Diagnosis driven Anomaly Detection for CPS

Diagnosing system failures, a process that identifies the root causes of malfunctions, is a critical task in many Cyber-Physical Systems (CPS) applications. The growing complexity of CPS has made it increasingly important to develop diagnostic approaches to ensure their robustness and reliability. Consistency-Based Diagnosis (CBD) has become the state-of-the-art for complex CPS when limited or no information about possible faults is available [Reiter, 1987, Diedrich and Niggemann, 2022]. CBD requires models that represent the normal working behavior of the CPS, typically formulated using propositional logic, comprising symbols for individual components within the system. Furthermore, CBD needs discrete health states of the system's components, known as observations. These health states are often generated through anomaly detection methods [Jung et al., 2018, 2016]. However, diagnosis and anomaly detection are often treated separately in the literature. Current research in anomaly detection for multivariate time series often employs deep learning methods to identify anomalies at the system level or for individual signals [Garg et al., 2022].