QSIM-Guided Weak Supervision for Fault Detection in Cyber-Physical Systems

Abstract

This paper presents an initial investigation into the combination of qualitative simulation and machine learning for fault detection in cyber-physical systems. In recent years, the detection and diagnosis of faults in cyber-physical systems has become increasingly dependent on machine learning. However, supervised machine learning approaches rely on large labelled datasets, which are rarely available during the initial deployment phase. In systems such as smart buildings, faults occur rarely, and even the nominal behaviour of a system can vary across structurally identical installations. We therefore propose a framework that addresses this issue by generating weak supervision labels using qualitative reasoning. Given a qualitative model that captures the system’s nominal behaviour, we use a conformance check to determine whether the system traces are likely to be faulty or not. These binary pseudo-labels are then used to train a downstream classifier. Unlike statistical anomaly detection or self-supervised techniques, the generated supervision signal is based on physical consistency rather than deviations from a learned baseline. We evaluate our approach on four benchmarks relating to electrical, fluid, mechanical and thermal dynamics, comparing our methodology with unsupervised anomaly detection and supervised learning with limited labels. Our initial experiments suggest that, while qualitative conformance checks can provide usable training signals when labelled data is unavailable, their effectiveness depends on the discriminative structure of the qualitative model.