Satellite reaction wheels fault detection based on normalized conformal prediction intervals over symbolic regression

Anomaly detection is a critical component in the monitoring of industrial processes. This work focuses on detecting friction anomalies in satellite reaction wheels (RAW) using a Conformal Anomaly Detection
(CAD) framework. Our approach is based on Normalized Inductive Conformal Prediction (NICP), combined with Symbolic Regression (SR) and Multilayer Perceptron (MLP) models. RAW friction and its expected nominal
behavior are used as a baseline for identifying deviations across 12 distinct anomaly types. To support real-time monitoring, we implement an alarm-based detection system that leverages a sliding window technique
for processing streaming data. Our method addresses and resolves certain limitations of CAD in outlier detection by focusing the evaluation windows on the anomalies.