PhD position in "A transformation language for simultaneous data integration and inference in digital twins" - MSCA Cofund SEED programme

Digital Twins (DTs) integrate in real time heterogeneous data sources, coming from sensors, devices, and subsystems of the physical counterpart. This integration is especially important in large-scale DTs, from the domain of manufacturing systems up to nation-wide DTs. However, not all data needed by DTs is directly measurable in the physical system, e.g. because of the costs of the corresponding sensors. Hence, besides integrating observed data, DTs need also to infer non-observed data. Given the heterogeneity of the possible data sources, and the difficulty in inferring meaningful data, these phases are generally very costly.

In the state of the art, data integration and data inference for DTs are performed in separate steps by different tools. Model Transformation (MT) languages have a natural application in the data integration step. Machine Learning (ML) techniques are instead prominently used for data inference, after training on historical traces of the system or simulations. Executing these two steps sequentially has however important drawbacks: 1) if inference is performed before integration, then it may effectively infer only local data about a device or subsystem; 2) if integration is performed before inference, then the information loss during integration may negatively impact the inference potential.

We argue that data integration and inference for DTs should be defined and performed at the same time. However, defining the integration+inference logic as a single step is not trivial, since it involves two different sets of competences. We aim at a linguistic solution to this problem, by designing and implementing a single language for the simultaneous specification of integration and inference of sensor data into the DT model. The language will be obtained by extending a MT language with ML capabilities. Users will be able to define transformation rules for data integration, coupled with ML tasks that are aware of the integration logic, avoiding both information loss and restriction to local views. This would allow DT engineers to reduce the costs of the data integration and inference phase. This will also increase the applicability of the partners’ tools in DTs of several domains.