Unsupervised Fault Detection in Sensor Networks for Extraterrestrial Habitats

Abstract

Various sensors, such as accelerometers, temperature sensors, and pressure sensors, are employed in deep space habitats to monitor operating conditions. The health management system (HMS) depends on accurate sensor data to make reliable decisions. However, the harsh environment of space, characterized by micrometeorite impacts, fire, radiation, dust, and other extreme conditions, can lead to sensor malfunctions. Consequently, the collected data may be corrupted by anomalies, significantly impairing the performance of the HMS. Detecting sensor anomalies and excluding faulty data from decision-making processes is therefore essential. Unsupervised learning has been widely adopted for anomaly detection, as it can train models using unlabeled datasets. This paper presents a novel unsupervised learning approach based on convolutional autoencoders (CAEs) to detect faults in temperature and pressure sensors within the habitat. The proposed method is thoroughly evaluated using a habitat simulator (HabSim), and its capabilities and limitations are discussed.