Arkitektonisk design och verifiering/validering av system med maskininlärning komponenter

Forskningsprojekt, 2020 – 2024

Machine Learning (ML) has shown large potential for solving problems intractable by classical methods. These include pattern recognition in unstructured data such as images, a key enabling technology for e.g., automated driving. It has also been shown to be a useful tool for other types of various automotive pattern recognition tasks and for the development of predictive models.

Due to their stochastic nature, using ML components requires specific approaches related to the architectural design and verification/validation of the vehicle’s electrical system. On the one side, suitable architecture containing both stochastic and traditional deterministic components shall be defined which assures safety, robustness, fault tolerance, and the exchange of data between ML components and the rest of the vehicle. On the other side, the nature of ML algorithms makes them difficult to test and analyze from a safety perspective, which requires the implementation of specific verification and validation methods.

In order to address the above-mentioned needs, this project aims to:
• Define and evaluate different architectural styles and patterns for developing systems with ML components and exchange of obtained knowledge between their independent subsystems.
• Develop methods for validation and verification, enabling the use of ML for safety-relevant vehicle systems with high accuracy and availability requirements.

In the short term, the project is expected to contribute to the development of safe, robust, and fault-tolerant electrical systems with ML components in vehicles. In the long run, the project is expected to contribute to the continuous deployment of ML components to different autonomous subsystems in cars and their safe and reliable utilization in the transportation ecosystem.