Sensor2EventLog: Bridging Continuous IoT Data and Process Mining through Eventization

Process Mining (PM) provides a powerful means to uncover, monitor, and improve real-world processes by extracting knowledge from event data. Applying PM to cyber-physical systems (CPS) offers unique advantages (such as enabling discovery, conformance, and predictive monitoring), directly from sensor data, particularly from Internet of Things (IoT)-enabled devices. However, most sensor data streams are continuous, heterogeneous, and lack explicit process semantics, while PM requires discrete, case-related event logs. Existing abstraction approaches are heavily based on fixed thresholds or rigid time segmentation strategies, which often degrade the quality of the resulting event logs and restrict the range of PM tasks that can be reliably performed. This paper presents Sensor2EventLog, a knowledge-based framework that transforms raw sensor data into process-aware event logs incorporating Machine Teaching (MT) principles. The core of the framework is a modular abstraction layer embedded in an interactive teaching loop (planning, explaining, and reviewing) where human experts map sensor behaviors corresponding to process states (i.e., process activities). The resulting event logs conform to standard formats, enabling the full range of PM tasks (including predictive process monitoring) without relying on fixed time windows. Experimental results show that, by incorporating domain knowledge into this process, the proposed approach achieves 20–30% higher state abstraction accuracy than purely data-driven methods.