Statistical analysis of internal activations: application to autoencoder in space domain

This thesis addresses the problem of confidence estimation in neural networks, with a specific focus on anomaly detection for satellite telemetry data. The proposed framework explores the extraction of statistical features from intermediate layers of an Au- toencoder to assess how good it performs the task of anomaly detection. The goal is to provide not only an anomaly score but also a measure of how confident the model is in its decision, improving interpretability and trust in satellite monitoring systems that rely on data-driven learning mechanisms. The study focuses on multivariate time-series data representing telemetry signals acquired from sensors distributed across satellite subsystems. An autoencoder architecture is trained on nominal operational data and evaluated on synthetic anomalies of different types and magnitudes. The anomaly detection performance is measured through reconstruction-based scores, while the internal activations of the network are analyzed to derive a quantitative expression of model reliability, referred to as the confidence score. A statistical analysis of intermediate activations is conducted to identify meaningful correlations between the model’s internal state and its predictive performance. Different statistical descriptors such as variance, entropy, and layer-wise distance measures are investigated to quantify uncertainty directly from the network’s internal layers behavior. Preliminary results show that confidence indicators derived from intermediate layers can effectively discriminate between reliable and uncertain predictions, offering an additional layer of information for anomaly detection frameworks. This work lays the foundation for more interpretable and trustworthy neural network models for space telemetry applications and can be extended to other safety-critical systems where reliability assessment is essential.