This paper presents a use-case-driven, model-based solution for predictive maintenance (PM) in cyber-physical production systems. In contrast to off-the-shelf partial solutions, the aim is to adopt an approach that starts with behavior and requirements to map generic properties, and then systematically translates these down to the components involved, including digital twins. System Modelling Language (SysML) was used as the modelling language. It effectively captures the heterogeneity of hardware and software in production, and various other modelling tools can be incorporated. The paper demonstrates how modelling with SysML works for such a system and supports system development in general. Starting from a use case, stakeholder requirements for the system are derived and translated into specifications. From these, an architecture for a suitable solution is derived. SysML is used to create a platform-independent model for PM. A view is used to illustrate how stakeholders are connected to the system solution. Furthermore, the model shows how specific requirements relate to solution components. By presenting different views of the system, individual aspects such as modularity and security can be examined in greater detail. Depending on its suitability for the specific case, the model representation may be, for example, a component view or a behavioral view. Furthermore, the paper demonstrates how, with the aid of the model and artefacts derived from it, a service-oriented architecture (SOA) can be implemented systematically and more easily than before, using administrative shells (AAS) with their sub-model Predictive Maintenance [1]. The paper concludes with insights gained during the modelling process, highlights the further potential of the model-based approach, and discusses possible disadvantages and limitations.