LLM-Based Approach for Extracting OCEL from Incident Data: An Experimental Study

Incident monitoring in industrial environments requires comprehensive and structured event logs to support advanced analytics such as process mining and incident prediction. While traditional equipment logs are often limited to single-object perspectives, valuable information is also embedded in unstructured pre- and post-incident textual records, which frequently describe complex interactions involving multiple entities. The ObjectCentric Event Log (OCEL) standard enables the representation of such multi-object events, but generating OCELs from unstructured data remains a significant challenge. In this work, we present and compare two LLM-based extraction strategiesrecursive and two-stepfor transforming unstructured incident-related texts into object-centric event logs. We further introduce a robust LLM-based event matching methodology and introduce a unification step to both event types and object types. This unification step is crucial for enhancing the consistency and interpretability of the resulting logs, ensuring their suitability for downstream tasks such as process mining and incident prediction. Extensive experiments on real-world data from the data center domain demonstrate the effectiveness of our approach and provide detailed insights into extraction quality, type unification, and event attribute classification.