Detector of Anomalies Based on Regularized Autoencoder for Orbital Solar Array Telemetry Data

This thesis addresses the problem of anomaly detection in satellite Electrical Power Systems (EPS), with a specific focus on solar array telemetry data. An unsupervised, data-driven framework is developed using Convolutional Autoencoders (CNNAE), trained on proprietary dataset to identify anomalous patterns in multivariate time-series. Both regularized and non-regularized model configurations are investigated. A preliminary step consists in a clustering-based inspection and preprocessing phase, called Inspection 0, which combines Mean Shift and Dimensionality Reduction via Principal Component Analysis (PCA) to explore data variability and identify representative patterns. Multiple CNNAE configurations are trained and then evaluated. The detection framework is validated with the injection of synthetic White Gaussian noise on test data. The performance are evaluated using the Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) across multiple anomaly scores, including reconstruction error (MSE), Mahalanobis distance, latent L2 norm and regularized loss. Additionally, a PCA-based baseline is implemented for comparative analysis. Experimental results confirm that CNNAE show strong anomaly detection capabilities in discriminating between normal and anomalous behaviors, particularly when the latent space regularization is applied. The latent representation also enables encoder-only onboard deployment for fast, resource-efficient anomaly screening, while the full architecture can be implemented off-board for long-term monitoring. This work demonstrates the feasibility and flexibility of the proposed deep learning solution in real operational scenarios and lays the foundations for future developments in autonomous satellite health monitoring, including integrations with prognostics and diagnostic modules. The framework can be extended to other satellite subsystems, such as thermal control or batteries.