous spacecraft health-monitoring systems.