Predictive maintenance on event logs: Application on an ATM fleet
This work provides a new dataset and evaluation framework for researchers and practitioners working on predictive maintenance, particularly for systems where only event logs are available, which is an incremental contribution to the field.
This paper addresses predictive maintenance using event logs instead of sensor data, a common scenario when sensor data is unavailable. It introduces a new public dataset of event logs from 156 ATM machines and proposes an evaluation framework that incorporates business constraints to assess predictive maintenance solutions.
Predictive maintenance is used in industrial applications to increase machine availability and optimize cost related to unplanned maintenance. In most cases, predictive maintenance applications use output from sensors, recording physical phenomenons such as temperature or vibration which can be directly linked to the degradation process of the machine. However, in some applications, outputs from sensors are not available, and event logs generated by the machine are used instead. We first study the approaches used in the literature to solve predictive maintenance problems and present a new public dataset containing the event logs from 156 machines. After this, we define an evaluation framework for predictive maintenance systems, which takes into account business constraints, and conduct experiments to explore suitable solutions, which can serve as guidelines for future works using this new dataset.