LearnTwins is a national Austrian research project funded by FFG (Österreichische Forschungsförderungsgesellschaft).
The project has a duration of three years: from December 2020 to November 2023. Three partners contribute to the project: two research partners and one industrial partner.
Digital twins are very useful for answering or completing inquiries about the past or future behaviour of a complex cyber-physical system (of systems), which is either not yet fully implemented or is used remotely, whereby the creation of a physical local copy is not possible or economical. Digital twins are relatively inexpensive to create when they can be derived and simulated directly from artefacts created during development. However, these optimal conditions are often not (yet) given in practice. An automated creation, e.g. by learning methods, would remedy this. In addition, for successful use, the applied models and the insights derived from them must sufficiently reflect the properties of the real system. If this is not the case, they are even counterproductive as they lead to wrong conclusions. Therefore, one must be able to trust digital twins. Hence, they must be correct and reliable and at the same time cost-effective in their creation and maintenance.
The use of a new technology is often accompanied by doubts about its reliability and concerns about possible side effects. If digital twins are created automatically, e.g. by learning methods, the process behind it is not easy to understand for the user. For acceptance of the technology by the affected user group, instruments must be available to correctly assess the reliability, traceability and limitations of digital twins and to establish trust.
LearnTwins addresses these mentioned challenges by using a combined learn-based testing method. This is based on the insight that the properties of complex systems often cannot be captured concisely in a single model (type). Therefore, the project aims to combine different learning methods to create the digital twin (automatic learning, classical machine learning and deep learning). In addition to already existing data sources, learning data will be gained by executing test cases on the real system, whereby the test cases, in turn, will be created automatically from the (learned) digital twin.
The technical work will be embedded in a foresight process. For this purpose, the involvement of stakeholders is planned, who actively work out desired futures and strategies regarding the developed technology.
The results of the project will enable the faster and more economical creation of high-quality and reliable digital twins and accelerate the necessary digital transformation of product artefacts. The results on the understandability of automatically learned models should contribute to a higher acceptance and a more focussed use of learning-based methods. The developed methods will be tested and evaluated in three realistic use cases from different domains.
LearnTwins will establish a method kit for automated learning of trustable digital twins of cyber-physical systems (CPS).
Purpose of the learned digital twins is to reliably predict the behaviour and dependability of a CPS through simulation or characterization, to gain additional insights into the CPS, and to support its validation, verification and testing.
The method kit will provide new forms of model learning by combining automata learning, classical machine learning, deep learning, and explainable AI and will include automated test-case generation (for quality evaluation and learning data derivation), taking advantage of existing artefacts like log data, tests, and partial models.
A special focus of the learning approaches will be the explainability and understandability of the learned digital twins and a correct perception of their capabilities and limitations, supported by the involvement of psychologists and usability researchers in the team.
Practicability, generalizability and user acceptance of the method kit will be ensured by a user-centric research process and demonstrated and evaluated by applying it to three use cases.