Uncertainty-Aware Model Learning for Fault Diagnosis
TL;DR
This paper presents two novel algorithms (NDDGD-S and NDDGD-C) that learn Grangercausal propositional logic models from uncertain time-series data by jointly exploiting lower-bound and upper-bound observations of the operating range.
Abstract
Fault diagnosis in Cyber-Physical Systems (CPSs) is increasingly challenged by inevitable uncertainties such as sensor noise, incomplete system knowledge, and non-deterministic component interactions. These uncertainties obscure the boundary between nominal and faulty behaviour, complicating both fault detection and root-cause isolation. In this paper, we present two novel algorithms (NDDGD-S and NDDGD-C) that learn Grangercausal propositional logic models from uncertain time-series data by jointly exploiting lower-bound and upper-bound observations of the operating range. NDDGD-S learns causal relationships unsupervised directly between observations, while NDDGD-C learns supervised relationships between annotated component health states and observations. Both algorithms output propositional logic system descriptions suitable for consistency-based fault diagnosis. We evaluate our approach on a coupled four-tank system and demonstrate that incorporating lower- and upperbound time-series data yields more complete system descriptions compared to methods trained on a single time-series, thereby improving diagnostic coverage under uncertainty.