Heterogeneous Dataset Inspection and Semi-Supervised Methods for Anomaly Detection

With the increasing complexity of modern space missions, continuous monitoring of spacecraft health has become a critical challenge. Satellite telemetry data, typically collected as high-dimensional multichannel time series with heterogeneous sampling and complex temporal dependencies, represents the primary source for monitoring spacecraft subsystems. However, traditional rule-based or threshold-based monitoring approaches often fail to detect subtle anomalies. This thesis investigates semi-supervised anomaly detection in spacecraft telemetry using the European Space Agency Anomaly Detection Benchmark (ESA-ADB) dataset. As a first step, a structural analysis of telemetry data is conducted to better understand signal characteristics and subsystem relationships. This analysis includes examination of temporal signal behaviour, cross-channel correlations, power spectral density (PSD), and sampling interval patterns across mission subsystems. These analyses provide insights into telemetry heterogeneity and support informed channel selection for anomaly detection experiments. Building on this analysis, two advanced deep learning algorithms reported as effective for telemetry anomaly detection in the ESA context are implemented and evaluated: Telemanom, a forecasting-based approach using Long Short-Term Memory (LSTM) networks, and DC-VAE (Dilated Convolutional Variational Autoencoder), a reconstruction-based generative model designed to learn latent representations of normal telemetry behaviour. Experiments are performed within the TimeEval benchmarking framework in a Docker-based environment to ensure reproducibility and consistent evaluation. The experimental study demonstrates how telemetry structure and subsystem dynamics influence anomaly detection behaviour. The results highlight the importance of combining structural data analysis with machine-learning-based detection models to interpret telemetry behaviour and support reliable spacecraft health monitoring.