An unsupervised framework for semantics driven causal explanations for anomalies?

Abstract

Explainability for anomaly detection from multivariate time series sensor data procured from large assets in Industry 4.0 is a challeng- ing and relatively unexplored problem. Apart from the temporal nature of time series data itself, another challenging aspect of this problem is the necessity of making the explainer context aware by infusing seman- tics which may originate from a different data modality. To address this problem, we present a workflow for the first-of-its-kind semantics-driven causal explainer for time series which uses Bayesian Network Structure learning techniques and tailors them for the anomaly explanation prob- lem by simultaneously leveraging the knowledge of asset semantics from a graph model. We also present explanatory insights obtained from in- vestigating a mechanical vibration anomaly for a steam turbine which was validated in our engagement with a large European energy company.