Interpretable Anomaly Detection in Space Systems Using Physics-Informed Clustering

Space systems generate vast quantities of multivariate time-series data from numerous sensors, reflecting the complex interlinking of physical parameters within a dynamic system. As the volume, complexity, dynamism, and multivariance of this data increase, conventional anomaly detection methods become suboptimal for accurately identifying and labeling outliers. In physics-guided approaches, physics is typically built into the loss function as a single equation, or the data are generated through simulation of a physical system. The novelty of this framework, however, lies in coding all equations governing the system as relationships between parameters. These relationships are captured through a multimetric approach and encoded into an image. Subsequently, interpretability for anomaly detection is achieved through dynamic relationship graphs drawn from this image. The strength of this physics-informed, zero-shot training framework is demonstrated by its ability to identify anomalous regions, classifying systems based on physical properties, and it also serves as a diagnostic framework for the digital twin of on-orbit systems. Indeed, this concept can be applied to any complex scientific system to highlight violations of scientific laws, whether they stem from sensor failure, subsystem failure, or system failure. Furthermore, it has been validated using Tennessee Eastman Process data, providing confidence in its effectiveness.