On the Relevance of Graphical Causal Models for Failure Detection for Industrial Machinery

Assessing the reliability of industrial machinery is an important aspect within maintenance processes in order to maximize productivity and efficiency. In this paper we propose to use graphical models for fault detection in industrial machinery within a condition-based maintenance setting. The contribution of this work is based on the hypothesis that during fault free operation the causal relationships between the observed measurement channels are not changing. Therefore, major changes in a graphical model might imply faulty changes within the machine's functionality or its properties. We compare and evaluate four methods for the identification of potential causal relationships on a real world inspired use case. The results indicate that sparse models (using L 1 regularization) perform better than traditional full models.