Inferring Failure Processes via Causality Analysis: from Event Logs to Predictive Fault Trees

Abstract

In the current Artificial Intelligence era, the integration of the Industry 4.0 paradigm in real-world settings requires robust and scientific methods and tools. Two concrete aims are the exploitation of large datasets [1] and the guarantee of a proper level of explainability, demanded by critical systems and applications [2]. Focusing on the predictive maintenance problem, this work leverages causality analysis to elicit knowledge about system failure processes. The result is a model expressed according to a newly introduced formalism: the Predictive Fault Trees [3]. This model is enriched by causal relationships inferred from dependability-related event logs. The proposed approach considers both fault-error-failure chains between system components and the impact of environmental variables (e.g., temperature, pressure) on the health status of the components. A proof of concept shows the effectiveness of the methodology, leveraging an event-based simulator [4]