Process mining relies on event logs extracted from information systems. Manual work on the factory shop floor leaves few digital traces and therefore remains largely outside its reach. Egocentric video, recorded by head-mounted cameras, captures this work in detail, but raw footage is not event data. This paper presents a fully automated pipeline that derives standard event logs from egocentric factory videos using foundation models. The pipeline transcribes each video into a timestamped textual narration, induces a controlled vocabulary of processes and activities for each factory, and converts every transcription into a sequence of events constrained to this vocabulary. Applied to six factories of the Egocentric-10K corpus, it produces event logs covering more than one hundred hours of manual work without task-specific training or human annotation. We evaluate the resulting logs in two steps. A validity assessment confirms that the extracted events are structurally sound and faithful to the transcribed footage. An analysis-oriented assessment then investigates which questions the logs can answer in practice: how worker time is distributed over value-adding work, material handling, transport, and rework; how often value-adding work is interrupted; and how workers performing the same task compare. In several cases, the logs surface improvement opportunities that qualitative reporting missed, while also revealing that the induced vocabulary determines what can and cannot be measured. Pipeline, prompts, and all derived artifacts are publicly available.
原文链接
分享
引用
