Single- and multi-task linear models for ATMs fault classification in human-centered predictive maintenance

The recirculator, a complex component within Automated Teller Machines (ATMs) responsible for handling banknotes, poses a challenging task for fault diagnosis due to its intricate nature, which renders it impractical to integrate dedicated sensors and potential multiple faults. This paper presents advanced single-task (STL-LR) and multi-task (MTL-LR) logistic regression models explicitly designed for capturing specific and similar discriminative patterns of multiple faults. Our approach focuses on maintaining the expert human operator at the center of the model checking and development process (human-in-the-loop approach). This objective has been achieved by including training data extracted from the intervention management platform, which collects the annotations of human operators. By leveraging this data, our STL and MTL models enhance generalization performance, especially in cases where discrepancies exist between machine-reported errors and technician-observed anomalies. The results illustrate the potential of the STL-LR and MTL-LR models as the main core of PdM DSS to aid technicians in accurately pinpointing fault-prone areas. This research contributes to Industry 5.0 by presenting a novel predictive maintenance approach that evolves task-specific learning to the generalization advantages of MTL. This evolution holds promise for fostering more efficient and effective maintenance strategies in complex equipment environments. • Introduced a human-centered DSS for predictive maintenance in ATMs. • Integrated human expertise with automated diagnostics for PdM. • Collected and analyzed real-world operational data from ATMs. • Developed STL and MTL models enhancing fault detection accuracy. • Proposed edge computing for real-time fault diagnosis and response.