LearnTwins - Learning Digital Twins for the Validation and Verification of Dependable Cyber-Physical Systems

The LearnTwins project aims to support the creation of digital twins. Different machine learning techniques are used to contribute the missing parts or aspects for the desired digital twin.

Short Description

Digital twins are realistic digital representations of real-world systems and are mainly used in the industrial sector. They allow system behaviour to be analysed and predicted without the need for access to the system itself. This can be helpful during development when the overall system does not yet exist, and for customer systems that are not accessible, or where downtime should be kept to a minimum.

Digital twins are relatively inexpensive to create where they can be derived and simulated directly from artefacts created during development. In many cases, however, this is not fully possible – e.g., where no artefacts are available for purchased components, or no models have been created for specific physical parameters during development.

The project will develop new learning methods and recombine existing learning algorithms to address, in particular, the different characteristics of the different system aspects:

  • discrete behaviour, which is based on countable system states (automata learning), and
  • continuous behaviour, which is typical of physical parameters such as temperature, response time or power consumption (classical machine learning, deep learning).

The training data include data from operation as well as results from tests carried out specifically for the learning process.

The trained digital twins will be used for quality assurance, safety and security analyses and for predicting the effects of changes such as new functions.

Three use cases have been defined for evaluation: 

  1. industrial measurement systems for the project partner AVL,
  2. electronic systems in modern passenger cars, and 
  3. smart home control systems.

The key aspects in the development of learning methods are reliability, understandability, and user acceptance of the trained digital twins.

Technical work is therefore embedded in a foresight process. The plan is to involve stakeholders who actively work out desired futures and strategies for the technology. The first such workshop, with 45 participants from different sectors of society, was held in April 2021 and has led to a number of insights which will inform the technical work.

The results of the project will allow high-quality and reliable digital twins to be created faster and more economically, and accelerate the digital transformation of product artefacts. The results relating to the understandability of automatically learned models should encourage greater acceptance and more targeted use of learning-based methods.

Project Partners

Consortium lead

AIT Austrian Institute of Technology GmbH

Other consortium partners

  • Graz University of Technology (TU Graz), Institute of Software Technology
  • AVL List GmbH

Contact Address

AIT Austrian Institute of Technology GmbH
Dipl.-Ing. Rupert Schlick
Giefinggasse 4
A-1210 Vienna