On simple baselines for domain shift in condition monitoringPanić, Branislav (Avtor) Pasic, Mirza (Avtor) Nagode, Marko (Avtor) Oman, Simon (Avtor) condition monitoringdomain shiftbearing fault identificationdeep learningbaseline studytransfer learningMethods for handling domain shifts in condition monitoring have proliferated, yet their evaluation often lacks rigorous baseline comparisons and systematic isolation of individual shift factors. This paper proposes a task-focused methodology that structures domain shift studies as a pipeline from data collection through task definition, dataset construction, representation, normalization, model specification, and training. Using bearing fault identification as a case study, we introduce the Relative Performance Drop (RPD) metric and conduct over 600,000 evaluations across the Case Western Reserve University and Paderborn University bearing datasets. Our results reveal that domain shift severity depends strongly on which physical factors vary: rotational speed causes substantial degradation (42.8% RPD), load and force have negligible impact, while shifts in fault type or severity suggest the task itself may require reframing rather than more sophisticated algorithms. Design choices often treated as implementation details can influence cross-domain performance as substantially as model architecture, with optimal choices reversing across collections. A difficulty taxonomy clusters domain pairs by their mean and variance of RPD, distinguishing shifts addressable by simple pipeline design, those requiring dedicated adaptation methods, and those indicating ill-posed task definitions. The accompanying open-source implementation enables application of this methodology to new collections and tasks with full reproducibility.20272026-07-22 15:46:55Članek v reviji185103UDK: 62ISSN pri članku: 1879-0836DOI: 10.1016/j.ress.2026.113041COBISS_ID: 285742339sl